Datasets:

Daniel-P-Gonzalez

/

OCD

like 0

# Introduction Molecular features of solid tumours become central in tailoring targeted therapies, but the accessibility to tumour tissue may be sometimes limited due to the size of bioptic samples or the unavailability of biological material, particularly during patients' follow up. In this context cancer-derived cell- free DNA in blood (cfDNA) represents a promising biomarker for cancer diagnosis and an useful surrogate material for molecular characterization. The two classes of alterations detectable in cfDNA from cancer patients include quantitative and qualitative abnormalities. Concerning the former aspect, it is now evident that cancer patients have a higher concentration of cfDNA than healthy individuals (see ref. 2 for a review). The concentration of cfDNA is influenced by tumor stage, size, location, and other factors. On the other hand, increased plasma DNA level is not a specific cancer marker, as it is observed also in patients with premalignant states, inflammation or trauma. Total cfDNA concentration has been proposed as a marker for early cancer detection, but the studies conducted so far showed a scarce discriminatory power between patients and controls as well as limited sensitivity and specificity, not allowing one to reach any final conclusion on the diagnostic impact of this parameter. Several studies report a prognostic value of total cfDNA, while conflicting results have been obtained in testing this marker for therapy monitoring. The reduced specificity of this quantitative test leads us to evaluate additional biomarkers reflecting qualitative alterations in cfDNA. A higher specificity in cancer diagnosis can be achieved by detecting tumor specific alterations in cfDNA, such as DNA integrity, genetic and epigenetic modifications. Blood cfDNA in cancer patients originates from apoptotic or necrotic cells. In solid cancers, necrosis generates a spectrum of DNA fragments with variable size, due to random digestion by DNases. In contrast, cell death in normal blood nucleated cells occurs mostly via apoptosis that generates small and uniform DNA fragments. It has generally been observed that in patients affected by several neoplastic diseases plasma DNA contains longer fragments than in healthy subjects – reflected by the increase of DNA integrity index. The above mentioned parameters can obviously be considered as non-specific biomarkers, since the increase of cfDNA concentration and integrity is common to the large majority of human solid cancers. When cfDNA is used to detect genetic and epigenetic modifications in a specific tumor, it is necessary to select definite molecular targets that are expected to be altered in affected patients. In cutaneous melanoma, the oncogene *BRAF* is frequently mutated. *BRAF* is a serine–threonine protein kinase involved in the RAS–RAF–MEK–ERK pathway which regulates cell growth, survival, differentiation and senescence. The oncogene *BRAF* is frequently mutated in other human cancers constitutively activating the MAPK pathway. The most common *BRAF* mutation, which accounts for more than 90% of cases of cancer involving this gene, is the T1799A transversion, converting valine to glutamic acid at position 600 (V600E). *BRAF* somatic mutations have been reported in 66% of malignant melanomas and are likely to be a crucial step in the initiation of melanocytic neoplasia, as they are found also in melanocytic nevi. *BRAF* mutations are an attractive target for therapeutic interventions, as they represent an early event in melanoma pathogenesis and are preserved throughout tumor progression. Specific inhibitors of mutant *BRAF*, such as PLX4032, were developed and tested in clinical trials showing response rates of more than 50% and improved rates of overall and progression-free survival in patients with metastatic melanoma with the *BRAF^V600E* genetic variant. *BRAF^V600E* mutation has been investigated as a marker in cfDNA from melanoma patients by Daniotti et al. and Yancovitz et al.. Finally, it is widely demonstrated that a limited number of genes is epigenetically disregulated in cutaneous melanoma. *RASSF1A* (Ras association domain family 1 isoform A) is a tumor suppressor gene, which regulates mitosis, cell cycle and apoptosis. It is inactivated mostly by inappropriate promoter methylation in many types of cancers. *RASSF1A* promoter is methylated in 55% of cutaneous melanomas. Methylation of *RASSF1A* increases significantly with advanced clinical stage, suggesting that inactivation of this gene is associated with tumor progression. *RASSF1A* promoter hypermethylation has been detected in cfDNA from melanoma patients – in association with a worse response to therapy and reduced overall survival. Previous studies assessed the diagnostic performance of each of the above mentioned biomarkers singularly considered in selected case-control comparative surveys. The aim of the present study was to identify a sequential multi-marker panel in cfDNA able to increase the predictive capability in the diagnosis of cutaneous melanoma in comparison with each single marker alone. To this purpose, we tested total cfDNA concentration, cfDNA integrity, *BRAF^V600E* mutation and *RASSF1A* promoter methylation associated to cfDNA in a series of 76 melanoma patients and 63 healthy controls. # Materials and Methods ## Patients and samples Seventy six patients (32 females and 44 males, median age 63, range 23–94 years) affected by cutaneous melanoma were enrolled at the Department of Dermatological Sciences of the University of Florence. The series included: 12 patients with in situ melanoma (4 females and 8 males; age range:39–80 years, median 60 years), 49 patients with local disease (22 females and 27 males; age range:23–88 years, median 60.9 years), 5 patients with regional metastatic disease (1 females and 4 males; age range:53–88 years, median 69.4 years) and 10 patients with distant metastatic disease (5 females and 5 males; age range: 28–94 years, median 50 years). For additional baseline and clinical characteristics of invasive melanomas see. As a control group 63 healthy subjects with less than 50 melanocytic nevi (median age 62, range 25–79 years) were enrolled in the study upon a dermatological examination to exclude the presence of melanoma and to provide the number of nevi. Blood samples (5 ml) were collected in EDTA tubes during the dermatologic examination and before surgery. The research protocol was approved by the review board of the University of Florence and all the patients signed an informed consent. Plasma was separated from blood in EDTA tubes, within three hours from blood draw by two centrifugation steps at 4°C for 10 min: at 1600 rcf and 14000 rcf, respectively. Plasma aliquots (505 µl) were stored at −80°C. DNA was extracted from 500 µl of plasma within 3 months from collection, by the QIAamp DSP Virus Kit (Qiagen, Italy) according to the manufacturer's instructions. RNAse digestion was included in the procedure to prevent RNA interference during the subsequent qPCR reactions. ## Molecular biomarkers in cfDNA All the cfDNA samples from melanoma patients and healthy controls were submitted in duplicate to the four qPCR assays targeting the chosen biomarkers, for a total of about 1000 determinations. All the qPCR reactions were performed using the 7900HT Fast Real-Time PCR instrument (Applied Biosystems). All the methods described in the following section have been previously developed or optimized for cfDNA by our laboratory using plasma samples from different case studies. The total amount of cfDNA as well as the DNA integrity index were determined by two qPCR assays targeting respectively a 67 bp and a 180 bp sequence on the single copy gene *APP* (Amyloid Precursor protein, chr. 21q21.2, accession NM_000484), as already reported. The primers and the hydrolysis probe for the 67 bp amplicon were previously reported , while for the 180 bp amplicon a different reverse primer was designed on the same target sequence. The shorter amplicon (67 bp) was used to quantify total cfDNA, while the ratio between the absolute concentration of the longer amplicon (180 bp) and the shorter one (67 bp) defined the integrity index 180/67, which was used to assess the fragmentation of cfDNA. An integrity index close to 1 indicates that all the cfDNA molecules are at least 180 bp in length in the *APP* gene. An integrity index of less than 1 means that cfDNA contains fragments below 180 bp in the same target sequence. CfDNA that is more intact will be closer to a value of 1 for the integrity index. The reactions were carried out in a 12.5 µl mix containing 1× Quantitect® Probe PCR Master Mix (QIAgen), 300 nM primers, 200 nM probe and 1 µl sample. The thermal profile of the amplification was the following: 95°C for 10 min and 45 cycles of PCR at 95°C for 15 s, 60°C for 60 s. For cfDNA quantification we used an external reference curve ranging from 10 to 10⁵ pg/tube, obtained by serial dilutions of genomic DNA extracted from a blood pool of healthy donors and measured spectrophotometrically (Nanodrop ND1000, Nanodrop, USA). Circulating cell-free DNA bearing the mutation *BRAF^V600E* was quantified by an allele-specific qPCR assay, as already reported. The specificity for the mutated allele was conferred by the forward primer and the LNA probe. cfDNA (0.5 ng) was amplified in a reaction mixture containing 1× Quantitect® Probe PCR Master Mix (QIAgen), 200 nM primers and 200 nM probe in a final volume of 20 µl. The thermal profile of the reaction included a denaturation step at 95°C for 10 min and 50 cycles of PCR at 95°C for 15 s, 64°C for 60 s. *BRAF^V600E* percentage was calculated by referring to a standard curve obtained by mixing DNA from mutant (SKMEL28) and wild type (MCF7) cell lines in the following proportions: 100%, 50%, 20%, 10%, and 1% mutated alleles. The presence of the *BRAF^V600E* mutation was excluded in the MCF7 human breast adenocarcinoma cell line and confirmed in the SKMEL28 human melanoma cell line by High Resolution melting followed by sequencing (data not shown). Subsequently *BRAF^V600E* concentration was expressed in nanograms per ml plasma by multipling this percentage for absolute DNA concentration determined by the qPCR assay for *APP*. The methylated form of *RASSF1A* promoter was quantified in plasma after digesting unmethylated DNA by a methylation-sensitive enzyme: 5 µl of plasma DNA were treated with 10 units of Bsh1236I (Fermentas, Canada) in a reaction volume of 25 µl at 37°C for 16 hours. Subsequently, 5 µl of enzyme-treated DNA underwent a qPCR assay for *RASSF1A* promoter, in a final volume of 25 µl, according to the protocol already described by Chan et al.. A reference curve obtained by serial dilutions of genomic DNA was used to quantify the methylated alleles. Results were expressed as genomic equivalents (GE, each corresponding to 6.6 pg DNA) per ml plasma. ## Statistical Analysis All the considered biomarkers were analysed as continuous variables in their original scale or after an appropriate transformation. Comparison of biomarkers distribution in cases and controls overall as well as according to stage of disease was performed by using the Kolmogorov-Smirnov test. The relationship between each biomarker and the disease status was investigated by resorting to a logistic regression model in both univariate and multivariate fashion. In the logistic regression model, each regression coefficient is the logarithm of the odds ratio (OR). Under the null hypothesis of no association, the value of OR is expected to be 1.00. The hypothesis of OR = 1 was tested using the Wald Statistic. For each model the biomarker that was statistically significant (alpha = 0.05) in univariate analysis was considered in the initial model of multivariate analysis. A final more parsimonious model was then obtained using a backward selection procedure in which only the variables reaching the conventional significance level of 0.05 were retained (final model). The relationship between each biomarker and disease status was investigated by resorting to a regression model based on restricted cubic splines. The most complex model considered was a four-nodes cubic spline with nodes located at the quartiles of the distribution of the considered biomarker. The contribution of non-linear terms was evaluated by the likelihood ratio test. We investigated also the predictive capability (ie diagnostic performance) of each logistic model by means of the area under the ROC curve (AUC). This curve measures the accuracy of biomarkers when their expression is detected on a continuous scale, displaying the relationship between sensitivity (true- positive rate, y-axes) and 1-specificity (false-positive rate, x-axes) across all possible threshold values of the considered biomarker. A useful way to summarize the overall diagnostic accuracy of the biomarker is the area under the ROC curve (AUC) the value of which is expected to be 0.5 in absence of predictive capability, whereas it tends to be 1.00 in the case of high predictive capacity. To aid the reader to interpret the value of this statistic, we suggest that values between 0.6 and 0.7 be considered as indicating a weak predictive capacity, values between 0.71 and 0.8 a satisfactory predictive capacity and values greater than 0.8 a good predictive capacity. Finally the contribution of each variables to the predictive capability of the final model was investigated by comparing the AUC value in the model with that of the same model without the variable itself. All statistical analyses were performed with the SAS software (Version 9.2.; SAS Institute Inc. Cary, NC) by adopting a significance level of alpha = 0.05. # Results The box-plots reported in, panel A–D, describe the distribution of each biomarker in case and controls. reports some descriptive statistics of these distributions. Using the Kolmogorov–Smirnov test, we found that the difference of the distributions of each biomarker in cases and controls was statistically significant (p-value \<0.05). As reported in supplemental, the same results were observed when this comparison was performed according to the stage of disease for cfDNA and integrity index 180/67. Conversely these findings were not observed within stage I–II for methylated *RASSF1A* and within stage 0 and stage III–IV for *BRAF^V600E*. For all the biomarkers considered in the logistic regression model we found that a linear relationship between the log odds and their values on the original (methylated *RASSF1A*) or logarithm (total cfDNA, integrity index 180/67 and *BRAF^V600E*) scale was appropriate. As reported in, disease status was significantly associated with all the biomarkers in the logistic univariate analysis. Consequently the initial model of the logistic multivariate regression analysis was built by including all four biomarkers. As reported in, total cfDNA, integrity index 180/67 and methylated *RASSF1A* retained a statistically significant (p-value \<0.05) association with disease status in the multivariate final logistic model. The AUC values computed for each biomarker (univariate logistic model) indicated a weak/satisfactory level of predictive capability by ranging between 0.64 (*BRAF^V600E*) to 0.85 (total cfDNA). Of note for all the considered biomarkers the 95% Confidence Interval (95%CI) of the AUC fails to include the 0.5 value (i.e. absence of predictive capability). Overall, a good predictive capability was observed for the final logistic model with an AUC of 0.95 (95% CI: 0.91–0.98). The contribution of each variable of the final model to the diagnostic performance is shown in and graphically described in. The highest predictive capability was given by total cfDNA (AUC:0.86, 95%CI: 0.80–0.92) followed by integrity index 180/67 (AUC:0.90, 95%CI: 0.85–0.95) and methylated *RASSF1A* (AUC:0.89, 95%CI: 0.84–0.95). As shown in the supplemental a comparable predictive capability was observed for each considered biomarker (univariate analysis) according to the stage of disease. Only for BRAF^V600E within the stage 0 and stage III–IV the 95% CI of the AUC includes the 0.5 value. # Discussion The analysis of cfDNA may have the potential to complement or replace the existing cancer tissue and blood biomarkers in the future. In order to reach this goal, specific and sensitive analytical procedures must be developed and optimized to compute proper circulating target molecules showing differences between patients and healthy subjects. It is now widely accepted that a single biomarker cannot fully distinguish between controls and patients and consequently an approach based on different markers would be preferable in order to achieve a stronger predictive ability. It has been demonstrated that in prenatal screening, a combination of multiple markers, each with limited sensitivity and/or specificity, can lead to a more powerful screening test. Similarly, Schneider and Mizejewski suggest to develop a multi-marker screening approach for cancer diagnosis. Unfortunately this strategy has been proven unsuccessful, notwithstanding the high number of new biomarkers reported in the literature, even if some examples on prostate ovarian and colorectal cancer clearly showed that multi-marker screening can have its place in early cancer detection. The study presented here tests the diagnostic potential of four markers associated to cfDNA in identifying melanoma patients. Particular efforts were dedicated to the technical aspects of the methods adopted for each single parameter allowing to reach accurate and reproducible measurements. We evaluated total cfDNA concentration by a qPCR assay for the single copy gene *APP*, as well as DNA fragmentation represented by the integrity index 180 bp/67 bp. On the other hand, tumour contribution to cfDNA was assessed by quantifying *BRAF^V600E* mutated alleles and *RASSF1A* promoter methylation. These markers have been used in a panel in all patients, thus representing a simple model potentially adoptable by any laboratory. Following the standard approach for the clinical validation of biomarkers for early detection the next step will be focused on the assessment of the impact of these biomarkers on clinical practice including the identification of the most suitable thresholds to use for the early detection of melanoma by clinicians. Our preliminary results show that by jointly considering the panel of biomarkers here investigated the highest predictive capability is given by total cfDNA followed by integrity index 180/67 and methylated *RASSF1A*. According to these results, an approach based on the simultaneous determination of the three biomarkers (total cfDNA, integrity index 180/67 and methylated *RASSF1A*) could be suggested to improve the diagnostic performance in melanoma. Alternatively, as reported in, a more parsimonious sequential approach could be adopted using pre-selection by cfDNA, followed by further selection using integrity index 180/67 and/or methylated *RASSF1A*. We plan to evaluate the prognostic role of both these approaches as soon as the follow-up time of our case study will be adequate (5 years). However preliminary data (not shown), obtained in a subgroup of patients submitted to an additional blood draw 2 weeks after surgery, show a decrease of the four biomarkers, suggesting the potential role of these test as useful tools for monitoring patients after initial diagnosis/surgery. Even though each biomarker investigated in the present work is not exclusively associated with melanoma, their combination reveals a high specificity for melanoma detection. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: CO PP. Performed the experiments: FS. Analyzed the data: PV CMC. Contributed reagents/materials/analysis tools: DM MP. Wrote the paper: PP. Patients enrollment: VDG MG.

# Introduction Epstein-Barr virus (EBV), also referred to as human herpesvirus 4 (HHV-4), is the first identified human virus with documented involvement in carcinogenesis. Many cells of the non-Hodgkin lymphomas (NHL) show the presence of the monoclonal form of EBV genome (i.e. the EBV-positive phenotype). This finding points to the origin of the malignancy from a single infected cell and involvement of EBV in its pathogenesis. A number of hypotheses have been proposed to explain the etiology of NHL in individuals from the general population with inadequately controlled EBV infection. One postulated mechanism is the EBV-induced proliferation of B lymphocytes and a resultant increase in the number of B-cells at risk of oncogenic mutations and clonal expansion. Despite its chronic character by definition, chronic lymphocytic leukemia (CLL) is characterized by marked heterogeneity. Only 30% of patients survive up to 10–20 years after diagnosis. The remaining CLL patients develop terminal phase within 5–10 years, despite mild onset of the disease. The individuals with the aggressive form of CLL survive no more than 2–3 years after diagnosis. The reasons for such heterogeneous natural history of the condition remain unclear. Potential involvement of EBV in the clinical course of CLL is still unexplained. Latent EBV infection is controlled by a cell-mediated immune response in healthy carriers. This immune response is impaired in CLL patients and might result in poor control of reactivation and replication of the virus. Since EBV may activate B cells, stimulate their proliferation, and inhibit their apoptosis, we hypothesized that it could contribute to unfavorable clinical course of CLL and may be one of the reasons for the observed disease heterogeneity. The aim of this study was to define a role of EBV in the etiopathogenesis of CLL. The detailed objectives included the determination of the EBV-DNA copy number in mononuclear cells and isolated B lymphocytes from peripheral blood of CLL patients and healthy individuals and the analysis of association between this parameter and the established prognostic factors, stage of the disease, and its clinical manifestation. # Methods ## Characteristics of CLL patients and healthy volunteers The study included peripheral blood samples from 115 previously untreated patients with CLL (54 women and 61 men). The control group comprised 40 healthy subjects (16 women and 24 men). Neither the CLL patients nor the controls used immunomodulating agents or hormonal preparations, showed signs of infection within at least 3 months prior to the study, underwent blood transfusion, or presented with autoimmune condition or allergy. Moreover, none of the controls had a history of oncological therapy or prior treatment for tuberculosis or other chronic conditions that could be associated with impaired cellular or humoral immunity. The diagnosis of CLL was established on the basis of diagnostic criteria included in the IWCLL guidelines of the American National Cancer Institute (NCI). Detailed characteristics of patients and controls are presented in and. This study was approved by the Ethics Committee of the Medical University of Lublin (decision no. KE-0254/227/2010). Written informed consent was obtained from all patients with respect to the use of their blood for scientific purposes. ## Examined material Peripheral blood (20 mL) from the basilic vein of CLL patients and healthy controls was collected into EDTA-treated tubes (15 mL) and into tubes containing clot activator (5 mL) (aspiration and vacuum systems Sarstedt, Germany). Immediately after collection, the samples were used for immunophenotyping of lymphocytes, isolation of mononuclear cells for the EBV-DNA copy number determination, serum collection for the determination of specific anti-EBV antibodies concentration, and cytogenetic studies. ## Isolation of mononuclear cells and serum Peripheral blood was diluted with 0.9% buffered saline (PBS) without calcium (Ca²⁺) and magnesium (Mg²⁺) (Biochrome AG, Germany) in 1: 1 ratio. The diluted material was built up with 3 mL of Gradisol L (specific gravity 1.077 g/ml; Aqua Medica, Poland), and centrifuged in a density gradient at 700 × g for 20 min. The obtained fraction of peripheral blood mononuclear cells (PBMCs) was collected with Pasteur pipettes and washed twice in PBS without Ca²⁺ and Mg²⁺ for 5 min. Subsequently, the cells were suspended in 1 mL of PBS without Ca²⁺ and Mg²⁺, and either counted in the Neubauer chamber or tested for viability with trypan blue solution (0.4% Trypan Blue Solution, Sigma Aldrich, Germany). Viability below 95% disqualified the cells from further analyses. Serum was obtained from the samples collected into the tubes containing clot activator, aliquoted, and stored at –80°C for enzyme-linked immunosorbent assay (ELISA) test. ## Isolation of DNA and determination of the EBV copy number DNA from 5 × 10⁶ PBMCs was isolated manually with the QIAamp DNA Blood Mini Kit (QIAGEN, Germany). The procedure for isolation followed the manufacturer’s protocol, with a modified volume of DNA elution. Concentration and purity of the isolated DNA were verified with the BioSpec-nano spectrophotometer (Shimadzu, Japan). The EBV-DNA copy number in PBMCs was determined with the ISEX variant of the EBV PCR kit (GeneProof, Czech Republic). Qualitative and quantitative diagnostics of EBV was performed using the Real Time Polymerase Chain Reaction (RT-PCR). Specific conservative DNA sequence of a single-copy gene for the EBV nuclear antigen 1 (EBNA-1) was amplified in the course of the PCR process. The number of viral DNA copies/*μ*L of the eluent was adjusted for the efficiency of the DNA isolation process, and then it was expressed as the viral DNA copy number/*μ*g DNA. All the samples were examined in duplicates. A negative control, i.e. the pure buffer used for DNA elution, was amplified in every case. As the sensitivity of the system amounts to 10 copies/μL, all the samples with the EBV-DNA copy number below this detection threshold were considered EBV-negative \[EBV\]. The PCR was performed with the 7300 Real Time PCR System (Applied Biosystems). The reaction was conducted on MicroAmp® Optical 96-Well Reaction Plates (Life Technologies) with MicroAmp® Optical Adhesive Film (Life Technologies). ## Assessment of activated T and B cells A standard, whole-blood assay with erythrocyte cell lysis was used for preparing the peripheral blood specimens. The cells were phenotypically characterized by incubation (20 min in the dark at room temperature) with a combination of relevant fluorescein isothiocyanate (FITC)-, phycoerythrin (PE)-, and CyChrome- labelled monoclonal antibodies (MoAbs). Immunofluorescence studies were performed using a combination of the following MoAbs: CD3 PE, CD19 PE, CD5 FITC/CD19 PE, CD4 PE, CD8 PE, CD8 FITC/CD4 PE, CD25 CyChrome, and CD69 CyChrome, purchased from BD Biosciences (USA). Finally, cells were washed and analyzed by flow cytometry, performed on a BD FACSCalibur System. Five data parameters were acquired and stored: linear forward and side scatter (FSC, SSC), log FL-1 (FITC), log FL-2 (PE), and log FL-3 (PE-Cy5). For each analysis, 20,000 events were acquired and analyzed using CellQuest Pro software. Isotype-matched antibodies were used to verify the staining specificity and as a guide for setting the markers to delineate positive and negative populations. Mean fluorescence intensity (MFI) and the percentage of cells expressing surface markers were analyzed. ## Analysis of CD38 and ZAP-70 expression in CLL cells CLL cells were stained for CD38 antigen and ZAP-70 protein expression (as described previously by Hus et al.) and analyzed using flow cytometry. The cut- off point for CD38 and ZAP-70 positivity in leukemic cells was ≥30% and ≥20%, respectively. ## I-FISH analysis PBMCs were cultivated for 24 hours in RPMI 1640 medium without mitogen stimulation. After hypotonic treatment and methanol—acetic acid 3: 1 fixation, cell suspensions were dropped onto microscopic slides and used directly for I-FISH. The commercially available Vysis probes (Abbott Molecular Europe, Wiesbaden, Germany) LSI ATM SpectrumOrange/CEP 11 SpectrumGreen Probe, and LSI TP53 SpectrumOrange/CEP 17 SpectrumGreen Probe were used. At least 200 nuclei were analyzed for each probe. The cut-off levels for positive values for normal controls were 2.5% (mean ± SD). ## Anti-EBV immunoassay Commercial enzyme-linked immunosorbent assay (ELISA) kits, purchased from IBL International (Germany) for a quantitative determination of specific anti-EBV antibodies in human serum were used. Protocols followed were in accordance with the manufacturer’s recommendations. The ELISA Reader Victor TM3 (PerkinElmer, USA) was used. ## Statistical analysis Normal distribution of continuous variables was tested using the Shapiro-Wilk test. Statistical characteristics of the continuous variables were presented as medians, minimum and maximum values, as well as arithmetic means and their standard deviations (SD). The Student *t*-test was used for independent variables, and the Mann-Whitney *U*-test was used for intergroup comparisons. The power and direction of relationships between pairs of continuous variables were determined on the basis of the values of Spearman’s coefficient of rank correlation (*R*). The distributions of discrete variables in the studied groups were compared with the Pearson’s Chi-square test or the Fisher’s exact test. The survival curves were constructed with the Kaplan-Meier method, and the proportions of survivors within the studied groups were compared with the log- rank test. Univariable and multivariable Cox proportional hazard regression models were used to determined association between patients characteristics and time to first treatment. For the multivariable model, stepwise variable selection was used. Significant variables (p\<0.1) were tested for inclusion in the regression models, and nonsignificant variables were removed sequentially until only those significant at p\<0.05 remained. Receiver operating characteristic (ROC) curves were generated for significant predictor variables of EBV(+) CLL patients. Areas under the ROC curves (AUCs) were calculated for each parameter and compared. All the calculations were carried out with Statistica 10 (StatSoft®, USA) package, with the level of significance set at *P* \< 0.05. # Results The PBMCs of 115 CLL patients and 40 controls were tested for the presence of EBV-DNA with an aid of real-time PCR. A total of 62 (53.91%) CLL patients presented with a large number of EBV-DNA copies (more than 10 copies/*μ*L). The following three groups were identified on the basis of this criterion: (a) CLL patients whose PBMCs showed the presence of the EBV-DNA, i.e. the EBV(+) group, (b) CLL patients whose PBMCs lacked the EBV-DNA, i.e. the EBV group and (c) healthy controls. In univariable analysis, among clinical and biochemical parameters of CLL patients both traditional and new prognostic factors associated with shorter time to first treatment turn out to be statistically significant. We found that categorical variables such as the splenomegaly, trisomy 12, CD38, and positive anti-EBV EA IgA were associated with time to first treatment. Among continuous variables lower hemoglobin, higher beta-2 microglobulin and LDH, higher CD19+CD25+ cells, CD19+CD69+ cells, CD3+CD69+ cells, CD19+CD5+CD38+ cells and CD19+CD5+CD23+ cells had significant association with time to treatment. Higher level of anti-EBV: EA IgG, EBNA IgG and VCA IgG were associated with shorter time to first treatment. Interestingly, EBV(+) patients had more than 7 times shorter time to first treatment comparing to EBV patients. This phenomenon was confirmed during multivariable analysis. The multivariable Cox analysis identified the higher beta-2-microglobulin (HR = 1.43; p = 0.0033), enlarged spleen (HR = 4.70; p\<0.0001), EBV presence (HR = 23.39; p\<0.0001) and lower anti-EBV EBNA-1 IgM level (HR = 0.98; p = 0.0089) as independent predictors for shorter time to first treatment. Since the Cox proportional hazard regression model showed highly statistically significant association of EBV(+) with progression of illness, a ROC curve was drawn to test the variables predictive validity in EBV(+) CLL patients. As the area under the curve shows , CD19+CD25+ cells \[%\] in the peripheral blood parameter was the most sensitive and specific to determine EBV(+) (AUC = 0.854). presents a Kaplan-Meier curve illustrating the time to first treatment depending on the EBV-DNA copy number/*μ*g DNA isolated from PBMCs and a Kaplan-Meier curve illustrating the probability of lymphocyte doubling-free survival depending on the EBV-DNA copy number/*μ*g DNA isolated from PBMCs. presents the comparison between clinical and laboratory parameters in CLL EBV(+) patients, CLL EBV patients and the study group. presents an assessment of anti-EBV antibody concentrations in relation to the presence or absence of EBV-DNA copies in PBMCs of CLL patients and control group. presents statistically significant correlations between activated T CD3+ and B CD19+ cells in CLL patients. # Discussion A total of 20 CLL patients had more than 1000 EBV-DNA copies/*μ*g of PBMC DNA; the EBV-DNA copy number of the remaining patients ranged between 100 and 1000 (*N* = 22) or between 10 and 100 (*N* = 20). These results are similar to the data reported by Kimura et al., according to whom the EBV-DNA copy number per *μ*g DNA extracted from PBMCs of patients with EBV-related lymphoproliferative disorders and infectious mononucleosis ranges between 10 and 10000. In contrast, the EBV-DNA copy number in most healthy controls and immunocompromised patients after organ transplantation did not exceed 10 copies/*μ*g DNA. Both Kimura et al. and Stevens et al. postulated that more than 100–1000 EBV-DNA copies/*μ*g DNA isolated from examined material are associated with clinical signs of EBV infection, manifesting as the chronic active EBV disease (CAEBV) or post- transplant lymphoproliferative disease. These findings and the results of our study point to a potential presence of a CLL subtype being associated with EBV infection. Moreover, an increase in the EBV-DNA copy number was documented in most of our patients during approximately 2-year follow-up. We revealed the presence of EBV-DNA in PBMCs and isolated B lymphocytes in more than a half of our CLL patients. To the best of our knowledge, no previous studies distinguished between the CLL forms being associated with EBV infection or unrelated to this virus. The EBV-associated form of CLL seems to be characterized by more aggressive phenotype. We showed that the EBV-DNA copy number in PBMCs of patients with hepatomegaly or thrombocytopenia and individuals who required earlier implementation of treatment was significantly higher than that in the remaining individuals. A number of previous studies documented the role of EBV in induction of thrombocytopenia. While the presence of EBV in patients with infectious mononucleosis is usually associated with a slight decrease in platelet count, in the case of CAEBV, it can be associated with severe thrombocytopenia, anemia (usually of autoimmune origin), and splenomegaly (resulting from lymphocyte infiltration) or even liver failure. Moreover, we showed that the EBV-DNA copy number correlated significantly with serum concentrations of beta-2-microglobulin and LDH. As early as 1981, Ibsen et al. revealed that the level of beta-2-microglobulin is at its highest during initial stages of infectious mononucleosis, and subsequently, within 3 weeks to 3 months after recovery, it normalizes to its baseline level. The fact that concentration of beta-2-microglobulin constitutes an established predictive factor in CLL patients may suggest that the elevated level of this protein is associated with EBV infection in at least some of the cases. Furthermore, we revealed significant associations between other negative prognostic factors such as high cytoplasmic expression of ZAP-70, surface expression of CD38 in leukemic cells, surface expression of CD23, CD25, and CD69, as well as unfavorable genetic mutations, and EBV-DNA copy number. Tsimberidou et al. reported that 38% of CLL patients had evidence of EBV infection by in situ hybridization for EBV EBER1, a small noncoding RNA species. Tarrand et al. reported that LMP1 mRNA levels were higher in CLL patients than in healthy subjects (14% vs. 1% of healthy controls), suggesting that EBV late gene expression occurs at least in a subset of CLL cells. We demonstrate significant associations between viral load of EBV-DNA and various clinical and pathologic variables among CLL patients, including associations with progression and time to treatment. These findings are in line with conclusions made by Visco et al. who postulated that EBV-DNA load at presentation is an independent predictor of overall survival in patients with CLL. # Conclusions In conclusion, more than a half of CLL patients presented with CLL EBV-DNA in their PBMCs, whereas no detectable amounts of genetic material for this pathogen were found in healthy controls. Greater EBV-DNA copy number was associated with shorter overall survival and time to progression in CLL patients. Positive correlation between EBV-DNA copy number and established unfavorable prognostic factors of CLL implies that increased EBV load in peripheral blood may predict poor clinical outcomes of CLL. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JR EG MP IKG MM AS AG ASG SG MZ AM PG ES. Performed the experiments: JR EG MP IKG MM AS AG ASG SG MZ AM PG ES. Analyzed the data: JR EG MP IKG. Contributed reagents/materials/analysis tools: JR EG MP SG. Wrote the paper: JR EG.

# 1. Introduction Temporal lobe epilepsies (TLEs) are classified into three types based on the region of occurrence: the limbic system, mesial TLE (MTLE); neocortex, lateral TLE; and mixed TLE. MTLE involves parts of the limbic system (such as the amygdala and hippocampus) and is thought to be an independent epileptic syndrome associated with unique epileptic seizures and different treatment outcomes. In 1954, Penfield proposed that MTLE is part of a disease group caused by hippocampal sclerosis (HS). The epileptogenic region in MTLE is located within the limbic system (the hippocampus, amygdaloid body, uncus, and parahippocampal gyrus) of the temporal lobe, whereas that for lateral TLE is located on the lateral surface of the temporal lobe neocortex. Patients with MTLE experience some distinct symptoms, such as cacosmia, flashback, uneasiness, a sense of fear, and compound visual and auditory hallucinations, perceived as aura. After aural sensation, symptoms of epileptic seizure, such as staring, arrest of movement, and various types of automatisms (such as fumbling for clothes, cheek-biting, and fidgeting), may occur. Lightheadedness persists for 5 to 10 minutes after attack symptoms. The signs of an attack in lateral TLE are acousma (simple sounds such as ringing, but not words) and dizziness. The most common cause of TLE is HS. HS is diagnosed based on the presence of unilateral hippocampal atrophy and hyperintensities on the T2-weighted magnetic resonance imaging (MRI) or fluid-attenuated inversion recovery (FLAIR) images. In addition, HS is rarely observed in both hippocampi. Seizure control with antiepileptic drug treatment is possible in 70%–80% of patients; however, 20%–30% of patients have drug-resistant intractable epilepsy. Therefore, approximately 10% of all epileptic patients require surgical treatment. The surgical treatment of TLE results in an 80% reduction in epileptic seizures; however, surgical treatment is less effective against other types of epilepsy. Surgical methods such as anterior temporal lobectomy (ATL) and selective amygdalohippocampectomy are used for treating TLE. Moreover, the surgical resection of mesial structures (hippocampus and amygdala) was first described in the 1950s. Although removal of the lateral temporal cortex was performed during earlier times, the selective removal of the epileptogenic lesion has only recently become mainstream. In ATL, approximately 35–45 mm of tissue from the tip of the temporal lobe cortex toward the inferior horn of the lateral ventricle is removed, and thereafter, the hippocampus and amygdala are resected. ATL provides a relatively wide surgical field for observing the structures. However, this method has certain limitations. Complications such as memory disturbances and upper-quarter homonymous hemianopsia commonly occur after ATL. Moreover, resection on the dominant side may lead to the impairment of the language center located at the tip of the temporal lobe. Hence, neurosurgeons often hesitate to further resect the lateral regions of the temporal lobe cortex. Furthermore, diagnosing HS is crucial as the hippocampus is important for cognitive functions such as memory and spatial discrimination. In addition, diagnosis of HS can be difficult, even for neurologists or neurosurgeons. Neuropathological conditions of HS are mainly characterized by hippocampal neuronal loss. In 2013, the International League Against Epilepsy proposed a classification system for identification of pathological subtypes of HS based on the patterns of neuronal loss. Atrophy and/or high signal intensity of hippocampi on T2-weighted imaging (T2WI) and FLAIR are the most reliable findings of HS. These findings reflect the neuropathological features of HS. In Japan, diagnosis based on neuroimaging is usually performed by neurosurgeons. The 2018 report of Labor Standards Bureau, Ministry of Health, Labor and Welfare, revealed that the number of radiologists in Japan was 6,813, whereas the number of doctors in Japan was 311,963. Therefore, Japanese doctors should be able to reach a diagnosis using neuroimages by themselves. In this study, we present an alternative diagnosis method using artificial intelligence (AI), which involves automatic detection of the hippocampus and diagnosis of MTLE with the hippocampus as the epileptogenic area. Here, we demonstrated that the deep learning-based AI program successfully detects the hippocampus and therefore, epileptic attacks because of MTLE with the hippocampus as the epileptogenic area based on MRI, with an accuracy of \> 90%. # 2. Methods ## 2.1. Patients The human and animal studies were approved by the Ethics Committee of Sapporo Medical University Hospital. The study was conducted in accordance with the ethical principles of the Declaration of Helsinki (1964) and its later amendments. The need for consent was waived by the ethics committee. This study included all consecutive patients diagnosed with epilepsy at our hospital between January 2016 and December 2019 who had not been previously treated. A total of 259 patients (128 men and 131 women) were enrolled and analyzed retrospectively. Forty-six patients were diagnosed with MTLE, whereas 12 patients were diagnosed with MTLE with the hippocampus as the epileptogenic area. Patients diagnosed with MTLE with the hippocampus as the epileptogenic area were selected because they had hippocampal atrophy, hyperintensities on FLAIR MRI and spikes in the mesial temporal lobe on intracranial electroencephalography (EEG). Patients who were found to be seizure-free in a presurgical evaluation with subdural electrodes underwent resection surgeries at the Sapporo Medical University. Two or three continuous coronal sections of T2WI at the level of the anterior commissure from patients with MTLE and those with other diseases such as psychogenic nonepileptic seizures (control subjects) were prepared for AI analysis. MR images were allocated to two datasets: learning and test datasets. The learning data were used for deep learning, whereas the test data were used for analyzing the diagnostic accuracy (i.e., validation). The test data included 30% of all data. MRI data of patients with MTLE with the hippocampus as the epileptogenic area included four images on T2WI (25.0%) and six images on FLAIR (21.4%). MRI data of control subjects included 77 images on T2WI (36.5%) and 77 images on FLAIR (27.7%). After the deep learning process, the diagnostic accuracy was calculated using the test data from T2-weighted images and FLAIR images. All data were fully anonymized before we accessed them. ## 2.2. MRI examination The basis for pre-processing was T2WI of the head using a clinical 1.5- or 3.0-Tesla magnetic resonance scanner (Signa HDxt 3.0 Tversion16®, GE Healthcare, Connecticut, USA). The imaging parameters for the T2-weighted fast-spin echo imaging were as follows: flip angle, 90°; repetition time, 5000 ms; echo time, 102.0 ms; bandwidth, 50.0 kHz; field of view, 200 mm × 200 mm; scan thickness, 4.0 mm; slice gap, 1.0 mm; number of slices, 26–30; matrix, 352 × 256; number of signals averaged, 1; and imaging time, 1 min 05 s. The imaging parameters of FLAIR imaging were as follows: flip angle, 90°; time of repetition, 5000 ms; echo time, 102.0 ms; bandwidth, 50.0 kHz; field of view, 200 mm × 200 mm; scan thickness, 4.0 mm; slice gap, 1.0 mm; number of slices, 26–30; matrix, 352 × 256; number of signals averaged, 1, and imaging time = 1 min 05 s. ## 2.3. Data processing We used an HP Z840 workstation (Hewlett-Packard Company, Palo Alto, California, USA) with a Core Xe-6700 K 4.00 GHz (Intel, Santa Clara, CA, USA) central processing unit, 64 GB of random access memory, and a GeForce GTX 1080 (NVIDIA, Santa Clara, CA, USA) graphics processing unit for the training phase of deep learning. The image sets were processed using a code written in Python 3.7.4 ([http://www.python.org](http://www.python.org/)) and Pillow 3.3.1 (<http://pypi.python.org/pypi/Pillow/3.3.1>), which is a python imaging library. We used the Open Computer Vision (OpenCV) 4.5.1 (<https://docs.opencv.org/4.5.1/>) and Keras (version 2.3.0), a framework for neural networks and a part of the Tensorflow (version 1.14.0) platform. OpenCV is a library of programming functions that is free for use under an open-source BSD license (OpenCV. Open Source Computer Vision Library. 2015.). OpenCV was used for hippocampal detection. MRI images were obtained of 113 T2WI and 148 FLAIR cases with a diagnosis of epilepsy from January 2016 to December 2019. On both the left and right sides, 226 T2WI and 296 FLAIR images with a resolution of 96 × 96 pixels were used for hippocampal images, and 327 T2WI and 367 FLAIR images with a resolution of 96 × 96 pixels were used for non-hippocampal images after amplification using augmentation. Haar-like features and local binary patterns (LBPs) were used as features. Hyperparameters of the created cascade classifier, mainly scaleFactor, minNeighbors, and minSize, were adjusted for optimization. Image processing was performed separately for the training and test image sets to prevent overlearning during the training phase. Data augmentation was performed for the training data sets; thus, the number of images from patients with MTLE with the hippocampus as the epileptogenic area and from control subjects increased from 12 to 1134 and 134 to 7344 for T2WI, and 22 to 1975 and 201 to 8635 for FLAIR, respectively. In addition, data augmentation was performed for the validation data sets; thus, the number of images from patients with MTLE with the hippocampus as the epileptogenic area and from control subjects increased from 4 to 392 and 77 to 5359 for T2WI, and 6 to 580 and 77 to 5369 for the FLAIR images, respectively. After data augmentation, the total number of images in the training dataset for T2WI was 8478, and for FLAIR was 10610, and that in the validation dataset for T2WI was 5751 and for FLAIR was 5949 (Tables). The image data generator created varied images using rotation, width and height shift, horizontal and vertical flip, zoom, and shear. This type of data augmentation is commonly used in deep learning (the code is available at <https://keras.io/>) for small datasets such as those of medical diseases. The authors selected the VGG16 model because it has been trained on big data and its weights from the pre-training can be used by end-users from the internet to analyze data. The output data were compared with the initial teacher data (two categories–MTLE with the hippocampus as the epileptogenic area and controls). The deep learning model, which was a fine-tuned VGG16 model, comprised several layers (six convolutional layers, three maximum pooling layers, and three fully connected layers). Details of the neural network used in the model are shown in. The number of epochs used for each study was 20. We used SGD as an optimizer for the deep learning model. ## 2.4. Visualization of deep learning Explanation of the output of a deep network remains a challenge. The classification of images with AI is presented in the black box. Convolutional neural networks include multiple layers. An image classifier identifies pixels that have significant influence on the decision making. Therefore, we attempted to visualize the image processing using deep learning neural networks and determine the mechanisms used to classify or diagnose based on an image in the validation dataset. Here, we focused on the last layer to visualize the area of interest used by AI. One such visualization method for deep learning is the VGG16-based gradient-weighted class activation mapping (Grad-CAM). We used the Grad-CAM method to visualize the areas of interest to distinguish between MTLE and control using the AI program. ## 2.5. Diagnosis by six board-certified neurosurgeons The analysis and diagnosis of MTLE based on T2 and FLAIR were performed by six independent investigators who were board-certified neurosurgeons. HS was diagnosed based on the presence of following characteristics: hippocampal atrophy, hippocampal hyperintensity on T2WI, and other hippocampal signal alterations such as loss of internal architecture of the hippocampus. Three epilepsy specialists independently confirmed the diagnosis of MTLE with the hippocampus as the epileptogenic area based on MRI and EEG data. ### 2.5.1. Selected images We selected 20 MRI images for T2WI and FLAIR (10 images of patients diagnosed with MTLE with the hippocampus as the epileptogenic area and 10 patients diagnosed with other epilepsies). These 20 images represented the data of the group. ### 2.5.2. Patients diagnosed with MTLE Retrospectively, 46 patients were diagnosed with MTLE at our institution. However, we excluded cases with no imaging, cases in which surgery was performed, and cases for which no ECoG recording with subdural electrodes existed. Finally, T2WI was validated in 28 cases and FLAIR in 34 cases. ## 2.6. Statistical analysis Data are presented as mean ± standard deviation or standard error. Statistical analysis was performed using IBM SPSS 22.0 (IBM Corp. Armonk, NY, USA). The Shapiro–Wilk test was used to confirm whether the data followed a normal distribution.When the data distribution was not normal, statistical differences were assessed using the Mann-Whitney U test. We also performed a cross tabulation. A p value of \<0.01 was considered statistically significant. # 3. Results ## 3.1. Stratification of the patients This study enrolled 12 patients who were diagnosed with MTLE with the hippocampus as the epileptogenic area based on MRI and EEGs by three epilepsy specialists. To assess the diagnostic accuracy, the patients were divided into two groups: those with MTLE with the hippocampus as the epileptogenic area and control subjects. Imaging data of both the groups were further divided into the learning and validation datasets. The validation dataset comprised approximately 30% the data of both the groups (patients with MTLE with the hippocampus as the epileptogenic area and control subjects). The validation dataset included data of randomly selected patients. Patients’ data in the validation and learning datasets were statistically matched for age and sex. Bilateral hippocampi were automatically and successfully detected by the AI program, with an accuracy of 96%–99% using T2WI and of 89% using FLAIR images. Haar-like feature was selected because it is more accurate than LBP when the hyperparameters are changed to optimize the detection rate and reduce false positives. FLAIR had a 89% detection rate and 0.25 false positives per picture. While checking the detection rate and false positive rate, we examined the effect of changing mainly the following parameters as optimization factors: scaleFactor, minNeighbors, minSize, and maxSize. For the T2WI classifier, scaleFactor was set to 1.06, minNeighbors to 3, minSize to (60,60), and maxSize to (130,130). Similarly, for FLAIR, we set the scaleFactor to 1.1202, minNeighbors to 3, minSize to (60,60), and maxSize to (130,130). We compared the diagnoses of MTLE with the hippocampus as the epileptogenic area between six board-certified neurosurgeons and AI for each MRI sequence (T2WI and FLAIR). The diagnostic accuracy of AI using T2WI data was 94% and using FLAIR data was 95%. When the diagnostic accuracy was measured using only the original images as the test data, the diagnostic accuracy using T2WI data was 98%, and that using FLAIR data was 95%. However, this result is considered to be insufficiently evaluated because of the small number of original images. We performed a cross-validation analysis, and the system showed 90–93% accuracy for the T2WI data and 89–95% accuracy for the FLAIR data. For MTLE with the hippocampus as the epileptogenic area, 3 out of 9 T2 cases and 10 FLAIR cases were selected as validation data. Verification was performed five times by replacing the data. Allocation was done using a random number table. The relatively high accuracy from the beginning of the epoch may be due firstly to the fact that VGG16 was a very good model for MTLE differentiation, and secondly to the fact that the image size was small (96 × 96 pixels) and the accuracy was high. However, considering that diagnostic accuracy increased with each successive epoch, we believe that there was a learning effect. ## 3.2. Selected images The disease (MTLE with the hippocampus as the epileptogenic area) probability in the extracted hippocampus was found to be 92.5% based on T2WI and 82.5% based on FLAIR using the AI program. On the other hand, the disease (MTLE with the hippocampus as the epileptogenic area) probability was 74.1% based on T2WI and 74.2% based on FLAIR according to the six board-certified neurosurgeons. Those judged to have a 50% or greater likelihood of hippocampal sclerosis, according to the AI diagnosis, were treated as correct answers. Those that were not hippocampal sclerosis were treated as correct if the likelihood of hippocampal sclerosis was judged to be less than 50%. Interestingly, diagnosis with AI was good in all, except one, cases; one of the 20 images showed false positives on T2WI and seven of the 20 images on FLAIR. Diagnoses by the board-certified neurosurgeons based on T2 and FLAIR were not different. The images for which MTLE with the hippocampus as the epileptogenic area was misdiagnosed as normal, or *vice-versa*, always showed a tendency towards hyperintensity on T2WI. Diagnoses by AI were significantly more accurate than those by board-certified neurosurgeons based on T2WI images (p = 0.0001, effect size 0.6152); however, the difference in diagnostic accuracies of AI and neurosurgeons based on FLAIR was not statistically significant (p = 0.0520, effect size 0.3072). However, the sensitivity, specificity, and F-value of AI for T2WI were 0.8000, 0.9667, and 0.8421, respectively, and those of physicians for T2WI were 0.5000, 0.5667, and 0.5172, respectively. Similarly, for FLAIR, the sensitivity, specificity, and F-value for the diagnosis by AI were 0.3077, 1.0000, and 0.4706, respectively, whereas those by physicians were 0.5833, 0.5833, and 0.5833, respectively. Although the sensitivity, specificity, and F-value of AI for T2WI were high, the sensitivity for FLAIR was low. ## 3.3. Patients diagnosed with MTLE The Shapiro–Wilk test showed that the data did not follow a normal distribution (p\<0.001); therefore, Mann–Whitney U test was performed. The diagnostic accuracy for MTLE by AI was higher than that by physicians for both T2WI and FLAIR (T2, p = 0.0034; effect size 0.3917, FLAIR, p = 0.0006; effect size, 0.4147). AI diagnosis was superior in the cross tabulation (p\<0.001) as well. However, the sensitivity, specificity, and F-value of AI for T2WI were 0.5556, 0.9362, and 0.5582, respectively, and those of physicians for T2WI were 0.5000, 0.6754, and 0.4576, respectively. Similarly, in FLAIR, the sensitivity, specificity, and F-value for the diagnosis by AI were 0.4, 1.0000, and 0.5714, respectively, whereas those by physicians were 0.6833, 0.6181, and 0.5256, respectively. Although the F-value for AI was high, the sensitivity was low. ## 3.4. Diagnostic site of MTLE with the hippocampus as the epileptogenic area AI focused on the color gradient of the area of interest for diagnosing the epileptogenic area in the hippocampus. The focus was expected to be on the outer area as well as on the surface of the hippocampus based on the usual pathology. # 4. Discussion The occurrence of MTLE with HS is rare. It originates from the limbic system in the mesial temporal lobe, particularly the hippocampus, amygdala, and parahippocampal gyrus, as well as their connections. HS is characterized pathologically by prominent neuronal loss and gliosis in the hippocampus and amygdala. It is crucial to determine whether the seizure originates from the medial, lateral, or multifocal epileptogenic points because the diagnosis is difficult and therapeutic methods for each condition differ. The curative effect of the surgical treatment is well-established because the attack resolution rate after medical therapy is only 8% per year, whereas that after surgical treatment is 58% in the case of intractable TLE. In addition, it is necessary to localize the epileptic focus preoperatively in several cases to prevent unnecessary brain resection. After the detailed localization of an epileptogenic focus through cortical mapping, certain invasive examinations such as cortical mapping with subdural electrodes may be required before moving on to resection as a second- stage surgery. However, it would be remarkable if a noninvasive examination could guide the diagnosis and localization of an epileptogenic focus. If HS can be diagnosed preoperatively using noninvasive modalities such as MRI, a temporal lobectomy can be performed in a single stage. Jack et al. (Mayo Clinic, 1990) noninvasively showed hippocampal atrophy and sclerotic change in an MRI coronal section; thereafter, assessing etiology, clinical course characteristics, EEG data, convalescence, surgical outcomes, and prognosis has become easy. In 1993, Wieser indicated that MTLE is equivalent to an epileptic syndrome or as a disease unit. Although MTLE is intractable, even with treatment using multiple anticonvulsant agents, it can be well-distinguished from their epileptic syndromes such as lateral TLE since they show surgical outcomes. Therefore, differential diagnosis is very important. The epidemiological cause of MTLE is HS, accounting for approximately two-thirds of all cases. Other causes include neoplastic lesions (such as ganglioglioma, glioma, and dysembryoplastic neuroepithelial tumor) and vascular lesions (such as arteriovenous malformations, cavernous angiomas or venous angiomas, and cortical dysplasia). Furthermore, MTLE associated with HS predominantly occurs among individuals aged is 4–16 years. MTLE due to HS generally occurs earlier than MTLE due to other causes. MTLE is often associated with anamneses such as febrile convulsion (particularly complicated), brain hypoxia, an infectious disease, and/or head injury. Hippocampal dysfunction initiates epileptogenic seizures and impairs cognition in patients with MTLE. Recently, the use of AI in the medical care field has increased at a tremendous pace. For example, a patient was diagnosed with leukemia using the AI program within only 10 minutes after learning from more than 20 million cancer research articles, which proved lifesaving (Watson, IBM co.). Several studies have presented a machine learning-based method for identifying the epileptic seizure onset zone. The reason for the improvement is technical skillfulness and precise evaluation of the epileptogenic zone in the brain alongside functional mapping. Recently, many scholars have been engaged in hippocampus segmentation algorithms. A number of methods have been proposed, including cluster methods, using map and feature embedding models or iterative local linear mapping models, atlas-based and label fusion methods, statistical machine learning techniques combined with KNN and SVM models, probabilistic modeling frame methods, nonlinear image registration algorithms, energy minimization models, using local Gaussian distribution fitting energy with level set function and local mean and variance as variables, and bias-corrected distance regularized level set method with MR image contrast enhancement. Based on the aforementioned example, AI is suitable for processing a large amount of medical data (such as diagnostic imaging data), comparing case data with previous research data, and organizing data. Moreover, AI is a developing technology in the medical field. Previous data on rare diseases such as neurosurgical diseases are limited; hence, the accuracy may not be reliable because AI diagnosis requires a large amount of data. Here, we demonstrated an example of a diagnostic imaging technique for epilepsy. As shown in, AI focuses on the CA1 and CA2 regions of the hippocampus. In hippocampal sclerosis, pathological findings include not only neuronal atrophy but also neuronal loss and gliosis. Although not beyond the realm of conjecture, it is possible that AI can identify tiny differences in MRI signal values that cannot be corroborated by the physician’s naked eye. We believe that the results of this study are helpful to specialists as well as non-specialists in community medicine for aiding in the diagnosis of epilepsy. # 5. Limitations Recently, deep learning technology has advanced rapidly; however, its ability to simultaneously diagnose several cases is suboptimal. In addition, the possibility of misdiagnosis because of bugs or malfunctioning of the software is inevitable. AI in medical care remains on the periphery of diagnostic science. Further research is required to learn and push the boundaries of the field. # 6. Conclusion We present a highly accurate deep learning-based AI program that can diagnose MTLE with the hippocampus as the epileptogenic area better than some board- certified neurosurgeons. Almost all AI programs that are big hit worldwide are not all-purpose and perform limited functions such as extract and reproduce human specific data on the computer. However, AI is a powerful tool and can sometimes overcome the biases shown by humans. In addition, AI has the advantage of generating stable results. In the medical field, doctors have harsh working environments, which include overwork due to lack of personnel, frequent calls at night, and insufficient holidays; this may account for false judgments sometimes. On the contrary, AI can maintain a certain level of output accuracy and be a reliable assistant to doctors. 10.1371/journal.pone.0282082.r001 Decision Letter 0 Rajamanickam Yuvaraj Academic Editor 2023 Yuvaraj Rajamanickam This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 24 May 2022 PONE-D-22-00437Deep Learning for the Diagnosis of Mesial Temporal Lobe EpilepsyPLOS ONE Dear Dr. Sakashita, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Jul 08 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Yuvaraj Rajamanickam, Ph.D Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf>. 2\. Please provide additional details regarding participant consent. In the ethics statement in the Methods and online submission information, please ensure that you have specified (1) whether consent was informed and (2) what type you obtained (for instance, written or verbal, and if verbal, how it was documented and witnessed). If your study included minors, state whether you obtained consent from parents or guardians. If the need for consent was waived by the ethics committee, please include this information. If you are reporting a retrospective study of medical records or archived samples, please ensure that you have discussed whether all data were fully anonymized before you accessed them and/or whether the IRB or ethics committee waived the requirement for informed consent. If patients provided informed written consent to have data from their medical records used in research, please include this information Once you have amended this/these statement(s) in the Methods section of the manuscript, please add the same text to the “Ethics Statement” field of the submission form (via “Edit Submission”). For additional information about PLOS ONE ethical requirements for human subjects research, please refer to <http://journals.plos.org/plosone/s/submission-guidelines#loc-human-subjects- research>. 3\. We note that the grant information you provided in the ‘Funding Information’ and ‘Financial Disclosure’ sections do not match. When you resubmit, please ensure that you provide the correct grant numbers for the awards you received for your study in the ‘Funding Information’ section. 4\. PLOS requires an ORCID iD for the corresponding author in Editorial Manager on papers submitted after December 6th, 2016. Please ensure that you have an ORCID iD and that it is validated in Editorial Manager. To do this, go to ‘Update my Information’ (in the upper left-hand corner of the main menu), and click on the Fetch/Validate link next to the ORCID field. This will take you to the ORCID site and allow you to create a new iD or authenticate a pre-existing iD in Editorial Manager. Please see the following video for instructions on linking an ORCID iD to your Editorial Manager account: <https://www.youtube.com/watch?v=_xcclfuvtxQ> \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: No Reviewer \#2: Partly \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Firstly, I take this opportunity to congratulate the authors on their successful submission of their paper for publication. Initially, I felt the paper can be recommended for publication with reasonable modifications. However, once I saw Fig. 3, I can no longer recommend the paper since there is something fundamentally wrong while training the deep learning model. The validation accuracy does not change even after training the model. This infers either the validation accuracy depicted in Fig. 3 is erroneous or the VGG16 model can detect MTLE even without re-training the model. Kindly rectify this issue since this is the main result of the paper. Comments: • The patient flow chart is well described. However, the number of patients in the training/validation set needs to be specified separately. Also, please indicate the number of images/augmented images from each patient. As of now, the data information in lines 114, 125, 157, 161, etc is confusing. • Please specify how the training and validation data was selected. Was the dataset split at the patient-level or the image-level. The images from the same patient should not be included in both the training and validation set. • I also recommend performing a cross-validation to verify the robustness of the methods prescribed in the manuscript. • I am also concerned why 20 patients were specifically chosen to compare between the AI system and the clinicians. Ideally, the comparison should be performed for all the 46 MTLE patients. Please do not use a ‘convenient’ sample for comparisons. • I also have concerns regarding the statistical testing: o T-tests are performed if the data distribution is ‘normal’. I am not sure if the data points fall into a normal distribution as the sample size is too low. Please perform a ‘normality’ testing before applying a t-test. Alternatively, you need to apply Man-Whitney U test or Wilcoxon rank sum test. o Kindly report the effect sizes along with p-values. o In Table 3, for the first 10 patients, the ground truth is ‘1’ and for 11-20 patients the ground truth is ‘0’. Therefore, while performing the test, you will end up with two p-values (one for 1-10 patients and one for 11-20). Since you have reported a single p-value please specify how this was performed. • Data augmentation is widely used while training the deep learning models. You should not apply them while computing the validation results as this will skew the accuracy since a single data point (patient) is considered as multiple observations. The validation/test results should be reported considering each patient as a single data point. • Finally, if possible try to explain why the AI system was capable of diagnosing better than clinicians using Fig. 4. In other words, what did the AI system detect in the images that were missed by the clinicians. Good luck. Reviewer \#2: 1. The authors failed to compare their segmentation results with hippocampal segmentation papers in the literature 2\. The authors didn’t validate their segmentation accuracy through the deep learning cross validation methods 3\. Authors didn’t explain about the architecture of deep learning in the manuscript 4\. Authors didn’t compare their results with other segmentation algorithms or standard tool box 5\. Authors didn’t explain the challenges of segmentation of hippocampus through deep learning algorithm 6\. Authors gas to validate their results through other metrics like sensitivity, specificity, f1 score etc to verify the performance of algorithm 7\. Authors has to explain the optimization of parameters of deep learning model 8\. How MRI out performs the less cost EEG data in epilepsy diagnosis 9\. Why T2 and FLAIR images are used instead of T1 weighted images? 10\. Why there is huge difference between diagnostic accuracy of AI and neurosurgeon in table 3? Whether results can be validated with multiple surgeons results? \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: **Yes: **Jac Fredo Agastinose Ronickom \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0282082.r002 Author response to Decision Letter 0 1 Aug 2022 Reviewer \#1: • Initially, I felt the paper can be recommended for publication with reasonable modifications. However, once I saw Fig. 3, I can no longer recommend the paper since there is something fundamentally wrong while training the deep learning model. The validation accuracy does not change even after training the model. This infers either the validation accuracy depicted in Fig. 3 is erroneous or the VGG16 model can detect MTLE even without re-training the model. Kindly rectify this issue since this is the main result of the paper. RESPONSE: Thank you for your valuable comments. We completely agree with reviewer \#1. The following section has been amended in the manuscript (p.12, line 276-281): The relatively high accuracy from the beginning of the epoch may be due firstly to the fact that VGG16 was a very good model for MTLE differentiation, and secondly to the fact that the image size was small (96 × 96 pixel) and the accuracy was high. However, considering that diagnostic accuracy increased with each successive epoch, we believe that there was a learning effect. • The patient flow chart is well described. However, the number of patients in the training/validation set needs to be specified separately. Also, please indicate the number of images/augmented images from each patient. As of now, the data information in lines 114, 125, 157, 161, etc is confusing. RESPONSE: Thank you very much for your valuable suggestion. We have added a table (Table 1) and organized the data. (p.7, line 177) • Please specify how the training and validation data was selected. Was the dataset split at the patient-level or the image-level. The images from the same patient should not be included in both the training and validation set. RESPONSE: Thank you for your valuable question. We have now described how we split the training and validation datasets. • I also recommend performing a cross-validation to verify the robustness of the methods prescribed in the manuscript. RESPONSE: Thank you for your valuable comment. We have now performed a cross-validation analysis, and the system showed a 90–93% accuracy for T2WI and an 89–95% accuracy for FLAIR. We have included this information in line 287 (p.12) of the text. • I am also concerned why 20 patients were specifically chosen to compare between the AI system and the clinicians. Ideally, the comparison should be performed for all the 46 MTLE patients. Please do not use a ‘convenient’ sample for comparisons. RESPONSE: Thank you for your comment. We have amended the text regarding this issue as follows (p.10, line 235): Retrospectively, 46 patients were diagnosed with MTLE at our institution. However, we excluded cases with no imaging, cases in which surgery was performed, and cases for which no ECoG recording with subdural electrodes existed. Finally, T2WI was validated in 28 cases and FLAIR in 34 cases. And as follows (p.13, line 316): The diagnostic accuracy for MTLE by AI was higher than that by physicians for both T2WI and FLAIR (T2, p = 0.0034; effect size 0.3917, FLAIR, p = 0.0006; effect size, 0.4147). However, the sensitivity, specificity, and F-value of AI for T2WI were 0.5556, 0.9362, and 0.5582, respectively, and those of physicians for T2WI were 0.5000, 0.6754, and 0.4576, respectively. Similarly, in FLAIR, the sensitivity, specificity, and F-value for the diagnosis by AI were 0.4, 1.0000, and 0.5714, respectively, whereas those by physicians were 0.6833, 0.6181, and 0.5256, respectively. Although the F-value for AI was high, the sensitivity was low. • I also have concerns regarding the statistical testing: o T-tests are performed if the data distribution is ‘normal’. I am not sure if the data points fall into a normal distribution as the sample size is too low. Please perform a ‘normality’ testing before applying a t-test. Alternatively, you need to apply Man-Whitney U test or Wilcoxon rank sum test. o Kindly report the effect sizes along with p-values. o In Table 3, for the first 10 patients, the ground truth is ‘1’ and for 11-20 patients the ground truth is ‘0’. Therefore, while performing the test, you will end up with two p-values (one for 1-10 patients and one for 11-20). Since you have reported a single p-value please specify how this was performed. RESPONSE: Thank you for your comments. The Mann-Whitney U test was used to validate the data. (p.11, line 243) The following text has also been amended in the manuscript (p.12, line 288): Those judged to have a 50% or greater likelihood of hippocampal sclerosis, according to the AI diagnosis, were treated as correct answers. Those that were not hippocampal sclerosis were treated as correct if the likelihood of hippocampal sclerosis was judged to be less than 50%. • Data augmentation is widely used while training the deep learning models. You should not apply them while computing the validation results as this will skew the accuracy since a single data point (patient) is considered as multiple observations. The validation/test results should be reported considering each patient as a single data point. RESPONSE: Thank you for your salient comments. The following text has been amended (p.11, line 271): When the diagnostic accuracy was measured using only the original images as the test data, the diagnostic accuracy using T2WI data was 98%, and that using FLAIR data was 95%. However, this result is considered to be insufficiently evaluated because of the small number of original images. • Finally, if possible try to explain why the AI system was capable of diagnosing better than clinicians using Fig. 4. In other words, what did the AI system detect in the images that were missed by the clinicians RESPONSE: Thank you for your comment. The following text has been amended in response (p.16, line 391): As shown in Fig. 5, AI focuses on the CA1 and CA2 regions of the hippocampus. In hippocampal sclerosis, pathological findings include not only neuronal atrophy but also neuronal loss and gliosis. Although not beyond the realm of conjecture, it is possible that AI can identify tiny differences in MRI signal values that cannot be corroborated by the physician's naked eye. Reviewer \#2: 1\. The authors failed to compare their segmentation results with hippocampal segmentation papers in the literature RESPONSE: Thank you for your valuable comment. The following sentences have been added to the text (p.15, line 376): Recently, many scholars have engaged in hippocampus segmentation algorithms. Several methods have been proposed, including cluster methods (Pang S), using map and feature embedding models or iterative local linear mapping models (Shumao P), atlas-based and label fusion methods (Carmichael O T), statistical machine learning techniques combined with KNN and SVM models (Hao Y), probabilistic modeling frame methods (Platero C), nonlinear image registration algorithms (Zhu H), energy minimization models (Lijn F V D), using local Gaussian distribution fitting energy with a level set function and local mean and variance as variables (Xiaoliang J, Wang L), and bias-corrected distance regularized level set method with MR image contrast enhancement (Selma T). 2\. The authors didn’t validate their segmentation accuracy through the deep learning cross validation methods RESPONSE: Thank you for your valuable comment. We performed a cross-validation analysis, and the system showed 90–93% accuracy for T2WI and 89–95% accuracy for FLAIR data. We have included this information in line 288 (p.12) of the text. 3\. Authors didn’t explain about the architecture of deep learning in the manuscript RESPONSE: Thank you for your appropriate comments. We have included this figure. Details of the neural networks of the model are shown in Fig. 2. (p7, line 189) 4\. Authors didn’t compare their results with other segmentation algorithms or standard tool box RESPONSE: Thank you for your important remarks. We have included the information in the text, on page 11, starting at line 259, as below: Haar-like feature was selected because it is more accurate than LBP when the hyperparameters are changed to optimize the detection rate and reduce false positives. FLAIR had a 89% detection rate and 0.25 false positives per picture. While checking the detection rate and false positive rate, we examined the effect of changing mainly the following parameters as optimization factors: scaleFactor, minNeighbors, minSize, and maxSize. For the T2WI classifier, scaleFactor was set to 1.06, minNeighbors to 3, minSize to (60,60), and maxSize to (130,130). Similarly, for FLAIR, we set the scaleFactor to 1.1202, minNeighbors to 3, minSize to (60,60), and maxSize to (130,130). 5\. Authors didn’t explain the challenges of segmentation of hippocampus through deep learning algorithm RESPONSE: Thank you for your important remarks. We have included this information in line 156 (p.5) in the text, as follows: OpenCV was used for hippocampal detection. MRI images were obtained of 113 T2WI and 148 FLAIR cases with a diagnosis of epilepsy from January 2016 to December 2019. On both the left and right sides, 226 T2WI and 296 FLAIR images with a resolution of 96 × 96 pixels were used for hippocampal images, and 327 T2WI and 367 FLAIR images with a resolution of 96 × 96 pixels were used for non- hippocampal images after amplification using augmentation. Haar-like features and local binary patterns (LBPs) were used as features. Hyperparameters of the created cascade classifier, mainly scaleFactor, minNeighbors, and minSize, were adjusted for optimization. 6\. Authors gas to validate their results through other metrics like sensitivity, specificity, f1 score etc to verify the performance of algorithm RESPONSE: Thank you for your valuable comment. The following sentences have been added to the text: (p.12, line 298): Diagnoses by AI were significantly more accurate than those by board-certified neurosurgeons based on T2WI images (p = 0.0001, effect size 0.6152); however, the difference in diagnostic accuracies of AI and neurosurgeons based on FLAIR was not statistically significant (p = 0.0520, effect size 0.3072) (Table 4). (p.12, line 303): However, the sensitivity, specificity, and F-value of the AI for T2WI were 0.8000, 0.9667, and 0.8421, respectively, and those of physicians for T2WI were 0.5000, 0.5667, and 0.5172, respectively. Similarly, for FLAIR, the sensitivity, specificity, and F-value for the diagnosis by AI were 0.3077, 1.0000, and 0.4706, respectively, whereas those by physicians were 0.5833, 0.5833, and 0.5833, respectively. Although the sensitivity, specificity, and F-value of AI for T2WI were high, the sensitivity for FLAIR was low. (p. 13, line 317): The diagnostic accuracy for MTLE by AI was higher than that by physicians for both T2WI and FLAIR (T2, p = 0.0034; effect size 0.3917, FLAIR, p = 0.0006; effect size, 0.4147). However, the sensitivity, specificity, and F-value of AI for T2WI were 0.5556, 0.9362, and 0.5582, respectively, and those of physicians for T2WI were 0.5000, 0.6754, and 0.4576, respectively. Similarly, in FLAIR, the sensitivity, specificity, and F-value for the diagnosis by AI were 0.4, 1.0000, and 0.5714, respectively, whereas those by physicians were 0.6833, 0.6181, and 0.5256, respectively. Although the F-value for AI was high, the sensitivity was low. 7\. Authors has to explain the optimization of parameters of deep learning model RESPONSE: Thank you for your important remarks. We have now added the following sentence in the text (p.7, line 190): We used SGD as an optimizer for the deep learning model. 8\. How MRI out performs the less cost EEG data in epilepsy diagnosis RESPONSE: Thank you for your very important question. We believe that AI will be useful for screening. We do not mean, of course, that EEG data are unnecessary, nor do we mean that MRI is superior to EEG. We hope that both MRI and EEG data will be used for more accurate diagnosis of MTLE in the future. 9\. Why T2 and FLAIR images are used instead of T1 weighted images? RESPONSE: Thank you for your very important question. In general clinical practice, physicians often diagnose hippocampal sclerosis using T2 rather than T1. This is thought to be because atrophy and sclerosis are more coordinated in T2. Therefore, T2WI and FLAIR were used in this study. 10\. Why there is huge difference between diagnostic accuracy of AI and neurosurgeon in table 3? Whether results can be validated with multiple surgeons results? RESPONSE: Thank you for your very important question. We have addressed this in the following lines (p.16, line 391): As shown in Fig. 5, AI focuses on the CA1 and CA2 regions of the hippocampus. In hippocampal sclerosis, pathological findings include not only neuronal atrophy but also neuronal loss and gliosis. Although not beyond the realm of conjecture, it is possible that AI can identify tiny differences in MRI signal values that cannot be corroborated by the physician's naked eye. 10.1371/journal.pone.0282082.r003 Decision Letter 1 Rajamanickam Yuvaraj Academic Editor 2023 Yuvaraj Rajamanickam This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 14 Sep 2022 PONE-D-22-00437R1Deep Learning for the Diagnosis of Mesial Temporal Lobe EpilepsyPLOS ONE Dear Dr. Sakashita, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Oct 29 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Yuvaraj Rajamanickam, Ph.D Academic Editor PLOS ONE \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: (No Response) Reviewer \#2: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Partly Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Firstly, I take this opportunity to congratulate the authors on their successful re-submission of their paper for publication. Comments: 1\. “Initially, I felt the paper can be recommended for publication with reasonable modifications. However, once I saw Fig. 3, I can no longer recommend the paper since there is something fundamentally wrong while training the deep learning model. The validation accuracy does not change even after training the model. This infers either the validation accuracy depicted in Fig. 3 is erroneous or the VGG16 model can detect MTLE even without re-training the model. Kindly rectify this issue since this is the main result of the paper.” The relatively high accuracy from the beginning of the epoch may be due firstly to the fact that VGG16 was a very good model for MTLE differentiation, and secondly to the fact that the image size was small (96 × 96 pixel) and the accuracy was high. However, considering that diagnostic accuracy increased with each successive epoch, we believe that there was a learning effect. Please elaborate the statement that the “diagnostic accuracy is increased in a successive epoch”. As observed from the Figure 4, the validation accuracy is decreasing or remaining the same throughout the training. The training accuracy is expected to increase and saturate overtime. The validation accuracy typically represents the results on an unknown test set. Therefore, if the validation accuracy is high without any training, it means that the model without any training is can perform classification task. Ideally, the validation accuracy should increase with training and later saturate or decrease due to overfitting. 2\. The patient flow chart is well described. However, the number of patients in the training/validation set needs to be specified separately. Also, please indicate the number of images/augmented images from each patient. As of now, the data information in lines 114, 125, 157, 161, etc is confusing. Thank you very much for your valuable suggestion. We have added a table (Table 1) and organized the data. (p.7, line 177) Thank you for adding a Tables 2 & 3 regarding patient information. However, please combine these two tables since I found that most of the patients were overlapping between the two analyses. The Table 1 is interesting. Please include details regarding the ‘validation’ set. Also, I find it consuming why data augmentation was performed on the test set. Data augmentation is performed on the training set to increase the number of observations and to make the system robust to noise. If it was applied on the test set, you are typically skewing the performance metrics. I understood you have presented both the results, but the results based on data augmentation is skewed. 3\. I also recommend performing a cross-validation to verify the robustness of the methods prescribed in the manuscript. Thank you for your valuable comment. We have now performed a cross-validation analysis, and the system showed a 90–93% accuracy for T2WI and an 89–95% accuracy for FLAIR. We have included this information in line 287 (p.12) of the text. Please specify the details of cross-validation: How many folds? How was the hyper parameters optimized? etc. 4\. I am also concerned why 20 patients were specifically chosen to compare between the AI system and the clinicians. Ideally, the comparison should be performed for all the 46 MTLE patients. Please do not use a ‘convenient’ sample for comparisons. I also have concerns regarding the statistical testing: T-tests are performed if the data distribution is ‘normal’. I am not sure if the data points fall into a normal distribution as the sample size is too low. Please perform a ‘normality’ testing before applying a t-test. Alternatively, you need to apply Man-Whitney U test or Wilcoxon rank sum test. Kindly report the effect sizes along with p-values. In Table 3, for the first 10 patients, the ground truth is ‘1’ and for 11-20 patients the ground truth is ‘0’. Therefore, while performing the test, you will end up with two p-values (one for 1-10 patients and one for 11-20). Since you have reported a single p-value please specify how this was performed. Those judged to have a 50% or greater likelihood of hippocampal sclerosis, according to the AI diagnosis, were treated as correct answers. Those that were not hippocampal sclerosis were treated as correct if the likelihood of hippocampal sclerosis was judged to be less than 50%. The reply to this question is incomplete. Please specify whether you have performed normality testing of the feature distribution to choose Mann-Whitney U test. Also, specify the type of d-values presented. Now, the modification on p 12, line 288 have given more concerns regarding the statistical testing. Mann-Whitney U test is performed to assess if two distribution as significantly different. However, since the authors claim that the prediction values were converted to binary as correct or wrong, then this statistical testing cannot be applied. Reviewer \#2: Authors has addressed all the comments of the reviewer. I recommend the manuscript for publication. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0282082.r004 Author response to Decision Letter 1 26 Oct 2022 Thank you for your thoughtful and constructive feedback regarding our manuscript titled “Deep Learning for the Diagnosis of Mesial Temporal Lobe Epilepsy.” We also appreciate the time and effort you and each of the reviewers have dedicated to providing suggestions, which have strengthened our paper. Thus, it is with great pleasure that we resubmit our article for further consideration. We have incorporated changes that reflect your suggestions. We hope that our edits and responses satisfactorily address all of the noted issues and concerns. Reviewer \#1: • Please elaborate the statement that the “diagnostic accuracy is increased in a successive epoch”. As observed from the Figure 4, the validation accuracy is decreasing or remaining the same throughout the training. The training accuracy is expected to increase and saturate overtime. The validation accuracy typically represents the results on an unknown test set. Therefore, if the validation accuracy is high without any training, it means that the model without any training is can perform classification task. Ideally, the validation accuracy should increase with training and later saturate or decrease due to overfitting. RESPONSE: Thank you for your valuable comment. Although it is difficult to understand which epoch the horizontal axis is referring to in the graph as the number of iterations increases within 1 epoch, accuracy increases to about 0.8 within the 1st epoch. In this learning, the accuracy almost reaches a plateau at 2 epochs, making this difficult to reflect it in the graph. • Thank you for adding a Tables 2 & 3 regarding patient information. However, please combine these two tables since I found that most of the patients were overlapping between the two analyses. The Table 1 is interesting. Please include details regarding the ‘validation’ set. Also, I find it consuming why data augmentation was performed on the test set. Data augmentation is performed on the training set to increase the number of observations and to make the system robust to noise. If it was applied on the test set, you are typically skewing the performance metrics. I understood you have presented both the results, but the results based on data augmentation is skewed. RESPONSE: Thank you for your valuable suggestion. We have compiled the data previously presented in Tables 2 and 3 in the new Table 3. The data in Table 1 has been revised, with the amplified validation data deleted. We have added the details regarding the validation data in the new Table 2. Cross-validation was performed using three of these cases as validation data. • Please specify the details of cross-validation: How many folds? How was the hyper parameters optimized? etc. RESPONSE: For MTLE with the hippocampus as the epileptogenic area, three out of nine T2 cases and 10 FLAIR cases were selected as validation data. Verification was performed five times by replacing the data. We obtained results of 90–93% accuracy for T2WI and 89–95% accuracy for FLAIR. • The reply to this question is incomplete. Please specify whether you have performed normality testing of the feature distribution to choose Mann-Whitney U test. Also, specify the type of d-values presented. Now, the modification on p 12, line 288 have given more concerns regarding the statistical testing. Mann-Whitney U test is performed to assess if two distribution as significantly different. However, since the authors claim that the prediction values were converted to binary as correct or wrong, then this statistical testing cannot be applied. RESPONSE: Thank you for your valuable comment. The Shapiro–Wilk test showed that the data did not follow a normal distribution (p\<0.000); therefore, Mann–Whitney U test was performed. Thank you again for giving us the opportunity to strengthen our manuscript with your valuable comments and suggestions. We have worked hard to incorporate your feedback, and hope that our revised manuscript is suitable for publication. Yours sincerely, 10.1371/journal.pone.0282082.r005 Decision Letter 2 Rajamanickam Yuvaraj Academic Editor 2023 Yuvaraj Rajamanickam This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 22 Nov 2022 PONE-D-22-00437R2Deep Learning for the Diagnosis of Mesial Temporal Lobe EpilepsyPLOS ONE Dear Dr. Sakashita, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Jan 06 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Yuvaraj Rajamanickam, Ph.D Academic Editor PLOS ONE Journal Requirements: Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: (No Response) Reviewer \#2: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Partly Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Firstly, I take this opportunity to congratulate the authors on their successful re-submission of their paper for publication. Comments: RESPONSE: Thank you for your valuable comment. Although it is difficult to understand which epoch the horizontal axis is referring to in the graph as the number of iterations increases within 1 epoch, accuracy increases to about 0.8 within the 1st epoch. In this learning, the accuracy almost reaches a plateau at 2 epochs, making this difficult to reflect it in the graph. Question: This is exactly where I am confused. Within 2 epochs, the model saturated and the validation accuracy started decreasing. This just shows that the VGG16 model can detect MTLE even without re-training the model. This just shows that the entire training process is redundant in this study. RESPONSE: Thank you for your valuable suggestion. We have compiled the data previously presented in Tables 2 and 3 in the new Table 3. The data in Table 1 has been revised, with the amplified validation data deleted. We have added the details regarding the validation data in the new Table 2. Cross-validation was performed using three of these cases as validation data. RESPONSE: For MTLE with the hippocampus as the epileptogenic area, three out of nine T2 cases and 10 FLAIR cases were selected as validation data. Verification was performed five times by replacing the data. We obtained results of 90–93% accuracy for T2WI and 89–95% accuracy for FLAIR. Question: Please elaborate on this procedure of replacing the data. Since this cross-validation results are the key results of the paper it needs to be well detailed. RESPONSE: Thank you for your valuable comment. The Shapiro–Wilk test showed that the data did not follow a normal distribution (p\<0.000); therefore, Mann–Whitney U test was performed. Question: The answer to this question is incomplete. “The reply to this question is incomplete. Please specify whether you have performed normality testing of the feature distribution to choose Mann-Whitney U test. Also, specify the type of d-values presented. Now, the modification on p 12, line 288 have given more concerns regarding the statistical testing. Mann-Whitney U test is performed to assess if two distribution as significantly different. However, since the authors claim that the prediction values were converted to binary as correct or wrong, then this statistical testing cannot be applied.” Reviewer \#2: Reviewer has given his decision to accept the paper in the previous revision. Editor can take the final decision regarding the publication of the draft in the journal. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0282082.r006 Author response to Decision Letter 2 29 Dec 2022 Reviewer \#1: Question: This is exactly where I am confused. Within 2 epochs, the model saturated and the validation accuracy started decreasing. This just shows that the VGG16 model can detect MTLE even without re-training the model. This just shows that the entire training process is redundant in this study. RESPONSE: Thank you for pointing this out. We apologize for not being able to give a clear answer. To check whether the VGG16 model can detect MTLE, we performed 20 diagnoses that were also used in the paper with a 1-epoch learning model. The result was not limited to whether MTLE was true or false, and the probability of MTLE was between 0.10 and 0.41 for FLAIR and between 0.01 and 0.55 for T2, which were not diagnostically useful. Although it is not a direct answer, it shows that it is necessary to repeat the number of times of learning. Question: Please elaborate on this procedure of replacing the data. Since this cross-validation results are the key results of the paper it needs to be well detailed. RESPONSE: A random number was used to assign which cases were to be trained and which were to be tested. (p.11, line 281) Question: The answer to this question is incomplete. “The reply to this question is incomplete. Please specify whether you have performed normality testing of the feature distribution to choose Mann-Whitney U test. Also, specify the type of d-values presented. Now, the modification on p 12, line 288 have given more concerns regarding the statistical testing. Mann-Whitney U test is performed to assess if two distribution as significantly different. However, since the authors claim that the prediction values were converted to binary as correct or wrong, then this statistical testing cannot be applied.” RESPONSE: Thank you for pointing this out. I apologize for my lack of knowledge about statistics. The Shapiro–Wilk test showed that the data did not follow a normal distribution (p\<0.000); therefore, Mann-Whitney U test was performed. (p.13, line 322) A cross tabulation was performed to compare the number of correct and incorrect numbers for AI and Dr. Results have shown that AI is superior in diagnosis, and we believe that we have demonstrated the effectiveness of AI. (p.13, line 325) 10.1371/journal.pone.0282082.r007 Decision Letter 3 Rajamanickam Yuvaraj Academic Editor 2023 Yuvaraj Rajamanickam This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 8 Feb 2023 Deep Learning for the Diagnosis of Mesial Temporal Lobe Epilepsy PONE-D-22-00437R3 Dear Dr. Sakashita, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Yuvaraj Rajamanickam, Ph.D Academic Editor PLOS ONE Additional Editor Comments (optional): I understand that there are mix opinios on this paper. Overall, the data interpretation and results in this manuscript are useful and interesting to the PLOSONE readers. The conclusions are well suppoted with the data. Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: (No Response) \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: No \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: No \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Dear Authors, I still find the answers to my reviews incomplete and therefore, I cannot recommend the manuscript for publication. 1\. I did not understand this answer: "The result was not limited to whether MTLE was true or false, and the probability of MTLE was between 0.10 and 0.41 for FLAIR and between 0.01 and 0.55 for T2, which were not diagnostically useful". My question was since from the accuracy figure (Fig. 4) the VGG16 model performs better on the validation data without any training. Why do you then add additional training and make the performance inferior? 2\. Cross-validation need to be systematically performed in this study. 3\. I still did not understand how the statistical testing was performed. I find multiple p values all over the paper. Also, the specific type of d-value used in the manuscript needs to be defined. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No \*\*\*\*\*\*\*\*\*\* \<quillbot-extension-portal\>\</quillbot-extension-portal\> 10.1371/journal.pone.0282082.r008 Acceptance letter Rajamanickam Yuvaraj Academic Editor 2023 Yuvaraj Rajamanickam This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 10 Feb 2023 PONE-D-22-00437R3 Deep Learning for the Diagnosis of Mesial Temporal Lobe Epilepsy Dear Dr. Sakashita: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Yuvaraj Rajamanickam Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Breast cancer is both the leading cancer and cancer-related death in women, with nearly 1.7 million new cases diagnosed and over half a million deaths reported globally in 2012. The same year, China alone accounted for nearly 190,000 cases and roughly 48,000 deaths. While prevalence in the US has been decreasing since the 2000’s, the breast cancer incidence has been steadily increasing in Asia since the 1980’s. Thanks to technological advancements and improved screening methods, more cases are being diagnosed at earlier stages, and early detection is directly correlated with an increased chance of survival. Despite this, the staggering incidence indicates that further screening, therapeutics, and preventative measures are necessary to reduce the rate of breast cancer and improve the prognosis of the disease. There are a variety of factors which contribute to the development of breast cancer, the most significant of which being gender and old age. Additional etiologic agents include race, hormones, tobacco and alcohol consumption, obesity, lack of childbearing, and a combination of environmental and genetic factors. Genetics are estimated to be the primary causal factor in 5–10% of breast cancers, while all others develop spontaneously with an accumulation of genetic and epigenetic changes. Hereditary breast–ovarian cancer syndrome is the familial tendency to develop these cancers. The best characterized of these hereditary mutations are in BRCA1 and BRCA2 genes, which can interfere with repair of DNA cross links and DNA double strand breaks. These inherited mutations pose a lifetime risk of developing breast cancer between 40% and 80%, indicating cancer is not inevitable for carriers of these mutations. However, only 2 to 3% of breast cancers have mutations in BRCA genes, and an estimated 75–80% of hereditary breast cancers involve unknown genes. Additionally characterized on the breast cancer cells are three important receptors: estrogen receptor (ER), progesterone receptor (PR), and ERBB2 (Her2), and the presence of these receptors can influence prognosis and treatment. Despite ongoing efforts to improve screening and treatment of breast cancer, further research is needed to determine other unknown genetic mutations which are involved in the progression of the disease. Due to the variety of complex interactions between genetic and environmental factors, each tumor potentially exhibits a unique gene mutation profile. By profiling an individual’s cancer genome it becomes possible to distinguish the oncogenic mechanisms that regulate the cancer. As such, there is accumulating evidence which suggests that individualized, tailored therapies are necessary for effective treatment against cancers. Until recently, individual genome sequencing for personalized medicine was impractical due to the cost and lengthy assay times; however, new semiconductor-based sequencing called Ion Torrent sequencing is tackling many of these issues associated with other sequencing methods. In this study, we have used Ion Torrent sequencing to analyze 105 clinical breast cancer samples to identify the genetic mutations in 737 loci of 45 known cancer-related genes. # Results ## Breast Cancer Mutation Spectrum in Chinese Patients We analyzed 105 breast cancer samples from Chinese patients ranging from 21–100 years of age. The patients were categorized based on their age, menopausal states, receptor status (ER, PR, and Her), and AJCC/TNM cancer staging system (**–**). This Personalized Cancer Mutation Panel is designed to target 737 mutations in the following 45 key cancer genes: ABL1, AKT1, ALK, APC, ATM, BRAF, CDH1, CDKN2A, CSF1R, CTNNB1, EGFR, ERBB2, ERBB4, FBXW7, FGFR1, FGFR2, FGFR3, FLT3, GNAS, HNF1A, HRAS, IDH1, JAK3, KDR, KIT, KRAS, MET, MLH1, MPL, NOTCH1, NPM1, NRAS, PDGFRA, PIK3CA, PTEN, PTPN11, RB1, RET, SMAD4, SMARCB1, SMO, SRC, STK11, TP53, and VHL. The mean read length was 76 bp, and the average sequence per sample was approximately 24 Mb. With normalization to 300,000 reads per specimen, there was an average of 1639 reads per amplicon (range: 28 to 4732), 176/189 (93.1%) amplicons averaged at least 100 reads, and 168/189 (88.9%) amplicons averaged at least 300 reads. Of the 45 oncogenes and tumor suppressor genes sequenced in the 105 breast cancers, only PIK3CA (35.2%), TP53 (15.2%), and ERBB2 (1%) incurred missense mutations. Immunohistochemical staining revealed different states of mutation in the ERBB2 (v-erb-b2 avian erythroblastic leukemia viral oncogene homolog 2), ER (Estrogen), and PR (Progesterone) receptors of these patient samples in addition to their above incurred mutations (**–**). Frequencies of these incurred mutations at different stages of breast cancer development observed in our sample set according to AJCC Cancer Staging is shown in. The detailed list of missense, point mutations, insertions and deletions profiled on the 737 loci of 45 tumor suppressor and oncogenes in 105 breast cancer samples is listed in the. ## Missense Mutation Distribution in the Exons and Functional Domains of PIK3CA PIK3CA mutations were identified in 35.2% of 105 tumors and most of these mutations were focused in exons 4 (2.6%), 9 (42.0%), and 20 (55.3%). These exons encode the helical and kinase domains, and mutations in these domains are associated with increased lipid kinase activity and oncogenicity. The phosphatidylinositol 3-kinase (PI3K) pathway has been identified as an important player in cancer development and progression. Upon receptor tyrosine kinase activation, the PI3K kinase phosphorylates inositol lipids to phosphatidylinositol-3,4,5-trisphosphate. PI3K is a heterodimeric enzyme composed of a p110α catalytic subunit encoded by the PIK3CA gene and a p85 regulatory subunit encoded by the PIK3R1 gene. Phosphatidylinositol-3,4,5-trisphosphate activates the serine/threonine kinase AKT, which in turn regulates several signaling pathways controlling cell survival, apoptosis, proliferation, motility, and adhesion. Immunohistochemical staining revealed different states in the ERBB2, ER, and PR receptors and the frequency of PIK3CA mutations differed markedly at different states of these receptors (**–**). The frequencies of PIK3CA mutation occurring at different receptor states and in pre- and post- menopausal women is illustrated in. For example, 37.7% of post-menopausal women carrying PIK3CA mutations were 50% ER+ and PR+, and 33.3% of post-menopausal women were Her+; however, PIK3CA mutation frequencies in women positive for these receptors were slightly less in pre-menopausal women who were ER+ and PR+. Not only that, women carrying PIK3CA mutations and who were ER+ were 43.9% PR+, 57.1% Her+; who were PR+ were 48.9% ER+ and 100% Her+; who were Her+ were 57.1% ER positive and 100% PR+. Also, PIK3CA mutations associate with older age at diagnosis with 66.7% of those in the age range of 81–100 years, 25.0% in the age range of 61–80 years, 36.4% in the age range of 41–60 years, and 26.3% in the age range of 21–40 years. ## Missense Mutation Distribution in the Exons and Functional Domains of TP53 The p53 tumor suppressor gene is located on 17p13 chromosome and spans 20 kb genomic DNA encompassing 11 exons that encodes for 53 KD phosphoprotein. The phosphoprotein is a transcription factor which regulates apoptosis, genomic stability, and angiogenesis. Functional loss of p53 can lead to defective DNA replication and malignant transformation, common in the dysplasias of breast cancers. The p53 gene exhibits numerous genetic alterations in patients with breast cancer. This was indeed true with our sample set as well, constituting several missense mutations throughout the p53 coding region. The incidence of p53 abnormalities varies with the degree of dysplasia and patients features. This highlights the need for the administration of effective treatments such as cell-cycle inhibitors in the form of target therapies and combinatorial target therapies against the wide range of p53 mutations accumulated in that locus. Most TP53 mutations detected in our sample set by Ion Torrent sequencing were missense (16/105, 15.2%), one of the frequently occurring mutation in TP53. The mutations were along exons 4–10, encoding the DNA-binding and oligomerization domain; specifically, the missense mutations were concentrated along the domain required for interactions with FBX042, HIPK1, AXIN1, the DNA major groove, and the domain that contains the nuclear export signal. The frequency of TP53 mutations varied widely at different states of the hormone receptors (**S1–3**). The frequencies of TP53 mutation occurring at different receptor states and in pre- and post- menopausal women is illustrated in. For example, 15.1% of post-menopausal women carrying TP53 mutations were 11.5% ER+, 9.1% PR+, and 33.3% Her+; however, TP53 mutation frequencies in women positive for these receptors were slightly more in pre-menopausal women. Also, TP53 mutations associate with older age at diagnosis with 53.67% in the age range of 81–100 years, 34.50% in the age range of 61–80 years, 15.8% in the age range of 41–60 years, and 34.9% in the age range of 21–40 years. ## Multiple Mutations and Mutation Hot Spots in Human Breast Cancers Clinical success with individualized combination therapy relies on the identification of mutational combinations and patterns for co-administration of a single or combination of target agents against the detected mutational combinations. Some of the mutations detected in our tumor group through sequencing analysis were not only recurrent and frequent but also occurred in combination with other mutations. Breast cancers in our sample set contained the following: 72.6% of samples had at least one or more missense mutations, 34.0% had at least two or more missense mutations, 7.5% had at least three or more missense mutations, 1.9% had at least four or more missense mutations, and 27.4% of samples incurred no deleterious mutations in any of the screened 737 loci of the potential tumor suppressor and oncogenes. # Discussion In this study we have performed a high-resolution genomic sequencing on 105 breast cancers in Chinese patients using the high throughput Ion Torrent sequencing technology. We mainly identified mutations focused along two hotspot loci, PIK3CA and TP53 in the breast cancer genomes of our sample set. In comparison with traditional Sanger sequencing and other sequence analysis methods, our analysis was at much faster rates and were of reduced sequencing costs per base. The cost and complexity associated with the 4-color optical detection used in all other NGS platforms is evaded through the use of Post Light sequencing technology employed in the Ion Torrent sequencing. Despite these benefits, there is less awareness about the use and availability of this platform; however, it’s starting to reach clinical investigators in recent times. In this study, we identified PIK3CA mutations in 35.2% and TP53 mutations in 15.2% of breast tumors. Previous studies have identified PIK3CA mutation in 10.3–37.5% of the HER2-positive breast cancer cases – and p53 mutations in 18%–25% of primary breast carcinomas. TP53 mutations generally have a poor prognostic power, which may be due to the screening approach used. TP53 mutations are commonly detected through IHC, which detects only mutations that induce protein accumulation, missing frameshift, nonsense, and splice mutations. The Ion Torrent sequencers helped us detect robustly coding, silent mutations, insertions, and deletions more precisely. Not only that, breast cancer is a heterogeneous disease. There is a high degree of diversity between and within tumors as well as among cancer-bearing individuals, and all of these factors together determine the risk of disease progression. Due to these various levels of heterogeneity, generalized treatments may be less effective. Instead targeted therapy, which involves the usage of specially designed drugs to selectively target molecular pathways correlated with the malignant phenotype of breast cancer cells, may be more useful. This indicates the necessity of sequencing individual human breast cancers in order to match the use of a single targeted drug or two or more targeted drugs in combination against individual breast cancer-specific mutations. It is also critical to examine the biological features associated with each of these individual tumors to assign an appropriate treatment response. For example, trastuzumab, a humanized monoclonal antibody targeting the extracellular domain of the HER2 receptor that blocks the ligand-independent HER2 signaling, is affected by the status of PIK3CA gene mutations. Similarly, prognostic studies focusing on breast cancer in the absence of p53 mutations predicts longer survival following primary therapy. However the clinical course of metastatic breast cancers and p53 mutations have not been thoroughly investigated and it remains somewhat controversial whether p53 has any significance in prediction of the clinical outcome of breast cancer. Our results were somewhat consistent to the above studies in terms of the observed mutation frequencies in the PIK3CA and TP53 loci, however none of the previous reported studies compared the correlation of these mutations at different intensities and in the presence and absence of all three hormone receptors in pre- and post-menopausal Chinese patients. In this study we have evaluated and compared the relationships between TP53 and PIK3CA status with different subgroups of Chinese patients. Our current study is more of a feasibility test aiming to validate the applicability of this advanced tool in categorizing the breast cancers into different subgroups based on the identified features in their cancer genome and proteome. We further aim to use this information in a prospective clinical study to test the response of a personalized treatment regimen in different subgroups of Chinese patients. The Ion Torrent sequencing platform helped us identify distinct mutation combinations as listed in **,** and, rendering potential possibilities for developing personalized combinatorial therapies. For example, depending on patients’ mutated loci, their accompanied ER, PR, and Her receptor states, combined with the knowledge of patients menopausal status, personalized combinatorial therapies can be rendered as a more effective and specific treatment option over those that are currently available. Breast cancer treatment options include surgery, radiation therapy, and chemotherapy, but often depend on the stage of the disease. Targeted therapies, including anti- hormone therapies have become standard treatment for breast cancers expressing the targets of these drugs. Clinical trials are underway to evaluate targeted therapies and investigate how best to use these drugs in combination with each other and with other standard therapies. As there is increasing information about the changes in breast cancer cells in recent times, newer drugs that specifically target these changes have been developed. These targeted drugs either work synergistically with the chemo drugs or by themselves with less toxicity due to a selective effect to a more systemic modulation of proteins associated with oncogenesis. Briefly, trastuzumab, as mentioned earlier, is a humanized monoclonal antibody targeting the extracellular domain of the HER2 receptor that blocks the ligand-independent HER2 signaling. It was initially approved by the FDA for metastatic breast cancer in 1998. Lapatinib is a dual EGFR/ErbB2 reversible tyrosine kinase inhibitor blocking both HER1 and HER2 and consequently the downstream pathways of MAPK/Erk1/2 and PI3K/Akt pathways. Two different types of antihormone therapies are used to treat women with ER-positive breast tumors: selective estrogen receptor modulators and aromatase inhibitors. Aromatase inhibitors are used mainly in postmenopausal women because they do not work very efficiently in premenopausal women, whose ovaries make too much aromatase. Anti-hormone therapies such as tamoxifen and anastrozole can be used to treat most stages of breast cancer. Women with early-stage breast cancer are usually treated with surgery followed by antihormone therapy and radiation therapy, and sometimes chemotherapy. Antihormone therapies can also be used to reduce the risk of developing breast cancer. Women at high risk for the disease are usually treated with Tamoxifen and another selective estrogen receptor modulator, raloxifene, and several other aromatase inhibitors. In this study we have used the Ion Ampliseq Cancer Panel to sequence 737 loci in 45 cancer-related genes, mainly oncogenes and tumor suppressor genes, of 105 human breast cancer samples. Having gained more knowledge and experience through next generation technologies, it is necessary to expand our understanding of specific mutations to enhance individualized therapies. Therefore, gathering a complete profile of mutations in breast cancers for the application of personalized and tailored targeted therapy is critical to develop future cancer treatments. We believe a faster and more cost effective genotyping tool such as Ion Torrent sequencing technology will be greatly beneficial to assign such specific therapeutics for breast cancer patients in the near future. # Materials and Methods ## Ethics Statement The study has been approved by the Human Research Ethics Committee of the People’s Hospital of Shan Xi Province, Xian, China. For Formalin fixed and paraffin embedded (FFPE) tumor samples from the tumor tissue bank at the Department of Pathology of the hospital, the institutional ethics committee waived the need for consent. All samples and medical data used in this study have been irreversibly anonymized. ## Patient Information Tumor samples used in the study were collected from the People’s Hospital of Shan Xi Province, Xian, China. A total of 105 FFPE tumor samples from female breast cancer patients were analyzed. Patients were classified by age ranges as follows: 19 were between 21–40 years, 55 were 41–60 years, 28 were 61–80, 3 were 81–100, and one patient was of unknown age. AJCC/TNM cancer staging is as follows: 24 patients at stage 1, 29 at stage 2a, 14 at stage 2b, 20 at stage 3a, 8 at stage 3c, and 2 of unknown stage. Tumor samples were also analyzed for immunohistochemical status of estrogen receptor (ER), progesterone receptor (PR), and ERBB2 (**–**). ## DNA Preparation Formalin-fixed, paraffin embedded (FFPE) tissue samples were deparaffinized in xylene and 3–5 µm thick sections were extracted. DNA was then isolated using the QIAamp DNA Mini Kit (Qiagen) following manufacturer’s instructions. ## Ion Torrent PGM Library Preparation and Sequencing An Ion Torrent adapter-ligated library was constructed with the Ion AmpliSeq Library Kit 2.0 (Life Technologies, Part \#4475345 Rev. A) as per manufacturer’s protocol. Briefly, 50 ng of pooled amplicons were end-repaired, and Ion Torrent adapters P1 and A were ligated with DNA ligase. Adapter-ligated products were then purified with AMPure beads (Beckman Coulter, Brea, CA, USA), nick- translated, and PCR-amplified for a total of 5 cycles. The resulting library was purified with AMPure beads (Beckman Coulter), and the concentration and size of the library was determined by Agilent 2100 BioAnalyzer and Agilent BioAnalyzer DNA High-Sensitivity LabChip (Agilent Technologies). Sample emulsion PCR, emulsion breaking, and enrichment were performed using the Ion PGM 200 Xpress Template Kit (Life Technologies, Part \#4474280 Rev. B), according to the manufacturer’s instructions. Briefly, an input concentration of one DNA template copy/Ion Sphere Particles (ISPs) was added to emulsion PCR master mix and an IKADT-20 mixer (Life Technologies) was used to generate the emulsion. Next, ISPs were recovered and template-positive ISPs were enriched for use with Dynabeads MyOne Streptavidin C1 beads (Life Technologies). The Qubit 2.0 fluorometer (Life Technologies) was used to confirm ISP enrichment. 316 chips were used to sequence barcoded samples on the Ion Torrent PGM for 65 cycles, and an Ion PGM 200 Sequencing Kit (Life Technologies, Part \#4474004 Rev. B) was used for sequencing reactions, as per the recommended protocol. ## Variant Calling Data from the PGM runs were processed initially using the Ion Torrent platform- specific pipeline software Torrent Suite to generate sequence reads, trim adapter sequences, filter, and remove poor signal-profile reads. Initial variant calling from the Ion AmpliSeq sequencing data was generated using Torrent Suite Software v3.2 with a plug-in “variant caller v3.2” program. In order to eliminate errors in base calling, several filtering steps were used to generate final variant calling. The first filter was set at an average depth of total coverage of \>100, an each variant coverage of \>20, a variant frequency of each sample \>5, and P-value\<0.01. The second filter was employed by visually examining mutations using Integrative Genomics Viewer (IGV) software (http//[www.broadinstitute.org/igv](http://www.broadinstitute.org/igv)) or Samtools software SAMtools software (<http://samtools.sourceforge.net>), as well as by filtering out possible strand-specific errors, such as a mutation detected in either “+” or “−” strand, but not in both strands of DNA. The third filtering step was set as variants within 727 hotspots, according to the manufacturer’ instructions. The last filter step was eliminate variants in amplicon AMPL339432 (PIK3CA, exon13, chr3∶178938822–178938906), which is not uniquely matched in human genome. From our sequencing runs using the Ion Ampliseq Cancer Panel, false deletion data were generated from the JAK2 gene locus and thus the sequencing data from this locus were excluded from further analysis. ## Bioinformatical and Experimental Validation We used the COSMIC3 (version 64), MyCancerGenome database (<http://www.mycancergenome.org/>) and some publications to assess mutations reappearing in lung cancer. Additionally, some detected missense mutations were confirmed by Sanger’s sequencing. ## Statistical Analysis We selected reappearing somatic missense/insertion-deletion mutations of breast cancer to perform the statistical analysis. # Supporting Information We would like to thank Rong Shi at the Wu Jieping Foundation, Dr. Haibo Wang, Zhi Yu, Ying Li, and other members of San Valley Biotechnology Inc. Beijing for their assistance in sample and data collection. We would also like to thank the staff at the Beijing Military Hospital for their generous support for DNA sequencing and data collection. [^1]: Authors Hua Ye, Chuanning Tang, Feng Lou, Dandan Zhang, Hong Sun, Haichao Dong, Guangchun Zhang, Zhiyuan Liu, Zhishou Dong, Baishuai Guo, He Yan, Chaowei Yan, Lu Wang, Ziyi Su, and Yangyang Li are employees of San Valley Biotechnology, Inc. This does not alter the authors’ adherence to all the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: XB EZ SYC JG. Performed the experiments: HY ZW LC CT JL HL WZ WH FL DZ HS HD GZ ZL ZD ZD BG HY CY LW ZS YL. Analyzed the data: XB EZ CT SYC JG. Contributed reagents/materials/analysis tools: XB EZ SYC JG. Wrote the paper: XB EZ VN CT LJ XFH SYC JG. RS HW ZY YL.

# Background Preterm birth (PTB) is the leading cause of neonatal mortality and morbidity, with an incidence estimated between 5–12%, depending on the geographic region. PTB may be the consequence of spontaneous preterm labor with intact membranes (sPTB), preterm prelabor ruptured membranes (PPROM), or indicated PTB (for fetal and maternal reasons). Although PPROM often leads to sPTB, it is recognized that PROM and sPTB can present as separate entities. A better understanding of the similarities and differences between the respective pathways would open new avenues for research and treatment. Genetic predisposition to PTB has been demonstrated in numerous studies –. Among twin pregnancies, the genetic contribution to PTB is estimated to be approximately 30%. In singletons, genetic susceptibility to PTB is based on the evidence of familial aggregation, measures of heritability, identification of disease-susceptibility genes and racial disparity in PTB rate – that may be related to differences in risk-predisposing allele frequencies. PTB rates are higher in sisters of women with a history of PTB compared to their sisters-in- law (16% vs. 9%). Several studies have confirmed a two-fold increase in risk of PTB for black American women compared to white American, women even after controlling for socio-economic factors associated with PTB. Single nucleotide polymorphisms (SNPs) are the most thoroughly investigated genetic markers in relation to PTB. Many SNPs in maternal or fetal genomes have been reported to be associated with spontaneous preterm birth, but often the results have not been replicated in subsequent studies. New systems biology approaches allow integration of functional genomics related to a phenotype or a disease in a connectivity network. Such convergent genomics allow identification of factors that affect downstream or upstream regulation of many candidate genes. Multiple relevant factors may converge on a similar pathway or gene, or conversely, one upstream stress condition/transcription factor may be responsible for regulation of several different pathways. Pathways or genes identified through this process may not be detected in individual genetic association studies. This ‘convergent genomics’ approach has been undertaken in the area of preterm birth in a recent study evaluating the impact of genetics on gestation age at delivery. Pooling SNP data from the published genome wide association study (GWAS) on preterm birth \[<https://www.genevastudy.org/Publications>\] and data from the custom database on preterm birth (dbPTB) (<http://ptbdb.cs.brown.edu/dbPTBv1.php>) the authors analysed genetic contribution to preterm birth in three gestation categories: less than 30 weeks, less than 34 weeks and less than 37 weeks, but they did not take into account phenotypic differences. To our knowledge, the comparative analysis of PPROM and sPTB, while taking ethnicity into account, has not yet been undertaken at this methodological level. We have, therefore, performed a systematic literature review of studies that investigated genetic factors involved in PTB to assess whether maternal candidate genes and SNPs previously identified in association with preterm birth can be used to: 1\) identify maternal genetic markers for stratification of preterm birth due to PPROM or to spontaneous PTB with intact membranes 2\) to differentiate between and clarify pathophysiological mechanisms involved in spontaneous PTB with intact membranes and PPROM # Methods ## Search strategy We searched Medline (Ovid) from 1^st January, 1990 to 1^st January, 2013, ISI Web of Knowledge, SCOPUS and databases including the Preterm Birth Genetics Knowledge Base, PTB Gene (<http://ric.einstein.yu.edu/ptbgene/index.html>), KEGG ([www.genome.jp/kegg](http://www.genome.jp/kegg)), DAVID ([david.abcc.ncifcrf.gov](http://david.abcc.ncifcrf.gov)), GO ([www.geneontology .org/GO.database.shtml](http://www.geneontology.org/GO.database.shtml)) and Ingenuity Pathway analysis (IPA) (<http://www.ingenuity.com/products/ipa>) for studies assessing maternal genetic polymorphisms in association with preterm birth. The publically accessible Database for Preterm Birth (dbPTB) (ptbdb.cs.brown.edu/dbPTBv1.php) is a web-based tool which contains genes, genetic variants and pathways involved in preterm birth and any additional information on related literature. The search terms used to identify studies included “preterm birth” OR “preterm labor” OR “PPROM” combined with “polymorphisms” OR “genetic” OR “genes”, restricted to human. We also searched reference lists of study reports and review articles. Studies published only as abstracts were also included. No language restrictions were applied. ## Selection of studies Three review authors (AC, AA and SM) independently assessed studies for inclusion in the review, using the criteria above. Any discrepancies were resolved by discussion or by consulting an additional author (ZA). ## Inclusion criteria Studies were included if they were assessing single nucleotide polymorphisms (SNPs) in women who had sPTB or PPROM before 37 weeks gestation. Definitions for sPTB and PPROM were used as described in the studies. Spontaneous preterm birth was defined as spontaneous preterm labor with intact membranes, followed by preterm birth before 37 weeks. Preterm PROM was defined as membrane rupture diagnosed by vaginal pooling of fluid and a positive nitrazine test, before 37 weeks gestation and at least one hour prior to the initiation of regular contractions. Only SNPs that were reported to be significantly associated (p value \<0.05) with sPTB or PPROM were included in pathway analyses. ## Exclusion criteria Studies were excluded if they investigated medically indicated preterm births for pregnancy complications such as pre-eclampsia, placenta praevia, fetal anomalies, and gestational diabetes, or if we were unable to separate the data for sPTB and PPROM. Review articles were also excluded. ## Data Extraction and management Data were extracted from each included study, entered into Excel spreadsheets, and double checked for accuracy. To prevent bias from multiple publications from the same cohort, only the largest number reported and/or the latest manuscript was used to calculate population size. Data were collected on the gene name and official symbol, and the unique SNP identifier rs number, was obtained from the databases dbSNP, OMIM and PTB Gene. The ethnicity of the population was noted. The genotype and allele association was recorded and expressed in terms of p-value, for each SNP. ## Pathway analysis We used the Ingenuity Pathway Analyses (IPA) (Ingenuity Systems, Inc., Redwood City, CA, USA) to examine biological networks and disease functions associated with SNPs in maternal genomes that were significantly associated with either sPTB or PPROM. The Ingenuity Pathway Analysis (IPA) software package (<http://www.ingenuity.com>) uses the Ingenuity Knowledge Base, a manually curated database which contains information from the published literature as well as many other sources, including several gene expression and gene annotation databases such as IntACT, BIND, MiPs et al.. IPA also includes interaction data to assign genes to different groups and categories of functionally related genes. It measures the associations of genes that are entered into the software, termed “focus genes”, with other molecules, such as proteins, genes, cells, tissues, diseases or medications. Focus genes are defined in the Ingenuity Pathway Analysis as the important (associated or above a cut-off) genes which interact with molecules in the Global Molecular Network. A gene cannot be considered a focus gene if there are no known molecular interactions involving that gene in the Ingenuity Knowledge Database. Ingenuity calculates single p-values for the enrichment of each gene category using the Fisher's exact test, taking into consideration both, the total number of molecules from the analyzed data set and the total number of molecules linked to the same gene category in the Ingenuity Knowledge Base reference set. For each gene category, correction for multiple comparisons is calculated using the Benjamini-Hochberg method and corrected p values of enrichment are provided. Graphical presentation of networks is available as well as a score for each network. The score represents an approximate interaction between the focus molecule and each network. We have also used IPA to identify functionally related genes that correspond to specific canonical pathways from a collection of 200 well-characterised metabolic and cell-signaling cascade pathways manually curated by IPA scientists from journal articles, text books, and KEGG ligand database. The Fisher’s exact test is used to calculate the probability that the association between the genes in the dataset and the canonical pathway can be explained by chance alone. Finally, we used the IPA upstream regulator analysis to identify transcriptional factors that may control genes and their pathways identified in relation to preterm birth, and to further examine how they may regulate their targets in order to provide testable hypotheses for gene regulatory networks (<http://www.ingenuity.com/wp-content/themes/ingenuitytheme/pdf/ipa/feature_high light_upstream_downstream.pdf>). ## Datasets Maternal genes or SNPs found to be significantly associated with sPTB and PPROM were uploaded separately into the IPA analysis tool. We examined the lists of all genes investigated in the included studies and compared them with the genes that were associated with each condition for similarities/differences. We related the genes to the postulated pathophysiological mechanisms of sPTB and PPROM. The software analysis mapped these focus genes to biological networks and disease functions. Comparative analyses were carried out on IPA for the two groups, sPTB and PPROM. The data generated were compared with regards to: i) the number and type of biological networks involved in each group, ii) genes/molecules common between PPROM and sPTB, and molecules unique to each group, iii) differences in the top disease functions within the networks, iv) differences in the top disease functions in the canonical pathways and v) any differences in the top transcriptional regulators involved in PPROM compared to sPTB. The number of links between molecules eligible for each network has been set to 140, which is the maximum number of connections in any network allowed by the IPA software. Links relevant only for mammalian species are used to reconstruct networks. # Results A total of 1,424 reports were retrieved using our search strategy. Seven hundred and thirty nine duplicates were removed, two hundred and forty two studies were excluded after screening the titles and abstract because they were not primary studies, not related to preterm birth, did not report maternal genes or SNPs or were not associated with either sPTB or PPROM. Full text articles were retrieved for the remaining 100 potentially eligible reports (the list of all studies is available on request). Fifteen studies met the criteria and were included in the analyses (10 on sPTB with intact membranes, – and 5 on PPROM,. A total of 3,600 cases and controls matched for ancestry (White European, African American, South American (Latino) and Japanese) were included. Of these, a total of 878 women experienced PTB (N = 572 with sPTB and N = 306 with PPROM) and 2,722 women had term births defined as\>37weeks. A total of 2175 SNPs in 274 genes were investigated; 2169 SNPs in 266 genes were included in the studies on sPTB and 780 SNPs in 190 genes were genotyped in the studies on PPROM. 188 genes were investigated in relation to both sPTB and PPROM. A total of 248 SNPs in 102 genes were found to be statistically significantly associated (p\<0.05) with spontaneous preterm birth with intact membranes (sPTB), and 39 SNPs in 32 genes were found to be statistically significantly associated with PPROM (**)**. These genes were uploaded into the IPA for network analyses. Most of the genes analyzed for sPTB and PPROM phenotypes were studied in the Chilean population. In addition, 30 genes from the significantly associated sPTB gene set were studied in African American population, but only in relation to sPTB and not PPROM. ## Ingenuity Pathway analysis ### Network analysis We performed the IPA analysis of the PPROM and integrated sPTB datasets based on ontology classification. Results from the comparative analyses between sPTB and PPROM are presented in. The focus genes uploaded for PPROM were mapped to 3 networks with scores ranging from 2.3 to 51, however, only one network reached a score above 3 (Network A). For sPTB, focus genes were mapped to 10 networks with scores ranging from 2 to 115, and only two networks had a score above 3 (Networks B and C). Each network included a number of “new” molecules (not from the input gene list) that were assigned to the network by the IPA algorithm (Network A = 111, Network B = 69, Network C = 123). ### Function analyses IPA algorithm identified the top ranked functions of these 3 networks: cellular movement and immune cell trafficking represented the first two functions in Network A and Network B, while the top ranked network function involved in Network C was cellular growth and proliferation. The top four molecular and cellular functions identified by IPA were exactly the same for the sPTB and PPROM. These were: i) cell to cell signaling and interaction, ii) cellular movement, iii) lipid metabolism and iv) small molecule biochemistry. Although by general ontological description the data sets looked similar, more detailed insight in the defined functional ontologies has uncovered the distinct characteristics specific for each phenotype. sPTB appears to be associated with glucocorticoid signaling pathway with the following downstream affected genes: ADRB2, AGT, CCL2, HSPA4, HSPA6, HSPA1B, HSPA1L, IFNG, IL5, IL6, IL8, IL10, IL1B, IL1R2, IL1RN, IL4, MMP1, NFKB1, NFKBIB, NR3C1, POMC, SELE and SERPINE1. Apart from glucocorticoid signaling, the identified markers were clearly inflammatory, matrix degrading or collagen metabolism related markers along with NF-kB related molecules, which indicate an inflammatory component. The only two genes from the list representing glucocorticoid signaling pathway that overlap with PPROM are AGT and MMP1. AGT encodes angiotensinogen, a precursor of angiotensin involved in controlling arterial pressure, which has been associated with preterm-birth regardless of etiology. MMP1 encodes matrix metalloproteinase 1 which is higher in women delivering preterm compared to those delivering at term. In contrast, two genes, TNF and NOS2A, were only detected in PPROM but not sPTB. TNF encodes tumor necrosis factor α, a pro-inflammatory cytokine found in the amniotic fluid and in blood of mothers and fetuses with PTB. NOS2A encodes nitric oxide synthase 2A, a P450 type protein which is found in the foeto- placental unit that may be leading to the reduced placental blood flow and increased resistance in the foeto-maternal circulation. In addition, NOS2s are important in cytokine signaling and innate immunity and are involved in an antimicrobial pathway. Regarding lipid metabolism, sPTB and PPROM are characterized by different functions. sPTB shows specific functional enrichment in eicosanoid metabolic pathways and functions. Eicosanoids are signaling molecules derived from fatty acids, involved in inflammation and immunity. In contrast, in PPROM dataset ‘Lipid metabolism’ ontological grouping, metabolism of eicosanoids does not present at a high significance level. Although the PPROM list of functional annotations is shorter and shows less significant p-values compared with sPTB, it has a distinct enrichment of functions related to lipid metabolism (triglycerides and fatty acids) and the lipid metabolic disorders such as hyperlipidemia, hypercholesterolemia (genes represented include HDL, LIPC and FAS). There are also differences in connective tissue functions between sPTB and PPROM, such as altered fibroblast proliferation (TNFA and IGF1vs NOS2 and NOS3), blood pressure and kidney dysfunction (PLAT). In addition, connective tissue disorders have different components in sPTB and pPROM. In PPROM, structural and mechanical tissue damage-related functions (IGF1, IL1A, TNF, TNFRSF1B) and weak autoimmune component (ACE, ANG, CSF1 (includes EG:12977), IL1A, IL6R, MMP1 (includes EG:300339), MMP10, NOS2, TNF, TNFRSF1B) are identified. In sPTB in contrast, the cell-mediated-immune response was one of the dominating categories. A strong autoimmune component is represented with the large number of immune-related genes (ACE, ADRB2, CCL2, CCL8, COL3A1, CTLA4, CYP19A1, DHFR, EPHX2, HSPA1A/HSPA1B, HSPA1L, IFNG (includes EG:15978), IL10, IL10RA, IL12B, IL18 (includes EG:16173), IL1A, IL1B, IL1R1, IL1R2, IL1RN, IL2RA, IL2RB, IL4 (includes EG:16189), IL4R, IL5, IL6, IL6R, IL8, LTF, MMP1 (includes EG:300339), MMP10, MMP16, MMP2, MMP3, MMP8, MMP9, NFKB1, NOD2, NR3C1, POMC, PTGS1, SELE (includes EG:20339), SLC6A4, TIMP3, TLR2, TLR7, TNFRSF1A, TNFRSF1B, VEGFA). ### Disease analysis The top ranked disease functions involved with PPROM were inflammatory processes while the top disease functions involved with sPTB were connective tissue disorders. Although both sPTB and PPROM have inflammatory and connective tissue disorders among the top 5 ranked diseases, their inflammatory categories vary greatly as demonstrated in. sPTB is associated with rheumatic disease and autoimmune diseases (including neurological and psychiatric diseases), endometriosis and particularly strongly linked with infection (bacterial, parasitic or viral origin). Cell mediated immune response in sPTB is associated with the following functions: IFNG (includes EG:15978), IL10, IL12B, IL15 (includes EG:16168), IL1B, IL2RA, IL2RB, IL4 (includes EG:16189), IL6, TLR7 for T cells homeostasis, and IFNG (includes EG:15978), IL12B, IL1B, IL2RA, IL2RB, IL4 (includes EG:16189), IL6, TLR7 for differentiation of helper cells. In PPROM however, inflammation component was related mainly to local organ inflammation with the highest scores p = 10⁻¹⁴ to p = 10⁻⁷ with the functions of ACE, AGT, CD14, COL4A3, COL4A4, FAS, IFNGR2, IGF1, MMP1 (includes EG:300339), NOS2, NOS3, REN, TNF. Further diseases linked to PPROM were nephritis or glomerulonephritis (ACE, COL4A3, COL4A4, IGF1, NOS2, REN, TNF). In addition, PPROM inflammatory response is supported mainly by macrophages and especially phagocyte migration, activation and proliferation (AGT, CSF1 (includes EG:12977), FAS, TIMP2 (includes EG:21858), TNF, IL6R, CD14, TNFRSF1B). These have been previously associated with early onset sepsis in very low birthweight infants, and chorioamnionitis-associated preterm delivery. Cell mediated immune response specific for PPROM also includes differentiation of helper T lymphocytes (P = 1.35×10⁻¹²), T cell homeostasis (p = 5.04×10⁻¹¹), differentiation of T lymphocytes (p = 6.53×10⁻¹¹). Similarities between sPTB and PPROM include the ACE and AGT genes which might be responsible for hematologic/coagulation function disorders linked with markers (STAT1, MMP1, TNF, IGF1, MMP9). They are involved in collagen metabolism and matrix degradation, an effector of PPROM. Functional SNPs in these genes and their action may be exaggerated by coagulopathies (ACE, AGT) or by apoptosis (STAT1). In PPROM, an increased synergistic effect on vasculature and heart dynamics may be associated with several additional genes TNFA, IGF1, NOS2 and NOS3. Body mass index (BMI) was associated with both sPTB and PPROM and also two top canonical pathways “Hepatic fibrosis/hepatic stellate cell activation” and “Atherosclerosis Signaling”, although the content of the functions grouped by these categories was different as mentioned above. Both, low and high BMI has been associated with PTB previously. ### Network analysis After analysing individual functions and diseases related to sPTB and PPROM, we overlayed the networks reconstructed from the functions with the corresponding ontological categories to illustrate the dominant associations. We reconstructed dense networks with connected functions, which were not present in the starting lists of genes. We also refined one network for each condition. All molecules in the three networks which showed a significant association (score\>3) are presented in. . illustrates connectivity between genes associated with sPTB and their relation to few dominant functional categories. We used IPA design tool to re-arrange the nodes in each network so that the functions associated with each category are located close to the corresponding category labels. Only the NFκB TF is associated with all the main categories and with the majority of the genes. A number of interleukins and growth factors specific for sPTB: IL18, IL12B, IL8, IL4, IL10, PTGS1 or shared between two subtypes: F3, TNFRSF1B, TNFRSF1A, are linked to all functional categories. Top ranked PPROM-associated functional network is shown in. Vasculogenesis can be observed in association with PPROM-specific functions. Blood pressure category, although more profound for the PPROM set, is also associated with the sPTB list of genes. ### Upstream transcriptional regulation We used IPA upstream regulator analysis to predict the potential TFs responsible for differential gene expression in sPTB and pPROM. Regulatory interactions for the sPTB and pPROM gene lists suggested by the IPA largely overlap, although the association with sPTB shows lower p-values. The NfKB complex is highly ranked for both, SPTB and PPROM through the network connectivity, therefore the fraction of genes common for preterm birth conditions is likely regulated by the inflammatory TF NFκB. Our predicted TFs are shown in together with the downstream genes that they may regulate. Nearly all PPROM-specific genes are linked to a signal transducer and activator of transcription 1 (STAT1). STAT1 is a 91 kD protein, member of the STAT family of transcription factors originally identified as the mediators of the cellular response to interferon alpha (IFN) and estrogen receptor 2 (ESR2). Genes unique to sPTB are regulated mainly by three TFs; glucocorticoid receptor (NR3C1), peroxsome proliferator activated receptor γ (PPARG) and interferon regulating factor 3 (IRF3). Interestingly, the NR3C1 polymorphysm shown to be associated with sPTB, is included in the dataset and appears in the top network reconstructed for sPTB. # Discussion Based on our analyses, we propose autoimmune/hormonal regulation axis for sPTB, whilst pathways implicated in the etiology of PPROM include hematologic/coagulation function disorder, collagen metabolism, matrix degradation and local inflammation. We could clearly separate integrated datasets/pathways for sPTB and PPROM and identified genes encoding transcription factors specific for sPTB (NR3C1, PPARG and IRF3) and PPROM (STAT1 and ESR2), or common for both subtypes (NFκB complex). Earlier genetic studies achieved only limited success in explaining heritability of PTB. They have been hampered by a relatively small number of participants with variable ethnic background, lack of well characterized phenotypes, lack of replication of genetic associations in subsequent studies and up until recently, lack of unbiased genome-wide association studies. It is likely that variants of many genes contribute to the final PTB with only a modest effect. These are difficult to identify and will often be missed in small scale association studies. In addition, these candidate gene association studies are not designed to investigate gene-gene interactions. By integrating the genomic data on a large number of individuals into networks, we have overcome the low statistical power of individual studies. The analysis also permitted inclusion of different ethnic groups. Our network analysis has allowed us to expand gene selection and include some, as yet unexplored candidates. It is noteworthy that potential for false positive associations is usually small in the core network connections where several interconnected functions may suggest potential targets which could be subsequently validated. In addition, the IPA software allows correction for false positive discovery when multiple comparisons are involved. We included the data from multiple gene/polymorphisms association analyses related to PTB in different ethnic groups. For sPTB, we included three populations: 1) South American (Chilean), 2) African-American and 3) Caucausian. However, data for African-American and Caucasian women with PPROM were not available. Interestingly, we found a difference in functional description among different ethnicities. As shown in, only data for the African-American ethnic group point to a role of hormonal regulation in preterm labor. It remains unclear whether the hormonal regulation of sPTB is indeed ethnicity specific, although large variability in the incidence of preterm labor in different ethnicities support the association. We hypothesized that although different ethnic groups may bear different polymorphisms in relation to PTB, they are tightly linked to the network of ethnicity invariant functions associated with sPTB. Hormonal deregulation may comprise an upstream or a downstream component in the chain of events leading to sPTB. It will be interesting to find out if an ethnic group which is at higher risk for PTB, has higher frequency of specific polymorphisms in a number of genes (Cpla2, Vegf, VEGFA, CYP19A1, HSPA1A and A1B, PGRMC1) in the proximity of LH and FSH functions. We recognized that the hormone-regulatory sub-network which includes FSH, estrogen, progesterone, LH and their connected functions is central to sPTB. This was also confirmed in African Americans where FSH is linked strongly with the network reconstructed from the sPTB associated genes. Our findings are in keeping with a recent phylogenetic analysis of potential association between genes involved in human birth acceleration and preterm birth risk. The study revealed a pivotal role of the defined ‘hormonal’ sPTB network. Screening of 8,400 SNPs in 150 human accelerated genes for association with preterm birth in 165 Finnish preterm and 163 control mothers showed that 8 of the 10 most significant SNPs were identified in the FSHR gene. The hormone-regulatory component of the sPTB gene group network connects all other sPTB-associated functions in the network, indicating that this may be crucial for the upstream pathways such as the infection/inflammation pathway. The most densely connected to sPTB is the inflammatory/infection related network that can be present upstream as well as downstream of associated hormonal perturbations. The predicted TFs, such as NR3C1 and PPARG are known modulators of hormonal regulation of immunity and lipid metabolism and have been implicated in preterm birth infants previously. We identified PPARG as one of the leading regulators for sPTB related to modulation of expression of lipid metabolism genes and hormone regulation functions. Endogenous glucocorticoids influence fetal development and regulate glucose and fat metabolism, cardiovascular system and immune functions. Several polymorphisms in the glucocorticoid receptor gene (NR3C1) are linked to altered function of glucocorticoid receptor. SNPs may increase glucocorticoid sensitivity *in vivo*, and have been associated with birth weight in preterm neonates. STAT1 which is identified as PPROM-specific regulator, belongs to a family of transcription factors. STAT1 was initially identified as an interferon α (IFNα) mediator and a major component of the cellular response to IFNγ. STAT1 regulates the immune system, cell differentiation, cell growth inhibition and apoptosis. It is interesting that STAT1 was recently shown to be associated with renin- angiotensin system and may be a regulator of the genes associated with a ‘blood pressure’ category specifically enriched in the PPROM-associated dataset. STAT1 is also involved in the negative regulation of angiogenic factors such as matrix metaloproteinases (MMP9). MMPs play a central role in cell proliferation, migration, differentiation, angiogenesis, apoptosis and host defences. The interaction of STAT1 with other transcription factors varies with the cellular system and it can induce the constitutive expression of a subset of genes involved in immune regulation. Mutations in the STAT1 gene leading to its reduced activity are associated with infectious disease. In contrast, a high level of expression of STAT1 stimulates the TNF-α apoptotic pathways and inhibits NF-κB. Increasing concentration of TNFα in umbilical cord was associated with an increased risk of preterm birth. Another transcription regulator linked to PPROM is estrogen receptor 2 (ESR2), a member of the family of estrogen receptors and superfamily of nuclear receptor transcription factors. Enhanced estrogen receptor activity is involved in the proinflammatory cascade leading to parturition. Furthermore, estrogen receptor activation facilitates labor by enhancing transcription of genes encoding uterine contraction-associated proteins including oxytocin and COX-2, which catalyzes the production of prostaglandins. Distinct functions associated with the studied gene group and network connectivity of predicted TFs are indicative of diverse etiology for PPROM and sPTB. A relatively small number of studies and genes was included in the analyses of PPROM, therefore, further work on PPROM is required using an unbiased genome wide approach to elucidate potential novel mechanisms. One of the weaknesses of this study is reliance on candidate gene association studies. Lack of GWAS studies specific to sPTB and pPROM results in biases in candidate gene selection. As many of the authors of candidate gene studies have been involved in infection/inflammation area of research, that pathway dominates in gene selection. However, the unbiased GWAS approaches such as the GENEVA GWAS study, failed to generate any genome-wide significant results, potentially because spontaneous PTB cases were analysed together with all indicated PTB cases. Future GWAS studies should eliminate the biases in candidate gene selection and examine further diverse etiology of sPTB and pPROM. In addition, gene-gene and gene-environment interactions cannot be assessed through IPA analysis. These, together with comorbidity such as coagulopathies and collagen disorders, and their interaction with inflammatory triggers that may produce discrete phenotypes, will have to be incorporated into clinical trial design and data collection of future studies. We did not control for methodological quality of included studies. It was evident that quality of data from different studies varied; while comprehensive information on a large number of markers was available for Chilean and African American groups, in Caucasians, only limited information was available from several small association studies often reporting a single gene analysis. Several authors did not report on quality control processes, and did not include information on genotype call frequency, validation cohorts or statistical power calculations. In addition, no formal genetic analysis has been conducted in these studies to test for population stratification. Future work should include unbiased whole genome approaches in large, phenotypically well characterized cohorts of women with PTB and controls matched for ancestry. In order to calculate the required number of cases and controls in a genome-wide association study, which would allow to detect clinically relevant difference in allelic frequencies (OR≥3), approximately 350 PTB cases and 1400 controls would be required to achieve 80% power. We assumed an additive model of inheritance, 1∶4 ratio of cases: controls and genome wide significance (p = 1.00E-08) for alleles at\>5% frequency. Such numbers of participants can be achieved only through international collaborations on PTB. In conclusion, we predicted transcriptional regulators that may be related to etiology of PPROM and sPTB, and identified genetic markers that may help in the assessment of their potential risk factors. Heterogeneity in the natural history of preterm birth exists and therefore, a one-size-fits-all prevention strategy may not be appropriate. Prevention of preterm birth requires more research into the etiology of preterm birth subtypes and this study has contributed to our understanding of the molecular mechanisms and regulation of genes/pathways implicated in PPROM and sPTB. Our hypothesis generating study has identified several new candidate genes and their potential role in the pathogenesis of PPROM and sPTB, which should be explored further and validated in carefully planned experiments with large numbers of patients. The study represents the first step and foundation for future research, which will be conducted through access to the international biorepositories of blood or tissues from large numbers of pregnant women with well-defined and standardized phenotypes. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: AC ZA SM OV AA. Performed the experiments: SM AC OV. Analyzed the data: OV AA AC ZA. Contributed reagents/materials/analysis tools: OV ZA. Wrote the paper: AA AC ZA OV SM.

# Introduction For many decades, gender differentiation has been studied as an interdisciplinary topic and within a variety of fields including psychology, social science, anthropology, history, and biology. Existing studies have explored the nature of the existing gender differences, their origin, and impact on individuals’ lives. How to interpret the observed deviation between women and men is subject to debate among scholars. It is, however, universally accepted that behavioral differences are rooted in the different biological roles, and are reinforced by a society’s values and cultural beliefs. Previous research has shown that gender-specific inequalities might originate from biological predispositions (e.g. hormones, brain structure), as well as the organization of the hunter-gatherer societies in which humans initially evolved. This differentiation is subsequently aggravated by cultural/societal expectations, which are likely to lead the two genders to develop and maintain their social ties in different ways. The study of social networks is essential for understanding how gender role influences the nuances observed in the structure and evolution of these social interactions. Although it does not provide answers regarding the origins of the gender differences in social behavior, it can help identify and understand these discrepancies to a larger extent. Below, we first explore individual-level characteristics, specifically the psychological traits and mobility behavior within the cohort, noting that both these aspects have been found to relate strongly to social behavior. Next, we focus on social traits. For the two gender-groups, we evaluate similarities and differences with respect to social network role. Our analysis of social networks is based on longitudinal data describing person-to-person interactions (physical proximity using Bluetooth scans), calls and text messages, and online friendships (based on Facebook communication activity). Finally, we use classification models to quantify the extent to which a person’s gender can be inferred from their observed characteristics and behavior. # Data The basis of this paper is the *Copenhagen Network Study* (CNS), a study focusing on nearly 800 freshmen at the Technical University of Denmark who volunteered to donate data via Nexus 4 smartphones. The bulk of data collection was behavioral data from from the smartphones, supplemented with data from online questionnaires and 3rd party APIs, such as the Facebook Graph API. The derived datasets include: - Friendship graph and interactions (comments on wall posts) from Facebook, - Person-to-person proximity events, measured using Bluetooth, - Telecommuncation data (call and text message logs; only metadata, no content), - Location records (based on GPS and WiFi), - Questionnaires (responses to personality questionnaires, described in detail below). This work is based on data collected between September 2013 and May 2014. The number of active participants and the quality of their data varies over the duration of the observation. To eliminate the effect of missing data on statistics, we calculate all indicators and network properties on a weekly basis and average for each individual. Participants with three active weeks or less during the nine-month period are excluded from the analysis. After this filtering, this dataset consists of 166 female and 601 male students. In order to avoid the difference in population sizes affecting the standard deviations, we apply subsampling over the male and female population separately and calculate the distribution over the mean of the random sub-samples. Below, unless otherwise specified, we use the following strategy to compare the two (female/male) classes. We we draw 1000 random subsamples, each equal to the half of the original class size from each class. Then, we perform pairwise comparisons between subsamples. We test the null hypothesis that the means of the two sampling distributions are identical (a two-tailed test). In order to compare results across domains (personality, mobility, social interactions, etc) we measure the differences in distributions between the two genders using **effect size** *r*, defined as the ratio between the means of each distribution *x*₁, *x*₂ and the pooled standard deviation *σ*_*p*: $$\begin{array}{r} {r = \frac{\mu\left( x_{1} \right) - \mu\left( x_{2} \right)}{\sigma_{p}\left( x_{1},x_{2} \right)},} \\ \end{array}$$ where *σ*_*p*(*x*₁, *x*₂) is defined via $$\begin{array}{r} {\sigma_{p}\left( x_{1},x_{2} \right) = \sqrt{\frac{{(|}x_{1}{| - 1) \cdot}\sigma^{2}\left( x_{1} \right) + {(|}x_{2}| - 1) \cdot \sigma^{2}\left( x_{2} \right)}{|x_{1}| + |x_{2}\left| - 2 \right.}}} \\ \end{array}$$ and *σ*²(*x*) is given by $$\begin{array}{r} {\sigma^{2}(x) = \frac{1}{n - 1}\mspace{180mu}\sum\limits_{i = 1}^{n}\mspace{180mu}\left( x_{i} - \mu(x) \right)^{2}.} \\ \end{array}$$ # Personality In this section, we investigate how gender differences are expressed through personality metrics. Data from responses to personality questionnaires show that although there are considerable variations within a gender, differences between males and females exist in a number of traits and at every age. As part of the CNS study, we consider the following dimensions of personality, which are listed below along with the central gender-related results pertaining to that measure. 1. **Big Five.** The Big Five Inventory (BFI) is a widely used method for assessing human personality using five broad factors: *openness*, *extraversion*, *neuroticism*, *agreeableness*, and *conscientiousness*. To measure big five, we use the questionnaire developed in Ref.. Previous work has consistently found women to be more neurotic and agreeable than men. There is less of a consensus with respect to gender differences in the remaining BFI attributes. For instance, some studies report higher conscientiousness and openness among women, while others find men as more conscientious. Detailed description of each personality trait and reference to additional literature are provided in. 2. **Self-esteem.** We use the definition that self-esteem is a feeling of self-worth and use Rosenberg’s 10-item instrument to measure it. Feingold found that males have slightly higher self-esteem than females, and Kling et al. showed that this effect increases considerably in late adolescence. However, other studies exist that show no significant difference between males and females with respect to self-esteem. 3. **Narcissism.** Narcissism has been previously found to be positively correlated with self-esteem. Here we assess Narcissism using the Narcissistic Admiration and Rivalry Questionnaire (NAR-Q), which integrates two distinct cognitive and behavioral aspects of narcissism: the tendency to approach social admiration through self-enhancement, and the tendency for an antagonistic self-defense (rivalry). The literature is consistent here: men tend to be more narcissistic than women, regardless the age and income. 4. **Stress.** Several studies have been conducted to measure stress levels among students in higher education, reporting that female students tend to have more stress (and more stressors) than male students, regardless of the instrument used for measurement. In this study, we measure stress with the widely used Perceived Stress Scale (PSS). 5. **Locus of control.** Locus of control reflects the extent to which a person perceives a reward or reinforcement as contingent on his own behavior (internal locus) or as dependent on chance or environmental control (external locus). We measure locus of control using a simplified, 13 item scale proposed by Goolkasian (see: <http://www.psych.uncc.edu/pagoolka/LC.html>). A lower score corresponds to internal locus, whereas a higher score indicates external locus. In general, the two genders have not been found to differ with respect to this psychological trait. Lefcourt argued that those who are classified as having an internal locus of control not only perceive but also desire more personal control than individuals with an external locus and found a that females desired greater internal control than males. However, women have been found to favor external control in items related to academic achievements. 6. **Satisfaction with life.** Satisfaction with life constitutes a judgment of one’s life in which the criteria for judgment are up to the person. We use the satisfaction with Life Scale (SWLS) instrument, which has been widely used to assess subjective well-being within various groups of population. The SWLS includes five generic statements, in which a subject must respond with a 1-7 scale, indicating the degree of agreement or disagreement. Results regarding gender have been shown to be highly dependent on age. Specifically, adolescent and elderly males have higher life satisfaction than females, while no observable difference is found among young adults. 7. **Loneliness.** We measure loneliness using the UCLA Loneliness Scale, a 20-item scale, in which a subject must indicate how often they feel an item characterizes them. Male college students have been found to be more lonely than female students. It has also been shown that men are less willing to acknowledge feelings of loneliness, due to their more pronounced negative consequences of admitting to this feeling. In summary, women tend to score higher with respect to negative emotionality (such as neuroticism and stress) than men, but it has been argued that this may be due to females more readily admitting to or perceiving such intense feelings. Individualism also plays an important role in personality differences between the two genders. In the present study, we analyze the gender effect on the aforementioned personality traits in an environment where females are the numerical minority, and within a highly specific group of individuals (students at a technical university). The diverse dataset, however, allows us to combine the results from the questionnaires with the participants’ behavior in a natural setting. ## Results We test the null hypothesis that the two samples have equal means. We start with the Big Five Inventory, and measure effect sizes between the two genders. shows the normalized difference (i.e., the effect size) observed between males and females with respect to neuroticism, conscientiousness, agreeableness, extraversion, and openness. Each histogram represents the distribution of the difference in means normalized by the pooled standard deviation, and the mean in a subsample of females is subtracted from the mean in a subsample of males (for details on the effect size, see). The horizontal bars denote 5 and 95% percentiles. Neuroticism exhibits the largest deviation, positioned far to the left from the zero baseline (with a mean of *d*_neu = −0.635), indicating that women score significantly higher in this personality characteristic than men. We also find significant, albeit less pronounced, differences with respect to conscientiousness (*d*_con = −0.436) and agreeableness (*d*_agr = −0.259). Finally, we do not find statistical significance in the average values of extraversion (*d*_ext = −0.118) and openness (*d*_ope = 0.143). depicts the results describing the remaining personality measures. Stress is significantly higher among women (*d*_str = −0.451), while it is clear that men score higher in self- esteem (*d*_se = 0.423). Overall, narcissism is higher among male students (*d*_nar = 0.349), but mainly due to rivalry (*d*_riv = 0.334), which is its antagonistic aspect and less because of admiration (*d*_adm = 0.241), which constitutes the assertive aspect of narcissism. We find that women score higher in I-E Rotter Scale (*d*_loc = −0.157), indicating a greater average sense of external locus of control. Women also score higher with respect to satisfaction with life (*d*_sat = −0.149). Finally, we find no statistically significant difference for loneliness (*d*_lon = 0.095). # Mobility In this section we verify whether there are observable differences in mobility traces between the participants of the two genders. We begin by providing a brief overview on the literature discussing differences between male/female mobility patterns, then discuss findings from our cohort. There is a general consensus that mobility patterns are not gender neutral and womens’ mobility through urban space is distinguishable from mens’. Differences between men and women in their mobility have been ascribed to various components of the gender role, such as gender-related tasks, distinct family roles, and labor market position. Men and women are assumed to perform a similar number of trips, but with distance traveled and the mode of transportation differing between them. Specifically, surveys conducted in Western countries in the ’90s have demonstrated that women travel fewer kilometers than men and make more trips as pedestrians and using public transportation. Moreover, the purpose of travel tends to differ, with women traveling most frequently for household errands and men making a majority of trips to work. Other studies explain the shorter commuting distances of women as a result of their weaker position in the labor market. Interestingly, females have been observed to travel longer distances and explore larger areas in foraging tribes, the reason for this difference is argued to originate from the fact that women are expected to return home more frequently while gathering than men are while hunting. Recent studies based on mobile phone records, however, have not observed substantial differences in travel distances, regularity, and predictability of movements between male and female commuters. However, a study using travel diary data collected in Portland reports higher levels of activity among part-time employed women than those of part-time employed men throughout the day. In conclusion, despite of recent advances in studying mobility behavior in detail based on high resolution observational data, gender-based differences are rarely observed. ## Results We follow the same procedure as in personality-related measurements: we apply subsampling to obtain equal sample sizes, calculate the effect size, and test the null hypothesis that the means of the two distributions are equal. A common quantitative description of mobility behavior is given by the distribution of unique locations visited by an individual over some time period, e.g., using *P*_*u*(*l*), which is the relative frequency of visiting location *l* by individual *u*. Relative frequency is given by the relative time the individual spent at some location on a weekly basis. We analyze location data obtained by periodically collecting the position estimate from the location sensor of the students’ phones. The list of unique locations that characterize an individual is extracted as a list of clusters of location measurements a DBSCAN-based algorithm developed in Ref. and validated in Ref.. To further quantify individual mobility patterns, we measure the heterogeneity of the locations visited over time using entropy. Entropy is a measure of uncertainty or predictability of a distribution. Here we use entropy to capture the heterogeneity of an individual’s time spent across unique locations. Using *P*_*u*(*l*), it is defined as $$\begin{array}{r} {H_{u} = - \mspace{180mu}\sum\limits_{l \in L(u)}\mspace{180mu} P_{u}(l)\log P_{u}(l),} \\ \end{array}$$ where *L*(*u*) denotes the set of locations for user *u*. Individuals distributing their time more evenly are characterized by higher entropy. The effect sizes measured in the location related metrics are shown in. We find that women both visit more unique locations over time, and they have more homogeneous time distribution over their visited locations than men, indicating that time commitment of females is more widely spread across places. # Networks and interactions Now we turn our attention towards social interactions among the students. We begin by providing a brief overview on the literature discussing differences between male/female network structures. We then discuss findings based on our cohort. Previous work suggests that the sizes of real-world ego networks of the two genders are drawn from similar distributions. In contrast, women tend to have more friends online, as seen in multiplayer games or social networking services. A study based on Facebook data describing around 1800 U.S. college students found that females show higher social activity and have greater betweenness centrality in their Facebook network compared to males. Social networks display high gender homophily, both offline and online. The extent of preferring same gender friends varies with age, with e.g. girls forming smaller, more homogeneous groups than boys at young ages. As soon as adolescents begin forming romantic ties during puberty, women start to invest more heavily in opposite-sex relationships; but they shift preferences to younger women (presumably daughters) as they age. Men, on the other hand, are shown to increase their female contacts as they get older and particularly at the end of their life cycle. Interestingly, heterophily between genders is prominent among the strongest ties. For instance, calls and text conversations are both more frequent and longer among mixed-sex pairs of individuals. Homophily has been studied as a function of transitivity (a measure of the probability of two individuals being connected provided they are both connected to the same alter). In this case, structural factors, such as network proximity, have been found to have a stronger effect on triadic closure compared to homophily: a high number of shared contacts is a better indicator of triadic closure than sharing an attribute. A study based on data from several U.S. elementary schools reports that females form more triads than males and that dyads consisting of females are more likely to be in triangles. Kovanen et al. studied temporal gender homophily in 3-motifs using a large dataset of mobile phone records. They find that female-only motifs are over-represented compared to a reference model, whereas male-only motifs are under-represented. Contradicting the aforementioned findings, however, a study based on data from the Spanish social networking site *Tuenti*, found high levels of homophily in females’ dyads but a higher tendency of male users to form same-sex triangles. Women have not only been found to be more actively engaged in online interactions, but also to spend more time engaged in phone conversation. In a review paper, Smoreda and Licoppe report that women tend to disclose more information to correspondents (especially about intimate topics) and are more expressive than males, which results in longer conversations, whereas men communicate mainly for instrumental purposes. In addition, other studies have shown that calls to a woman are longer than calls to a man regardless the gender of the caller. Circadian rhythms in call patterns have revealed further differences between men and women, with women making longer phone calls in the evenings and during the night, and mainly when the recipient is a man (which indicates an emphasis to romantic relationships). Likewise, young women have been reported to send a greater number of text messages, especially if the receiver belongs to the opposite gender. In summary, previous studies found clear differences in the way men and women engage with their social networks. However, most of the studies focus on a single channel of interactions (e.g. online communications, behavior in an organization), failing to capture a potential persistence or deviation of the characteristics across different settings. Here we use the CNS data to compare communication across a number of different channels. ## Results We consider three types of communication: physical proximity (i.e., person-to- person) interactions, Facebook activity, and mobile phone communication (calls and text messages). Previous studies have shown that each channels may describe different aspects of social ties and potentially corresponding to different levels of connection intensity. To illustrate these differences, in we report the fraction of active links in each communication channel over time. Note that the vertical scale is logarithmic, indicating an increased presence of proximity links (purple pentagons), a moderate level of active Facebook connections (red circles) and a comparably low level of active telecommunication links (bottom curves). visualizes a single-day snapshot of the three networks in this study. The community structure revealed in the person-to-person network is a fingerprint of the classes the students attend together. Less structure is observed in the Facebook feed network, which shows a single large component; the call network is the most fragmented of all of the networks. The high level of homophily in the proximity network is evident from the frequency of orange and blue lines, which represent the female-female and male-male connections respectively, as well as cliques that contain nodes of the same color. The call network shows the highest fraction of mixed gender connections, a possible indicator of couples. When investigating the networks between the study participants, we apply a different approach to accounting for the imbalance male/female subjects than in case of personality and mobility. Here, subsampling would alter the network structure and, thus, render e.g. measurements of homophily and other network metrics meaningless. Instead, we use the following reference model: we randomly permute genders between participants with uniform probability and then perform the calculations. Overall, we produce 2*E* network realizations, where *E* is the number of edges in the network. To approve or reject the null hypothesis that the network is independent of gender homophily, we calculate the *z*-score, given by: $$\begin{array}{r} {z = \frac{x - \mu\left( \widetilde{x} \right)}{\sigma\left( \widetilde{x} \right)}.} \\ \end{array}$$ Here, *x* is an indicator, $\mu\left( \widetilde{x} \right)$ and $\sigma\left( \widetilde{x} \right)$ are the mean and standard deviation of the indicator in the reference model. The *z*-score is expected to be zero if the null hypothesis is true. To test the null hypothesis of no difference between the two gender groups, we draw the permutation distribution of the differences between the two genders and measure where this distribution falls relative to the mean difference of the empirical data. The *p*-values then are calculated by dividing the number of permuted mean differences that are larger/smaller than the one observed in the empirical data, by the number of items (2*E*) in the permutation distribution. We explore the influence of gender homophily on formation of friendships in the various networks among the participants. To do this, we first identify the fraction of same-gender friends out of all friends an individual has. shows the *z*-scores of various network connections obtained by comparison with the permutation model (see also). Women have remarkably more same gender friends than the ones measured with the reference model in online interactions and person-to-person interactions (*z*-score is 13.10 and 12.24 respectively). On the other hand, men also show a preference for forming homophilous ties through mobile communications, though to a less extent. To study whether men and women tend to form closed triangles with same-gender alters, we count the various motifs in each network. Results are shown in (color bars): male only (blue), female only (orange) and mixed (brown). Furthermore, we compare the results with the respective distribution of the expected motifs found in the reference model for the Facebook network. We find that male-only triads are insignificantly underrepresented compared to the reference model, and that there are more female-only motifs than what we would observe by chance (*z* = 13.101, *p* \<.0001). Whereas a similar pattern is observed in person-to-person interactions, same-gender motifs are overrepresented for both genders in mobile communications. We conclude that women prefer other women for both their dyadic and triadic relationships in every form of interaction, while homophily is noticeable among males only in their trusted interactions through the phone. We find that women tend to have a significantly higher number of contacts than men in both online and mobile networks, whereas the size of the person-to-person networks are similar. shows how degree varies over time in terms of mobile communication (calls); females have more contacts during nearly the entire period of interest. We measure betweenness centrality (see for the definition) of each individual to investigate whether one gender tends more prominently positioned in a network than the other. We find that women consistently show higher betweenness indices, regardless of the mode of interaction. Next we study the entropy of interactions. Similarly what we did for mobility distribution in, we calculate the entropy of the distribution of interactions over the contacts: $$\begin{array}{r} {S_{u} = - \mspace{180mu}\sum\limits_{i \in N(u)}\mspace{180mu} P_{u}(i)\log_{2}P_{u}(i),} \\ \end{array}$$ where *P*_*u*(*i*) is the probability that user *u* interacts with their *i*-th contact in his ego-network *N*(*u*). The value of *S*_*u* is estimated by the corresponding number of interactions relative to all interactions performed by user *u*. Individuals who interact equal amounts with many friends will have high entropy (and therefore can be characterized by lower predictability, whereas those who limit the vast majority of their interactions to a small set of others are expected to have low entropy (more predictable). In, we plot the distribution of entropy effect sizes measured between males and females for the three interaction networks. We observe a significant difference for Facebook and calls, with women displaying higher entropy than men, indicating that females distribute their interactions with friends considerably more homogeneously. In addition, females exchange remarkably more text messages than males (*p* \<.001). With respect to time spent on social interactions, we find that in our study, women are described by significantly longer conversation times during phone calls than men, regardless of the initiator of the call (*p* \<.0001). The longest calls (on average) are observed on ties where a male initiates contact to a female (with an average duration of 117 seconds), with second longest average call-durations observed between females (an average duration of 114.56 seconds). The shortest average duration (71.52 seconds) is measured between pairs of males. # Gender prediction Based on the findings presented above, we consider the classification problem of predicting gender based individual and social characteristics. In the literature, there have been attempts to predict gender based on Call Detail Record (CDR) data using semi-unsupervised techniques and deep-learning algorithms. De Montjoye et al. found that gender is a strong predictor of neuroticism, a trait that is seen in the literature to be consistently higher among women. In this study, we combine the questionnaire data, mobility patterns, as well as social interaction habits of each participant, to build a dataset that offers adequate complexity for achieving a good performance in the gender-inference problem. Additionally, the machine learning process also provides insight into the question: What are the most predictive behavioral indicators of gender. ## Results We use the behavioral measures calculated above as features to train four different models: *logistic regressor*, *AdaBoost*, *support vector classifier* (SVC), *random forest*, and *gradient boosting classifier* implemented in the scikit-learn Python package. Each models is evaluated using 10-fold stratified cross-validation. Each of the models underwent a hyper-parameter fine tuning procedure described in detail in the Methods section. Men constitute 78% of the study participants. This poses a significant imbalance in the data and therefore, we measure performance using the area under receiver operating characteristic curve (ROC-AUC) which is robust against imbalance, as well as using the *F*₁ score that is sensitive to the imbalance. The value of ROC-AUC can be interpreted as the probability that the classifier is able to identify the female in of a male/female pair. The *F*₁ score is the harmonic mean of precision (what fraction of people identified as women are actually women) and recall (what fraction of women are identified as women) at a selected threshold. Results are summarized in for each classifier along with the corresponding values of the random classifier based on the imbalance present in the data. All classifiers after the hyper-parameter fine tuning procedure perform similarly well, with ROC-AUC values of 0.86 and *F*₁ scores of 0.5 and higher (compared to a random classifier with ROC-AUC of 0.5 and *F*₁ of 0.22). Next, we investigate the question of which behavioral features are most informative regarding gender. To do this, we use the feature importance obtained by fitting a random forest model to the data. We find that that a tendency towards gender homophily in the social networks is the most important behavioral feature; this is true for all three types of interactions that we consider. Some aspects of personality are also important. Within the big five traits, neuroticism and conscientiousness are most predictive, while narcissism and self-esteem are most powerful among the remaining personality tests. High on the list, we also find various communication related network characteristics. With respect to feature types, network indicators are the most important ones, occupying five of the top six indicators. # Discussion ## Conclusion In this work we have studied gender differences within a population of freshmen. We have been able to identify a number of gender differences in personality traits (measured via questionnaires), mobility patterns, as well as social network behavior based on person-to-person, telecommunication, and online social networks. *Personality*. Starting from gender differences with respect to personality, our findings are in accordance with observations in the psychology literature on gender differences. Discrepancies (or differences that are not significant) correspond to personality traits that, according to previous research, display ambiguous behavior over genders. *Mobility*. With respect to mobility behavior, our results are not consistent with findings in the literature. Previous work has found a restricted travel space for women, but we find that women travel more than men on average. *Networks*. Humans use multiple channels when we communicate: real-world conversations unfold when we meet person-to-person, we call each other and send text messages, we engage with Facebook posts and write comments, send email, and use other instant messaging platforms. Based on the communication channels we have access to here, we find that each of these channels plays a slightly different role with respect to gender similarities and differences. First, within all networks, we observe differences with respect to gender homophily, specifically an over-representation of female-only dyadic and triadic connections. This over-representation is mostly emphasized in weak links, which is consistent with the literature. Males show a lower level of homophily. For stronger social ties (that is, contacts require more effort to maintain), both genders show similar level of homophily. Second, we find that women tend to simply communicate more. On average, females maintain more contacts than males in the population, they exchange significantly more text messages, and talk longer on the phone. This is expressed both via a larger number of contacts as well as a higher entropy of neighbors. Furthermore, in agreement with a previous study on Facebook, we find that, in general, women have more central positions in the network, as expressed through a higher average betweenness centrality. *Predicting gender*. Finally, we use the features discussed above to predict gender based on personality and behavioral patters. The prediction task is based on combined personality, mobility and network features for each individual study participant, allowing us reveal the relative importance of each feature in predicting gender. We find that personal characteristics and social behavior can be used to identify the gender of an individual with high performance (*AUC* = 0.87). We find that network features are highly revealing, followed by personality test scores. ## Limitations Our results point out significant differences in various aspects of social behavior between males and females, based on a population of nearly 800 freshmen at a large European university. However, in order to have a clear understanding of the results, it is important to note the limitations of the dataset as well as the methods applied, which we address in the following paragraphs. ### Population sizes In our dataset, male population is around four times larger than the female population. This skewed female/male ratio presents may present certain biases simply because, in addition to gender differences, women may behave like a minority in some cases (e.g. with respect to homophily in the social networks). The female/male ratio also presents methodological challenges. We describe our approach to mitigate these issues, both at the individual level as well as for the network analysis in the Methods section. ### Demographics The cohort of the CNS experiment consists of Danish and international freshmen at a technical university. Population impose strong constraints on demographics, with respect to age and social embeddedness. Furthermore, we do not have detailed demographic information regarding the contacts the participants made outside the experiment. Although demographic information is necessary to understand the results, and individuals located in Denmark indeed display different personal behavior (for example, a low overall level of neuroticism), our results regarding the comparison of males and females are in agreement with existing literature on personality traits. ### Non-binary gender identification In this work, as in Copenhagen Networks Study in general, the participants reported their gender through a questionnaire that only offered two options: female and male. This limiting distinction might have contributed to additional noise in the measurement of differences as well as to lowering the performance of the models in the gender-inference task. # Materials and methods ## Ethics statement The Danish Data Protection Agency has approved the overall project structure (data collection, anonymization, and storage), as well as the content of the current study, cf Journal: 2012-41-066. The project complies with both local and EU regulations. All participants in the study have provided written informed consent. The data obtained from Facebook was collected in accordance with the Terms of Service. ## Network metrics We construct three types of networks representing the various interactions among the participants: physical proximity, Facebook, and call networks. We then aggregate them over time windows of one week. Only consenting participants of the CNS are included in the networks, since we do not possess complete information (e.g, gender and social activity) about the other contacts of the students. However, extending the ego networks of the students with individuals outside the experiment (for instance, in the call networks), we can extract additional descriptors about the participating students, such as their total number of contacts or distribution of mobile phone conversation times. In the present study we show statistics over different network metrics that are related to the local structure of the graphs and the position and role of participants in the global network. In the following, we provide a detailed description of the applied network metrics. 1. **Degree.** The number of contacts *k*_*i* an individual has in their respective social network. We calculate the degree in two different settings. First, by considering all contacts of a student, we can infer the total degree (without referring to the gender of the contacts). Second, by limiting the interactions to the participants, we calculate the degree describing same-gender contacts. 2. **Betweenness centrality.** This measure quantifies the importance of an individual with respect to information flow on the network, when the shortest paths are taken into account. It is defined as: $$\begin{array}{r} {C_{B}(i) = \sum\limits_{j \neq k \neq i}\mspace{180mu}\frac{n_{jk}^{\ell}(i)}{n_{jk}^{\ell}}.} \\ \end{array}$$ Here, $n_{jk}^{\ell}$ denotes the number of shortest paths between individuals *j* and *k* among which $n_{jk}^{\ell}(i)$ number of paths go through individual *i*. Therefore, betweenness effectively measures the fraction of shortest paths that pass an individual, which is a precursor of their relevance in case of any flow on the network (rumor propagation, spread of an infectious disease, etc). ## Imputation of missing values Due to the method of data collection some fraction of students has missing data in various channels. Overall, 21.5% of the participants exhibit missing data in at least one channel. To address the problem, we first remove participants with missing features in more than two of the five feature categories (personality, location, call, Facebook, and person-to-person interactions). We then apply a KNN based imputation to the remaining data, described as follows. For each user we find their *k*-nearest neighbors (with *k* = 7) by calculating the average difference of non-missing features with other users. We only use features that are present in the potential neighbor’s feature set, that is, if *L*_*uv* = *F*_*u* ∩ *F*_*v* denotes the set of overlapping features of users *u* and *v*, the distance between the users is given as: $$\begin{array}{r} {d_{uv} = \frac{1}{|L_{uv}|}\mspace{180mu}\sum\limits_{i = 1}^{n}\mspace{180mu}\left| x_{i}^{(u)} - x_{i}^{(v)} \right|} \\ \end{array}$$ where $x_{i}^{(u)}$ and $x_{i}^{(v)}$ are the values of the *i*-th feature for users *u* and *v* respectively, and \|*L*_*uv*\| is the size of the overlap set. Once the *k*-nearest neighbors of all users are determined, for each student we impute their missing values by the average of the corresponding non-missing feature values of their neighbors. If there is a single neighbor, their value is assigned. ## Fine tuning of the machine learning models We fine-tuned each of the models used in the gender prediction task. Through a grid search with cross validation we found the set of hyper parameters for which each model achieved the highest harmonic mean between *F*₁ score and ROC-AUC on previously unseen data. lists the parameter values in the grid. Optimal values are bold. ## Personality traits We consider eleven personality traits in the main paper, with five traits forming the Big Five Inventory. lists all the personality traits along with their definition and references to the corresponding literature. [^1]: The authors have declared that no competing interests exist.

# Introduction Industrial hemp (*Cannabis sativa* L.) has a rich history in the civilization of humans because it can provide both phytochemicals and lignocellulosic biomass. This crop originated in Eurasia and is useful all over the world, largely as a fiber crop. Following the emergence of more economically helpful fiber crops, the demand for hemp reduced, and the purpose shifted to usage as a food additive. Hemp seed contains essential fatty acids and proteins and gamma- linolenic acid, which has many health benefits. Hemp seeds and oils are also used to produce nutritional supplements and cosmetics. Recently, attention has been focused on its rich repertoire of pharmaceutical compounds. Hemp produces a diverse array of phytocannabinoids, terpenes, and phenolic compounds with prominent nutraceutical potential. Among them, phytocannabinoids are the most well-known phytochemicals. The predominant compound, cannabidiol (CBD) and tetrahydrocannabinol (THC) followed by cannabigerol (CBG) and cannabichromene (CBC) are highly promising compounds to improve the quality of human health. They act as therapeutic agents for central nervous system diseases such as epilepsy, inflammation, anxiety, and neurodegenerative disorders such as Parkinson’s, Huntington’s, Tourette’s syndrome, and Alzheimer’s. Terpenes present an array of pharmacological properties, including anxiolytic, antibacterial, anti-inflammatory, and sedative effects on human diseases. The increasing popularity of hemp-based phytochemicals has spurred the comprehensive analysis of genome and gene expression studies. These analyses are necessary to identify the genes involved in secondary metabolite pathways and have helped to discover transcription factors which are the key proteins that positively or negatively control the synthesis of secondary metabolites. Accurate gene expression studies such as Northern blotting, ribonuclease protection assay (RPA), serial analysis of gene expression (SAGE), and quantitative real-time PCR (qRT-PCR) are essential to confirm genomic and transcriptomic data. Among these methods, qRT-PCR is the most frequently used for gene expression analysis because of its high sensitivity, specificity, accuracy, reproductivity, and relatively low cost. qRT-PCR also requires a minimal amount of RNA compared to hybridization-based methods. The assessment of different samples in the same parameter is evaluated by qRT- PCR. The analysis is used to detect changes in the expression of genes of interest relative to a reference gene. Because of the variances incurred during RNA extraction, DNase treatment, and cDNA synthesis, the reliability of gene expression results can be affected by sample size, RNA degradation, reverse transcription efficiency, and cDNA quality. To provide accurate and reproducible results of gene expression profiles, researchers use reference genes as internal controls. Expression levels vary depending on different environmental conditions, making it critical to identify appropriate and reliable reference genes for each experimental set-up in the respective plant tissue and genotype to prevent biased or misinterpreted data. Mangeot-Peter et al. (2014) identified suitable reference genes in hemp stem tissue for accurate expression profiling of cell wall synthesizing genes. Subsequently, Guo et al. (2018) studied seven reference genes in various hemp tissues, such as root, stem, leaf, and flower. To our knowledge, there has been no reports on the suitability of reference genes for normalization of gene expression in hemp under different experimental conditions. The *F-box* gene is used for hemp gene expression analysis. However, its stability under stress conditions has not been analyzed leading to inaccurate normalization of qRT-PCR analysis. This study aims to evaluate stable reference genes under different abiotic stresses/hormone stimuli in hemp. # Materials and methods ## Plant material, greenhouse conditions, generation of clones, growth, and care Industrial hemp and medical marijuana plants share *Cannabis sativa* as their common scientific name. Therefore, in this paper, the authors referred to industrial hemp as “hemp”, to distinguish it from medical marijuana. The hemp strain, Thunderbird, was grown following the approved guidelines for industrial hemp provided by the Pennsylvania Department of Agriculture—Bureau of Plant Industry under the regulated permits IH-16-P-2017 and IH-17-P-2017. Greenhouse conditions were maintained at 25°C with a 14-hour light photoperiod at 25–40μEm^-2s^-1. Hemp clones were achieved by collecting a 3-inch segment containing two axillary buds and coating the 45-degree cut with Clonex Rooting gel (Hydrodynamics International, Inc. Lansing, MI). The explant was placed in Root Riot plugs (Hydrodynamics International, Inc. Lansing, MI) and maintained under propagation domes for two weeks at which point they were transferred to four-inch pots containing high porosity soil, HP Mycorrhizae from Pro-Mix (Rivière-du-Loop, Québec, Canada). Genetically identical clones of similar size were obtained by vegetative cuttings from the same female mother plant. Clones were kept under 24-hour light under propagation domes and 12-hour light during the pre-flowering and flowering periods. The temperature was maintained at 25°C. The humidity for rooting clones was maintained at 65% and decreased gradually to 45% once the clones started to flower. Lost Coast Plant Therapy (Plant Protector, Inc. Loleta, CA) was applied to the clones biweekly at a dilution of 30mL per 4 liters to control pests. ## Plant stress treatments All treatments except UV light treatment were performed in the greenhouse on the same day. Four-week-old, cloned plants grown in small pots were soaked in water including 100mM of mannitol (drought stress), 100mM of NaCl (salt stress), 200μM of CuSO₄, 100μM CdCl₂, or 100μM of Pb(NO₃)₂ and 200μM of ZnSO₄ (heavy metal stresses), and100μM abscisic acid (ABA), 100μM of methyl jasmonate (MeJA), 1mM of gibberellic acid (GA₃), or 100μM of salicylic acid (SA) (hormone treatments) for eight hours. For UV treatment, hemp cloned plants were exposed to UV-C radiation for 10 minutes. After each stress treatment, the 3^rd and 4^th leaves from the top of the plant were sampled and immediately frozen in liquid nitrogen and stored in the -80°C freezer until total RNA was extracted. All the treatments were performed in three biological replicates. For the mock plants, distilled water was used to soak the hemp plants. ## Total RNA extraction and cDNA synthesis Total RNA was extracted from 100mg of each plant sample using the Spectrum^™ Plant Total RNA kit (Sigma Aldrich, St. Louis, MO, USA). RNA concentration and absorbance ratios (A260/280 and A260/230) were measured using a NanoVue Plus spectrophotometer (General Electric Healthcare Limited, UK) to measure the quantity and quality of the total RNA. After treatment with DNase I (TaKaRa Bio, Dalian, China) to remove genomic DNA contamination, 2μg of total RNA was used to synthesize cDNA using the high-capacity cDNA reverse transcription kit (Applied Biosystems, Foster City, CA) according to the manufacturer’s protocol. ## Candidate reference genes selection, primer design, and PCR reaction We identified 13 candidate reference genes and two target genes using a BLAST search from NCBI (<https://www.ncbi.nlm.nih.gov/>) and the *Cannabis sativa* genome browser gateway (<http://genome.ccbr.utoronto.ca/cgi-bin/hgGateway>). The *Cannabis sativa* genome browser gateway is based on the Purple Kush strain of medical marijuana. Primers were designed based on the sequences of 13 genes using Primer3 Plus (<http://www.bioinformatics.nl/cgi- bin/primer3plus/primer3plus.cgi>) with the criteria: amplicon size 80–200bp, primer size 18–24bp, Tm 60°C, GC content 45–60%. All primer sequences are listed in. ## qRT-PCR amplification PCR was performed using cDNA as the template to confirm the specificity of the primers to the target genes. Using a 2% (w/v) agarose gel, all PCR products were analyzed using electrophoresis to confirm a single band of the expected size for each of the primer pairs. To test the PCR amplification efficiency, the regression coefficient (R²) for each gene was calculated using a standard curve generated from a fivefold dilution series of cDNA (1, 1/10, 1/100, 1/1000, and 1/10000) for each primer pair. Based on the slopes of the standard curves, the PCR efficiency of each gene was determined from the respective logarithm of the cDNA dilution and plotted against the mean threshold cycle (Ct) values. The PCR efficiency was calculated using the equation: $$\text{E}\mspace{720mu}(\%) = \left( {10 - \left( {1/\text{slope}} \right) - 1} \right) \times 100,$$ where E is the efficiency, and the slope is the gradient of the best-fit line in the linear regression. qRT-PCR was performed with 5μL of SYBR Select Master Mix (Applied Biosystems, Waltham, MA, USA) in a 10μL total reaction mixture containing 400nM of the gene-specific primers and 1μL of cDNA. PCR reaction was performed using a Bio-Rad CFX96 system (Bio-Rad, Hercules, CA, USA) under the following reaction conditions: Initial denaturation at 95°C for 10 minutes, 35 cycles of 95°C for 10 seconds, and 60°C for 1 minute. Three technical replicates were used for each biological replicate and average Ct was used for data analysis. As a negative control, water and total RNA were used instead of cDNA to confirm that there was no amplification from contaminated DNA or hemp genomic DNA. ## The stability of reference genes and statistical analysis Boxplots of quantitative cycle (Cq) values for the 13 candidate reference genes were depicted in all leaf samples with every treatment using the boxplot R package to show the variation of each gene expression. The expression of 13 reference genes was analyzed under 11 different stresses using four algorithms, geNorm, NormFinder, BestKeeper, and RefFinder to rank the stability of the candidate reference genes. The pairwise variation (Vn/Vn+1) between two sequential normalization factors was calculated with geNorm to determine the optimal number of candidate reference genes for accurate normalization. ## Validation of identified reference genes Two cannabinoid pathway genes, *CBDAS* and *AAE1*, were used as target genes to validate the reliability of the selected reference genes using the most stable candidate reference genes and the least stable reference genes. Primer design and calculation of PCR amplification efficiency for these genes was performed as described above. Relative gene expression levels of *CBDAS* and *AAE1* were calculated using the 2^−ΔΔCt method (28). Statistical analysis was performed using a paired t-test (a = 0.05) (28). # Results ## PCR specificity and amplification efficiency of the candidate reference genes Thirteen reference genes (*18S*, *40S*, *CHAL*, *UBE2*, *EF-1α*, *F-box*, *GAD*, *PCS1*, *PP2A*, *SAND*, *TATA*, *TIP41*, and *TUB*) were identified from NCBI and the *Cannabis sativa* genome browser gateway based on a homology search with *Arabidopsis* genes. Primers were designed and used to confirm their specificities based on their amplification efficiency and specificity. Single bands were amplified in agarose gel electrophoresis for all the gene primers with predicted sizes. For the qRT-PCR amplification, the PCR efficiency (%) ranged from 91.22 to 113.87, and the regression coefficient (R²) varied from 0.9893 to 0.9994. ## Ct values of candidate reference genes Transcript abundances of 13 candidate reference genes were assessed by qRT-PCR for each gene, tested in triplicates across all 11 treatments and a control, which was 36 biological samples. A majority of the candidate reference genes Ct values ranged from 20 to 30. The lowest expression level with the highest Ct values between 27.4 and 32.1 was *PCS1*. The *EF-1α* gene showed the highest expression level, with the lowest Ct values ranging from 18.9 to 25.3. The *CHAL* gene displayed the highest difference among all 36 samples tested, with a minimum Ct value of 22.3 and a maximum Ct value of 30.9. These Ct value analyses showed that the transcription levels of candidate reference genes are unstable under different stress conditions. ## Analysis of reference genes by geNorm The geNorm was used for evaluating the expression stability of the 13 candidate reference genes. Data analysis was calculated based on individual 11 different treatments and three different groups of treatments such as osmotic stress (OS: mannitol, NaCl), heavy metal stress (HM: CdCl₂, CuSO₄, PbNO₃, ZnSO₄), and hormonal stimuli (PH: ABA, GA₃, MeJA, SA). The total ranking was also shown by combining all 11 treatments together. This algorithm evaluated the stability of reference genes (M) based on the average pairwise variation of all tested genes. In this analysis, the lower the M value, the more stable the gene expression. A reference gene that has an M value less than 1.5 is used for qRT-PCR. *PP2A* and *TIP41* were the most stable reference genes with the lowest M value (0.46) whereas *CHAL* had an M value of 1.07 and was ranked as the least stable gene. Individually, *EF-1a* and *SAND* were the most stably expressed genes under osmotic stresses with an M value of 0.22 while *F-box* and *TATA* were the least stably expressed genes. The *TUB* and *TATA* genes showed the lowest M values of 0.16 among all of the heavy metal stressed clones. Exposure to hormonal stimuli resulted in *PP2A* and *F-box* to be the most stable with an M value of 0.27 and *CHAL* to be the least stable with an M value of 0.75. *F-box* was ranked as the second least stably expressed gene under both osmotic and heavy metal stresses, but it was ranked as the first and second most stable gene under UV and plant hormone treatments, respectively. The *TATA* gene was least stably expressed under osmotic stresses but was among the top two and three under heavy metal stress and plant hormone stimulus, respectively. *CHAL* was the least stably expressed in response to UV light application. geNorm can determine the minimal number of reference genes that should be used to obtain an accurate normalization. The optimal number of reference genes was determined based on the pairwise variation (Vn) between two normalization factors (NFn) composed of an increasing number of reference genes. The threshold value (Vn/Vn+1 = 0.15) indicates if the number of reference genes less than or equal to the value of n is sufficient to use as a reference gene. As shown in, the pairwise variation value V2/V3 of all experimental samples was less than 0.15, demonstrating that two reference genes should be sufficient for normalization under all conditions tested. ## Analysis of reference genes by NormFinder NormFinder is a quantity-model-based software and uses complex statistical models to compute the variation between the expression of genes across different biological groups. The lowest expression stability value represents the most stable reference genes. Results from the NormFinder analysis are summarized in. The EF-1α and TUB genes were the most stably expressed in all samples and were ranked as fourth and third by geNorm, respectively. The F-box, TATA, and CHAL were ranked as the three least stable genes both by NormFinder and geNorm. The TUB, PCS1, and TATA genes were the most stably expressed under osmotic stress, heavy metal stress, and plant hormone stimuli, respectively. Compared to geNorm, TUB, PCS1, and TATA were ranked as third, sixth, and third positions in each category, respectively. The least stably ex-pressed reference genes under osmotic stress (CHAL, F-box, TATA), heavy metal stress (CHAL, F-box, TATA), and plant hormone stimuli (40S, GAD, CHAL) had similar rankings when compared to geNorm rankings. The GAD and F-box genes were found to be the most stable reference genes under UV stress while PCS1 and CHAL were the least stable. A similar trend was observed in the geNorm analysis. ## Analysis of reference genes by BestKeeper The BestKeeper program is an excel-based algorithm and uses standard deviation (SD) and coefficient of variation (CV) data of the average Ct values for specific treatments. Lower CV ± SD values represent higher stability. When using the BestKeeper algorithm, genes with an SD value \> 1 are undesirable reference genes. When all samples were taken into consideration, *TATA* (SD = 0.74), *40S* (SD = 0.79), *PCS1* (SD = 0.84), EF-1a (SD = 0.90), and TUB (SD = 0.99) were determined to be reliable reference genes. *TATA* showed the lowest SD among all 13 reference genes in all samples and the SD values were greater than 1 in osmotic stress (1.29) and mannitol (1.82). The 40S gene was ranked as the second most stable candidate in all samples tested, but the SD value of 40S under NaCl and *PbNO*_*3* stresses were 1.09 and 1.36, respectively. *PCS1* was ranked at the third position in all samples tested and SD values were below 1 in any individual treatment and the three treatment groups. The *CHAL* gene displayed the highest SD value with 1.95 in all samples displaying that this gene is not ideal for gene expression normalization. The *CHAL* gene exhibited an SD value less than 1 only under GA₃ (0.71), ABA (0.27), CdCl₂ (0.58), and CuSO₄ (0.98) treatments. ## Analysis of reference genes by RefFinder RefFinder is a web-based tool for comprehensive analysis that integrates geNorm, NormFinder, Delta Ct, and BestKeeper approaches. The reference genes were ranked from the most stable (M value is the lowest) to the least stable expression (M value is the highest) using RefFinder. Among them, the most stable candidate was the EF-1α gene, followed by the TUB gene in all samples. The EF-1α and TUB genes were also ranked in third and first places under osmotic stress conditions, respectively. The TATA gene was most stably expressed under heavy metal and plant hormone treatments while this gene was the least stable under osmotic stress. The CHAL gene was ranked as the least stable gene in all samples tested. The GAD and CHAL genes were the most and least stably expressed genes respectively under UV application, which was the same findings as to the NormFinder software. ## Validation of selected reference genes To validate the selected reference genes, gene expression levels of AAE1 and CBDAS were measured. Each of the two most stable reference genes, EF-1α and TUB, a combination of these two stable reference genes (EF-1α+TUB), and the least stable reference gene (CHAL) were used as internal controls. AAE1 expression was significantly reduced under drought (Mannitol) and salinity (NaCl) stresses. EF-1α, TUB, and a combination of EF-1α and TUB were used for normalization of qRT-PCR analysis. There was no significant difference in the AAE1 expression between the mock treatment and osmotically stressed samples (Mannitol and NaCl) when CHAL was used as an internal control. The expression of CBDAS was also reduced under osmotic stresses when expression data was normalized with EF-1α, TUB, and a combination of EF-1α and TUB unless the CBDAS expression under NaCl stress was normalized with the TUB gene. When CHAL was used as a reference gene, CBDAS gene expression was reduced under mannitol treatment but there was no difference between mock and NaCl treatments. # Discussion Industrial hemp is from the plant species *Cannabis sativa* and has gained importance as a medicinal crop because of its potential to produce secondary metabolites such as cannabinoids, terpenes, and phenolic compounds. According to Schluttenhofer and Yuan (2017), hemp was cultivated for commercial or research purposes in at least 47 countries in 2017 and the global hemp market doubled from the year 2016 to 2020. Recently, a comprehensive gene expression analysis is aimed at elucidating the metabolic pathways for cannabinoids and terpene synthesis to improve hemp traits. To validate this data, qRT-PCR analysis is suitable, however, appropriate hemp reference genes for accurate gene expression analysis have not been well established. In this report, we evaluated 13 hemp reference genes under 11 different stress conditions. Research in other plant species has revealed that different environmental conditions would require unique reference genes to accurately interpret expression levels. Eleven different conditions including osmotic stresses, heavy metal stresses, plant hormone stimulus, and UV light application were reported to affect the cannabinoid synthesis. The results obtained from geNorm, NormFinder, BestKeeper, and RefFinder were not consistent, particularly BestKeeper which was much more distinct from the other software methods (Tables –). This finding was expected because the BestKeeper algorithm evaluates data differently when compared to the three other programs. To rank the most suitable reference genes across all treatments, there was no unanimity when compared to four different algorithms, which represented the combined results obtained from four programs. In most cases, one candidate gene was ranked as the most stable gene by two or three programs, which showed that it might be a good reference gene under various treatments. Based on the combined rankings of the four programs used in our study, the overall results showed that the most stable genes varied while the least stable genes were almost the same. Across all plants tested, both NormFinder and RefFinder determined EF-1α as the most stable gene in all samples tested. In previous reports, EF-1α was demonstrated to be the most stable gene under different stresses in a variety of crops such as tobacco, maize, soybean, potato. Interestingly, this gene was not the most stable in any of the three groups (OS, HM, PH). The TUB gene appeared to be best the candidate under osmotic stresses because this gene was ranked as the most stable by both NormFinder and RefFinder which is consistent with the results obtained in Parsley under abiotic stresses. Under heavy metal stress, *TATA* was ranked as the most stable gene by BestKeeper and RefFinder and the second most stable gene by geNorm. This gene was also ranked as the best reference gene in hormone stimuli by NormFinder and RefFinder. Interestingly, *TATA* was the least stable gene under osmotic stresses by geNorm, NormFinder, and RefFinder. *TUB* was the most stably expressed gene under osmotic stresses, whereas *TATA* was ranked as the best stable gene under both heavy metal stress and hormone stimuli. Unlike most stable genes, *CHAL* was found to be the least stable gene in most of the rankings with all samples and the three treatment groups (OS, HM, PH) when analyzed by all four programs. According to Wang et al. 2015, candidate genes showing a high level of variation of Ct values should be avoided as internal controls. Our results showed that variation of the Ct value in *CHAL* was highest among all 13 reference genes, which is consistent with the fact that *CHAL* was ranked as the least stable by all four programs used in this study. In previous *Cannabis* qRT-PCR studies, the *F-box* gene has been used as an internal control for qRT-PCR. Mangeot-Peter et al. (2016) performed the reference gene analysis in hemp stems and concluded that the *F-box* gene was ranked as one of the most stable genes and *Histone 3* as the least stable gene among 12 reference genes tested under normal conditions. In this study, however, the *F-box* gene was the second least stable gene by RefFinder and the third least stable gene determined by both the geNorm and NormFinder programs when all samples were analyzed. Based on our group rankings (OS, HM, PH), *F-box* was ranked the second least stable genes by geNorm, BestKeeper, and NormFinder under both osmotic and heavy metal stresses. The *F-box* gene was stably expressed under normal conditions in hemp leaves and relatively stable under hormone stimuli as evident by its second position as ranked by both geNorm and RefFinder. These results show that *F-box* may not be a suitable reference gene for hemp qRT-PCR analysis under osmotic and heavy metal stresses. However, it could be acceptable as a reference gene under normal and plant hormone treatments. Overall, our study suggests that the *F-box* gene may not be the best reference gene for *C*. *sativa*, particularly in plant stress-related studies. Guo et al. (2018) have studied the stability of reference genes in different hemp tissues/organs. They ranked ubiquitin and *EF-1α* as the most stable genes in leaf samples at different stages, and *PP2A* as the least stable gene in different organs. Notably, *EF-1α* was the most stable reference gene in our global ranking, showing that *EF-1α* is most stable under the normal condition and different abiotic stresses and hormonal stimuli. Many studies have proved that the use of more than one reference gene enables the possibility of avoiding variations and achieving more accurate normalization of qPCR data. To assess the optimal number of reference genes for the normalization of qRT-PCR data, we used the geNorm program to perform a stepwise calculation of the pairwise variation (Vn/Vn+1) between sequential normalization factors. In this analysis, a Vn/Vn+1 \< 0.15 indicates that introducing an additional reference gene for normalization is unnecessary. Under all treatments, V2/V3 values were less than 0.15, which indicated that two reference genes were enough for the normalization of the real-time PCR data under any treatments in this study. To validate the reliability of the selected reference genes, we measured the relative expression of two cannabinoids pathway genes using *EF-1α* and *TUB* as the most stable reference genes and *CHAL* as the least stable reference gene. Since CBDA content is decreased by the influence of osmotic stress, we measured the expression of *AAE1* and *CBDAS* genes that are involved in the rate- determining enzymatic reactions leading to CBDA synthesis under drought and salinity stresses. The expression of these two genes was significantly reduced under drought and salinity stresses when qRT-PCR data were normalized by *EF-1α*, *TUB*, and the combination of *EF-1α* and *TUB*. Notably, the expression level of both genes was normalized by *CHAL* under salinity stress and did not show a significant difference when compared with mock plants. These results suggest that *EF-1α* and *TUB* genes individually or in combination are suitable reference genes for hemp under osmotic stresses. Our validation study demonstrated the effectiveness of the ranking of reference genes by the programs used geNorm, NormFinder, and RefFinder. To the best of our knowledge, this study is the first report that performed a systematic analysis of hemp reference genes under different abiotic stresses and hormonal stimuli. The knowledge obtained in this study could contribute to enhancing future hemp research related to the elucidation of mechanisms involved in the synthesis, transport, and accumulation of abundant secondary metabolites in hemp. # Supporting information We would like to thank Dr. Yuka Imamura, Ph.D. of Penn State College of Medicine Genome Sciences and Bioinformatics Facility for data analysis services. The authors would like to thank Ms. Jessica Wolf for proofreading the manuscript. [^1]: The authors have declared that no competing interests exist.

# Introduction Very long chain fatty acids (VLCFAs) are fatty acids (FA) with an acyl chain longer than 18 carbons. They are components of a large variety of plant lipids like the membrane lipids phospholipids and sphingolipids, the storage lipids triacylglycerol and the hydrophobic lipid barrier comprising cuticular waxes and suberin. VLCFAs are elongated in the endoplasmic reticulum (ER) by the elongase complex that sequentially adds two carbons to long chain acyl-CoAs (16 or 18 carbons) originating from *de novo* synthesis in the plastids. The elongase complex includes four enzymes starting with, the 3-keto-acyl-CoA synthase (KCS) that condensates the acyl-CoA with a malonyl-CoA to form a 3-ketoacyl-CoA intermediate that is in a second step reduced by the β-ketoacyl-CoA reductase (KCR) in 3-hydroxyacyl-CoA. The 3-hydroxyacyl-CoA dehydratase (HCD) then dehydrates the 3-hydroxyacyl-CoA in trans-2,3-enoyl-CoA that is finally reduced by the fourth enzyme, the trans-2,3-enoyl reductase (ECR). The acyl-CoA elongated by two carbons can re-enter an elongation cycle to eventually produce VLCFAs ranging from C18 to C32 in Arabidopsis. In Arabidopsis, 21 *FAE-like/KCS* genes grouped in 8 distinct subclasses encode the condensing component of the elongase complex. The different KCS are characterized by their substrate (acyl chain length) and tissue specificities. The three other elongase subunits show a much lower gene diversity in Arabidopsis. Two genes, *KCR1* and *KCR2* are homologous to yeast *KCR YBR159*. However, only KCR1 is able to restore elongase activity in *ybr159* yeast mutant. Similarly, Arabidopsis genome presents two genes with similarity to yeast 3-hydroxyacyl-CoA dehydratase *PHS1*, *PASTICCINO2* (*PAS2*) and *PROTEIN TYROSINE PHOSPHATASE-LIKE* (*PTPLA*). Like *KCR1* with *ybr159*, only *PAS2* was able to complement null yeast *phs1* mutant. Finally, ECR is encoded by *CER10* that complements the *tsc13* yeast mutant. Beyond these models model species for fungi and plants, HCD-encoding genes are important for human and dog health or basidiomycete survival. VLCFAs are essential lipids since all the mutations in yeast and plants preventing acyl-CoA elongation result in lethality. *A*. *thaliana kcr1* and *pas2* loss of function mutants led to global decrease of the VLCFA in the different lipid pools and to embryo lethality. Silencing of the tobacco ECR in leaves leads to necrotic lesions and epidermal cell ablation. Cell death could also be observed in plants with ectopic expression of seed specific KCS *FAE1* in the epidermis indicating that the nature and amount of VLCFA are important for cell viability. Likewise, enhancement of VLCFA levels altered plant development as illustrated by the KCS FAE1 or yeast PHS1 overexpression in Arabidopsis. VLCFA were directly involved in cell differentiation and lateral root organogenesis by promoting polar auxin transport in the *pas1* mutant. Sphingolipids are most likely involved in polar auxin transport since ceramide synthase mutants *loh1/loh3* also showed a reduced auxin-dependent lateral root initiation. Defective development associated with unbalanced VLCFA/LCFA ratio was often observed with some membrane defects. Reduced VLCFA elongation impaired cell elongation and division especially membrane trafficking during cell plate formation, but also altered Fts-Z assembly during plastid division. Specific depletion of VLCFA in sphingolipids induced also membrane trafficking and cytokinesis defects that could be related to enhanced stability in endosomal transient interactions (Markham et al., 2011, Molino et al., 2014). *In vitro* experiments directly demonstrated the importance of acyl chain length of the sphingolipid glucosylceramide in liposome fusion. Apart from being structural components of membrane lipids, VLCFAs have also other functions in plant development. VLCFA are essential components of cuticular and epicuticular waxes that were responsible, when missing, for post- genital organ fusion. Interestingly, reduced FA elongation by mutation of cytosolic acetyl carboxylase PAS3 or the VLCFA dehydratase PAS2 was correlated with cytokinin hypersensitivity and cell proliferation and recently VLCFA were described as potential non-cell autonomous regulators of plant development by repressing cytokinin synthesis. In yeast, partial inactivation of FA elongation led to biochemical and cytokinesis defects similar to those observed in Arabidopsis. Yeast *phs1* and Arabidopsis *pas2-1* mutants showed reduced Acyl-CoA elongation, associated with 3-hydroxyacyl-CoA accumulation, and an increase in free sphingoid base like phytosphingosine (PHS). To identify new components of FA elongation, we took advantage of these similarities to carry out a suppressor screen of a leaky *phs1* strain (*Tet-PHS1)* with an *A*. *thaliana* cDNA library. We identified PTPLA as a suppressor of the *Tet-PHS1* yeast strain that was able to restore both the *Tet-PHS1* yeast growth and the FA elongation defects. PTPLA was however not able to rescue the developmental defects of the Arabidopsis *pas2-1* mutant but could further enhance FA elongation in presence of an active PAS2. The loss of *ptpla* function was characterized by 3-hydroxyacyl-CoA accumulation as expected for a FA elongase dehydratase but surprisingly also led to the accumulation of VLCFA. The non-overlapping expression pattern between the two dehydratases led us to propose the existence of a second elongase complex associated with PTPLA that was involved in repressing the activity of the major elongase complex comprising PAS2 dehydratase. A plant like Arabidopsis would thus have two different elongase complexes functionally interacting in adjacent cell tissues. # Results ## Arabidopsis PTPLA rescues the *S*. *cerevisiae acyl-CoA dehydratase Tet-PHS1* mutant To identify new genes able to suppress VLCFA depletion defects, yeast *Tet-PHS1* mutant was transformed with an Arabidopsis cDNA library. Since null *phs1* mutation is lethal, an inducible strain was used *(Tet-PHS1)*. A total of 698 clones growing on selective medium were selected, sequenced and confirmed in a second screen. Two cDNAs were identified as strong suppressor of *Tet-PHS1*. As expected, *PHS1* ortholog *PAS2* (AT5G10480) corresponded to the majority of the yeast clones (457 clones) but a second related cDNA, *PTPLA* (AT5G59770) was also identified in 35 clones. PTPLA is closely related to PAS2 and PHS1 (respectively 32% and 35% of identity). The PHS1 protein has six transmembrane domains, a C-terminal retention signal to the ER and a dehydratase domain that has been shown to have three essential amino acids necessary for PHS1 dehydratase activity. PTPLA and PAS2 proteins showed respectively five and four putative transmembrane domains. Both proteins presented also a retention signal to the ER (KXKXX or KKXX) and the three conserved and essential amino acids required for dehydratase activity. A previous study demonstrated that PAS2 was able to complement a null-*phs1* mutant while PTPLA could not, suggesting different activity between the two proteins. The absence of complementation of null-phs1Δ strain by PTPLA was confirmed with the clones isolated in TET-PHS1 screen. The phenotype of the PTPLA complementation of *Tet-PHS1* mutant was thus more carefully evaluated. First, PTPLA was able to restore the growth of the *Tet-PHS1* strain in presence of doxycycline (*Tet-PHS1*+DOX) to levels comparable to *Tet-PHS1*+DOX strain transformed with *PHS1* or *PAS2* cDNA albeit the kinetics of growth was slower. The absence of the dehydratase PHS1 blocked fatty acid elongation and led to reduced VLCFA levels in yeast. As a corollary, phytosphingosine (PHS) level was enhanced since VLCFA are required for sphingolipids synthesis. *PTPLA* expression in *Tet-PHS1*+DOX was able to reduce PHS levels, and induce VLCFA elongation to levels similar to what was observed for *Tet-PHS1*+DOX expressing PHS1 or PAS2. For example, C26 amounts were increased by 5.4-fold in *Tet-PHS1*+DOX expressing *PTPLA* that is comparable to the ratio observed in *Tet-PHS1*+DOX expressing *PHS1* or *PAS2* (respectively of 5.2 and 3.7). Interestingly, a similar increase of VLCFA amounts was observed in wild-type R1158 yeast strain expressing *PTPLA* with more than a two-fold increase that was comparable to the effect of *PHS1* and *PAS2* expression in wild-type R1158 yeast strain ( and). Finally, the hallmark of acylCoA dehydratase deficiency in yeast and in plants is the accumulation of the precursors, the 3-hydroxyacyl-CoAs. The expression of PTPLA in *Tet- PHS1*+DOX strain led to the reduction of 3-hydroxy C20-CoA accumulation to the same extent as what was observed for *Tet-PHS1*+DOX strain expressing PHS1. All these data indicate that PTPLA was able to rescue PHS1 deficiency in the *Tet- PHS1*+DOX strain. The fact that PTPLA could not complement null-*phs1* strain would suggest that a minimal endogenous dehydratase activity was necessary for PTPLA suppressing activity. ## II PTPLA does not complement *A*. *thaliana pas2-1* mutant but enhances VLCFA levels The lack of complementation of a null allele of yeast *phs1* could be caused by some plant specific determinants of PTPLA activity. We thus evaluated if PTPLA was able to complement *Arabidopsis thaliana pas2-1* mutant, which has a reduced dehydratase activity associated with strong developmental defects. The disruption of VLCFA elongation in *pas2-1* mutant induces cell proliferation and abnormal cytokinesis leading to defective differentiation in the apical part and shorter primary root. These developmental defects were linked with reduced VLCFA levels in triglycerides, waxes, sphingolipids and phospholipids. Moreover, the complete loss of PAS2 function is embryo lethal. *PTPLA* was thus expressed in the heterozygous *pas2-1*/+ plant under the control of either the 35S or *PAS2* promoters. Segregation of *pas2-1/+* plants carrying either *35S*:*PTPLA* or *pPAS2*:*PTPLA* constructs showed around 25% *pas2-1/pas2-1* mutants in T2 progeny indicating an absence of complementation of *pas2-1* phenotype whereas *pPAS2*:*PAS2* totally rescue *pas2-1* phenotype. Correct *PTPLA* or *PAS2* expression and tissular localisation were controlled by quantitative RT-PCR (qRT-PCR) and by the observation of GFP-PAS2 or GFP-PTPLA fluorescence. *PTPLA* expression under *pPAS2* promoter did not increase VLCFA levels in *pas2-1* mutant while *pPAS2*:*PAS2*, completely rescued VLCFA deficiency. Interestingly, *pPAS2*:*PTPLA* or *p35S*:*PTPLA* expression in wild-type led to a significant increase of VLCFA content as seen in yeast even if no clear overexpression of *PTPLA* transcripts could be observed. These data suggest that ectopic expression of PTPLA was sufficient for enhancing VLCFA synthesis in wild-type context but was not able to functionally replace defective PAS2 dehydratase. ## III *PTPLA* is specifically expressed in root vascular tissues To better understand the difference between PTPLA and PAS2 in fatty acid elongation, expression patterns of several *pPAS2* and *pPTPLA* reporter fusions were compared in stable transgenic lines (promoter size and number of lines are described). *pPTPLA*:*GUS* staining was specifically localized in mature primary and secondary roots and restricted to the central cylinder. *pPAS2*:*GUS* staining was also present in mature roots and secondary roots but also expressed in the epidermis of cotyledons and leaves as previously described. Interestingly, in mature primary roots *pPTPLA*:*GUS* appeared to be restricted to vascular tissue while *pPAS2*:*GUS* expression profile was specific to the endodermis. Stable co-expression of *pPTPLA*:*mRFP1* and *pPAS2*:*GFP* in *Arabidopsis thaliana* showed clearly that the two genes have non-overlapping expression profiles. *pPTPLA*:*mRFP1* showed a continuous expression in vascular tissue from the differentiation zone of the meristem to the root-hypocotyl junction while *pPAS2*:*GFP* was only expressed in the endodermis, first in few cells leading to a patchy staining and eventually in every endodermal cell. The non-overlapping and specific expression patterns of *pPTPLA* and *pPAS2* suggested a spatial regulation of VLCFA synthesis. The condensing enzyme KCS, the first enzyme of the elongase complex is encoded by a large gene family which presents a different expression profiles. Several KCS transcripts are expressed in the roots and at least KCS2 and KCS20 were specifically expressed in root endodermis. In a similar way to the dehydratase, the 3-ketoacyl-CoA reductase is encoded by two genes (KCR1 and KCR2) but only KCR1 was able to complement yeast *ybr159* mutation. Our intention was to examine if *KCR1* and *KCR2* genes have similar expression profiles to *PAS2* and *PTPLA*. Analysis of GUS expression in stable transgenic lines expressing *pKCR1*:*GUS* and *pKCR2*:*GUS* showed different expression patterns in the root that matched those of *pPAS2*:*GUS* and *pPTPLA*:*GUS* respectively. *KCR1* and *PAS2* promoters showed expression in cotyledons, leaves and a specific staining in the endodermis of the roots. *pPTPLA*:*GUS* and *pKCR2*:*GUS* stained vascular tissues of mature primary and secondary root. Contrary to the *PTPLA* promoter, the *KCR2* promoter was also expressed in cotyledons, leaves and in the meristem of secondary roots. We then checked whether restricted *PTPLA* expression was associated with root development. The *ptpla* mutant was characterized by a T-DNA insertion in the seventh intron leading to at least 90% reduction of a truncated *PTPLA* mRNA ( and Figs). Contrary to *pas2-1*, which showed shorter primary roots compared to wild type, the length of the primary root as well as the number of lateral roots were not altered in the *ptpla* mutant. However, the primary root was slightly longer in the double *pas2-1/ptpla* mutant compared to the single *pas2-1* mutant suggesting that the absence of *PTPLA* partially rescued *pas2-1* root growth. The phenotype was observed at two developmental stages in two independent experiments in 10 and 14 day-old seedlings respectively. ## IV PTPLA associates with the elongase complex in the ER In plants, the elongase complex is localized in the ER. Subcellular localization of PTPLA was first characterized with the transient expression of *35S*:*mCherry-PTPLA* constructs in *Nicotiana benthamiana*. The subcellular distribution of mCherry-PTPLA showed the characteristic ER network which was confirmed by its colocalization with ER localized GFP-PAS2 fusion. To investigate whether PTPLA was directly associated with enzymes of the elongase complex, *in vivo* protein-protein interaction assays were carried out by Bimolecular Fluorescence Complementation (BiFC) experiments. PTPLA and different subunits of the elongase complex were fused with C_YFP- or N_YFP- at the N terminal of the proteins of interest to prevent potential interference with ER retention signal located at the C-terminal end and transiently expressed in *Nicotiana benthamiana* leaves. The different combinations of proteins are summarized in. Direct interaction of PTPLA with core elongase subunits KSC6, KCS10 and the reductase CER10 could be observed in the ER but also with, the potential elongase chaperone, the immunophilin PASTICCINO1. All interactions observed by confocal microscopy are shown in. PTPLA and PAS2 interacted with the same KCS like KCS5, 6, 8, 9, 10 and 18 but also showed specific association with KCS1 for PAS2 or KCS 11, 13, 15 and 17 for PTPLA. No interaction could be observed with 6 KCS including KCS2 for both PTPLA and PAS2 suggesting that this assay might be too stringent for evaluating some protein elongase associations. Most of the elongase subunits interacting with PTPLA and with a known expression profile were expressed in the root. Interestingly, PTPLA could also interact with itself and with PAS2 suggesting that acyl-CoA dehydratase could form homo- or hetero-dimers within the elongase complex in p35S:PTPLA or pPAS2:PTPLA transgenic lines. Interaction assay with KCR1 or KCR2 could unfortunately not be carried out since KCR fusion proteins were not correctly expressed with this transient system. In vivo split- luciferase assays confirmed that PTPLA homo or hetero-dimerization with PAS2 were as strong as the interaction with another complex subunit like CER10 indicating a direct involvement of PTPLA in the elongase complex. ## V *ptpla* mutant accumulates hydroxyacyl-CoA and VLCFA The PTPLA sequence is as divergent from PAS2 sequence as it is from yeast PHS1 but it has conserved essential amino acids for the catalytic activity, suggesting a potential function of PTPLA as a 3-hydroxyacyl-CoA dehydratase. The hallmark of the dehydratase deficient mutant is the accumulation of 3-hydroxyacyl-CoA precursors as characterized in the *pas2-1* mutant. Acyl CoA quantification by LC ESI-MS/MS in root extracts of the *ptpla* mutant confirmed the accumulation of C18, C20 and C22 3-hydroxyacyl-CoAs with respectively 3.0, 4.4, and 1.6 fold the levels observed in wild type. The 3-hydroxyacyl-CoAs accumulations in *ptpla* were however much lower to that observed in *pas2-1* mutant (inset). We then investigated if *ptpla* loss-of-function would also reduce VLCFA elongation as observed in *pas2-1*. But contrary to *pas2-1*, the *ptpla* mutant showed a modest but significative increase in VLC/LCFA ratio in seedlings roots but not in the apical part. Wild type levels of VLCFA were partially or totally restored in the *ptpla* mutant expressing *pPTPLA*:*PTPLA* construct confirming that the increase in VLCFA content was caused by the loss of PTPLA function. The degree of observable complementation was correlated with the level of *PTPLA* transcripts. The increase of VLCFA levels induced by the absence of PTPLA function was however dependent on the presence of PAS2 activity since the *ptpla* effect was abolished in *pas2-1/ptpla* double mutant. These data indicate that PTPLA is first, directly involved in fatty acyl elongation as a 3-hydroxyacyl-CoA dehydratase and second, is involved in regulating PAS2-associated fatty acid elongation. ## VI A specific elongase activity in vascular tissues regulates endodermal VLCFA elongation The fact that *PTPLA* and *PAS2* had non-overlapping roots expression profiles with respectively vascular tissue/pericycle and endodermis specificity suggests that two different elongase complexes are coexisting in adjacent root cells. To confirm that the vascular/pericycle PTPLA-associated complex was different from the endodermal PAS2-associated complex, we compared the effect of ectopic expression of PTPLA and PAS2 on VLCFA levels. When expressed in the endodermis (under the control of PAS2 promoter), both proteins induced an increase in VLCFA levels but when expressed in vascular tissue (under the PTPLA promoter) weaker difference with wild type could be observed. This result indicates that the elongase activity is stronger in the endodermis and that the low vascular/pericycle elongase activity was not caused by a less active PTPLA but rather by limiting quantity of elongase partners, since *PAS2* expression was not sufficient to induce higher VLCFA accumulation. The presence of a structurally and functionally different PTPLA-associated elongase complex is reinforced by the fact that like PTPLA, KCR2 could not complement the null yeast ortholog mutant *ybr159* and that *pKCR2*:*GUS* was also specifically expressed in vascular tissues in the root. Interestingly, the *bona* fide KCR1 was expressed in the endodermis like PAS2. Fatty acid analysis of *kcr2* mutants showed that VLCFA levels were also increased compared to wild type. In conclusion, two elongase complexes with potentially different catalytic properties coexist and more importantly functionally interact in adjacent cells. # Discussion In the present study, we identified Arabidopsis PTPLA as a new interacting component of the fatty acyl elongation complex. Several lines of evidence indicate that PTPLA is acting as a 3-hydroxyacyl-CoA dehydratase involved in VLCFA synthesis. First, PTPLA shares significant sequence identity with PAS2 and PHS1, both catalyzing the dehydratase activity required for the VLCFA elongation. These three proteins shared the three conserved amino acids determined as essential for the enzymatic activity of the yeast PHS1 protein and constituting the putative active site of the dehydratase. Substitution of Tyr-149 and Glu-156 residues in PHS1 resulted in a loss of growth restoration of *Tet-PHS1*+DOX cells and a complete loss of enzyme activity. Secondly, PTPLA was also interacting in the ER with several elongase subunits. Thirdly, PTPLA was able to restore VLCFAs elongation in the weak *Tet-PHS1*+DOX mutant and its overexpression increased VLCFAs levels in both yeast and plants. Finally, loss of *ptpla* function led to 3-hydroxyacyl-CoA accumulation, which is a hallmark of defective dehydratase activity. However, PTPLA was not able to restore growth of a null *phs1* strain, nor able to complement *pas2* developmental phenotype even when expressed under the control of *PAS2* promoter. PTPLA and PAS2 showed some difference in binding the different KCS enzymes. Ectopic expression of *PTPLA* increased VLCFA levels in wild type background, an effect that is dependent on PAS2 activity. The fact that both proteins were able to interact would suggest that PTPLA might stabilize or activate PAS2 or PHS1 activity. Altogether these results suggest that endogenous PTPLA activity is either very low or present narrower substrate specificity compare to PAS2-dependent elongase complex. Substrate specificity was previously associated with the first elongation step. In yeast, elegant genetic and biochemical experiments demonstrated the existence of a caliper-like mechanism able to monitor the acyl chain length. In plants, the specificity was associated with specific proteins of the KCS family. For example, *kcs2/daisy-1* and *kcs20* are required for the elongation of C22 VLCFA for cuticular wax and root suberin biosynthesis. KCS18 elongates 20 carbons acyl-chain substrates whereas KCS5 had a preference for 26 carbons acyl-chain substrates. In mammals, similar substrate specificity was also demonstrated for the ELOVL family (ELOVL1–7). For example, ELOVL2 and ELOVL5 catalyze the elongation of polyunsaturated acyl-CoAs with C20–C22 and C18–C20 specificity respectively. ELOVL1 and ELOVL4 are responsible for the production of saturated and monounsaturated VLCFAs with lengths respectively of C22-C26 and superior to C28. No substrate specificity was reported for the other subunits of the elongase complex. The 3-hydroxyacyl-CoA profiles revealed that a larger accumulation of C18-OH in *ptpla* compared to *pas2-1* mutant suggesting that PTPLA would preferentially use C16-CoA substrate. The specificity of the different elongase complexes is most probably determined by specific recruitment of KCS which are known for their different acyl-CoA preferences. The fact that PAS2 and PTPLA did not seem to show the same interactions pattern with the KCS confirmed the existence of different elongase complexes with most probably specific activities. Besides the substrate specificity, PTPLA differs from PAS2 by its specific expression in the vascular tissues in young seedling roots. Nobusawa et al., showed that PAS2 is present only in the epidermis of leaves and stems. We showed that *PAS2* was also specifically expressed in the endodermis of the primary and secondary roots, as well as at the epidermis of young secondary root tips. *PTPLA* and *PAS2* expressions could match those of several *KCS*. Joubes et al. have shown that KCS family are divided in 8 subclasses with different tissue specificity and at least seven *KCS* genes were found to be strongly expressed in roots. Strikingly, the first reductase genes *KCR1* and its homologue *KCR2* displayed expression patterns in the root reminiscent of those of *PAS2* and *PTPLA* respectively. Besides the similar expression pattern in the root, PTPLA and KCR2 were not able to complement loss of function of their yeast orthologous genes and led when mutated to higher levels of VLCFA in Arabidopsis. A more complete analysis of elongase gene coexpression at different developmental stages or upon different abiotic or biotic stresses could reveal different pattern of expression that would suggest more complex elongase enzyme associations. We thus propose that PTPLA would be a 3-hydroxyacyl-CoA dehydratase associated with specific elongase complex activity in the vascular tissue and pericycle cells. Contrary to PAS2-based elongase complex in the endodermis that provides most of root VLCFA, PTPLA-based fatty acyl elongation would have a lower activity most likely targeted toward C18-C20 fatty acids. PAS2-based fatty acyl elongation in the endodermis and epidermis would provide VLCFA for the main lipid pools of these tissues, respectively suberin and cuticular waxes. PTPLA- based fatty acyl elongation on contrary would rather promote the synthesis of a regulatory signal modifying PAS2 associated elongase activity. Recent work also identified a regulatory role of epidermal VLCFA by repressing cytokinin synthesis in vascular tissue in a non-cell autonomous. Our present work uncovered a similar regulatory role of VLCFA in the root. Indeed, in absence of PTPLA or KCR2, PAS2-dependent fatty acyl elongation was enhanced suggesting that PTPLA-based elongation in vascular tissue is repressing PAS2-based elongase activity in adjacent endodermal cells. From these data, we propose a model where vascular tissues express a specific elongase activity (PTPLA-associated) regulating endodermal VLCFA elongation (PAS2-associated). This regulatory elongase complex would most probably include KCR2 but further more in depth biochemical experiments are necessary to dissect elongase complex composition and stoichiometry. The model suggests the existence of a signal that would be the root counterpart of the model of a non-cell autonomous signal diffusing from the epidermis to the central cylinder of leaves and stems. The nature of the signal is still unknown but cytokinins would be a likely candidate. # Materials and Methods ## Yeast transformation and growth The Arabidopsis cDNA library was built in a *Saccharomyces cerevisiae* expression vector pFL61 from young *A*. *thaliana* seedlings (two leaves stage). The *Tet-PHS1* mutant strain was originating from the R1158 parental strain (*URA*::*CMV-tTA MATa his3-1 leu2-0 met15-0*) carrying the inducible TetO7 system upstream the *PHS1* gene and replacing the endogenous promoter. *PHS1* was shut off after addition of doxycycline in the medium as described previously. *URA3* gene was disrupted by a nourseothricin resistance cassette in *Tet-phs1* strain *(ura3-*, *nour*^*R**)* according to Janke *et al* (primers URA3 used for disruption, LG73 and LG74 see) to allow pFL61 selection. After transformation by the lithium acetate procedure, yeast were spread out on a SD medium without uracil but with doxycycline (10μg/mL) and 2% of glucose (SD- URA+DOX), then incubated at 28°C during 4–7 days. Among the 698 clones able to grow on the selective medium, PAS2 was expected as a potential suppressor of *Tet-phs1*, PAS2 primers (PAS2F and PAS2R primers) were used by PCR to eliminate these clones. Plasmid DNAs from the non-PAS2 remaining yeast colonies were extracted by zymolyase according to the cold spring harbor protocol (Deplancke, 2006) and PCR amplified using primer LG69 and LG70. All these PCR products were sequenced by GenoScreen and all the sequences were blasted into the TAIR database. PHS1 was cloned in pFL61 to use as positive control (primers CM22 and CM23). For kinetic growth, 0,3U.OD_600nm of the different strains (mix of 4 clones per strain) from a saturated preculture were inoculated in 50mL and 150mL of liquid SD-URA+DOX medium respectively and agitated at 28°C for 4 days. OD_600nm was monitored until the culture reached stationary phase. ## Plant material and construction of plant expression vector A T-DNA insertion mutant line for *PTPLA* (At5g59770) was identified using the Arabidopsis Gene Mapping Tool (<http://signal.salk.edu/cgi-bin/tdnaexpress>), and the seed stock (SALK_077395) was obtained from the NASC. Homozygous lines were selected after genomic DNA extraction and by PCR screening for the presence of a T-DNA insertion (CM14 and LB1.3 primers) and the absence of the *PTPLA* intact gene (LG79 and LG103 primers). The *pas2-1* mutant is an ethyl methane sulfonate allele in Col0 background that was maintained as heterozygous stocks. For all RT and lipid analysis, the homozygous *pas2-1 mutants* were selected based on the characteristic pepper-like shape of mutant seeds. The *pas2-1/ptpla* double mutant was generated by crossing the heterozygous *pas2-1* mutant with homozygous *ptpla* (SALK_077395). The F1 was genotyped by *bstnI* digestion of a specific *PAS2* PCR product (PAS2-1F and PAS2-1R specific primers) to detect *pas2-1* mutation and with the previous CM14/LB1.3 primers to detect *ptpla* T-DNA insertion. *PAS2* and *PTPLA* promoters used for the following constructs correspond to 2000 bp and 1250pb genomic sequence upstream the ATG codon of *PAS2* and *PTPLA* genes respectively. To generate the *pPAS2*:*GFP-PTPLA*, *pPAS2*:*GFP-PAS2*, *pPAS2*:*PTPLA*, *pPAS2*:*PAS2*, *pPTPLA*:*PAS2*, pPTPLA:PTPLA constructs, PAS2 (-2000 bp from the ATG) and PTPLA promoters (-1250 bp from the ATG) were first cloned into pB7WGF2 (*pPAS2*:*GFP- GTW*; *pPTPLA*:*GFP-GTW*) or pB2GW7 (*pPTPLA*:*GTW*) vectors by restriction enzymes to replace the 35S promoter. PCR amplifications of the promoters with primers containing restriction enzyme sites of HindIII (LG116) and SpeI (LG117) were used for pPAS2 cloning and restriction sites of SacI (LG123) and SpeI (LG124) were used for pPTPLA cloning. In parallel, the coding DNA sequence (CDS) of *PTPLA* was amplified with (primers CM02/CM03) or without (primers CM02/CM04) the stop-codon using the full-length cDNA (G61261: cDNA into pENTER223) and were transferred into pDONR207 by BP cloning and then recombined by LR into the previous vectors according to the Invitrogen^™ protocol. The same procedure was done with the *PAS2* CDS, the stop and non-stop versions already cloned in pDONR207. PTPLA G61261 was also recombined into 35S::gtw pGWB2, 35S::GFP-gtw pGWB6 and *35S*::*mCherry-gtw* to generate 35S:PTPLA, 35S:GFP-PTPLA and 35S:mCherry-PTPLA respectively. Finally, the *pPAS2*:*GFP-PTPLA*, *pPAS2*:*GFP-PAS2*, *pPAS2*:*PTPLA*, *pPAS2*:*PAS2*, *pPTPLA*:*PAS2*, *pPTPLA*:*PTPLA*, *35S*:*PTPLA* and *35S*:*GFP-PTPLA* generated constructs were transformed into heterozygous *pas2-1/+* by the floral-dip method (Clough and Bent, 1998). All the experiments described in this work were carried out on homozygous T3 plants. Concerning the *pPTPLA*:*mRFP1* and *pPTPLA*:*GUS* constructs, 1250bp PTPLA promoter was inserted into *pDONR207* to generate *pDONR207*:*pPTPLA* (LG94 and LG96 BP primers) and recombined into pGWB553 and pGWB3 upstream of the mRFP or GUS markers respectively by LR cloning. The *pPAS2*:*GFP* construct was obtained by LR recombination of a GFP-stop cDNA (pDONR207:GFP) into pB7FWG2 vector (*pPAS2*:*GTW-GFP*). Both *pPTPLA*:*mRFP1* and *pPAS2*:*GFP* constructs were transformed together into Col0 to observe the localization of *PTPLA* and *PAS2* expression. Finally, the *pPAS2*:*GUS* constructs was obtained by recombination of 1500bp PAS2 promoter into the pMDC162 vector. The *pKCR1*:*GUS* construct was obtained from Jerome Joubès, *pKCR2*:*GUS* construct from Ljerka Kunst and the *kcr2* mutant by Frédéric Beaudoin. ## Lipid Analysis Plants were vertically grown on MS Arabidopsis medium for 14 days. Roots were quickly sampled by cutting below the hypocotyl. Overexpressor and complemented yeast strains from a saturated preculture were inoculated in 50mL (LB-URA) or 150mL (LB-URA+DOX) respectively and 150 OD units were sampled at about 26h and 47h after inoculation respectively so that cells would be in exponential phase of the growth culture with an OD600 between 4 and 6. The cultures were centrifuged 10 minutes at 3000rpm and cells were washed with cold sterilized water. Both yeasts and roots were immediately frozen after sampling at -80°C for more than 1 hour and then lyophilized. For quantification of the fatty acids methyl ester (FAMES), around 2 mg of each dry sample were used according to the Li et al protocol by GC-MS. Three technical and two or three biological replicates per sampled were analyzed. Quantification of acylCoA of Arabidopsis and yeast material by LC ESI-MS/MS required 100mg of frozen fresh roots and 1 OD unit of frozen yeast. Roots and yeasts were grown as described before for the FAMES analysis. AcylCoA extraction and LC separation was carried out as reported by. The MS multi-reaction monitoring was done as described by Haynes *et al*.. Five replicates per samples were analyzed. ## Transient infiltration of *N*. *benthamiana* leaves For co-expression experiments and BiFC interactions, the vectors were transformed into *Agrobacterium tumefaciens* and inoculated in *Nicotiana benthamiana* leaves of one month-old plants. Agrobacterium carrying clones of interest were grown overnight at 28°C in 5 ml LB medium with appropriate antibiotics. Aliquots from the overnight cultures were resuspended in 10 mM MgCl₂ and 1 mM 2-(N-morpholine)-ethanesulphonic acid (MES) to obtain a final OD_600nm of 0,5 for tobacco leaf infiltration. The BiFC experiments were done twice. Split luciferase was carried out according to van Leene. ## Cytologic and microscopic analysis For analysis of the GUS (β-glucuronidase) activity, Arabidopsis seedling of 14 days were incubated at 37°C with 1mg/mL X-Gluc (5-bromo-4-chloro-3-indolyl-D- glucuronic acid) and in a GUS reaction buffer (100 mM sodium phosphate buffer, pH7.2, 10mM sodium EDTA, 0.1% Triton X-100, 1mM potassium ferricyanide, 1mM potassium ferrocyanide), after 2x5 minutes of vacuum infiltration. The stained seedling were cleared by successive ethanol washes from 30% until 70% and visualized under a Zeiss light macroscope (Axiozoom). Observations of XFP markers were carried out using a Leica SP5 AOBS confocal laser microscope using either a PL APO 20x0.70 NA or 63x1.20 NA water-immersion objectives. GFP and mCherry/mRFP1 fluorescence were respectively recorded after an excitation at 488 and 561 nm (Argon laser and laser diode respectively) and a selective emission band of 495–550 nm and 600–625 nm. YFP fluorescence was recorded after an excitation at 514 nm (Argon laser) and a selective emission band of 520–565 nm. Autofluorescence of the chloroplast was excited by the Argon laser (488 or 514 nm) and recorded with a selective emission band of 650–700 nm. Root length were measured with the segmented line tools of ImageJ software and the number of lateral root was counted under a binocular. ## Real-time RT-qPCR conditions and analysis The total RNAs were extracted from 14 day-old *Arabidopsis thaliana* roots using the RNeasy^® Plant Mini Kit (Qiagen) according to the manufacturer’s instructions. The Reverse Transcription was performed with reverse transcriptase (Fermentas) each reaction containing 1ug of total RNA. Quantitative PCR gene- specific primers were designed to span the introns. The qPCR primer sequences specific to PTPLA (CM35 and CM36), PAS2 (PAS2 Q-PCR1 F and PAS2 Q-PCR1 R) and AT2G28390 reference primers are resumed in. Quantitative PCR was performed on a CFX96 machine from Biorad under the following conditions: 3min at 95°C follow by 35 cycles of 10s at 95°C, 20s at 60°C and 20s at 72°C, and finally 95°C for 30s. The data were analyzed with the CFX96 manager Biorad 3.0 software. # Supporting Information We thank the Plant Observatory imaging platform of the IJPB institute for providing microscopy technical support and the chemistry platform for the FAMES analysis. We thank also Frédéric Beaudoin (Departement of Biological Chemistry and Crop Protection; Rothamsted Research; UK), Jérôme Joubes (Laboratoire de Biogenèse membranaire; CNRS UMR 5200, France) Ljerka Kunst (Department of Botany; University of British Columbia; Canada) and Sébastien Baud (IJPB, INRA, Versailles, France) for providing seeds and DNA as well as the European Arabidopsis Stock Center for providing Arabidopsis T-DNA insertion mutant. This work was funded by the Ministère de l’Enseignement Supérieur et de la Recherche (France) (doctoral fellowship to C.M.). The IJPB benefits from the support of the LabEx Saclay Plant Sciences-SPS (ANR-10-LABX-0040-SPS). DOX doxycycline ECR enoyl-CoA reductase ER endoplasmic reticulum FA fatty acid FAE *FIDDLEHEAD* gene FAMES fatty acid methyl ester GUS β-glucuronidase HCD 3-hydroxyacyl-CoA dehydratase HPLC high-performance liquid chromatography KCR β-Ketoacyl-CoA reductase KCS 3-keto-acyl-CoA synthase LCB long chain base NOUR nourseothricin PAS *PASTICCINO* genes PCR polymerase chain reaction PGK phosphoglycerate kinase promotor PHS phytosphingosine PHS1 yeast 3-hydroxyacyl-CoA dehydratase PTPLA protein tyrosine phosphatase like A Tet titrable promoter TetO7 VLCFA very-long-chain fatty acid [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** CM LG KH FB JDF. **Funding acquisition:** JN JDF. **Investigation:** CM YB LG FT CR RH FB. **Methodology:** CM YB LG FT RH FB. **Project administration:** JDF. **Supervision:** JDF. **Validation:** CM FT RH FB JDF. **Visualization:** CM LG. **Writing – original draft:** CM FT FB JDF. **Writing – review & editing:** CM FB JN JDF.

# Introduction Cell motility is a prerequisite for tumor progression and for invasive migration of carcinoma cells into surrounding tissue. In order to acquire a motile phenotype carcinoma cells undergo a dramatic morphological alteration, termed epithelial–mesenchymal transition (EMT), wherein they lose their epithelial characteristics and acquire the motility of mesenchymal cells. In the case of many carcinomas, EMT-inducing signals, such as HGF, EGF, PDGF, and TGF-β, emanate from the tumor-associated stroma and activate a series of EMT-inducing transcription factors, including Snail, Slug, zinc finger E-box binding homeobox 1 (ZEB1), Twist, Goosecoid, and FOXC2. These transcription factors pleiotropically orchestrate the complex EMT program. The loss of cell–cell contacts mediated by E-cadherin, an epithelial marker, is a typical hallmark of EMT. The down-regulation of E-cadherin is common in squamous cell carcinomas (SCC) and is associated with an enhanced ability of invasion and/or metastasis and with a poor prognosis, reflective of its critical role in tumor progression. It is widely believed that the down-regulation of E-cadherin occurs through the transcriptional repression mediated by binding of transcriptional repressors, such as Snail1 (*SNAI1*), to E-box sequences in the proximal E-cadherin promoter. The EMT program and the activation of Snail1 depends on a series of intracellular signaling networks and feedback loops involving ERK, MAPK, PI3K, and Akt signaling pathways. In contrast, little is known about the involvement of cyclic nucleotide-mediated signaling pathways in EMT. These pathways are implicated in many biological processes that cooperate in organ development and differentiation of epithelial cells. The effects of cyclic adenosine monophosphate (cAMP) via protein kinase A (PKA) on changes in cell motility and via exchange protein activated by cAMP (EPAC) on cell migration and integrin- mediated cell adhesion are particularly important for tumor invasion. Intracellular cAMP concentrations are regulated by adenyl cyclases (AC), which use ATP to produce cAMP, and by phosphodiesterases (PDEs), which catalyze the degradation of cAMP to AMP. Visinin-like protein 1 (VILIP-1, gene name *VSNL1*), a member of the family of neuronal calcium sensor proteins, modulates the levels of cyclic nucleotides, induces cell differentiation, and has recently been identified as a putative tumor migration suppressor gene. In esophageal cancer the reduced expression of VILIP-1 is correlated with invasive features, such as the depth of tumor invasion and local lymph node metastasis. In aggressive non-small cell lung carinoma cell lines and primary tumors the loss of VILIP-1 expression is associated with a poor survival. VILIP-1 is differentially expressed in chemically-induced murine skin squamous carcinomas of different degrees of aggressiveness. In an experimental model of murine SCC cell lines derived from these tumors it was demonstrated that the ectopic expression of VILIP-1 in two VILIP-1 non-expressing, high grade SCC lines increased cAMP levels, leading to a diminished MMP-9 and RhoA activity together with a significant reduction in the invasive properties of the carcinoma cells. VILIP-1 expression was further shown to decrease the expression of fibronectin-specific integrin subunits α5 and αv that contributed to cell adhesion, cell migration, and invasiveness of highly invasive SCC cell lines. Recently, we demonstrated that the tumor invasion suppressing effect of VILIP-1 in mouse skin SCCs exclusively depends on cAMP levels, but not on cGMP levels, and that both cAMP-effectors, PKA and EPAC, are involved in the reduction of the migratory ability of SCC cells. Here, we set out to investigate, whether and how VILIP-1-enhanced cAMP-signaling may be involved in EMT in SCC. # Materials and Methods ## Material FSK (adenylyl cyclase activator Forskolin), 8Br-cAMP, DDA (2′,5′-dideoxyadenosine, general AC inhibitor) EGF and TGFβ for cell stimulation experiments were obtained from Sigma (St. Louis, MO, USA), Tocris (Bristol, UK) and Calbiochem (San Diego, CA, USA). Cell culture reagents were obtained from Gibco-Invitrogen (San Diego, CA, USA). Unless otherwise specified, all other reagents were purchased from Sigma and Roth (Karlsruhe, Germany). ## Antibodies Rabbit polyclonal antibodies, raised against recombinant VILIP-1 protein, were affinity-purified on corresponding glutathion-S-transferase (GST)-tagged fusion proteins, immobilized on N-hydroxysuccinimide ester coupled agarose colums (Bio- Rad, Hercules, CA, USA) as previously described. Polyclonal rabbit anti E-cadherin (gp184) antibodies were kindly provided by Otmar Huber and described previously. Polyclonal rabbit anti integrin α5 antibodies were purchased from Chemicon (Temecula, CA, USA) and monoclonal antibodies against β-actin (sc-81178) and HRP-labeled secondary antibodies were purchased from Santa Cruz Biotechnologies (Santa Cruz, CA, USA). ## Cells and culture method Murine skin squamous cell carcinoma cell lines CC4A and CC4B, CH72 and CH72T3 were described previously. CC4A and CC4B were derived from the same tumor. When injected s.c. into nude mice, CC4A gave rise to a high-grade SCC or spindle cell carcinoma (or SCC IV), whereas CC4B gave rise to a well-differentiated, less aggressive, and low-grade SCC (SCCII). CH72 also gave rise to a low-grade SCC after s.c. inoculation, and CH72T3 is a subcloned cell line obtained by *in vivo* passaging of CH72 into nude mice, which resulted in a high-grade SCC. Cells were grown in DMEM (GIBCO) plus FCS (10%), L-glutamine (2 mM) and penicillin/streptomycin (100 µg/ml). ## Growth factor treatment CC4B and CH72 cells were plated in standard DMEM in 24-well or 6-well dishes, respectively. 24 h after plating and 8 h prior to treatment with EGF or TGFβ medium was exchanged to low FCS (1%) DMEM to basal the cells. Cells were treated for 72 h with the indicated concentrations of growth factors and afterwards lysed for Western blot or RT-PCR analysis. To compare morphological changes cells were fixed and images were taken with a Leica inverted microscope at a 200× magnification. The migratory capacity of the cells after growth factor treatment was analyzed in *in vitro* wounding assays over 24 h. In indicated cases agents increasing or decreasing cAMP concentrations were added 24 h before cell lysis or before wounding the cell monolayer. ## Transfection CC4A and CH72T3 were transfected with VILIP-1-GFP-vector or empty-GFP-vector whereas CC4B and CH72 were transfected with VILIP-1-siRNA or scrambled siRNA using Optimem and lipofectamin 2000 (Invitrogen) following the manufacturer's instructions. VILIP-1-siRNA (antiVILIP1_1: sense r(AGC CGU UAG UCU GAA UUA A)dTdT, antisense r(UUA AUU CAG ACU AAC GGC U)dAdA; antiVILIP1_2: sense r(CAA AGA UGA CCA GAU UAC A)dTdT, antisense r(UGU AAU CUG GUC AUC UUU G)dAdA; antiVILIP1_3: sense r(GUG CGA CAU UCA GAA AUG A)dTdT, antisense r(UCA UUU CUG AAU GUC GCA C)dAdA) was used as a cocktail of three siRNA oligos (150 ng of each per transfection) directed against the coding region of VILIP-1 and was purchased from Qiagen (Hilden, Germany). ## Western blot analysis Cultured cells were homogenized in an appropriate volume of homogenization buffer (25 mM Tris, 150 mM NaCl, pH 7.5, containing the protease inhibitors benzamidine (1 mM), phenylmethylsulfonylfluoride (0.1 mM)). Nuclei and debris were removed by centrifugation at 1.000 g for 5 min, protein concentrations were measured using BCA assay (Pierce, Rockford, IL, USA) and 40 µg protein of each sample was applied to 5–20% gradient SDS-PAGE. To analyze the expression level of VILIP-1, E-cadherin, integrin α5 and β-actin separated proteins were blotted on a PVDF membrane. The membrane was blocked with 5% milk powder in TBST (25 mM Tris, 150 mM NaCl, pH 7.5, 1% Tween 20) for 1 h at RT and afterwards incubated with the primary antibodies at 4°C overnight as previously described. After washing three times with TBST, secondary antibodies were applied for one hour at RT. Unbound antibodies were removed and the detected protein was visualized in a dark chamber using Western Lightning reagents (PerkinElmer Life Sciences, Boston, MA USA) and Hyperfilm (Amersham, UK). ## RT-PCR PCR primers were designed that selectively amplify cDNA encoding Snail1, VILIP-1 or GAPDH and synthezised by Invitrogen (Carlsbad, CA,USA) (Snail1: sense AGG ACG CGT GTG TGG AGT, antisense GGAGAATGG CTT CTC ACC AG; VILIP-1: sense ATG GGG AAR CAG AAT AGC AAA C, antisense TCA TTT CTG MAT GTC KCA CTG CA; GAPDH: sense ACC ACA GTC CAT GCC ATC AC, antisense TCC ACC ACC CTG TTG CTG TA; K, M, R indicate mixed bases used to obtain species-independent primers). RT-PCR experiments were performed 3 times using total RNA from SCC-lines. Total RNA was extracted using RNeasy Mini-Kit (Quiagen, Hilden, Germany) and reverse transcribed using Oligo(dT) primers and SuperScript III First-Strand-Kit (Invitrogen, San Diego, CA, USA). PCR was performed using 0.2 µM of each primer, PCR buffer, 0.2 mM dNTP-Mix, 2 mM MgCl₂, and 1 U taq polymerase (Invitrogen, San Diego, CA, USA) and DEPCH₂O in a 50 µl reaction mix. 35 cycles of amplification were performed for each sample. For each primer pair the reaction was also carried out in absence of reverse transcriptase to ensure that there is no DNA contamination. ## *In vitro* wound assay Cells grown in standard medium (2×10⁵ cells/well) were plated in 24-well plates. Cells were either grown in low FCS (1%) medium for 8 h and then treated with 10 ng/ml EGF in low FCS (1%) medium for 72 h before wounding or were transfected with VILIP-1-siRNA or the corresponding control 72 h before wounding and grown to confluence. Cells were placed in low FCS (1%) medium in order to basal the cells prior to growth factor treatment and to minimize cell proliferation. A wound was created by scratching the cell monolayer using a sterile 200 µl pipette tip. The wound was marked and 24 h after wounding cells were fixed and pictures were taken at a 200× magnification with a Leica inverted microscope and at least eight representative fields for each condition were analyzed. Cell migration was quantified by counting the number of cells/field. ## Statistical analysis Statistical analysis was performed using unpaired, two-sided Student's t-test for samples of unequal variance (Welch test). Values were accepted as significant when p was less than 0.05 (\*), less than 0.01 (\*\*) or less than 0.001 (\*\*\*). All error bars represent standard deviations. # Results ## Epithelial-mesenchymal transition (EMT) in squamous cell carcinoma (SCC) When we compared the morphology of cultures of the less aggressive, VILIP-1-positive skin cancer cells (CC4B and CH72) with the more aggressive, VILIP-1-negative skin cancer cells (CC4A and CH72T3), obvious morphological differences were noticed. Less aggressive, VILIP-1-positive skin cancer cells (CC4B and CH72) were well organized, tightly packed and formed clustered, cobblestone-like structures, typical of epithelial cells and suggestive of strong cell–cell adhesion. In contrast, aggressive, VILIP-1-negative skin cancer cells (CC4A and CH72T3) showed the mesenchymal morphological phenotype, including cell shape elongation and scattering of cells, which is suggestive of reduced cell–cell adhesion and increased cell motility. Since cell-cell adherens junctions of epithelial cells are formed by E-cadherin molecules, we assessed the cellular expression levels of E-cadherin in VILIP-1-negative and VILIP-1-positive SCCs. Immunoblotting showed that the expression of E-cadherin was down-regulated in VILIP-1-non-expressing cell lines CC4A and CH72T3, compared to VILIP-1-expressing cell lines CC4B and CH72. In contrast the integrin receptor subunit α5, mediating cell-matrix adhesion, was up-regulated in VILIP-1-non-expressing cells. The loss of the epithelial, cell type-specific morphology, the loss of E-cadherin expression and the associated reduction of cell-cell adherens junctions are hallmarks of EMT. These results suggest that aggressive SCCs must have undergone EMT while losing VILIP-1-expression. ## Growth factor-induced EMT: changes in cell morphology and expression of marker proteins in SCC Growth factors, especially TGFβ and EGF, have been shown to induce EMT along with the down-regulation of various epithelial markers, including E-cadherin, in SCC. We thus examined whether the stimulation of VILIP-1-positive CC4B and CH72 cells with different concentrations of TGFβ or EGF promotes EMT-like morphological changes and corresponding alterations in expression levels of VILIP-1, E-cadherin and integrin α5. When TGFβ-treated CC4B and CH72 cells were compared to untreated cells (control), they appeared rounded in cell shape (second panel) and immunoblotting revealed increased integrin α5 expression levels and slightly reduced VILIP-1 expression levels (lanes 2 and 3). In terms of E-cadherin protein levels, the induced alterations were not consistent between the two cell lines. Contrary to expectations, in CC4B cells E-cadherin protein levels cells seemed increased following TGFβ treatment ( lanes 2 and 3, upper panel) and correspondingly CC4B cells did not show scattering (upper row, second panel). However, CH72 cell responded in an inhomogeneous manner to TGFβ treatment. Stimulated CH72 monolayers exhibited areas of scattered cells (lower row, second panel) and stimulation with 0.1 ng/ml TGFβ slightly reduced E-cadherin protein levels in 2 of 3 repeats, whereas stimulation with 1 ng/ml TGFβ slightly increased E-cadherin expression (lower panel). In summary, TGFβ treatment had only moderate effects on VILIP-1 protein levels, did not alter or even tend to increase E-cadherin protein levels and did not lead to widespread cell scattering or cell shape elongation. Hence, TGFβ did not cause a shift from the VILIP-1-positive, less aggressive phenotype to the VILIP-1-negative, aggressive phenotype of SCC. By comparison, EGF treatment at 10 ng/ml resulted in a more obvious cell shape elongation and scattering of CC4B and CH72 cells ( third panel). Immunoblotting showed that EGF at 10 ng/ml caused down-regulation of E-cadherin in both carcinoma lines, which was consistent with the observed EGF-induced cell morphological changes. Interestingly, the expression of VILIP-1 was also clearly down-regulated in both cell lines in response to stimulation with 10 ng/ml EGF. Stimulation with 1 ng/ml EGF did produce a less pronounced down-regulation of E-cadherin and VILIP-1. Integrin α5 was up-regulated with increasing EGF concentrations in both cell lines (lanes 4 and 5). Collectively, the stimulation at the higher EGF concentration (10 ng/ml) induced clear EMT- like changes, and resulted in a shift in the morphology and the protein expression of VILIP-1-positive, less aggressive SCCs towards the phenotype of VILIP-1-negative, more aggressive SCCs shown in. ## Effect of the modulation of VILIP-1-expression on integrin α5 and E-cadherin expression In a previous study it has been shown that the knock down of VILIP-1-expression caused an increase in the expression level of integrin α5 and αv in skin SCC. Similarly, we found an increased expression of integrin α5, while VILIP-1 expression was down-regulated following EGF-stimulation. To determine whether the loss of VILIP-1 also affected the expression of E-cadherin, we transfected VILIP-1-negative SCCs with GFP-VILIP-1 or empty GFP-vector as control, and VILIP-1-positive SCCs with VILIP-1-specific siRNA or scrambled siRNA as control for 72 h respectively and assessed protein levels of E-cadherin. Immunoblotting confirmed that integrin α5 expression is inversely regulated by VILIP-1 (down- regulation in CC4A by 42%, in CC72T3 by 44%; up-regulation in CC4B by 37%, in CH72 by 55%). In contrast, no effect of either VILIP-1 overexpression in VILIP-1-negative SCCs CC4A and CH72T3 or VILIP-1 knock down in VILIP-1-positive SCCs CC4B and CH72 on E-cadherin expression was observed, indicating that E-cadherin and VILIP-1 are independently down-regulated by EGF during EMT. ## Expression levels of Snail1 mRNA in VILIP-1-positive and VILIP-1-negative SCC and effect of EGF Snail1 (*SNAI1*), a member of the slug/snail family of transcriptional repressors, is one of the several transcriptional factors that can suppress E-cadherin gene expression in squamous cell carcinoma and is a potent inducer of EMT. Accumulating evidence indicates that the EGFR family and its downstream signaling pathways, the PI3K–Akt- and MEK–ERK pathway, regulate the expression of Snail1, suggesting Snail1 as a candidate repressor for the down-regulation of VILIP-1 and E-cadherin expression in response to stimulation with EGF in mouse skin SCC. To verify this hypothesis we first determined the expression of Snail1 in the aggressive and less aggressive SCC cell lines. RT-PCR analysis showed that Snail1 mRNA is solely detectable in VILIP-1-negative aggressive SCC cell lines. However, following EGF stimulation and subsequent EMT-induction, Snail1 was up-regulated in VILIP-1-positive SCC cell lines. Interestingly, the induction of Snail1 expression in response to EGF was diminished in the presence of elevated cAMP following forskolin (FSK) stimulation, indicating a novel role of cAMP-signaling in EMT. Quantification of the RT-PCR showed that the EGF- induced increase of Snail1 mRNA was statistically significant compared to control (CC4B+EGF: p = 0.039, CH72+EGF: p = 0.029). Co-treatment with EGF and forskolin significantly attenuated the EGF-induced increase of Snail1 mRNA (CC4B+EGF+FSK: p = 0.04, CH72+EGF+FSK: p = 0.047). ## The effect of the modulation of VILIP-1-expression on Snail1-expression depends on cAMP-signaling Since the expression of VILIP-1 increases intracellular levels of cAMP in skin SCC, we analyzed the effect of VILIP-1 and cAMP-signaling on Snail1 mRNA levels. Following transfection of VILIP-1-negative SCCs with GFP-VILIP-1 or empty GFP- vector as control, and VILIP-1-positive SCCs with VILIP-1-specific siRNA or scrambled siRNA as control for 72 h, we found that knock down of VILIP-1-expression did not affect Snail1 mRNA expression. In contrast, ectopic expression of VILIP-1 in the aggressive, VILIP-1-negative cell lines CC4A and CH72T3 reduced Snail1 mRNA levels. The reduction of Snail1 mRNA was statistically significant (CC4A: p = 0.035, CH72T3 p = 0.037) and could be blocked by the application of the general adenylyl cyclase inhibitor DDA for 24 h before lysis ( lanes 2 and 4 versus 3 and 6, respectively), demonstrating that cAMP-signaling plays an important role for the VILIP-1 effect on Snail1 expression. ## Involvement of cAMP-signaling in the VILIP-1-siRNA- or EGF-induced migration of SCC cells To demonstrate the involvement of cAMP-signaling in the effect of EMT-induction and of VILIP-1-expression on the migratory capability of skin tumor cells, we performed *in vitro* wound closure assays. We either knocked down VILIP-1-expression by siRNA or applied EGF-stimulation leading to reduced VILIP-1-expression. Both the knock down of VILIP-1-expression by siRNA and EGF treatment resulted in a significantly increased migratory capability, documented by a higher number of migrating cells in the wound area after 24 h. In CC4B cells VILIP-1-specific siRNA enhanced the cell migration by 46% (p\<0.001) and EGF by 59% (p\<0.001). In CH72 cells following siRNA treatment 72% more (p\<0.001) and following EGF treatment 60% more (p\<0.001) migrating cells were observed. We have previously shown that VILIP-1-negative SCCs show greater migratory capability than their VILIP-1-positive counterparts, and that this effect depends on decreased cAMP levels. The application of 8Br-cAMP 24 h before wounding of siRNA or EGF treated cells prevented the enhancement of the migratory capability and significantly reduced the number of cells in the wound area (: p\<0.001 in all conditions). Following the additional 8Br-cAMP application, the number of migrating cells was significantly lower than in control conditions (: CC4B: p = 0.009, CH72: p = 0.002; C: CC4B p = 0.037, CH72 p = 0.039), confirming that increased motility induced by the loss of VILIP-1 or by EGF treatment is suppressed by cAMP-signaling. These results point towards a role of the putative tumor migration suppressor VILIP-1 and the associated cAMP pathway for EMT in SCC. # Discussion In this study, we examined the role of the putative tumor invasion suppressor VILIP-1 and cAMP-signaling during EMT in mouse skin tumor cell lines of different aggressiveness. When aggressive, VILIP-1-negative SCCs were compared to less aggressive VILIP-1-positive SCCs, distinct differences in morphology were observed. These differences resemble the change in cellular morphology during the transition of the epithelial to mesenchymal phenotype. This assumption was further supported by the results of the Western blot analysis, showing the loss of the EMT-marker E-cadherin in VILIP-1-negative CC4A and CH72T3 cells. In addition, a previous study revealed increased activity of two further EMT markers, RhoA and MMP9, in the aggressive, VILIP-1 negative SCCs. The spindle-like morphology, the loss of the epithelial marker gene E-cadherin together with the previously shown up-regulation in the activity of RhoA, MMP9 and of the protein level of integrin α5, as well as the enhanced migratory capability, indicate that VILIP-1-negative, aggressive SCCs underwent EMT, and that down-regulation of VILIP-1 might be related to EMT. To reproduce this process experimentally, we stimulated VILIP-1-positive CC4B and CH72 cells with EMT-inducing growth factors TGFβ and EGF. We found that stimulation with EGF induces SCC cells to undergo a transition from the epithelial to the spindle- like mesenchymal morphology. This was accompanied by the loss of E-cadherin and subsequent loss of cell-cell-contacts. Similar results have been obtained for several other carcinoma cells by authors of previous studies. In addition EGF treatment leads to the up-regulation of integrin α5, and most importantly to the down-regulation of VILIP-1 in CC4B and CH72 cells. This affirms the hypothesis that VILIP-1 is lost during EMT. EGF treatment of CC4B and CH72 cells induced an EMT-like phenomenon and caused VILIP-1-positive SCC cells to mimick the characteristics of VILIP-1-negative SCC cells, including the gain of increased migratory capability. In the literature TGFβ was also shown to induce EMT. However, in CC4B and CH72 cells TGFβ caused rounding of cells, but not cell shape elongation, and only slightly reduced E-cadherin in CH72 cells or even increased it in CC4B cells. Such increased E-cadherin levels following TGFβ treatment have also been observed in human trophoblasts. Another study revealed, that only 2 of 20 mouse cell lines treated with TGFβ responded with the induction of EMT. In keratinocytes it has been shown, that the induction of EMT by TGFβ depends on a hyperactive Ras-MAPK-pathway and that without this prerequisite only reversible morphological alterations are induced. However, the loss of growth control induced by TGFβ that occurs at a late stage of mouse skin carcinogenesis is independent of ras gene activation. These findings might explain the small effect of TGFβ observed in this study. VILIP-1-expression was not or only marginally affected by TGFβ treatment. Thus, EGF, rather than TGFβ is a key factor in malignant progression of squamous cell carcinoma lines. The observed down-regulation of E-cadherin and VILIP-1-expression during EGF- induced EMT might be caused either by a parallel transcriptional repression of both genes or by a serial mechanism, where reduced levels of VILIP-1/cAMP might contribute in a second step to the down-regulation of E-cadherin. We have previously shown that reduced VILIP-1/cAMP levels contribute to the up- regulation of integrin α5 in mouse skin SCC. Therefore, we analyzed the expression of Snail1, as a potent inducer of EMT and a transcriptional repressor of E-cadherin. Snail1 was detectable in untreated aggressive, VILIP-1- and E-cadherin-negative SCC cells and was inducible by EGF treatment in less aggressive, VILIP-1-positive cells. These results suggest the possible involvement of Snail1 in the repression of E-cadherin and VILIP-1-expression during EMT of mouse skin SCC. The inverse correlation of Snail1 and E-cadherin expression, together with the up-regulation of Snail1 during EGF-induced EMT are in line with findings from other studies investigating the characteristics of invasive SCC. It is widely believed that downstream pathways of the EGFR, particularly the PI3K-Akt and MAPK pathway, are involved in the initiation of Snail1 expression through the regulation of NF-κB and AP-1, which act as transcriptional activators of the Snail1 gene. It is noteworthy that in our study the enhancement of cAMP levels by application of forskolin in EGF treated cells repressed EGF-induced expression of Snail1, implicating cAMP-signaling in the regulation of Snail1. Ectopic expression of VILIP-1 in aggressive, VILIP-1-negative SCCs, which leads to increased cAMP levels, likewise decreased the expression level of Snail1. This effect could be blocked by the application of the general adenylyl cyclase inhibitor DDA. To our knowledge this is the first study showing that VILIP-1-dependent cAMP-signaling interferes with the expression of Snail1 and might thereby prevent the progression of EMT during tumor progression. Accumulating evidence indicates that enhanced cAMP-signaling counteracts the malignant progression of cancer cells. A few studies also report that cAMP-elevating agents block EMT. In melanoma cells cAMP regulates the NF- κB-mediated expression of EMT-associated genes. Among these genes were SIP1 and slug, two other repressors of E-cadherin expression. In the alveolar epithelial cell line A549 increased cAMP levels resulting from the inhibition of cAMP-PDE4 block TGFβ-induced EMT in a MAPK-signaling dependent manner. The two latter studies also describe a cAMP-mediated regulation of E-cadherin expression, whereas in other studies analyzing E-cadherin-mediated cell-cell-contacts and migration of cancer cells no cAMP-dependent effect on E-cadherin expression could be detected. Although we found a significant effect of VILIP-1 and cAMP- signaling on the expression level of the E-cadherin repressor Snail1, we could not detect any effect, neither of the knock down of VILIP-1 in the less aggressive SCCs, nor of the over-expression of VILIP-1 in the aggressive SCCs, on the expression of E-cadherin. Thus, we conclude that VILIP-1 is not necessary for basal expression of E-cadherin. Ectopic expression of VILIP-1 and subsequently increased cAMP levels seem not to be sufficient to abolish an established inactivation of the E-cadherin gene. Therefore, E-cadherin and VILIP-1 are rather subject to a parallel transcriptional repression during EGF- induced EMT in mouse skin SCC. E-cadherin silencing involves a high degree of complexicity with the cooperation of epigenetic mechanisms and different repressors acting at different stages of the malignant progression. Against this background it has to be considered that the Snail1-reducing effect of VILIP-1-cAMP might have an impact on the initial down-regulation of E-cadherin expression during the first steps of tumor progression, whereas in advanced stages the contribution of additional factors is necessary to reconstitute the E-cadherin expression. However, it is very likely that VILIP-1 and cAMP- signaling regulates other Snail1 repressor target genes during EMT. To understand this interesting phenomenon further studies are required to decipher the precise mechanism of the VILIP-1-cAMP-dependent Snail-1 regulation and its impact on gene repression. For instance, the reduction of integrin α5β1 signaling by VILIP-1/cAMP might be involved, since integrin α5β1 was shown to act in concert with the EGFR and via ILK-Akt-NF-κB signaling, which constitute two ways to influence the expression level of Snail1. Another way to interfere with the induction of the EMT program and Snail1 expression is the direct crosstalk of cAMP-signaling with the signaling cascades downstream of the EGFR, such as interfering with the MAPK cascade on the level of Raf or with PI3K pathways on the level of GSK3β and NF-κB activity. Further evidence for the EMT-suppressing role of VILIP-1-cAMP-signaling comes from the *in vitro* wound closure assays. The increase in the migratory capability of less aggressive, VILIP-1-positive SCCs caused by either siRNA knock down of VILIP-1 or EGF-treatment was eliminated by the application of 8Br- cAMP. Other studies analyzing the effects of EGF on cell migration, consistently describe an increase in the migratory capability following EGF treatment. As mentioned above the role of cAMP in tumor progression is controversial. For example, dibutyryl cAMP has been shown to slightly enhance collagen-mediated keratinocyte migration. In contrast, it has also been shown that cAMP inhibits growth factor-mediated matrix metalloproteinase 9 induction and keratinocyte migration. We have reported in a previous study that in mouse skin SCC enhanced cAMP-signaling reduced their migratory capability. Accordingly, the results of the present study showed that the migration-diminishing effect of cAMP-signaling counteracts the migration-inducing effect of EGF, suppressing a further hallmark of malignant tumors cells, which have undergone EMT. In summary, the present study shows the role of the putative tumor migration suppressor VILIP-1 in counteracting the induction of EGF-induced EMT. Our finding that VILIP-1 suppresses the expression of the EMT-related transcriptional repressor Snail1, and might thereby interfere with the induction of EMT in a cAMP-dependent manner, suggests a novel mechanism for the anti- invasive activity of VILIP-1-cAMP-signaling. Therefore, further investigation of the signaling networks involved in the VILIP-1-cAMP-mediated regulation of Snail1 and its targets in malignant tumors may help to identify novel anti- cancer strategies. [^1]: Conceived and designed the experiments: KS AJK KHB. Performed the experiments: KS. Analyzed the data: KS KHB. Contributed reagents/materials/analysis tools: AJK. Wrote the paper: KS KHB. [^2]: The authors have declared that no competing interests exist.

# Introduction The terrorist attacks that occurred on September 11, 2001 (hereafter 9/11) sent shockwaves across the United States and throughout the world. Nearly every aspect of human life was touched by these events and, therefore, it is not surprising that scientists have uncovered large scale shifts in human behavior following 9/11. Perhaps one of the most notable shifts to occur was the rise in levels of fear of terrorism among US residents. Gallup Poll results revealed, for instance, that only 24% of Americans were “very worried” or “somewhat worried” about being victimized by terrorism in April of 2000. This number jumped to 58% on the night of 9/11 and held steady around 50% for the next few months. Fear is a powerful human emotion originating in the evolutionarily ancient regions of the brain and corresponding to altered human behavior in a variety of ways. Indeed, fear response varies from person to person with some responding aggressively and others responding passively. Because 9/11 was shown to raise fear levels for many Americans, scholars have begun to utilize 9/11 as a natural experiment to examine collective fear response patterns. The result is that much is now known concerning the impact of 9/11 on collective behaviors such as governmental responses. What remains more elusive, however, is the degree to which 9/11 impacted individual-level fear-response behaviors (i.e., between-individual differences in behavior). Even though fear of terrorism tended to increase, on average, across the majority of US citizens, the various ways in which people respond to fear in general, and to the fear of terrorism following 9/11 specifically, are highly variable. One of the more common ways to reduce fear of terrorism is to engage in self-protective behaviors, such as refusing to fly on an airplane, not traveling internationally, and maybe even purchasing a gun. This latter option is particularly unique because it is available to virtually all US citizens and it does not interfere with day-to-day living. Criminological research has revealed that a common reaction to the fear of victimization in general is to purchase and carry a handgun for self-protection. Kleck and colleagues, for instance, reported that respondents who perceived more risk in their neighborhood were more likely to own a gun and were more likely to report intentions of purchasing a gun in the near future. Building on these results, we hypothesized that respondents would be more likely to report gun-carrying behaviors as a form of self-protection *after* 9/11 as compared to *before* 9/11 (hypothesis 1). Though gun acquisition and gun carrying may be a viable form of self-protection, not all people respond to fear by purchasing and carrying firearms. Thus, we should not expect all Americans to carry a weapon in response to 9/11 (or any other traumatic experience), but rather would expect individual-level traits and predispositions to partially moderate how people respond to fear of being victimized by a terrorist attack. A growing body of research has revealed one salient factor that has the potential to moderate how people respond to environmental stimuli is genotype. Gene-environment interactions (GxE), in general, run in one of two directions (other than a null result): a positive interaction or a negative interaction. A positive interaction—assuming a positive main effect of genotype on gun carrying—would suggest individuals with a certain genotype were more likely to carry a handgun after 9/11 as compared to before 9/11. A negative interaction—again assuming a positive main effect of genotype on gun-carrying—would suggest these individuals were more likely to carry a handgun before 9/11 and these differences were weakened, disappeared, or reversed after 9/11. A polymorphism in the *5-HTT* gene has been shown to interact with environmental factors, including those that fall within the parameters of a fearful or traumatic experience, to produce different phenotypic outcomes. A line of empirical research, for example, has suggested that *5-HTTLPR* interacts with stressful life events to predict depression (but see), substance use, and decision making. As such, we hypothesized *5-HTTLPR* would interact with 9/11 experience in predicting the self-protective behavior of gun carrying. It may be expected that individuals carrying the “risk” alleles for the *5-HTTLPR* polymorphism would be more likely to carry a handgun in response to 9/11 as compared to persons with other genotypes (hypothesis 2, a positive interaction). Note, however, that a negative interaction may also be expected; *5-HTTLPR* predicts handgun carrying prior to 9/11 but not after (hypothesis 3). The latter pattern of findings may emerge if 9/11 has an impact large enough to “overpower” any genetic predispositions (i.e., *all* people were more likely to carry a gun after 9/11, not just those with a genetic predisposition for this behavior). # Methods This research analyzes secondary data from the National Longitudinal Study of Adolescent Health (Add Health). Institutional Review Board approval to analyze the Add Health data was retained by all research team members from their respective institutions: The University of Texas at Dallas, Florida State University, and Sam Houston State University. The Add Health data have been described at length elsewhere. Briefly, the Add Health is a nationally representative longitudinal study of adolescents who were enrolled in middle or high school in 1995. Four waves of data have been collected with the most recent set of interviews being completed in 2008. Three key features of the Add Health study are important for the current analysis. First, wave 3 data collection took place between July 2001 and May 2002. Approximately 20% of all respondents were interviewed on or before 9/11/2001 (coded as 0) and all others were interviewed after 9/11/2001 (coded as 1). Sensitivity checks revealed that only eight total respondents (from the DNA subsample) were interviewed on 9/11/2001. The results of all analyses were identical to those presented here when these eight respondents were removed from the sample. None of these cases reported gun carrying. The second feature of the Add Health data is that all respondents were asked the following question during wave 3 interviews: “In the past 12 months, how often did you carry a handgun at school or work?” Responses were dichotomized so that 0 = *never* and 1 = *at least once*. A total of 32 (1.35%) respondents (in the DNA subsample) reported carrying a gun to work/school. In order to avoid deductive disclosure due to fewer than 50 respondents reporting gun carrying, all case counts gleaned from cross-tabulations will be reported as percentages. The third feature of the Add Health is that all twins and full siblings were genotyped during wave 3 interviews. Genotypic information was available for 2,574 respondents. After eliminating one twin from each monozygotic twin pair (to avoid artificially decreasing standard errors) and after eliminating cases with missing data, a final analytic sample of 2,350 was obtained. Respondents were genotyped for *5-HTTLPR*, which maps to 17q11.1-17q12. The *5-HTTLPR* polymorphism is the result of a 44 base pair (bp) VNTR in the 5′ regulatory region and two alleles have been identified: a short (S) allele (484 bp) and a long (L) allele (528 bp). The short allele has been associated with depression risk, lower life satisfaction, and with antisocial behavior. Building on this literature, it was hypothesized that the short allele would be associated with self-protective behaviors and, therefore, *5-HTTLPR* was coded co-dominantly so that 0 = *no short alleles* (i.e., LL homozygotes; *n* = 797 \[33.80%\]), 1 = *one short allele* (i.e., SL heterozygotes; *n* = 1,082 \[45.89%\]), and 2 = *two short alleles* (i.e., SS homozygotes; *n* = 471 \[19.97%\]). Less than one percent of the analytic sample (0.34%) was missing genotypic information for *5-HTTLPR* and was coded as missing. Only one of these missing cases reported gun carrying. Substantive conclusions from the analysis were identical when *5-HTTLPR* was coded recessively (i.e., 0 = *no/one short allele*, 1 = *two short alleles*). When *5-HTTLPR* was coded dominantly (i.e., 0 = *no short allele*, 1 = *one or two short alleles*), the pattern and direction of findings were similar but conventional significance levels were not reached (*p* = .196 for the interaction term). # Results Preliminary analyses revealed an association between the gun carrying variable and 9/11 in the full sample of respondents (*n* = 15,052). A logistic regression model indicated respondents interviewed after 9/11 were 53% more likely to carry a gun to work/school as compared to respondents interviewed on or before 9/11 (*b* = .426, odds ratio = 1.531, standard error for *b* = .194, *z* = 2.196, *p* = .028). Results gleaned from the DNA subsample were substantively similar but the association did not reach statistical significance (*b* = .278, odds ratio = 1.321, standard error for *b* = .432, *z* = .644, *p* = .520). Presented in are results from a logistic regression model where the gun-carrying variable was utilized as the dependent variable and the 9/11 indicator variable, *5-HTTLPR*, and an interaction between 9/11 and *5-HTTLPR* (i.e., 9/11 \* *5-HTTLPR*) were utilized as covariates (*n* = 2,350). Findings indicated 9/11 was associated with a marginally significant increase in gun carrying (*b* = 1.855, odds ratio = 6.392, standard error for *b* = 1.037, *z* = 1.788, *p* = .074), respondents with more short alleles on *5-HTTLPR* were more likely to carry a gun to work/school (*b* = 1.252, odds ratio = 3.499, standard error for *b* = .640, *z* = 1.958, *p* = .050), and the interaction between 9/11 and *5-HTTLPR* (i.e., 9/11 \* *5-HTTLPR*) exhibited a negative and statistically significant impact on gun carrying (*b* = −1.519, odds ratio = .219, standard error for *b* = .703, *z* = −2.161, *p* = .031). The interaction term was negative, indicating that the influence of *5-HTTLPR* on gun carrying was diminished for respondents who were interviewed after 9/11. The interaction between 9/11 and *5-HTTLPR* is plotted in, which reveals that the predictive influence of *5-HTTLPR* is contingent upon whether the respondent was interviewed before or after 9/11. As shown in the figure, *5-HTTLPR* positively predicted gun carrying for respondents interviewed prior to 9/11/2001 (*n* = 632 total, 0.42% of LL homozygotes carried a gun, 0.71% of SL heterozygotes carried a gun, 3.57% of SS homozygotes carried a gun, *b* = 1.252, odds ratio = 3.499, standard error for *b* = .640, *z* = 1.957, *p* = .050). Respondents who were interviewed after 9/11/2001 showed no association between *5-HTTLPR* and gun carrying (*n* = 1,718 total, 1.61% of LL homozygotes carried a gun, 1.50% of SL heterozygotes carried a gun, 0.84% of SS homozygotes carried a gun, *b* = −.266, odds ratio = .766, standard error for *b* = .290, *z* = −.919, *p* = .358). A coefficient difference test indicated that the effects of *5-HTTLPR* were significantly different across the two groups of respondents (*z* = 2.160, *p*\<.05). Three sensitivity analyses were conducted to test the robustness of the findings. First, the logistic regression model was re-estimated after controlling for the respondent's age, sex, and race. Given the association between racial/ethnic grouping and *5-HTTLPR* genotype, the latter control seemed particularly important to include (self-reported racial categories were as follows: 72.05% of respondents self-identified as White, 18.71% self- identified as Black, 8.22% self-identified as Asian/Pacific Islander, and 5.39% self-identified as Native American). When these effects were included in the logistic regression model (race was included as a dummy variable identifying the respondent as Black), the main effect for the 9/11 variable was not statistically significant (*b* = 1.664, odds ratio = 5.278, standard error for *b* = 1.053, *z* = 1.580, *p* = .114), the main effect for *5-HTTLPR* was marginally significant (*b* = 1.208, odds ratio = 3.347, standard error for *b* = .650, *z* = 1.858, *p* = .063), and the 9/11 \* *5-HTTLPR* interaction term was statistically significant (*b* = −1.523, odds ratio = .218, standard error for *b* = .706, *z* = −2.156, *p* = .031). Thus, the substantive conclusions of the interaction were largely unaffected by the inclusion of age, race, and sex controls. The second sensitivity analysis included a control variable for the respondent's arrest history. Specifically, respondents were asked whether they had ever been arrested or taken into custody by police (coded 0 = *no* and 1 = *yes*). This variable was important to consider because an alternative explanation for the effect of the 9/11 variable may be that the most antisocial respondents (i.e., those who are likely to have been arrested in the past) are those who are also most likely to be interviewed later by researchers due to transience on part of the respondent. The arrest variable predicted gun carrying (*b* = 1.395, odds ratio = 4.034, standard error for *b* = .403, *z* = 3.457, *p* = .001) in a bivariate model suggesting some of the reported behavior may be illegal gun carrying. Also, the arrest variable predicted the 9/11 indicator in a bivariate model (*b* = .395, odds ratio = 1.485, standard error for *b* = .176, *z* = 2.248, *p* = .025), indicating the importance of including this variable as a statistical control. Importantly, the substantive conclusions drawn from the other variables were unchanged when the arrest variable was included. To be sure, the effect of the 9/11 variable did not reach statistical significance (*b* = 1.699, odds ratio = 5.467, standard error for *b* = 1.044, *z* = 1.627, *p* = .104), the coefficient for *5-HTTLPR* attained marginal significance (*b* = 1.192, odds ratio = 3.293, standard error for *b* = .638, *z* = 1.866, *p* = .062), and the 9/11 \* *5-HTTLPR* interaction was statistically significant (*b* = −1.477, odds ratio = .228, standard error for *b* = .708, *z* = −2.086, *p* = .037). Virtually identical estimates were gleaned from a model that controlled for age, race, sex, and arrest history simultaneously. The final set of sensitivity analyses re-estimated each of the logistic regression equations using the rare-events logistic regression model. Because less than 2% of the sample reported gun carrying, it was important to re-analyze the associations outlined above with a statistical model that is able to account for low base rates on the dependent variable (i.e., limited 1 s as compared to 0 s). Overall, the substantive conclusions from the rare-events logistic regression models mirrored those outlined above. In terms of the 9/11 \* *5-HTTLPR* interaction term, the rare-events model produced a negative and statically significant interaction in the base model (i.e., a model including the 9/11 variable, the *5-HTTLPR* variable, and the 9/11 \* *5-HTTLPR* interaction) (*b* = −1.425, standard error for *b* = .702, *z* = −2.03, *p* = .042) and in a model with controls for age, race, sex, and arrest history (*b* = −1.434, standard error for *b* = .710, *z* = −2.02, *p* = .043). The split-sample models also were consistent with those reported above, but it is important to note the impact of *5-HTTLPR* was only marginally significant for the pre-9/11 cases in the rare-events logistic regression model (*n* = 632, *b* = 1.170, standard error for *b* = .638, *z* = 1.83, *p* = .067). Similar to the above, the impact of *5-HTTLPR* was not a statistically significant predictor of gun carrying for respondents interviewed post-9/11 (*n* = 1,718, *b* = −.254, standard error for *b* = .289, *z* = −.88, *p* = .379). A *z*-test indicated the difference in coefficients was statistically significant (*z* = 2.03, *p*\<.05). # Discussion Over the past decade, research has repeatedly shown genetic and environmental factors interact (GxE) in the prediction of human behavior. Working from this framework, we hypothesized that the events surrounding 9/11 would interact with a polymorphism in the *5-HTT* gene to predict gun-carrying behaviors (hypotheses 2 and 3). Results supported the *negative* interaction hypothesis (hypothesis 3) in two ways. First, a multiplicative interaction term between 9/11 and *5-HTTLPR* revealed a *negative* and statistically significant coefficient. Second, split sample models indicated *5-HTTLPR* was positively related to gun carrying before 9/11 but not after. Taken together, these results suggest a GxE between 9/11 and *5-HTTLPR* genotype and this pattern of findings indicates that certain persons (those carrying short alleles on *5-HTTLPR*) may have been more likely to carry a gun prior to 9/11 as a reflection of their greater likelihood to respond with self-protective behaviors to everyday situations. After 9/11, however, these differences were erased. These findings may be consistent with research indicating a greater impact of genetic factors on behavioral phenotypes in common or privileged environments as compared to disadvantaged or unpredictable environments, such as events that occur immediately following a terrorist attack. The importance of these findings is twofold. First, to our knowledge, this is the first study to show an association between 9/11 and gun-carrying behaviors (hypothesis 1). Thus, the current results open a new line of inquiry into the myriad ways that 9/11 (and perhaps other terrorist attacks) affected US residents. Second, and perhaps more important, is that the current study is the first to test for a GxE with a measured polymorphism (*5-HTTLPR*) and a terrorist attack (9/11). GxE researchers have long noted the limitations of extant research, namely that environmental factors may be contaminated with genetic influences (via gene-environment correlations). In other words, a test of GxE may actually reflect a gene-gene interaction (i.e., GxG). By utilizing a terrorist attack as an exogenous environmental influence, we removed all possibility that the GxE (i.e., 9/11 \* *5-HTTLPR*) was the result of a GxG. In short, this study presents what may be the purest way to test for GxEs. A related alternative would be to test for GxEs with natural disasters acting as the environmental influence. Limitations to the analysis must be considered. First, though the sample size was large (*n*\>2,000 for the DNA subsample), less than 40 respondents reported gun-carrying behaviors. This may have limited the statistical power of the analysis and most certainly inflated standard errors. Note that the conventional alpha level of.05 was utilized in order to limit the possibility of false- positives (Type I error) and that the rare-events logistic regression model produced a pattern of findings similar to those garnered from the standard logistic regression model. Nonetheless, some of the analytic cells had low case counts, meaning that the findings should be interpreted with due caution until replicated on an independent sample. The second primary limitation concerns the wording of the gun carrying question. Specifically, respondents were asked to report on whether they had carried a gun over the past 12 months. This means respondents interviewed *after* 9/11 may have reported on gun carrying prior to 9/11. It will be important for future work to replicate these findings using alternative samples not subject to this limitation to determine whether, and to what extent, this may have impacted the results. Additionally, it is important that scholars consider whether alternative explanations exist for why gun carrying should differ across the two groups of respondents. The most likely argument seems to be that respondents interviewed later in the year or in 2002 may have been systematically different than those interviewed earlier in the wave 3 process due to location issues and the difficultly of tracking down transient and perhaps antisocial respondents. We explored this possibility by including a control variable for prior arrest history. The substantive findings for the interaction were unchanged when this variable was included suggesting this alternative explanation is not viable. For now, the results of our study suggest that 9/11 is an exogenous environmental pathogen that can be used in future GxE research to examine a range of phenotypic outcomes. This research uses data from Add Health, a program project directed by Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by grant P01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Special acknowledgment is due Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Information on how to obtain the Add Health data files is available on the Add Health website (<http://www.cpc.unc.edu/addhealth>). No direct support was received from grant P01-HD31921 for this analysis. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JCB KMB. Analyzed the data: JCB. Contributed reagents/materials/analysis tools: JCB. Wrote the paper: JCB KMB BBB.

# Introduction Ciliates, eurychoric unicellular eukaryotes, are characterized by complexes of cilia and a nuclear dimorphism. In last 25 years, molecular phylogenetic analyses, especially based on small subunit rRNA (SSU rRNA) gene sequences, provided resolution of a number of important questions on the phylogenetic relationships within this group (for example, –). However, many questions remain open, mostly related to a number of spirotrichean lineages, either on the assignment of certain species to one or another group or, more importantly, on the phylogenetic relationships within certain orders/families that contain a large number of taxa. Among these, the order Urostylida is one of the most confused and diverse and is increasingly attractive for the researchers working on morphogenetic, taxonomic and molecular fields (for example, –). The most important apomorphy for urostylids is a zig-zagging ventral cirral pattern originating from more than six anlagen evolved possibly convergently for several times (for example,). There are more than ten studies that include details of interrelationships within this order (for example, –), which are mainly based on morphological/morphogenetic data, but none reaches the same conclusions as another. In his monograph of the Urostyloidea, Berger recognised 154 valid species, and assigned most of them to four families (Holostichidae, Bakuellidae, Urostylidae and Epiclintidae) using the frontal ciliature and the midventral complex as the main features. Recently, another systematic classification was proposed by Lynn, which also divided the order Urostylida into four families (Epiclintidae, Pseudokeronopsidae, Pseudourostylidae, Urostylidae). Between these two systems, there is only agreement over the classification of Epiclintidae. In order to investigate further the evolutionary relationships among the urostylids, molecular phylogenetic analyses based on SSU rRNA gene sequences have been increasingly used in recent few years,. Although these investigations undoubtedly show that Urostylida is a large group within the Hypotricha, the monophyly of this order is not yet certain, and relationships within it are still confused. Furthermore, molecular phylogenies based on other gene markers, albeit with sparse taxon sampling, have produced rather different results compared to SSU rRNA phylogenies. Comparison between different molecular trees is an essential step to reveal the evolution within investigated groups, even when independent datasets yield congruent results. The combined phylogenetic analyses of multiple genes have become popular due to the poor resolution of phylogenies based on single loci, and have successfully inferred better-resolved phylogenies within the major taxonomic groups, including animals, plants, fungi and bacteria. However, there are few ciliate phylogenies based on combined gene partitions. With the advent of multi-gene phylogenies, particular emphasis has been placed on congruence or combinability of independent and possibly heterogeneous datasets. To date, the only molecular urostylid phylogeny based on combined genes is that of Hewitt et al. who used SSU-5.8S-LSU rRNA. There are only three congruent phylogenies, based on different genes that include few urostylid taxa. The present study was initiated to improve our understanding of evolutionary relationships within the order Urostylida by extending the SSU rRNA gene, ITS1-5.8S-ITS2 region, and alpha-tubulin gene database. Moreover, molecular phylogenies are discussed with critical consideration of the taxonomic literature. In addition, statistical tests, i.e. incongruence length difference (ILD) test, Shimodaira-Hasegawa (S-H test) and partition addition bootstrap alteration (PABA) approach, are performed to detect incongruence among these three gene partitions. # Results ## Analyses of Sequences and Secondary Structures A total of one SSU rRNA gene, eight ITS1-5.8S-ITS2 regions, and 13 alpha-tubulin genes were sequenced in our analyses. The SSU rRNA gene had the most characters (1,635 bp unambiguously aligned), followed by alpha-tubulin (1,071 bp), then ITS1-5.8S-ITS2 (427 bp) for the 14-taxon datasets. The nucleotide sequences of all three genes among 14 urostylids share similarities of 90.59–99.26%, 52.77–94.03%, and 77.40–91.96%, respectively. It is noteworthy that alpha-tubulin amino acid sequences share similarities of 97.13–100.00% , so phylogenetic trees were constructed using alpha-tubulin nucleotide sequences instead of amino acid sequences in our analyses. Comparisons of the ITS2 region sequences as well as secondary structures show that there are two unique regions for *Pseudoamphisiella quadrinucleata*, and one for *Psammomitra retractilis*. As shown in, the main loop is divided into three parts (viz. I, II, and III) by Helix A and B, and there are 37 nucleotides in part I of *Pseudoamphisiella*, whereas there are only 31 ones in other species. Helix A in *Pseudoamphisiella* contains 19 nucleotides, whereas that of other species is constantly composed of 20 nucleotides. This is caused by one nucleotide deletion in the terminal loop of Helix A for *Pseudoamphisiella* (data not shown). Previous investigations – showed that for spirotricheans, 11 out of 12 paired nucleotides were identical in the labeled 15 nucleotides stretch of Helix A. However, our current analysis indicates that *Psammomitra* has rather different sequences and secondary structure in this region. ILD tests for all combined datasets (viz. Datasets 4, 5, 9–11) show that most of the partitioned datasets contain conflicting signal (*P* = 0.001), with only Dataset 9 being congruent (*P* = 0.256). In an attempt to further clarify the incongruence, each taxon was deleted in turn to determine if one or a few taxa were particularly problematic. However, in no dataset did this approach indicate that conflict is potentially caused by a specific taxon. ## Phylogenetic Analyses Inferred from Dataset 1 (SSU rRNA, 89 Taxa) In our analyses, the outgroup Protocruziidia is followed by Phacodiniidia and Euplotida, then the sister clade forming by Oligotrichia and Choreotrichia. Hypotricha seems to be paraphyletic: most species group together, and others cluster with Oligotrichia, Choreotrichia, and the core discocephalids, respectively. Though *Uroleptus* and *Paruroleptus* are assigned into the family Urostylidae according to Lynn, they are undoubtly classified out of the order Urostylida in our SSU rRNA gene tees, which is congruent with previous investigations. Considering exclusion of these two genera from the order Urostylida, all available SSU rRNA gene sequences of urostylids were included in our phylogenetic analyses, and they refer to 15 genera representing all four urostylid families (sensu Lynn) and four unclassified genera. In both analyses, the order appears to be always paraphyletic, and species fall into six clades, except for *Anteholosticha multistilata*, the position of which is unresolved. Clade I consists of two *Parabirojimia* species (family Urostylidae), which group with *Trachelostyla*, a non-urostylid genus. Clade II consists of three *Anteholosticha* species. Clade III is the “core” urostylid clade, and it is composed of seven genera which belong to the family Urostylidae (viz. *Metaurostylopsis*, *Urostyla*, *Diaxonella* and *Anteholosticha*), two genera of the family Pseudokeronopsidae (*Pseudokeronopsis*, *Thigmokeronopsis*), one genus of the family Pseudourostylidae (*Pseudourostyla*), three unclassified urostylid genera (*Apokeronopsis*, *Bergeriella* and *Nothoholosticha*), and the non-urostylid genus *Hemigastrostyla*. Clade IV has a closer relationship with Oligotrichia and Choreotrichia than with other urostylids, and consists of two genera of the family Holostichidae (viz. *Holosticha* and *Psammomitra*), and the type genus of the family Epiclintidae, *Epiclintes*. Clade V falls into the order Discocephalida, and consists of *Pseudoamphisiella* (family Holostichidae) and the unclassified genus *Leptoamphisiella*. Among the four urostylid families, the Epiclintidae is monotypic whereas the other three are multi- generic and paraphyletic. All species of Pseukeronopsidae and Pseudourostylidae fall into Clade III, and urostylid species appear in all six clades. Among 15 sequenced urostylid genera, species of *Anteholosticha* are the most diverse and representatives could be found in both Clades II and V. Of the other 14 genera, none have representatives in more than one clade. ## Phylogenetic Analyses Inferred from Dataset 2 (ITS1-5.8S-ITS2, 31 Taxa), Dataset 3 (Alpha-Tubulin, 26 Taxa) and Dataset 4 (Three-Gene Combined, 25 Taxa) As revealed in trees based on Dataset 1, analyses inferred from Datasets 2 and 4 also indicate that: (1) the outgroup Protocruziidia is followed by Euplotida, Oligotrichia, Choreotrichia; (2) Hypotricha is separated into several clades; (3) the core urostylid group contains only genera/species of Clade IV derived from Dataset 1, namely *Anteholosticha gracilis*, *A. manca*, *Bergeriella*, *Diaxonella* (absent from Dataset 4), *Metaurostylopsis*, *Thigmokeronopsis*, *Apokeronopsis* (which does not cluster with this group in trees based on Dataset 2), *Pseudokeronopsis*, *Pseudourostyla*, and *Nothoholosticha*; (4) *Pseudoamphisiella* is rather distant from other urostylids in Datasets 2, 4. However, the cluster pattern of species outside the core urostylid group is rather different among trees based on Datasets 1, 2, and 4. In analyses inferred from Dataset 3, the subclass Protocruziidia branches at the deepest level, however, compared to trees based on Datasets 1, 2, and 4, the clade comprising the euplotids is more closely related to the “core” Hypotricha. The monophyly of Choreotrichia is rejected. In addition, *Thigmokeronopsis* and *Pseudokeronopsis*, which belong to the core urostylids in analyses based on Datasets 1, 2, and 4, fall outside the core Urostylida. ## Comparison of Phylogenetic Analyses Inferred from 14-Taxa Datasets ML tree topologies inferred from seven 14-taxa datasets (Datasets 5–11) were not identical to each other. However, as revealed by trees based on Datasets 1–4, these analyses also strongly indicate that: (1) *Pseudoamphisiella* should be excluded from urostylids, and; (2) the core urostylid group contains *Anteholosticha manca*, *A. gracilis*, *Bergeriella*, *Metaurostylopsis*, *Thigmokeronopsis*, *Apokeronopsis*, *Pseudokeronopsis*, *Pseudourostyla* and *Nothoholosticha*. Using the S-H approach, out of 42 possible comparisons, 15 ones result in a *P* value above 0.05, signaling that congruence is not rejected, whereas 27 comparisons reject congruence (*P*\<0.05). Dataset 5 rejects all topologies inferred from other 14-taxa datasets, however, two topologies among them are not totally rejected. Conversely, topology based on Dataset 5 is only rejected by Dataset 7. Interestingly, all topologies obtained by datasets including alpha- tubulin are accepted by other datasets also including alpha-tubulin, but are rejected by all other datasets. Five, three, one, five, two and two of five nodes selected based on 14-taxa three-gene combined datasets could be found in trees inferred from Datasets 6–11, respectively. For Node 1, the addition of ITS1-5.8S-ITS2 region data causes the bootstrap values to decrease. For Nodes 2–4, the addition of alpha- tubulin gene data does the same thing. By contrast, the addition of SSU rRNA gene data always increases the support values. Considering all five nodes, the PABA approach also shows that bootstrap values tend to increase as more data or data partitions are added except when alpha-tubulin gene data is added as the second partition. # Discussion This study represents one of the few attempts to reconstruct generic-level relationships within Urostylida with molecular characters from multiple genes, and the only phylogenetic analysis that includes all four urostylid families. Though the phylogenetic results based on different datasets are mixed, and support values for some nodes are not high, some conclusions could be drawn following by comparison between our phylogenetic trees and system of Lynn. ## The Current Status of the Phylogenetic Relationships within the Order Urostylida Recent molecular phylogenetic investigations (for example, –), as well as the current work based on both single gene (Datasets 1–3) and multiple genes (Dataset 4) shows that the urostylid assemblage is not monophyletic and thus raises serious challenges to the classification of the order Urostylida,. This is consistent with the conclusion that there is a considerable amount of convergence in urostylid morphology which brings into question current classification scheme. In the present work, several datasets, with SSU rRNA, alpha-tubulin and ITS1-5.8S-ITS2 gene/region sequences for all known urostylid genera, were used in order to re-evaluate phylogenetic relationships within this assemblage and to make a comparison between molecular phylogeny and the system of Lynn which is mainly based on morphological/morphogenetic data. ## Classification of Four Unclassified Genera The systematic positions of four recently reported genera, namely *Bergeriella*, *Leptoamphisiella*, *Apokeronopsis* and *Nothoholosticha,* were not included in any of updated systems although they were putatively assigned to the order Urostylida based on either morphological/morphogenetic or molecular information in the original descriptions. Among them, a new family, Bergeriellidae, was erected for the type genus *Bergeriella*. In the present investigation, *Bergeriella* always falls into core urostylid group in all the trees, and is not closely related to any of the four urostylid families. Thus, according to both molecular and morphological/morphogenetic data, all the evidence supports the conclusion that *Bergeriella* should represent a distinct family within the order Urostylida. The results presented here show that the genus *Leptoamphisiella* is most related to *Pseudoamphisiella*, the type genus of the family Pseudoamphisiellidae, which is, however, assigned to the family Urostylidae in Lynn's system. Our analyses firmly support the conclusion that this family should be excluded from the order Urostylida, but rather belongs to a group of its own which clusters to the well-known discocephalids. Both *Apokeronopsis* and *Nothoholosticha* are confirmed as true urostylids belonging to the family Pseudokeronopsidae. ## Classification of the Family Urostylidae Nine genera included in our analyses (viz. *Anteholosticha*, *Diaxonella*, *Holosticha*, *Metaurostylopsis*, *Parabirojimia*, *Psammomitra*, *Urostyla*, *Pseudoamphisiella* and *Leptoamphisiella*), all of which are assigned to the family Urostylidae in Lynn's system, are distributed among Clades I–VI in the present analysis. As revealed in previous molecular and morphological investigations, , and in our SSU rRNA gene trees, *Uroleptus* and *Paruroleptus* should be removed from the urostylid family Urostylidae to the non-urostylid family Uroleptidae. Similarly, *Pseudoamphisiella* and *Leptoamphisiella*, two urostylid genera according to Lynn, should be placed in the suborder Discocephalina since they consistently cluster with Discocephalina. This is consistent with the results of previous studies based on molecular data, and supports the findings that some morphological/morphogenetic features of these genera, e.g. the cirri of the midventral complex are not arranged in the zig-zag pattern, and the general developmental process of the ciliary structure, are more similar to those of discocephalines than urostylids. The phylogenetic position of *Parabirojimia* is slightly variable according to different datasets, however, it always falls outside of the “core” urostylid group and does not have a robust relationship with any other typical urostylids. Considering the extremely unusual mode of development of the cortical structure during morphogenesis, especially the formation of the somatic ciliature, e.g. the transverse cirri, the right marginal rows, etc., it is reasonable to assign this genus/family to its own group, that is the suborder Parabirojimina, as suggested by Yi et al.. The genus *Metaurostylopsis* is only included in three systems. Among those genera included in the present investigation, Shi et al. considered that *Metaurostylopsis* has a close relationship with *Urostyla* and *Pseudourostyla*, Berger placed it together with *Parabirojimia* in family Bakuellidae, and five other (non-sequenced) genera, whereas Lynn suggested that *Metaurostylopsis* could be related to *Anteholosticha*, *Holosticha*, *Diaxonella*, *Parabirojimia*, *Psammomitra*, *Pseudoamphisiella*, and *Uroleptus*. However, among these hypotheses, only the sister relationship between *Metaurostylopsis* and *Pseudourostyla* is hinted by Dataset 3, indicating that none of the assignments of *Metaurostylopsis* in these three systems are reasonable. As noted by Berger, the systematic position of *Diaxonella* is complicated since the type species, *D. pseudorubra*, has been repeatedly reported under different generic and specific names (for example, –). This genus has only been included in two systems, since it was established by Jankowski. Based on the redescription of *D. pseudorubra* (as *D. trimarginata* by Shao et al.), it was assigned to the family Pseudourostylidae, thus as an urostylid species. This report also included a description of morphogenesis and the unusual mode of formation of left marginal rows, which has been reported in only another hypotrich genus, that is, *Pseudourostyla*. However, the present and previous molecular investigations, did not recover a close relationship between *Diaxonella* and *Pseudourostyla*, thus supporting Berger's hypothesis that this unusual morphogenetic process is very likely a result of convergent evolution and should not be regarded as a family level character as suggested by Eigner and Foissner. In addition, the placement of *Diaxonella* in family Holostichidae (sensu Berger) is also clearly rejected by the molecular data in both the present and previous investigations. This is consistent with the morphological finding that *Diaxonella* has more than two marginal rows, and is hence rather different from other holostichid genera (sensu Berger). According to Lynn, *Diaxonella* should be assigned into the family Urostylidae. However, only the connection between this genus and *Urostyla*, and *Anteholosticha manca* is accepted in the present work and previous investigations,. All this evidence indicates that *Diaxonella* is undoubtedly an urostylid, however its family-level assignment in both Berger's and Lynn's systems is highly questionable and needs to be re-evaluated. Of the final four genera, viz. *Holosticha*, *Psammomitra*, *Urostyla* and *Anteholosticha* which are also assigned to the family Urostylidae by Lynn, the first three are located in two separate clades in our trees. The relationship between *Holosticha* and *Psammomitra* hypothesized by Lynn and Berger was confirmed by both previous and present analyses except in trees based on single- gene datasets and in those based on datasets containing alpha-tubulin information with two genes combined. By contrast, the genus *Anteholosticha* appears to be heterogeneous and highly divergent, with species falling into different clades in all our trees. In addition, distinct from other genera, seven *Anteholosticha* species share no unique nucleotides at semi-conserved, parsimony-information sites in the alignment of SSU rRNA gene sequences. These findings indicate that *Anteholosticha* is probably a convergent assemblage of species as predicted also by Berger , and a revision of this genus is urgently needed. In summary, the family Urostylidae (sensu Lynn) seems to be a huge “melting pot” containing over 24 nominal genera, the monophyly of which is strongly rejected by the present analyses. Currently, a complete re-arrangement for its classification remains impossible partly because molecular information is lacking for too many taxa. Nevertheless, the following conclusions can be drawn based on our analyses: 1) as revealed in previous investigations,, *Parabirojimia*, *Psammomitra*, *Pseudoamphisiella*, *Leptoamphisiella*, *Uroleptus* and *Paruroleptus* should be removed from this family and the last four genera are not even members of the order Urostylida. 2) *Holosticha* should also be excluded from this family; 3) *Diaxonella* and *Urostyla*/*Parabirojimia* respectively might represent two isolated families; 4) the genus *Anteholosticha* is extremely diverse, polyphyletic and should be revised when more information becomes available. ## Classification of the Family Pseudokeronopsidae Two of the six genera in the family Pseudokeronopsidae (sensu Lynn), viz. *Pseudokeronopsis* and *Thigmokeronopsis*, and another two genera which should be included in this family, viz. *Apokeronopsis* and *Nothoholosticha*, group consistently into two clades in SSU rRNA trees , and into more then two clades in other trees. Thus, all of these analyses reject the monophyly of this family. Berger synonymised *Apokeronopsis begeri* as *Thigmokeronopsis crassa*, due to the genus *Apokeronopsis* was not erected then. However, phylogenetic trees based on Datasets 2, 3, 7, 8, 10, and 11, none of which contain SSU rRNA gene sequences except Dataset 10, failed to recover a close relationship between these two genera, although they did group together in other trees including SSU rRNA gene, which indicates that the connecting of *Apokeronopsis* with *Thigmokeronopsis* is probably due to inclusion of SSU rRNA. Considering the separation of these two genera is supported by some morphological/morphogenetic data, for example, presence or absence of thigmotactic cirri and the fusion pattern of macronuclear segments prior to division, the distinction of both genera is reliable but their systematic positions remain unresolved. Although *Thigmokeronopsis* and *Pseudokeronopsis* are placed into the family Pseudokeronopsidae by most investigators e.g., a sister relationship between these two genera is not revealed in any of our trees, nor in previous molecular phylogenetic analyses. The relationship between *Pseudokeronopsis* and *Nothoholosticha* is clearly supported both by morphological (viz. midventral pairs arranged in a zig-zag pattern, distinctly fewer transverse cirri than midventral cirral pairs, and one marginal row on each side of the body) and phylogenetic trees based on SSU rRNA gene and ITS-5.8S-ITS2 region sequences \[, in present investigation,\]. However, no close relationship is recovered in trees containing alpha-tubulin gene sequences. As a primary conclusion, it appears that the family Pseudokeronopsidae is not monophyletic although most of its members almost certainly belong to the core portion of urostylids. Very likely, some or most pseudokeronopsids should be transferred to the family Urostylidae, although a taxonomic revision of this group must await further data. ## Classification of Acaudalia and the Family Pseudourostylidae The family Pseudourostylidae comprises three genera, *Hemicycliostyla*, *Trichotaxis* and *Pseudourostyla* (sensu Lynn). SSU rRNA gene sequence data is available for only two pseudourostylids, viz. *Pseudourostyla franzi* and *P. cristata*. This classification is consistent with that of Berger. In our SSU rRNA gene trees, two *Pseudourostyla* species group with the *Pseudokeronopsis- Nothoholosticha* cluster, which is a sister group to other typical urostylids, e.g. *Anteholosticha*, *Metaurostylopsis*, *Apokeronopsis* etc.. And Chen et al. observed that, *Pseudourostyla* is morphologically similar to *Urostyla* and *Metaurostylopsis*, albeit with some minor morphological and morphogenetic differences. The latter two, however, were assigned to the family Urostylidae by Lynn. According to Berger, *Pseudourostyla*, *Thigmokeronopsis*, *Apokeronopsis* (syn. *Thigmokeronopsis*), and *Pseudokeronopsis* are included in the unranked higher taxon Acaudalia Berger, 2006. The monophyly of Acaudalia, however, is not recovered in any of our trees, and is rejected by AU tests (*P*\<0.05), which is consistent with several previous reports, although close relationships between *Thigmokeronopsis* and *Apokeronopsis*, and between *Pseudourostyla* and *Pseudokeronopsis*, were recovered in some trees. ## Classification of the Family Epiclintidae The family Epiclintidae (Wicklow & Borrow 1990) contains two genera, viz. *Epiclintes* and *Eschaneustyla*. Due to the absence of gene sequences for *Eschaneustyla*, however, the evolutionary relationships of these genera cannot be evaluated using molecular data. The phylogenetic position of *Epiclintes* is subject to a long and ongoing dispute due to its unusual cirral pattern. As referred in Berger , it has been historically assigned to the families Oxytrichidae,, Urostylidae, Amphisiellidae, Keronidae,, Spirofilidae, or as *incertae sedis* within the order Stichotrichida. Based on morphological and ultrastructural specializations, Wicklow and Borror established the family Epiclintidae for this genus, and supposed that *Epiclintes* is a specialized descendent from *Kahliella*-like stichotrichines. In a recent study, Hu et al. rejected the placement of *Epiclintes* in the families Oxytrichidae, Amphisiellidae, and Spirofilidae, or in the order Stichotrichida. Furthermore, several morphological and morphogenetic features of *Epiclintes* were found to be inconsistent with those of urostylids, including: (1) many oblique ventral rows originating from cirral anlagen but no zigzagic pattern formed, (2) a short row of frontal cirri deriving from UM-anlage, (3) partial replacement of the old adoral zone, (4) de novo formation of the oral primordium, the anlagen for marginal rows and dorsal kineties. The results of the present study are consistent with these findings and also reject a close relationship between *Epiclintes* and *Kahliella*. As a basal clade, it branches deeply from the assemblage of three *Holosticha* and one *Psammomitra* species. Thus, all the available evidence supports the separation of the Epiclintidae at family/suborder level as suggested previously. ## Congruence/Incongruence among Different 14-Taxa Datasets Seven phylogenies based on seven different datasets (Datasets 5-11) with same taxa were topologically incongruent, however, a “core” urostylid group is revealed in each tree. *Anteholosticha manca*, *A. gracilis*, *Bergeriella*, *Metaurostylopsis*, *Thigmokeronopsis*, *Nothoholosticha* and *Pseudourostyla*, always fall into this core group, whereas *Pseudokeronopsis*, *Holosticha* and *Psammomitra* only cluster within this group in some Datasets. Among the core group, five nodes are chosen to test congruence among partitions. In these seven 14-taxa analyses, ILD, S-H and PABA tests were used to detect congruence/incongruence among different partitions. The ILD test fails to show congruence among most datasets, and only Dataset 9 is suggested to be combined. By contrast, the S-H test shows that none of the tree topologies based on combined datasets (Datasets 5, 9, 10, 11) are totally rejected by all other datasets. Furthermore, the PABA approach revealed that, apart from the addition of alpha-tubulin gene as the second partition, all additions of partitions increase average BP over all five selected nodes. This is consistent with previous investigations, the ILD test appears to be too conservative, and should only used as a measure of heterogeneity between gene partitions rather than a measure for a combinability test. The ILD test indicates that SSU rRNA and ITS1-5.8S-ITS2 are congruent, and that alpha-tubulin is incongruent with them, whereas the S-H tests fail to pinpoint the cause of conflict. For the PABA approach, the mean bootstrap alteration values in suggest that in general the SSU rRNA gene contributed the most signal, followed by ITS1-5.8S-ITS2, and then alpha-tubulin. This is consistent with results of all five separated nodes, which shows that all partitions increase BP for Node 5, whereas ITS1-5.8S-ITS2 decrease BP for Node 1 , and alpha-tubulin decrease BP for the other three nodes. This is reasonable, considering that the SSU rRNA and ITS1-5.8S-ITS2 genes locate near each other, and SSU rRNA possesses most characters in our analyses. # Materials and Methods ## Selection and Identification of Ciliates The taxa in this study were selected to represent the morphological and morphogenetic diversity of Urostylida. Although the current taxon sampling does not cover all genera in Urostylida, representative taxa for each family were included. *Bergeriella ovata* (Liu et al. 2010), *Parabirojimia multinucleate* (Chen et al. 2010), and *Pseudoamphisiella quadrinucleata* (Shen et al. 2008) were collected from the coast near Guangzhou, southern China (22°42′N; 114°32′E). Other species and strains were collected from the coast near Qingdao, northern China (36°08′N; 120°43′E). All isolates were identified by the methods of Shao et al. and Li et al.. Terminology and systematic classification used in the current paper follow Berger and Lynn, respectively. ## Extraction and Sequencing of DNA Genomic DNA was extracted according to methods described in Yi et al.. Eukaryotic universal A (5′-AACCTGGTTGATCCTGCC AGT-3′) or 82F (5′-GAAACTGCGAATGGCTC-3′) and Eukaryotic universal B (5′-TGATCCTTCTGCAGGTTCACCTAC-3′) primers were used for amplification of the SSU rRNA gene by polymerase chain reaction (PCR). Cycling parameters for the SSU rRNA gene were as follows: 5 min initial denaturation (94°C), followed by 35 cycles of 1 min at 95°C, 1 min 30 s at 56°C, and 2 min at 72°C, with a final extension of 15 min at 72°C. A fragment of approximately 500 bp containing the ITS1, 5.8S ribosomal gene, and ITS2 was amplified using primers ITS-F (5′-GTAGGTGAACCTGCGGAAGGATCATTA-3′) and ITS-R (5′-TACTGATATGCTTAAGTTCAGCGG-3′), with the following cycling parameters: 5 min initial denaturation (94°C), followed by 35 cycles of 30 s at 95°C, 1 min at 56°C, and 1 min at 72°C, with a final extension of 15 min at 72°C. A fragment of approximately 1,000 bp comprising part of the alpha-tubulin gene was amplified using ciliate-specific primers Tub-1 (5′-AAGGCTCTCTTGGCGTACAT-3′) and Tub-2 (5′-TGATGCCTTCAACACCTTCTT-3′). Cycling parameters were as follows: 5 min initial denaturation (94°C), 35 cycles of 30 s at 94°C, 1 min at 56°C, and 1.5 min at 72°C, with a final extension of 15 min at 72°C. Purified PCR product of appropriate size was inserted into the pUCm-T vector (Shanghai Sangon Biological Engineering & Technical Service Company, China) and sequenced at the Invitrogen sequencing facility in Shanghai, China. ## Databases Selection Eight datasets were evaluated in our analyses: (1) SSU rRNA gene sequences including all available urostylid sequences plus some other spirotricheans (89 sequences in total); (2) ITS1-5.8S-ITS2 region sequences including all available urostylid sequences plus some other spirotricheans (31 sequences in total); (3) alpha-tubulin gene sequences including all available urostylid sequences plus some other spirotricheans (26 sequences in total); (4) three-gene combined dataset including all spirotrichean species available, and *Protocruzia adherens*, *Stylonychia mytilus* and *Sterkiella nova* for SSU rRNA and ITS1-5.8S-ITS2, and *Protocruzia contrax*, *Stylonychia lemnae* and *Sterkiella cavicola* for alpha-tubulin (25 sequences in total); (5) three-gene combined dataset including all available urostylid species (14 sequences in total); (6) SSU rRNA gene sequences including all taxa in Dataset 5; (7) ITS1-5.8S-ITS2 region sequences including all taxa in Dataset 5; (8) alpha-tubulin gene sequences including all taxa in Dataset 5; (9) two-gene combined dataset composed of Datasets 6 and 7; (10) two-gene combined dataset composed of Datasets 6 and 8; (11) two-gene combined dataset composed of Datasets 7 and 8. ## Secondary Structure Prediction and ITS2 Sequence Alignment The default settings of the mfold website (<http://frontend.bioinfo.rpi.edu/applications/mfold>) were used to produce the secondary structure and sequence in dot-bracket structural format of ITS2 RNA transcripts. The structures were edited for aesthetic purposes with RnaViz 2.0. The ITS2 sequences with the secondary structure format were aligned using the MARNA web server (<http://biwww2.informatik.uni- freiburg.de/Software/MARNA/index.html>), based on both the primary and secondary structures. ## Phylogenetic Analyses Sequences (except for ITS2 sequences) were aligned using the ClustalW implemented in Bioedit 7.0.0 and further modified manually using Bioedit. The final alignment of Dataset 1 included 1,607 positions, and the alignment is available from the authors upon request. A Bayesian inference (BI) analysis was performed with MrBayes 3.1.2 using the GTR+I+G model selected by MrModeltest 2 under the AIC criterion. Markov chain Monte Carlo (MCMC) simulations were run with two sets of four chains using the default settings: chain length 2,000,000 generations, with trees sampled every 100 generations. The first 5,000 trees were discarded as burn-in. The remaining trees were used to generate a consensus tree and to calculate the posterior probabilities (PP) of all branches using a majority-rule consensus approach. A Maximum Likelihood (ML) analysis was performed with PhyML V2.4.4 using the GTR+G+I model selected under the AIC criterion by Modeltest v.3.7. The reliability of internal branches was assessed using a non-parametric bootstrap method with 1,000 replicates. The following evolutionary models were selected by MrModeltest 2 for single datasets: GTR+I model for Datasets 2 and 7; GTR+I+G for Datasets 3, 6, and 8. Using these selected models, Bayesian trees for Datasets 2, 3, 4 were built as above. For Dataset 4, individual coding regions were treated as ‘unlinked’, so that separate parameter estimates as specified above were obtained for each gene partition for all runs. The following evolutionary models were selected by Modeltest v.3.7 for different datasets: GTR+I model for Datasets 2 and 7; GTR+I+G for Datasets 3– 6, 8–11. Using these selected models, ML trees for Datasets 2–4 were constructed as above. Phylogenetic trees were visualized with TreeView v1.6.6 and MEGA 4. ## Identifying of Congruence or Incongruence Congruence of different data partitions (in this case genes) was tested with both the incongruence length difference (ILD) test and Shimodaira-Hasegawa (S-H) test as implemented in PAUP\*4.0b. We excluded taxa with missing data in some gene partitions, and performed the ILD tests with Dataset 4 and Dataset 5, respectively. Six gene-by-gene comparisons were conducted based on 1,000 ILD replicates. In interpreting the results of ILD tests, recent studies have shown that the utility of the ILD test is limited as a measure of the incongruence among data partitions. Therefore, we used the ILD tests as a measure of heterogeneity between gene partitions and the results of ILD tests were not interpreted as a measure for a combinability test. In the case of S-H tests, variance estimations of the difference in the likelihood values of given topologies to the best topology were used to test whether the topology produced by a given partition was accepted or rejected by different data partitions. Therefore, the major-rule consensus topologies obtained by the 7 different 14-taxon datasets were compared to each other based on each of these datasets using the S-H test. RELL approximations with 1,000 replicates and ML methods described above were conducted. Because neither of these two approaches sufficiently described the source of possible incongruence and its influence in the dataset, the partition addition bootstrap alteration (PABA) approach was used to evaluate the influence of combining genes on nodal support of “core Urostylida”, five nodes with high supports in three gene combined tree were selected. # Supporting Information Our deepest gratitude goes to Dr. Alan Warren, Natural History Museum, UK, for his help in improving written English. Many thanks are due to Ms. Feng Gao and Jie Huang, Laboratory of Protozoology, OUC for gene sequencing. [^1]: Conceived and designed the experiments: ZY WS. Performed the experiments: ZY. Analyzed the data: ZY WS. Contributed reagents/materials/analysis tools: WS. Wrote the manuscript: ZY WS. [^2]: The authors have declared that no competing interests exist.

# Introduction Recent evidence has demonstrated that aspirin prophylaxis can reduce the incidence of preterm preeclampsia (PE) in asymptomatic pregnant women screened high-risk for preterm PE by 62%. As such, several professional organizations have recommended the first trimester PE screening by the “triple test” developed by the Fetal Medicine Foundation (FMF) for all singleton pregnancies to reduce PE associated maternal and perinatal mortality, morbidity, and their related costs. The first trimester triple test consists of maternal factors (MF), measurement of mean arterial pressure (MAP), uterine artery pulsatility index (UTPI), and placental growth factor (PlGF) achieved a superior screening performance for PE compared with the risk factor-based screening. Relative differences in detection rates (DRs) of the FMF triple test amongst different racial groups has been reported. At the same 10% false positive rate (FPR), we previously reported that DR in East Asian populations was only 64%, 11% lower than the 75% achieved in mixed European population. The lower DR in Asian women could be explained by the lower incidence of PE risk factors, nevertheless, there may be some other unknown disease modifying factors in which the current test needs to be improved. One way to enhance the performance of the FMF screening test would be the inclusion of additional biomarkers. Inhibin-A is a glycoprotein expressed during pregnancy by the placenta. Previous studies, assessing inhibin-A after 16 weeks’ gestation using enzyme-linked immunoassay (ELISA) assays whilst suitable for explorative research are less useful for routine screening, reported that maternal serum levels of inhibin-A in pregnancies developing PE are elevated compared to unaffected pregnancies. Meta-analysis has shown that aspirin prophylaxis needs to be initiated before 16 weeks to reduce the incidence of preterm PE. For inhibin-A to be considered as a potential PE screening marker it would have to be shown to be able to differentiate between PE and non-PE pregnancies before 16 weeks using standard immune-analyzers. Whilst case-control studies indicate that inhibin-A levels in pregnancies with established PE are significantly higher, there still remains a lack of consensus as to which temporal timepoint the difference in levels is evident; the additional improvement in DR that adding inhibin-A would result in when screening for preterm and term PE as opposed to early (\< 34 weeks) and late (≥34 weeks) onset PE; whether differences reported in earlier studies based on measurement of inhibin-A by ELISA are comparable to immunoassay measurements. The objectives of this study were firstly to determine whether inhibin-A levels in PE pregnancies determined using a standard immunoassay were increased at 11–13 weeks and secondly if so whether its use would improve the DRs for preterm PE if incorporated within the FMF triple test. # Materials and methods This was a nested case-control study using archived sera collected from 6,546 consecutive Chinese women with a singleton pregnancy, which resulted in a livebirth from a prospective cohort of pregnant women at 11–13 weeks’ gestation attending Down Syndrome screening program at our institution between December 2016 and June 2018 and participated in a Asian-based study for the validation of the FMF triple test preterm PE screening. All eligible women who agreed to participate were asked to provide written informed consent for storage and use of archived sera in future research at the time of screening. Ethical approval for the base-cohort study was obtained from the Joint Chinese University of Hong Kong—New Territories East Cluster Clinical Research Ethics Committee (CREC Ref. No. 2016.152). Excess serum was stored at -80°C for future research. Authors had access to information that could identify individual participants after data collection. The ethical approval for this retrospective study (CREC Ref. No.: 2021.258) was obtained from the Joint Chinese University of Hong Kong—New Territories East Cluster Clinical Research Ethics Committee. Data on pregnancy outcomes were collected from the hospital maternity electronic records. The diagnosis of PE was based on the definition of the International Society for the Study of Hypertension in Pregnancy. Preterm and term PE was defined as delivery with PE at \<37 and ≥37 weeks’ gestation, respectively. This work was supported by i. A start up grant from the Faculty of Medicine, The Chinese University of Hong Kong. ii. Grants from the Health and Medical Research Fund, Hong Kong SAR, China (HMRF-18190821) and iii. The Ministry of Science and Technology (MOST), China (No. 2021YFC2701600). The study sponsors had no role in the study design, collection, analysis, and interpretation of the data, or in the writing of the article. The primary aims of the study were to determine if the addition of inhibin-A would increase the DR and screening accuracy of the FMF triple test. Our previous studies indicated that before adjusting for aspirin use the DR and areas under the receiver operational curves (AUC) when screening for preterm PE using the FMF triple test were 59% and 0.85 respectively, whilst the DR and AUC when screening for any onset PE using maternal history, MAP and PlGF were 51% and 0.79, respectively. The sample size of the present study was therefore based on the assumption that the addition of inhibin-A to the screening test for any onset PE would detect an improvement in AUC from 0.8 to 0.85, a ΔAUC = 0.05. Assuming a correlation between screening tests in those screened high risk and low risk of 0.8 it would require 107 pregnancies with PE and a ratio between non-PE and PE pregnancies of 15 to achieve a power of 80% for a type 1 error of 5%. In the base-cohort study reported by Wah *et al*., 112 (1.7%) women developed PE (“PE”), 37 (0.6%) and 75 (1.1%) of whom had preterm and term PE. Fifteen unaffected pregnancies screened within ± 30 days of each PE affected pregnancy were therefore randomly selected from amongst the available screened pregnancies who did not develop PE to form an unaffected group for comparison of PE screening performance. ## Measurement of inhibin-A Archived sera were retrieved from -80°C storage and thawed in batches using a slow defrost protocol. Inhibin-A levels were determined using the BRAHMS KRYPTOR Gold Analyzer (ThermoFisher Scientific, Hennigsdorf, Germany). The inter- and intra-assay coefficient of variation of the analyzer inhibin-A immunoassay ranged from 2–6%. ## Adjusting inhibin-A levels for covariate fixed and random factors Inhibin-A levels of women with unaffected pregnancy were transformed to their equivalent log₁₀ values then assessed to determine whether their levels were independent of gestational age (GA) at time of blood draw, maternal age, weight, height, parity and history of PE in the previous pregnancy (nulliparous, parous with prior PE, parous without prior PE), smoking status, diabetic status, and method of conception (spontaneous, IVF \[*in vitro* fertilization\]) using univariate and multivariate regression analyses. Fixed and random factors with a significant impact (p\<0.05) on log₁₀ inhibin-A level were retained in the final model used to calculate the expected median values of log₁₀ inhibin-A. Inhibin-A levels were then transformed to their equivalent multiple of the expected median (MoM) values. Gestational age and weight were centered on 77 days and 69 kg, respectively, the same values as those previously used by Tan *et al*.. ## Estimation of preterm preeclampsia risk Individual women’s *a priori* risks of delivery with preterm PE and term PE were estimated based on maternal factors (MF) using the FMF competing risk model. The *posteriori* risks were determined using Bayes theorem-based approach by adjusting the *a priori* with the likelihood function of biomarker measurements. All risks for PE were estimated using our in-house laboratory PE risk calculation software and MAP, UTPI and PlGF were transformed to their equivalent MoM values using published transformation models. Expected levels and biomarker MoM distributions of MAP, UTPI and PlGF in unaffected and PE pregnancies used to estimate the *posteriori* risks were previously reported in the National Institute for Health Research SPREE study. Our earlier studies indicated that the overall median PlGF MoM in non-PE pregnancies in East Asians using the FMF PlGF MoM transformation model gave a distribution with a median of 0.84. Prior to calculating the PE risks, PlGF MoMs were corrected to recenter the PlGF MoM distribution to a median of 1 MoM in order to avoid overestimation of preterm PE risks. Gestational age at the time of screening used to derive the biomarker MoM values was determined from the fetal CRL measurement using a previously published Chinese dating formula. ## Data processing and statistical analysis Maternal demographic and biomarker characteristics were presented in median (interquartile range \[IQR\]). Comparisons were performed by Mann-Whitney U test for continuous variables and Chi-square test for categorical variables. Descriptive statistics was used to describe the central tendency and distribution of log₁₀ inhibin-A MoM in any onset PE, preterm PE, term PE and unaffected pregnancies. Independent samples t-test and ANOVA with Bonferroni correction for multiple comparison (p value\<0.016) were used to determine the differences in mean levels of log₁₀ inhibin-A MoM between outcome groups. Correlation between log₁₀ inhibin-A MoM and other biomarkers log₁₀ transformed MoM values in PE, unaffected as well as in all pregnancies was assessed. Linear regression analysis was performed to determine whether the level of log₁₀ inhibin-A MoM at 11–13 weeks was associated with gestational age at delivery in those with PE. ## Comparison of for preterm and term preeclampsia screening performance Receiver operating characteristic (ROC) curves were constructed and AUCs were determined. DRs at a 10% fixed FPR and corresponding risk level (1:XXX) were determined for each of the different screening biomarker combinations assessed. Differences between AUCs were tested for significance using the Delong test. McNemar’s test was used to determine the percentage change in the off-diagonal probabilities by adding or removing inhibin-A to the PE screening test biomarker at a fixed 10% FPR. Statistical Product and Service Solutions (SPSS) for Windows version 20 (SPSS, Illinois, USA) and MedCalc Statistical Software version 18.10.2 (MedCalc Software bvba, Ostend, Belgium; [http://www.medcalc.org](http://www.medcalc.org/); 2018) were used for statistical analyses. Tests were considered statistically significant if p-value \<0.05. # Results The maternal, pregnancy and screening biomarker characteristics according to pregnancy outcomes are summarized in. Women who developed PE showed the expected traits with regard to obstetric and medical history as well as levels of their biomarkers. Compared with the unaffected pregnancies, women with PE had higher maternal age and BMI, higher rates of IVF conception, nulliparity and previous history of PE, and lower rate of history of previous pregnancy without PE. There was no difference in the rates of smoking, chronic hypertension and SLE/APS between groups. Although there was a higher rate of pre-existing diabetes mellitus in the PE group, compared to the unaffected group, the total number of cases was small. Inhibin-A levels at 11–13 weeks in non-PE pregnancies were found to be significantly dependent on GA (p\<0.001), maternal age (p\<0.001), maternal weight (Wgt) (p\<0.001) and parous without prior PE (p\<0.001). The final model used to transform measured inhibin-A to it equivalent MoM value is reported in. The distribution of log₁₀ inhibin-A MoM in unaffected pregnancies was Gaussian with a mean of -0.0135 and a standard deviation (SD) of ±0.2108. Mean (SD) log₁₀ MoM inhibin-A in any-onset PE, preterm PE and term PE pregnancies were significantly higher than that of unaffected pregnancies \[0.0911 (±0.2571), 0.1593 (±0.2657), 0.0575 (±0.2477) with p-values of \<0.001, \<0.001 and 0.015, respectively\]. summarizes the observed covariance and correlation coefficients of inhibin-A with that of existing biomarkers. Levels of log₁₀ inhibin-A MoM in pregnancies that developed PE were negatively associated with GA at delivery. The correlation however failed to reach statistical significance (r = -0.132, p = 0.165). ## Preterm preeclampsia prediction For the prediction of preterm PE, the AUC when screening with MF only was 0.719 (95% confidence interval \[CI\], 0.63–0.81). Only screening by MF plus PlGF significantly increased the AUC when compared with MF only (p = 0.021). There were no significant differences among various two-biomarker combinations as well as among three-biomarker combinations. Tables and and summarize the overall screening performance of various combinations of biomarkers including adding inhibin-A or replacing PlGF with inhibin-A for the prediction of preterm PE. The best AUC was achieved by MF + (MAP, PlGF, inhibin-A) at 0.861 (95%CI, 0.802–0.919). Screening by MF + (MAP, UTPI, inhibin-A) had significantly better performance than screening by MF alone and any two-biomarker combinations. The screening performance using MF + (MAP, UTPI, inhibin-A) achieved lower AUC than MF + (MAP, UTPI, PlGF) but did not reach statistical significance. Screening by MF + (MAP, UTPI, PlGF, inhibin-A) significantly worsened the screening performance as compared with screening by MF + (MAP, UTPI, PlGF) (ΔAUC = -0.045, p = 0.001). At a fixed 10% FPR, a DR for preterm PE using MF only was 37.84% (95%CI, 24.32–56.76%) and screening by MF + (MAP, UTPI, PlGF) yielded the best DR at 64.86% (95%CI, 48.65–81.08%). McNemar’s test indicated that substituting PlGF with inhibin-A in combination with MAP and UTPI identified only 1 (2.7%) additional pregnancy but missed 5 (13.5%) pregnancies developing preterm PE that could have been identified by PlGF. Combining inhibin-A with maternal factors and the existing FMF triple test missed 4 (10.8%) pregnancies and did not identify any additional pregnancies developing preterm PE. ## Term preeclampsia prediction For the prediction of term PE, Tables and and summarize the overall screening performance of various combinations of biomarkers. Compared with the performance of MF alone for term PE prediction, which achieved a similar AUC as preterm PE prediction at 0.719 (95%CI, 0.66–0.78), only a combination of MF with PlGF could improve the screening performance (ΔAUC = 0.035, p = 0.037). Substituting PlGF with inhibin-A in the FMF triple test performed better than MF alone (ΔAUC = 0.050, p = 0.038) but did not improve screening performance of term PE prediction compared with the FMF triple test (ΔAUC = -0.016, p = 0.152). Combining inhibin-A to the FMF triple test did not yield any improvement in term PE screening performance over MF only (ΔAUC = 0.041, 0.098) and MF + (MAP, UTPI, inhibin-A) (ΔAUC = -0.008, p = 0.611), whilst it demonstrated inferior performance compared with the Triple test (ΔAUC = -0.024, \<0.001). McNemar’s test for term PE prediction demonstrated that substituting PlGF with inhibin-A in combination with MAP and UTPI identified 4 (5.3%) additional pregnancies but missed 10 (13.3%) pregnancies developing term PE that could have been identified by PlGF. Combining inhibin-A with maternal factors and the existing FMF triple test missed 6 (8%) pregnancies and identified 1 (1.3%) additional pregnancy with term PE. # Discussion ## Principal findings Our study has demonstrated that, although levels of inhibin-A at 11–13 weeks’ gestation are significantly higher in pregnancies that subsequently develop PE compared with unaffected pregnancies, replacing PlGF with inhibin-A in the existing FMF triple test does not improve the performance of preterm and term PE screening; while adding inhibin-A as the fourth biomarker decreases the screening performance. ## Results Inhibin-A levels in pregnancies with established PE have been reported as being significantly higher than those in unaffected pregnancies \[, –\] leading to the postulation that increased levels of inhibin-A in PE pregnancies is a consequence of trophoblast dysfunction allowing potential prediction of when and in which pregnancy PE could occur \[, –\]. Muttukrishna *et al*. conducted a longitudinal study and indicated that significant differences in inhibin-A levels, determined by ELISA, were only evident after 15–19 weeks of gestation in those who developed early onset PE. The authors however did not correct for independent factors such as GA and maternal weight, both of which would have been expected to impact on the determined inhibin-A level. In contrast, Akolekar *et al*., Spencer *et al*. and Poon *et al*., all using an ELISA test, reported significantly increased levels of inhibin-A at 11–13 weeks in plasma and serum after correcting for maternal and pregnancy factors. Poon *et al*. in a later study reconfirmed that inhibin-A levels were increased at 11–13 weeks in PE affected pregnancies using a standard immunoassay for the measurement of inhibin-A after correcting for maternal and pregnancy factors. Our data and analysis confirmed that inhibin-A levels determined by immunoassay at 11–13 weeks are similarly increased in PE affected pregnancies of East Asian women indicating that inhibin-A could be used for PE screening. We have further demonstrated that there are lower inhibin-A levels in parous women without prior PE, which reduces the risk for developing PE in the current pregnancy. This finding may indicate that women with a previous unaffected pregnancy have the ability to accomplish adequate maternal adaptation to tolerate pregnancy-related stress in the previous pregnancy, thus providing a protective effect against the development of PE in the current pregnancy. Using the gestational 90^th percentile of inhibin-A in normotensive pregnancies as a cut-off, Muttukrishna *et al*. reported screening sensitivities at 15–19 weeks for any and early onset PE of 28% and 67%, respectively, for a 12% FPR. Using a similar cut-off, Spencer *et al*. reported that inhibin-A had a DR of 35% for a 5% FPR which increased to 67% for the same FPR when combined with second trimester UTPI. Akolekar *et al*. reported similar DRs of 23% and 31% when screening by inhibin-A alone for 5% and 10% FPR and that DR increased to 85% and 89% at the same FPRs when combined with maternal history and UTPI. However, when screening for early onset PE in combination with PlGF and inhibin-A, Poon *et al*. reported that only PlGF and not inhibin-A remained as an independent predictor for development of early onset PE when constructing a multivariate logistic regression prediction model for early onset PE. The current approach to screening for PE has evolved from to one based on estimating marginal probability of an event in the presence of competing event combined with Bayes theorem allowing multiple biomarkers to be combined and assessment of screening performance. Using this approach Poon *et al*. in a recent study showed that inhibin-A did not improve the DRs of any-onset PE and preterm PE over the prediction provided by PlGF in a predominantly White, Black and South Asian population. Our analysis and findings in East Asian women would confirm Poon *et al*. study with regard to concurrent use of inhibin-A and PlGF to screen for preterm PE at 11–13 weeks. ## Clinical implications This study demonstrating that adding inhibin-A to or replacing a biomarker by inhibin-A in the currently used FMF triple test does not significantly improve the screening performance could be explained by the concept of diminishing marginal returns and the fact that inhibin-A levels significantly correlate with PlGF levels in the PE group. ## Research implications Our findings in this present study highlight that although a potential biomarker is associated with development of PE its inclusion in a risk estimation model for development of PE will not necessarily improve screening performance if it is significantly correlated with existing biomarkers or if its discriminatory performance, reflected by Mahalanobis distance, is on a par with existing biomarkers. Mahalanobis distance of inhibin-A in this study was 0.82 for preterm PE which is similar to the 0.88 estimated by Cuckle for early-onset PE. The MAP, UTPI and PlGF biomarker combination to screen for PE remains the best combination to use in the first trimester, and that PlGF will not be replaced by another placentally derived biomarker unless and until it has significantly higher Mahalanobis distance which exceeds that of PlGF and similar to that of MAP and UTPI or alternatively that any new biomarker incorporated reflects a different pathogenesis or dimension of the PE, such as maternal endothelial, cardiac and end-organ dysfunction related biomarkers. ## Strengths and limitations The strengths of our study were firstly, that inhibin-A was measured using a readily available immunoassay and immune-analyzer platform; secondly, that we assessed screening performance using the competing risk and Bayes based model, a de facto gold standard approach; and lastly that we assessed relative improvement or loss of screening performance on an individual case basis. A limitation of the current study was that we did not adjust the screening performance for aspirin prophylaxis. # Conclusions In conclusion, inhibin-A is significantly elevated at 11–13 weeks’ gestation in pregnancies that subsequently develop PE but has lower predictive performance for the screening of both preterm and term PE in the first trimester when compared with PlGF. Neither replacing PlGF by inhibin-A nor adding inhibin-A could enhance the screening performance of the FMF triple test for both preterm and term PE. # Supporting information We wish to thank the members of the Obstetrics Screening Laboratory, Maternal Fetal Medicine team, midwives, nurses, research students and assistants at the Prince of Wales Hospital in facilitating the performance of this study. 10.1371/journal.pone.0288289.r001 Decision Letter 0 Erez Offer Academic Editor 2023 Offer Erez This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 27 Mar 2023 PONE-D-22-29184Evaluation of first trimester maternal serum inhibin-A for preeclampsia screeningPLOS ONE Dear Dr. Moungmaithong, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, the manuscript received favorable reviews and requires minor revision to address the concerns raised by the reviewers. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by May 11 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Offer Erez, M.D. Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at  <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and  <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Please amend your ethics statement in the Methods section of your manuscript. Specifically, please state that the ethics approval for this retrospective study (CREC Ref. No.: 2021.258) was obtained from the Joint Chinese University of Hong Kong - New Territories East Cluster Clinical Research Ethics Committee. 3\. Thank you for stating the following financial disclosure:  This work was supported by a grant from the Health and Medical Research Fund, Hong Kong SAR, China (HMRF-18190821). The study sponsor had no role in the study design, collection, analysis, and interpretation of the data; or in the writing of the article. The base-cohort study “Prospective validation of prediction algorithms for preeclampsia in the first-, second- and third-trimesters of pregnancy” was supported by a grant from Department of Obstetrics and Gynaecology, Prince of Wales Hospital, The Chinese University of Hong Kong (CUHK), Hong Kong SAR. Please state what role the funders took in the study.  If the funders had no role, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript."  If this statement is not correct you must amend it as needed.  Please include this amended Role of Funder statement in your cover letter; we will change the online submission form on your behalf. 4\. Thank you for stating the following in the Competing Interests section:  LCP has received speaker fees and consultancy payments from Roche Diagnostics and Ferring Pharmaceuticals. LCP and DS have received in‐kind contributions from Roche Diagnostics, PerkinElmer, Thermo Fisher Scientific, and GE Healthcare. Other co-authors report no conflicts of interest. Please confirm that this does not alter your adherence to all PLOS ONE policies on sharing data and materials, by including the following statement: "This does not alter our adherence to  PLOS ONE policies on sharing data and materials.” (as detailed online in our guide for authors <http://journals.plos.org/plosone/s/competing-interests>).  If there are restrictions on sharing of data and/or materials, please state these. Please note that we cannot proceed with consideration of your article until this information has been declared.  Please include your updated Competing Interests statement in your cover letter; we will change the online submission form on your behalf. 5\. Please amend your manuscript to include your abstract after the title page. 6\. Please include your tables as part of your main manuscript and remove the individual files. Please note that supplementary tables (should remain/ be uploaded) as separate "supporting information" files. 7\. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: [http://journals.plos.org/plosone/s/supporting- information. ](http://journals.plos.org/plosone/s/supporting-information. ) 8\. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Not withstanding its acknowledged low dose aspirin issue limitation and retrospective design, this is a clearly described, statistically sound and biomedically important contribution to the current literature dealing with early pregnancy biomarkers that may help predict preeclampsia later in a pregnancy. Although the conclusion of the study is in one sense "negative" in so far as the study results do not support the use of inhibin as a biomarker in place of or in addition to currently used options (such as PlGF in particular), it is important to report such an outcome, if for no other reason than to minimise others wasting time and effort in pursing the same possibility. I see no value in nitpicking minor and/or trivial issues and recommend the manuscript's publication as submitted. Reviewer \#2: This is an interesting paper that aimed at investigating whether maternal serum inhibin-A at 11-13 weeks could be as an alternative to PlGF or as an additional biomarker within the FMF screening test for preterm PE. In a nested case-control study, the authors found that replacing PlGF by inhibin-A or adding inhibin-A as an additional biomarker in and to the FMF triple screening test for preterm PE does not improve screening performance and will fail to identify pregnancies that are currently identified by the FMF triple test. It is a nicely conducted study and well written manuscript. I suggest modifying the following issues before the publication of the paper: 1\) What can be the reason for lower inhibin-A levels in parous women with no previous history of PE? This could be shortly discussed. 2\) Table 1: by looking at the differences between groups in many comparisons, several in the term PE vs control comparisons (e.g. IVF, nulliparous, etc) could be significant. Please double check the numbers. 3\) It is stated that tests were considered statistically significant if the p-value was \<0.05. What was the threshold after Bonferroni correction? 4\) It is stated that “inhibin-A in the currently used FMF triple test does not significantly improve the screening performance could be explained by the concept of diminishing marginal returns and the fact that inhibin-A levels correlate with PlGF levels in PE groups.” However, there was no testing of any correlation. Please perform the test or remove the statement. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0288289.r002 Author response to Decision Letter 0 25 Apr 2023 Response letter to reviewers’ and editors’ comment Dear editors and reviewers, Thank you very much for your comments and suggestion. The manuscript was amended, and the following points were addressed. 1\) What can be the reason for lower inhibin-A levels in parous women with no previous history of PE? This could be shortly discussed Response: Thank you, the discussion based on this issue was addressed as following: “We have further demonstrated that there are lower inhibin-A levels in parous women without prior PE, which reduces the risk for developing PE in the current pregnancy. This finding may indicate that women with a previous unaffected pregnancy have the ability to accomplish adequate maternal adaptation to tolerate pregnancy-related stress in the previous pregnancy, thus providing a protective effect against the development of PE in the current pregnancy..” 2\) Table 1: by looking at the differences between groups in many comparisons, several in the term PE vs control comparisons (e.g., IVF, nulliparous, etc.) could be significant. Please double check the numbers. Response: Thank you, the description of baseline characteristics was expanded as following: “Compared with the unaffected pregnancies, women with PE had higher maternal age and BMI, higher rates of IVF conception, nulliparity and previous history of PE, and lower rate of history of previous pregnancy without PE. There was no difference in the rates of smoking, chronic hypertension, and SLE/APS between groups. Although there was a higher rate of pre-existing diabetes mellitus in the PE group, compared to the unaffected group, the total number of cases was small.” 3\) It is stated that tests were considered statistically significant if the p-value was \<0.05. What was the threshold after Bonferroni correction? Response: Thank you, ANOVA with Bonferroni correction with threshold (p value\<0.016) for multiple comparison was used. It was added to the manuscript. 4\) It is stated that “inhibin-A in the currently used FMF triple test does not significantly improve the screening performance could be explained by the concept of diminishing marginal returns and the fact that inhibin-A levels correlate with PlGF levels in PE groups.” However, there was no testing of any correlation. Please perform the test or remove the statement. Response: Thank you, S2 Table which demonstrated observed log10 MoM biomarker distribution standard deviation and inter-biomarker correlations in women with and without preeclampsia was added. It demonstrated significant correlation between inhibin-A and PlGF levels. Best regards, Sakita Moungmaithong 10.1371/journal.pone.0288289.r003 Decision Letter 1 Maged Ahmed Mohamed Academic Editor 2023 Ahmed Mohamed Maged This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 26 Jun 2023 Evaluation of first trimester maternal serum inhibin-A for preeclampsia screening PONE-D-22-29184R1 Dear Dr. Poon, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Ahmed Mohamed Maged, MD Academic Editor PLOS ONE Additional Editor Comments (optional): All reviewers comments were clarified Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#2: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The authors have adequately addressed the minor corrections/amendments suggested by the review process. Reviewer \#2: The authors have revised their manuscript according to the comments of the reviewers, therefore, now it is acceptable for publication. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* 10.1371/journal.pone.0288289.r004 Acceptance letter Maged Ahmed Mohamed Academic Editor 2023 Ahmed Mohamed Maged This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 2 Jul 2023 PONE-D-22-29184R1 Evaluation of first trimester maternal serum inhibin-A for preeclampsia screening Dear Dr. Poon: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Professor Ahmed Mohamed Maged Academic Editor PLOS ONE [^1]: LCP has received speaker fees and consultancy payments from Roche Diagnostics and Ferring Pharmaceuticals. LCP and DS have received in‐kind contributions from Roche Diagnostics, PerkinElmer, Thermo Fisher Scientific, and GE Healthcare. This does not alter our adherence to PLOS ONE policies on sharing data and materials. Other co-authors report no conflicts of interest. [^2]: ‡ AHK, AWT, NKW, MSL and JW also contributed equally to this work.

# Introduction Although the pathogenesis of inflammatory bowel disease (IBD), such as Crohn’s disease and ulcerative colitis in humans, still remains unclear, chronic epithelial permeability seems to be one of the mechanisms by which extensive inflammatory factors may be introduced into the irritated intestinal tissues. Therefore, it is believed that induction of mucosal healing is critical in the management of IBD. Furthermore, chronic inflammation is believed to associate with carcinogenesis, and prolonged duration of IBD likely also lead to colitis- associated cancer (CAC). Previous study had shown that activation of NF-κB in the inflamed tissue is strongly associated with carcinogenesis. In this regard, we have investigated the mechanism of NF-κB activation in the colonic epithelial cells using a murine model of IBD. We have previously reported that increased expression of tumor necrosis factor (TNF) in a murine IBD model is critical for the development of CAC. TNF is a pivotal cytokine associated with the continuous immune dysregulation in the inflamed tissue of IBD. In our previous study, the specific up-regulation of the type 2 receptor for TNF (TNFR2) was also observed in the inflamed intestinal epithelial cells. This observation seems logical since the cytoplasmic domain of TNFR2 can also activate NF-κB pathway, but it lacks association with the death domains (DD) like that of TNFR1. However, the specific role of such NF-κB activation in the inflamed epithelia via TNFR2 signaling in the context of CAC has not been elucidated. Myosin light chain kinase (MLCK) has also been reported to be expressed in the human intestinal tissue with IBD. MLCK is classically known to be required for the contraction of actomyosin via the phosphorylation of myosin light chain (MLC). It is also essential to the permeability of epithelial barrier according to in vitro and in vivo studies, and it is associated with the production of pro-inflammatory cytokine, such as TNF, in the inflamed intestinal tissues. In addition, several recent reports have implicated the role of MLCK in animal models of IBD. However, the association between MLCK and CAC development has not been reported. We hypothesized that one of the roles of epithelial NF-κB activation would be the induction of MLCK in the context of IBD. We therefore examined the role of MLCK in the development of IBD-associated carcinogenesis. # Materials and Methods ## Cell Culture Murine colonic epithelial cell line, MOC1, which was generated from ‘non-tumor’ colonic epithelia of BALB/c and transformed with SV40 large T antigen, was established by Dr. M. Totsuka (University of Tokyo, Japan) and maintained in RPMI 1640 (Sigma, St. Louis, MO) supplemented with 5% fetal bovine serum, 500 units/ml penicillin, 100 µg/ml streptomycin (Sigma) and 10 µg/ml insulin (Sigma) at 37°C in 5% CO₂. Cells were seeded at a density of 5×10⁴ cells/ml in 6-well plates 24–36 h prior to the experiments with or without recombinant (r) mouse interferon (IFN)-γ and/or r mouse TNF (Peprotek, London, UK). In some experiments, cells were also incubated in the presence of either blocking anti-mouse TNF monoclonal antibody (mAb) (MP6-XT22, rat IgG1b) (DNAX Research Institute, Palo Alto, CA) or MLCK inhibitor, ML-7 (Sigma). ## Animals Wild-type female C57BL/6 mice (6–8 wk old) were purchased from Japan Clea (Tokyo, Japan) and maintained under specific pathogen-free conditions in the Animal Care Facility of Tokyo Medical and Dental University (TMDU), Japan. Mice were used between 8–10 weeks of age. All animal experimentations were approved by the Animal Review Board of TMDU and were performed in accordance with institutional guidelines. ## Induction of Chronic Colitis and CAC Models Mice were randomized by body weight into three groups (n = 5) and given intraperitoneal (i.p.) 10 mg/kg of azoxymethane (AOM, Sigma) at day −7, followed by the administration of three cycles of 2.0% dextran sodium sulfate (DSS, molecular weight 10,000; Yokohama Kokusai Bio, Kanagawa, Japan) for 5 days and regular water for 16 days, and then injected i.p. with either the inhibitor against MLCK, ML-7 (2.0 mg/kg) or the vehicle control (2.0% ethanol) every 12 h from day 63 for 7 days. In some experiments, mice were injected i.p. with either 2.0 mg/kg of ML-7 every 12 h or 50 mg/kg of MP6-XT22, an anti-mouse TNF mAb, weekly starting at the end of first DSS treatment (day 5) (n = 10) and then euthanized at day 70. Mice were euthanized 11 weeks after the first injection of AOM, and colons were removed and immediately flushed with PBS. The isolated colonic epithelial samples from mice were prepared as previously described for assessment of protein and/or mRNA expression in the epithelia. ## Histological Scoring of Colitis and CAC Tissues from proximal and distal colons were removed for histologic assessment. For this, the tissue samples were fixed in 10% neutral-buffered formalin. Paraffin-embedded sections (5 µm) were stained with hematoxylin and eosin (H-E). The sections were analyzed without prior knowledge of the types of treatments. The histological scoring of colitis was determined according to the previously described system with minor modifications. Briefly, for extent of leukocyte infiltration in tissue layers, 0 points were assigned to normal appearance, or the presence of occasional inflammatory cells in the lamina propria; 1 point to increased numbers of inflammatory cells in the lamina propria; 2 points to confluence of inflammatory cells, extending into the submucosa; and 3 points to transmural extension of the infiltrate. For severity of leukocyte infiltration, 0 points were assigned to none to normal lymphoid aggregates; 1 point to mild infiltration; 2 points to moderate infiltration; 3 points to severe infiltration. For extent of leukocyte infiltration in colon, 0 points were assigned to none to occasional infiltration; 1 point to patchy (focal) infiltration; 2 points to intermediate infiltration; 3 points to diffuse (extensive) infiltration. For tissue damage, 0 points were assigned to no mucosal damage; 1 point to discrete lymphoepithelial lesions; 2 points to surface mucosal erosion or focal ulceration; and 3 points to extensive mucosal damage and extension into deeper structures of the bowel wall. The cumulative degree of these parameters was calculated as a total histological score ranging from 0 (no changes) to 6 (extensive cell infiltration and tissue damage). For crypt abscess, the assigned points were; 0, to no crypt abscess; 1, to the presence of crypt abscess. For CAC assessment, sections (5 µm) were cut stepwise (200 µm) through the complete block and stained with H-E. Tumor numbers were counted by trained individuals blinded to the treatment group. ## Western Blotting Western blotting was performed as previously described. Briefly, 10 to 100 µg of nucleic extracts or whole protein lysates from either the stripped epithelial samples or MOC1 cells were separated by 8–15% SDS-PAGE and each protein expressions were analyzed with following primary and secondary Abs: anti-mouse TNFR1 polyclonal Ab (pAb), anti-mouse TNFR2 pAb (R&D Systems, Minneapolis, MN), anti-phosphorylated (p)-p65 mAb at serine 536, anti-p65 pAb, anti-p-IκBα mAb at serine 32/36, anti-IκBα pAb, anti-p-MLC pAb, anti-MLC pAb (Cell Signaling Technology Inc, Beverly, MA), anti-MLCK mAb, anti-β-actin mAb (Sigma), anti- mouse IgG-HRP, anti-rabbit IgG-HRP (GE Healthcare Bio-Sciences, Piscataway, NJ) and anti-goat IgG-HRP (Santa Cruz). Signals were generated with ECL Western Blotting Detection System (GE Healthcare Bio-Sciences). ## Semi-quantitative PCR (q-PCR) Total cellular RNA was extracted from either whole colonic mucosa, isolated epithelial samples of non-tumor or tumor area removed from AOM/DSS-treated mice, or cultured MOC1 cells with RNA-Bee (Tel-Test, Inc, Friendswood, TX). Five micrograms of total RNA were subjected to reverse transcription using Superscript Reverse Transcriptase kit (Invitrogen, Carlsbad, CA). The cDNA samples were then applied for PCR with the following primer pairs: interleukin (IL)-1β, 5′-TTG ACG GAC CCA AAA GAT-3′ and 5′-GAA GCT GGA TGC TCT CAT CTG-3′; IL-6, 5′-GCT ACC AAA CTG GAT ATA ATC GGA-3′ and 5′-CCA GGT AGC TAT GGT ACT CCA GAA-3′; macrophage inflammatory protein (MIP)-2, 5′-AAA ATC ATC CAA AAG ATA CTG AAC AA-3′ and 5′-CTT TGG TTC TTC CGT TGA GG-3′; IFN-γ, 5′-CGA CTC CTT TTC CGC TTC CTG AG-3′ and 5′-TGA ACG CTA CAC ACT GCA TCT TGG-3′; TNF, 5′-GCC ATG AGG TCC ACC ACC CTG-3′ and 5′-CTA CTG GCG CTG CCA AGG CTG T-3′; glyceraldehyde-3-phosphate dehydrogenase (G3PDH), 5′-CTA CTG GCG CTG CCA AGG CAG T-3′ and 5′-GCC ATG AGG TCC ACC ACC CTG-3′, respectively. Real time PCR was performed with QuantiTect SYBER green PCR kit (Qiagen, Venio, Netherlands) using an ABI7500 real-time PCR system and 7500 system SDS software (Applied Biosystems, Foster City, CA). mRNA was shown as the relative amount normalized to that of G3PDH. ## Small Interfering (si) RNA Transfection MOC1 cells were transfected by lipofection (Lipofectamine RNAi MAX, Invitrogen, Carlsbad, CA) with the siRNA oligomers against either mouse MLCK (Mm Mylk 9988, Sigma), TNFR1 (MSS212008), TNFR2 (MSS238548) or non-targeting control (12935-112, Invitrogen) and incubated at a density of 5×10⁴ cells/ml in OPTI-MEM (Invitrogen) for 48 h. ## Transmission Electron Microscopy (TEM) TEM was performed as previously described. Briefly, colonic tissues from AOM/DSS-treated mice or MOC1 cells that were seeded on the Cell Disk® (Sumitomo, Tokyo, Japan) were fixed with 2.5% glutaraldehyde in 0.1 M phosphate buffer (PB) for 2 h. The samples were washed with 0.1 M PB, post-fixed in 1.0% OsO₄ buffered with 0.1 M PB for 2 h, dehydrated in a graded series of ethanol and embedded in Epon-812. Semi-thin sections were cut at 1 µm and stained with toluidine blue. Ultrathin sections, 90 nm, were collected on copper grids, double-stained with uranyl acetate and lead citrate, and then observed using transmission electron microscopy (H-7100, Hitachi, Hitachinaka, Japan). ## Statistical Analysis The results are expressed as the mean ± standard error of the mean (SEM). Statistical significance was determined using non-parametric Mann-Whitney *U*-test, and differences were considered to be statistically significant with p\<0.05. # Results ## Secretion of Cytokines that Support Tumor Growth is Associated with Epithelial MLCK Expression in the CAC Model We previously observed that cytokines such as IL-1β, IL-6 and MIP-2 were up- regulated in an animal model of colitis. This suggested that these factors may be essential for tumor development since each had been reported to support tumor growth in such model. However, it was unclear if these cytokines could be substantially up-regulated in the setting of CAC. In addition, it was also unclear if such elevated cytokines were mainly produced from the epithelia, since intestinal epithelial cells had been reported to be capable of expressing these molecules. Therefore we first assessed the expression levels of these cytokines in the mucosal tissues and the isolated epithelial cells in an animal model of CAC. Wild type C57BL/6 mice were administered AOM, then treated three times with DSS to induce chronic colitis and CAC. As seen in, q-PCR revealed significant up-regulation of IL-1β, IL-6 and MIP-2 in the inflamed colonic tissues when compared to the control mice. Associated with this, significant up- regulations of pro-inflammatory cytokines, such as IFN-γ and TNF, were also observed as seen in. In addition, the expressions of all of these cytokines were further up-regulated in the area of tumors compared to non-tumor area. Moreover, q-PCR also indicated that these cytokines were not much expressed by the epithelial cells. These results imply that such increase of these cytokines are presumably produced by infiltrating cells in the tissues such as granulocytes and macrophages and may be required for the progression of CAC. Given these results, we hypothesized that the increased pro-tumorigenic cytokines produced by the infiltrating cells in the inflamed tissues may be associated with the disrupted epithelial tight junctions (TJ), which results in bacterial translocation into the lamina propria. In addition, we have also observed that TNFR2 expression was specifically up-regulated in the inflamed epithelia. In this regard, it has been reported by Wang et al. that the up- regulated MLCK in human intestinal cells is required for TNF-induced collapse of epithelial TJ by TNFR2 signaling. Moreover, it has been reported by two groups that MLCK promoter activity is mediated by NF-κB activation. Therefore, we addressed the epithelial MLCK expression in the CAC model. As seen in, Western blotting revealed up-regulated MLCK expression in the epithelia from non-tumor area, and further up-regulation in that of tumor, in association with the up- regulated TNFR2 and p65. However, up-regulation of TNFR1 was not remarkable compared to that of TNFR2. These results indicate that the development of CAC may be associated with the disrupted TJ and elevated pro-tumorigenic cytokines, which are essentially induced by TNFR2 signaling and MLCK expression in the epithelia. ## MLCK Expression in the Epithelial Cells is Up-regulated by the Presence of Pro-inflammatory Cytokines in vitro Given the up-regulated TNFR2 and MLCK expressions in the epithelial cells associated with pro-tumorigenic and pro-inflammatory cytokines expressions in the mucosal tissues from chronic colitis and CAC model, it was surmised that each of these complicated phenomena may be associated with the mechanisms by which tumorigenesis is induced in the context of IBD. Especially, previous reports suggested that stimulation of human intestinal epithelial cell lines with IFN-γ results in the regulation of TNF receptors. Therefore, a murine colonic epithelial cell line (MOC1) derived from colonic epithelial cell was used to study the specific roles of TNFR2 and MLCK expressions in the epithelial cells in the context of inflammation. MOC1 cells were cultured in the presence of pro-inflammatory cytokines, rIFN-γ and rTNF. As seen in, the up-regulations of TNFR1 and 2 were induced by rIFN-γ alone, and this is consistent with previous observations with human T84 and Caco2 cells,. However, the increased TNFR2 level was more remarkable compared to that of TNFR1, and such TNFR2 expression in MOC1 cells was maximal at the concentration of 1.0 ng/ml rIFN-γ. Next, MOC1 cells were stimulated with different concentrations of rTNF together with rIFN-γ at 1.0 ng/ml. Western blotting showed that phosphorylations of p65 and IκBα were induced in these cells by the presence of rTNF in a dose-dependent manner. Moreover, the expressions of MLCK as well as TNFR2, but not much of TNFR1, were further up-regulated by the addition of rTNF in a dose-dependent fashion in collaboration with constant concentration of rIFN-γ, and these expressions were maximal at 2.5 ng/ml of rTNF. These results suggest that the pro-inflammatory cytokines, especially TNF, may be required for the induction of MLCK up-regulation via TNFR2 signaling in the colonic epithelial cells. ## MLCK Expression and Disrupted Intercellular Junctions are Induced by Up-regulated TNFR2, but not TNFR1 Given the observation of up-regulated MLCK expression in association with the up-regulated TNFR2 expression in MOC1 cells, we next pursued the specific linkage between TNFR2 and MLCK in vitro. To do so, gene expressions of either TNFR1 or TNFR2 in MOC1 cells were silenced by transfection with specific siRNAs. We first confirmed the efficacies of siRNA oligomers against either TNFR1 or TNFR2 in these cells. As expected, knocked-down expressions of either TNFR1 or TNFR2 were observed with each specific siRNA. It should be noted that the endogenous expression of MLCK in MOC1 cells was not affected by the silencing of either TNFR1 or TNFR2. We next assessed the effect of each silencing on the induction of MLCK up-regulation by pro-inflammatory cytokines in these cells. As seen in, MLCK expression was up-regulated by stimulation with rIFN-γ and rTNF, and this is consistent with the results seen in. In addition, the up-regulated MLCK expression in the presence of rIFN-γ and rTNF was not affected by the knocking-down of TNFR1 expression. However, such up-regulation of MLCK was remarkably suppressed by TNFR2 silencing. These results indicate that up- regulation of MLCK is specifically induced by TNFR2 signaling in the epithelial cells. It was still unclear whether such up-regulated MLCK expression substantially affects the intestinal epithelial cells. Therefore, we next studied the morphology of MOC1 cells in the presence of rIFN-γ and rTNF. As seen in, intercellular junctional complexes of these cells with intact TJ were observed using transmission electron microscopies (TEM) regardless of single stimulation with rIFN-γ. When stimulated with both rIFN-γ and rTNF, MOC1 cells showed collapsed intercellular junctional complexes with disappeared TJ, and the silencing of TNFR1 expression did not affect such features. However, the silencing of TNFR2 expression improved such disrupted intercellular junctions. Similar results were also observed using immunofluorescence microscopic studies (IFM) with anti-ZO1 pAb under confocal microscopies (data not shown). Taken together, these results suggest that the disrupted TJ among intestinal epithelial cells may be induced by TNFR2 signaling via MLCK up-regulation. ## The Blockade of TNF and the Suppression of MLCK Abrogate Disruption of Epithelial TJ in vitro Given the TNFR2 signaling-dependent up-regulation of MLCK, we next assessed the effect of blocking TNF signaling in the epithelial cells on the TJ, which may be disrupted by up-regulation of MLCK. MOC1 cells were cultured in the presence of rIFN-γ to induce TNFR2 expression followed by addition of rTNF. As seen in, the up-regulated MLCK expression was observed in these cells in association with remarkable TNFR2 up-regulation. In addition, such up-regulations of TNFR2 and MLCK were abrogated by the presence of MP6-XT22, a mAb against to TNF, in a dose-dependent fashion. It should be noted that slightly up-regulated TNFR1 expression by rTNF stimulation was also suppressed in the presence of MP6-XT22. Given the abrogation of MLCK up-regulation by MP6-XT22, we next studied the intercellular junctions among MOC1 cells under this condition. IFM (data not shown) and TEM studies showed restoration of TJ in the presence of MP6-XT22. These results suggest that TNF signaling in the colonic epithelia may induce disruption of TJ via up-regulated TNFR2 and MLCK expressions. Moreover, these results also suggest that neutralization of TNF may abrogate such TJ disruption. To confirm TJ disruption by MLCK up-regulation, rIFN-γ/rTNF-stimulated MOC1 cells were also incubated in the presence of ML-7, an MLCK inhibitor. As expected, the disturbed TJ, which was induced by the presence of rIFN-γ and rTNF, was restored by the inhibition of MLCK function. This result was originally considered to be the cause of suppressed MLCK function by ML-7. However, Western blotting showed that the up-regulated TNFR2 and MLCK expressions were interestingly suppressed by the addition of ML-7 in a dose- dependent fashion. It should be noted that the up-regulated TNFR1 expression induced by rIFN-γ and rTNF was not affected by the addition of ML-7. Given this result, MOC1 cells were also transfected with MLCK-specific siRNA. As seen in, up-regulation of TNFR2 induced by rIFN-γ and rTNF was suppressed by silencing MLCK expression. Taken together, these results indicate that the disrupted TJ among intestinal epithelial cells induced by TNFR2 signaling may be restored by the suppression of MLCK function. ## Suppression of MLCK Prevents the Development of CAC in Association with Restored TJ and Diminished Pro-tumorigenic Cytokine Productions in vivo Given the disrupted TJ via TNFR2 signaling and MLCK up-regulation in vitro, we finally assessed the effect of MLCK suppression in the setting of CAC. Mice induced with CAC by AOM treatment and three cycles of DSS were injected with either vehicle control, MP6-XT22 or ML-7, and later euthanized to assess colonic inflammation and CAC development. As seen in, the blockade of TNF in vivo, by weekly injection with MP6-XT22 throughout the entire experimental protocol, resulted in the reduction of CAC development. This result is consistent with our previous study. Moreover, suppression of MLCK, by the injection with ML-7 during the entire protocol resulted in even more reduction of CAC development (‘ML-7 entire’). In addition, administration of ML-7 at the recovering phase after the last DSS administration surprisingly revealed almost same degree of suppression on CAC development (‘ML-7 final’). These results were relatively associated with the histological assessment of colitis in these groups, although the significances of colitis suppression with either administration of MP6-XT22 or ML-7 were not remarkable compared to that of CAC suppression. Interestingly, q-PCR revealed that the expression levels of IL-1β, IL-6 and MIP-2 as well as IFN-γ and TNF (data not shown) in the colonic tissue, especially in the tumor area, were suppressed by the treatment with ML-7 as well as with MP6-XT22. In addition, the MLCK expression and MLC phosphorylation in the isolated epithelial cells from the non-tumor and tumor tissues were remarkably suppressed by the treatment with MP6-XT22 or ML-7. It should be noted that the epithelia isolated from tumor area in each groups showed relatively higher expressions of TNFR2, p-p65, p-IκBα, MLCK and p-MLC when compared to that in non-tumor epithelia. These results are consistent with the observations of the intercellular junctional complexes among cells under TEM, since chronic inflammation in the control group induced disruption of epithelial TJ, and treatments via both TNF blockade and MLCK suppression resulted in the restoration of the disrupted TJ in the non-tumor and tumor areas and in association with reduced CAC. These results indicate that the suppression of MLCK function may contribute to reduce CAC development via restoration of the disrupted TJ. # Discussion The etiology of IBD is considered to be associated with both epithelial permeability and dysregulated immune responses to luminal contents which include antigens derived from commensal bacteria in the gut. Regarding immune dysregulation, in patients with Crohn’s disease for example, excessive amount of Th1 and Th17 cytokines, such as IFN-γ and IL-17 respectively, are secreted predominantly by the infiltrating CD4⁺ effecter T cells in the intestinal lamina propria. TNF, a pro-inflammatory cytokine produced not only by such dysregulated effecter T cells but also by macrophages and granulocytes infiltrating the inflamed intestinal tissues, is involved in the dysregulated adoptive immune responses in CD and possibly UC as well. Current therapeutic approaches in neutralizing TNF using either chimeric or humanized Abs have provided some effective therapies in the management of CD, and to some extent UC as well. Our efforts in the study of pathogenesis of IBD have shown that TNF is expressed mainly by F4/80⁺ macrophages rather than CD4⁺ T cells in the DSS colitis model. It is known that myeloid cell-derived macrophages and granulocytes are also capable of expressing other cytokines such as IL-1β, IL-6 and MIP-2, a homolog of IL-8 in humans. These cytokines can also induce pro-tumorigenic activities such as angiogenesis and tumor proliferation , but at the same time, these cytokine may also be expressed by intestinal epithelial cells. In fact, we have observed in previous and current studies that the expression levels of these cytokines in colonic tissues are up-regulated in DSS colitis and further up- regulated in tumors. CAC development was suppressed by the treatment with either blocking anti-TNF mAb or MLCK inhibitor in association with reduced cytokine productions mentioned above. However, such cytokine expressions from the isolated colonic epithelial cells were not significant when compared to that of the entire colonic mucosa. Therefore, pro-tumorigenic cytokines most likely derived from the lamina propria where most macrophages reside. Others had also shown that NF-κB activation in the myeloid cells is critical for the induction of CAC, and that NF-κB activation in myeloid cells can be induced by TNF. DSS colitis is a commonly utilized murine model of IBD where a single administration of DSS can lead to acute colitis. In addition, three cycles of DSS administration can lead to chronic inflammation in the colon. Furthermore, AOM administration preceding three cycles of DSS treatment can lead to epithelial carcinogenesis and is a model of CAC. By using this CAC model, Greten and colleagues showed that epithelial NF-κB activation is critical for the development of CAC. However, the roles and mechanisms of epithelial NF-κB activation leading to CAC development had not previously been rigorously studied. In our previous and current studies, we demonstrated epithelial NF-κB activation in association with up-regulated TNF production by macrophages infiltrating the colonic lamina propria in the same CAC model. In addition, we showed that epithelial NF-κB activation was abrogated by neutralizing TNF with a specific mAb leading to the suppression of CAC. These results suggest that the NF-κB pathway in the epithelial cells is mainly activated by the up-regulated TNF production. Previous studies by others had suggested that insufficient immunological homeostasis due to TNF production in IBD was caused by sequential activation of macrophages and neutrophils dominantly expressing TNFR1 (TNFRSF1a/p55) in the inflamed tissue. However, we and other researchers had observed that intestinal epithelial cells express both TNFR1 and TNFR2 (TNFRSF1b/p75). There is spontaneous TNFR2 up-regulation in the epithelial cells in the setting of in vitro and in vivo inflammatory conditions. Increased expression of TNFR2 in the epithelia was associated with CAC development. These results suggest that TNF-induced epithelial NF-κB activation is associated with CAC development. Given the results of up-regulated TNFR2 expression and NF-κB activation in the inflamed epithelial cells, we hypothesized that this may somehow be associated with reducing the recruitment of the death domains, such as FADD and TRADD, which are involved in the caspase-dependent pathway of TNFR1 signaling, so that the epithelial cells may escape apoptosis. However, the role of the epithelial NF-κB activation via TNFR2 signaling associated with CAC induction was not known previously. Therefore, we investigated whether increased expression of TNFR2 in this model leads to CAC. Using the same CAC model, TNFR2 deficient mice were used. However, CAC development was not possible due to high mortality with DSS treatment which had also been observed previously. Therefore, we have utilized MOC1, which is a murine epithelial cell line derived from non-cancerous colonic epithelia transformed with SV40 large T antigen. This cell line likely possess more physiological functions compared to other cell lines derived from murine colon cancer such as CT26 cells. Furthermore, MOC1 cells can form confluent monolayers with epithelial TJ. Such characters of MOC1 cells allowed us to assess the collapsed TJ associated with MLCK up-regulation, because it was difficult to observe such morphologies with CT26 cells, which does not show either MLCK up-regulation or fine monolayers required for TJ observation. It had been shown that MLCK in human intestinal cells leads to TNF-induced collapse of epithelial TJ via TNFR2 signaling and that MLCK promoter activity is mediated by NF-κB. Therefore, in our study, MOC1 cells were transfected with the TNRF2 siRNA to determine MLCK activities in the presence of rTNF. This resulted in the abrogation of TNF-induced TJ disruption associated with suppressed MLCK up-regulation , and these results are consistent with previous reports. Moreover, we also observed that the TJ remained intact with anti-TNF treatment as well as inhibiting MLCK activity. In our current study, we also observed that TNFR1 expression in MOC1 cells was slightly up-regulated by addition of rTNF, and such up-regulated TNFR1 was suppressed by the presence of MP6-XT22. These results imply that presence of TNF may induce both up-regulations of TNFR1 and TNFR2 in inflamed epithelial cells as reported by Mizoguchi et al, suggesting that NF-κB activation in the epithelial cells we observed may be caused by both signaling. However, the degree of TNFR2 up-regulation was remarkably more than that of TNFR1. Moreover, TNFR1 silencing failed to restore irritated intercellular junctions in association with not down-regulated MLCK expression. These results indicate that TNFR2 signaling is the key factor for MLCK-mediated epithelial permeability, as consistent with the observations by Wang et al. These findings may also imply an important mechanism by which NF-κB pathway via TNFR2 signaling is regulated in the colonic epithelial cells in the setting of carcinogenesis. Thus, one potential mechanism by which TNFR2 expression in intestinal epithelial cells is associated with CAC development may be that the epithelial barrier dysfunction induced by MLCK activation via TNFR2 signaling leads to translocation of luminal bacteria into the lamina propria. Such epithelial permeability results in the stimulation of macrophages and granulocytes in mucosal tissue followed by cytokine production such as IL-1β, IL-6 and MIP-2 that are required for epithelial proliferation and tumor development. On the other hand, it should be noted that the impact of such pro- tumorigenic cytokines production, which is caused by disrupted TJ, on the severity of DSS colitis is controversial. Some recent studies have suggested that suppression of epithelial permeability by administration of MLCK inhibitors remarkably ameliorate DSS colitis. Our current results also suggest that colitis was suppressed by ML-7 treatment, however, the impact of such treatment on the suppression of colitis was not remarkably effective compared to that on the suppression of CAC development. One reason for the different observations may be that the severity of colitis in our studies was relatively mild compared to the others, and thus, we did not observe distinct abrogation of colitis by MLCK inhibition. In addition, another recent report showed that MLCK deficient mice revealed worse DSS colitis compared to wild type. These discrepant observations by others and ours suggest that another reason may be contributed by environmental factors such as the different microflora at the animal facilities that may induce different cytokine profiles. In our current study, we observed that IFN-γ and TNF-induced up-regulation of TNFR2, and possibly MLCK, in the epithelial cells were specifically suppressed by ML-7 treatment. Up-regulation of TNFR2, but not TNFR1, was also suppressed by silencing MLCK expression. These results suggest that TJ disruption, induced by MLCK expression and MLC phosphorylation in the epithelial cells, may result in further up-regulation of TNFR2 expression in vitro. This in turn causes a vicious circle on the epithelial permeability, and such condition may promote tumor growth in the context of inflammation. Such impact of the relationship between TNFR2 and MLCK on CAC development is also suggested by the fact that the administration of ML-7, even at the final phase of CAC induction, resulted in the remarkable reduction of tumor in mice. In fact, ML-7 injection after final cycle of DSS treatment seemed to result in almost the same effect as the injection during the entire protocol in terms of reduction of tumor number, shown in. It is suggested that colonic epithelial cells have undergone enough inflammation to initiate carcinogenesis until mice were injected with ML-7 at the final stage. One potential interpretation for this observation would be that undetectable microscopic tumors or aberrant crypt foci, which had not developed yet in the colonic epithelia, may still exist in mice treated with ML-7 only at the final phase. However, it is still suggested that the inhibition of MLCK function is absolutely effective at least for the suppression of tumor progression. In addition, we have examined whether such MLCK expression also induce up-regulation of endogenous IFN-γ and/or TNF expressions in MOC1 cells. However, neither expressions were affected regardless of MLCK expression (data not shown). Therefore, another mechanism may be involved in TNFR2 up-regulation in the irritated epithelial cells. Furthermore, our current study suggests that MLCK expression, which may be induced by TNFR2 signaling in the intestinal epithelial cells, can be a therapeutic target for the maintenance of continuous inflammation and prevention of CAC development in the setting of IBD. Here, we demonstrate TNFR2-mediated regulation of CAC development, which is due to tumorigenic cytokines induced by MLCK-induced disruption of the TJ. This is an important mechanism that helps us understand the regulation of mucosal immune response as well as IBD- associated epithelial tumorigenesis. We are grateful to N. Tsuge, Y. Kato, K. Mochida and Dr. Y. Saito for the technical support, Dr. H. Yagita for providing MP6-XT22 and Drs. M. Totsuka and T. Iwamoto for providing the MOC1 cells. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: T. Nagaishi MW. Performed the experiments: M. Suzuki T. Nagaishi MY MO YS SI. Analyzed the data: M. Suzuki T. Nagaishi MY MO. Contributed reagents/materials/analysis tools: TW MT SO RO M. Shimonaka HY T. Nakamura. Wrote the paper: M. Suzuki T. Nagaishi.

# Introduction Oilseed rape (*Brassica napus*) production has increased in many European Union (EU) countries since 2005 when the EU set a target for 20% of its energy to be from renewable sources by 2020. In the UK, oilseed rape production reached 652,000 hectares in 2015, increasing annually by an average 2% over the past ten years. UK oilseed rape is grown for vegetable oil, animal feed and biodiesel. It serves as a break crop with wheat production and accounts for about 80% of the UK combinable break crops. Oilseed rape has many arthropod pest and disease problems, prompting widespread use of chemical insecticides. Pesticide usage survey data collected by the UK’s Fera Science Ltd. (; methods and figures see) indicate that, between 1974 and 1988, pest control in oilseed rape in Great Britain was dominated by the use of organophosphate sprays and organochlorines as both foliar sprays and seed treatments. After 1990, pyrethroids quickly replaced these chemicals as foliar sprays, because of their higher efficacy in pest control and lower toxicity to humans, becoming the major insecticide group used from 1995 onwards. On average, pyrethroids represented 80% of total insecticidal weights, and 95% of total area treated for all foliar sprays used in oilseed rape in Great Britain for the past 25 years. Since the ban on organochlorines in 1999 (due to their high toxicity to users, and persistence and bioaccumulation in the environment), neonicotinoids have become the main seed treatments for oilseed rape in Great Britain, with 1,930 kg weight increase every two years. Three neonicotinoid active ingredients, clothianidin, thiacloprid and thiamethoxam, have been restricted in Europe since the end of 2013, and the 2014/15 season was the first without these seed treatments being available. The neonicotinoid restriction primarily resulted from the perceived negative lethal and sublethal impacts on both managed and wild pollinators. Since the start of the restriction, the evidence base has been strengthened around the effects of neonicotinoids on crop production, human health, pollinators and the wider environment. However, major gaps remain (e.g., concentration and toxicity of neonicotinoid metabolites in the environment, see). Strongly contrasting views about the risks of using this chemical group in agriculture exist, and little consensus has been reached, with an ongoing argument in UK between the National Farmers Union (NFU) and government regarding the lifting of the restriction. The neonicotinoid debate has resulted in uncertainties for the future of oilseed rape production and management for UK farmers, in terms of cropping areas, insecticide choices and impact on profits. Two new chemical groups have recently been adopted by UK growers (azomethines and oxadiazines, from around 2012), which could potentially act as alternatives to neonicotinoids. However, as with all currently available chemical groups, there is little publicly available information concerning their relative efficacy in controlling key pests, or their side effects on non-target species and the wider environment. Thus it is difficult to assess the overall impact of using a specific insecticide chemical on oilseed rape. In this study, based on experts’ opinions, we provide an insight into these matters by comparing the relative efficacies and side effects of insecticide chemical groups used in UK oilseed rape production. Experts’ opinions on the neonicotinoid restriction are also analysed. Because the survey finished in June 2015, the perceived inputs and yields are compared between 2013/14 and 2014/15 season. For future sustainable oilseed rape production in the UK, the potential alternative crop management options are thus discussed. # Materials and Methods ## Survey structure An online survey by Qualtrics was conducted to collect UK experts’ opinions. The online survey has been approved by the ethics committee of the School of Agriculture, Policy and Development, University of Reading. As stated in the ethical approval (B165) and cover letter, the data can only be shared in summary forms, where no individual will be identified, and raw data will be destroyed in March 2018. The survey had 28 questions (see. Survey questionnaire) divided into four parts: (i) general section (6 questions including occupations, experience of giving farmer advice, and knowledge of crops), (ii) wheat section (7 questions, including yields, main pests and damage, insecticide efficacy, and importance of arthropod natural enemies), (iii) oilseed rape section (7 questions, as in the wheat section) and (iv) neonicotinoid section (8 questions, including opinions and reasons for the restriction, expected yields and inputs, and alternative pest management strategies). This paper refers only to the oilseed rape and neonicotinoid sections. Respondents were also asked to give the regions they refer to when providing the information. To capture the degree of certainty, respondents were asked to give certainty scores on a 1–5 scale (5 being most certain) following several of the questions. ## Survey distribution In January 2015, a pilot study using 15 experts, mainly from universities, was conducted. Because only minor changes to the wording of a small number of questions were made based on their feedback, these responses were included in the analyses. The main nationwide survey was carried out between March and June 2015. Potential respondents were contacted by searching through directories of universities, research institutes, NGOs (non-government organizations), government sectors, consulting firms, and agri-chemical companies. Universities were filtered via the Guardian league table 2015 by subjects of ‘Agriculture, forestry & food’ and ‘Biosciences’, and then relevant researchers were selected by scanning through the related department staff pages. Independent consultants were contacted through the Association of Independent Crop Consultants (AICC). Most experts were contacted by email and some through general enquiry web pages. Snowball sampling methods (where a respondent may pass the survey to other related experts) were also utilized to gain more respondents. A reminder with a link to the survey was sent to individuals two weeks after the initial request, and one more reminder two weeks later if no reply was received. ## Statistical analyses ### Main arthropod pests and related damages in oilseed rape Respondents were asked to rank the relative importance of the three most widespread arthropod pests (one being most important of the three) in oilseed rape in the past five years (2009/10 to 2014/15). These scales were used as weights (i.e., weights for the 1^st, 2^nd and 3^rd most important pest were 0.5, 0.33 and 0.17 respectively, with a collective weight of 1) to produce a weighted average response for each pest. Answers were further categorised into different regions and the weighted average for each pest for each region was estimated. See for the regional distribution. Respondents were also asked to estimate, without insecticides, the direct feeding damage from each selected arthropod pest and the damage that Turnip yellows virus (TuYV; caused by *Myzus persicae*, peach—potato aphid) would have caused in the past five years, in terms of percentage yield loss (%), followed by the certainty scale (1–5, 5 being most certain). Average damage levels, standard deviations and certainty levels of the three most important pests were estimated. ### Efficacy of insecticides and arthropod natural enemies Information about the perceived efficacy of both chemical and natural pest control methods was requested in the survey. The perceived efficacy of each available insecticide chemical group in the 2013/14 season was requested for each of the main pests selected from the previous question, following a 0–6 efficacy scale as below: Pairwise efficacy comparisons among the available chemical groups were conducted by the Skillings–Mack and related post hoc tests, using the software R 3.2.5. Responses were omitted if a respondent answered for only one chemical group for a specific pest. Due to the asymmetrical distribution and the non-parametric analyses for these ordinal data, medians and interquartile ranges were used to summarize the efficacy level. The same efficacy and certainty scales were used to estimate the importance of arthropod natural enemies on pest control (without insecticides). Qualitative opinions on its importance for oilseed rape production were gathered by selecting from the following options: ‘Strongly agree’, \`Agree’, \`Neutral’,\`Disagree’, \`Strongly disagree’ and \`Not sure’. ### Side effects of insecticides on human health, environment and non-target arthropods Perceived negative influences on users’ health, natural enemies, pollinators, water and soil were compared between the two application methods for UK oilseed rape(seed treatments and foliar sprays), using a 0–5 scale (0 being no influence, and 5 being greatest influence). Sign tests were used to compare the perceived median differences between seed treatments and foliar sprays on these aspects. Responses were omitted if the respondent only answered for one of the two treatment types. Due to the asymmetrical distribution and the non-parametric analyses for these ordinal data, medians and interquartile ranges were used to summarize the side effect level. Potential risks of common foliar active ingredients to the non-target arthropods for UK oilseed rape were also analysed by the in-field Hazard Quotient approach (; see for details). Hazard Quotient is the required assessment for pesticide registration in the European Union. The method combines the laboratory acute test of two sensitive species (cereal aphid parasitoid *Aphidius rhopalosiphi* and predatory mite *Typhlodromus pyri*) and field application rate of the related foliar active ingredient to produce a Hazard Quotient. If the quotient of either species exceeds two, the threshold level, it indicates a potential hazard of the active ingredient towards non-target arthropods. Due to the limited publicly available test results, only one representative active ingredient (the one that was used most often in UK oilseed rape and had test results available) was chosen from each available chemical group. For the same reason, and also that *Typhlodromus pyri* would normally be applicable in orchard fruit crops, only *Aphidius rhopalosiphi* was used as the test species. ### Opinions on neonicotinoid seed treatment restrictions A five-point Likert scale was used to measure expert opinions on the current neonicotinoid restrictions (‘Strongly favour’ to ‘Strongly oppose’ including a ‘Not sure’ option). Different possible reasons for holding favourable and unfavourable opinions were displayed for respondents to select. Fisher’s exact test was used (with Monte Carlo simulated p value, 100,000 replicates) to explore whether or not experiences in farming services and the nature of the employing organization had an impact on respondents’ opinions. # Results ## Response rates and information about respondents In total, 455 online surveys were sent out, and 146 responses were received (43% effective response rate accounting for incorrect emails and declines), with 101 completed (70% of total responses). 90 people participated in the oilseed rape and neonicotinoid sections. The following analyses focused on these two sections. From 2010 to 2014, 43% of respondents (39 out of 90) had worked for independent consultants/ consulting organizations (Fig E.1). University staff accounted for 27%, government 17% and agri-chemical companies 16%. 65% of respondents had provided advice services to farmers. ## Main arthropod pests and related damage in oilseed rape (2009/10-2013/14) Based on a weighted average of expert ranking, the three main arthropod pests in oilseed rape in the past five years were perceived to be cabbage stem flea beetle (*Psylliodes chrysocephalus*) (weighted average response = 28), pollen beetle (*Meligethes aeneus*) (18), and peach—potato aphid (*Myzus persicae*) (8). Different regions varied slightly (e.g., for the North West region, the second most important pest was cabbage seed weevil, and the equal third were aphids and brassica pod midge) (Table E.1). Due to the limited number of respondents, and the fact that some respondents represented several regions (up to four), represents a broad visualization. The following analyses were conducted without categorising into regions. Without insecticide applications, the perception was that cabbage stem flea beetle would have caused the greatest direct damage, with a mean yield loss of 24% (certainty level 3). However, expert opinion on this also had the largest variation, with a standard deviation of 26%. Second was pollen beetle damage, with a mean value of 19%, and standard deviation of 20% (certainty level 3). Perceived peach—potato aphid direct damage averaged 13%, with a standard deviation of 10% (certainty level 3). TuYV loss without insecticides was perceived to be 15%, with a standard deviation 11%, and a certainty level of 2. ## Efficacy of insecticides in oilseed rape (2013/14) In terms of efficacy of the insecticide groups for oilseed rape protection, because of the potential complex annual variation in efficacy, only the available chemical groups used for the 2013/14 season were studied. It should be noted that some chemical groups are not applied against some pests due to certain phenological and pharmacological factors. For example, neonicotinoid seed treatments are not normally used against pollen beetles, because they are primarily applied to autumn sown crops, whereas pollen beetles pose a threat in spring. Neonicotinoid sprays and oxadiazines are mainly used for this pest. Carbamates (mostly pirimicarb) are used mainly for aphid control, and azomethines are applied mainly against aphids and pollen beetles. This information was reflected in by the perceived relatively low efficacy levels for the non-targeted pests. Based on the pairwise comparisons among different chemical groups, neonicotinoids and the two newly introduced chemical groups (oxadiazines and azomethines) were perceived to have higher efficacy than carbamates and pyrethroids for peach–potato aphid (81–90% and 1–20% respectively) and pollen beetle control (51–80% and 21–50% respectively). Neonicotinoid seed treatments were perceived to have more efficacy than pyrethroids against cabbage stem flea beetle (81–90% and 21–50% respectively). ## Side effects of insecticides in oilseed rape Comparing the two main insecticide application methods for UK oilseed rape, foliar sprays were thought to have significantly more negative impacts than seed treatments on users’ health, natural enemies, pollinators, water and soil. The median influence levels (from 0, no influence, to 5, greatest influence) across all categories were perceived to be 3 for sprays and 1 for seed treatments. Focusing on the Hazard Quotients for foliar sprays, different active ingredients have large differences in terms of the potential risks to non-target arthropods. Dimethoate has much higher potential hazard levels, with lambda-cyhalothrin and thiacloprid following. In comparison, pymetrozine, indoxacarb and pirimicarb pose lower potential risks towards non-target arthropods. ## Uncertain future–neonicotinoid debates From the information gathered, 72% (65 of 90) participants either opposed or strongly opposed the neonicotinoid restrictions in the UK while only 10% (9 of 90) supported or strongly supported this policy. 13 respondents took a neutral stance, and three respondents were not sure about this proposition (Fig E.2). In terms of organizations (NGOs and food industries were omitted, due to low respondent numbers), an average of 51% respondents from the government, universities, and private research institutes (Group A) opposed the neonicotinoid restriction, compared with 90% of respondents from the agri- chemical companies, commercial/independent consultants, and growers (Group B). This difference was driven less by respondents who supported the restriction (an average of 20% from Group A compared with 7% from Group B), but more by those who held ‘Neutral’ and ‘Not sure’ opinions (30% compared with 3%). To compare this more quantitatively, we regrouped the six options in two ways: one was *oppose* (including ‘Oppose’ and ‘Strongly oppose’) versus *favour* (including ‘Favour’ and ‘Strongly favour’); the other was having an opinion (*oppose*/ *favour*) versus ‘Neutral’/ ‘Not sure’. Fisher’s exact test was used to test differences among organizations regarding their opinions on this issue. Results showed that there was no clear difference among organizations for whether to favour or oppose neonicotinoid restriction, but more people from Group A than Group B had a neutral/ not sure proposition (Table E.2). Fisher’s exact test also showed a clear division between Group A and B regarding whether a respondent had provided advice services to farmers in the past five years (Table E.3). An average of 41% from Group A and 85% from Group B had done so. Although only 10% (9) of respondents favoured the neonicotinoid restriction in the UK, all of them expressed concerns about the negative impacts neonicotinoids could have on pollinators and the wider environment. Four respondents believed that farmers can adjust to the management accordingly, and two suggested that the arthropod pest problems were not severe. Of the 65 respondents that opposed the restriction, the two most widespread reasons were also around environment and pollinators (95% and 89% respectively). 71% of respondents worried that other products are not as efficient as neonicotinoids, and 60% thought that oilseed rape production will be greatly reduced. ## 2014/15 VS 2013/14 oilseed rape growing season For the 2014/15 season (compared with 2013/14), 64% of respondents (48 out of 75) felt that more time had been spent by agronomists inspecting crops since the restrictions. 70% of respondents (54 out of 78) indicated that oilseed rape farmers spent more money on insecticide products during the 2014/15 season, but 37% (29 out of 78) thought that just a bit more had been spent. Only one respondent thought that farmers spent a bit less than 2013/14, while 15 were unsure. Most of the 76 respondents expected that the winter oilseed rape harvest in 2014/15 season would be between 0–1 t/ha less or about the same as the previous year (36% and 24% respectively). 22% of respondents were not sure about the answer, which was to be expected since yield changes are influenced by many factors in addition to pesticide use. Similar patterns occurred regarding the spring oilseed rape harvest. As the survey finished in June, yields for the year 2014/15 were estimated by the respondents directly. ## Alternative pest control options In a hypothetical situation where neonicotinoids would be permanently banned in the UK, respondents’ suggestions towards alternative pest management strategies differed: 76% respondents (66 out of 87) would choose to use new insecticides if available. Other options such as ‘Grow oilseed rape less often’, ‘Use new oilseed rape varieties’, and ‘Grow a smaller area of oilseed rape’ were similar in terms of support (46%, 45% and 43% respectively). The use of currently available insecticides was mentioned by 30% of respondents. For the ‘Others’ option, 65% (13 out of 20) respondents mentioned the use of biocontrol/ IPM (integrated pest management) as an alternative approach. Focusing on natural pest control, 67% (56 out of 83) of respondents agreed that natural enemies are important for oilseed rape production. Without insecticide treatments, 57% of respondents thought that natural enemies could exert 1–20% control on oilseed rape pests, and 32% suggested 21–50% control. # Discussion ## Main arthropod pests and related damage in oilseed rape (2009/10-2013/14) The most important arthropod pests in UK oilseed rape based on perceptions from the survey (cabbage stem flea beetle, pollen beetle, and peach—potato aphid) were also indicated by on a European scale, except for the peach—potato aphid, which was considered a minor pest in Europe. However, peach—potato aphid is an important pest in the UK, especially as a vector of TuYV ( showed an average 15% yield loss by this virus in untreated crops). The large variation in levels of perceived pest damage suggests that pest damage is likely to vary greatly under different contexts. suggested a 1% untreated yield loss for cabbage stem flea beetle, 0.5% for pollen beetle, and 3% for aphids carrying TuYV. In 2010, found varied pollen beetle damage on different study sites in England (0–6% yield loss). The low certainty levels provided by experts (an average of 3 out of 5) suggest that uncertainties and research gaps remain in this area. ## Efficacy of insecticides in oilseed rape (2013/14) Due to the commercial confidentiality and expensive trials, it is difficult to obtain data on insecticide efficacy which can differ among active ingredients within one chemical group. Efficacy levels are also difficult to evaluate and compare because many factors can affect them both temporally and spatially, including insecticide resistance (Table E.4), application methods, plant growth stages, etc.. The perceived high efficacy of neonicotinoids against insect pests and related diseases has been demonstrated in several studies (: cabbage stem flea beetle;: pollen beetle;: peach—potato aphid), although others have shown limited efficacy (: peach—potato aphid). For the relatively new groups, oxadiazines and azomethines, the perceived relatively high efficacy against pollen beetle was illustrated in a recent German study. However, limited efficacy by these two groups was also found. It should be noted that without careful management (e.g., rotating insecticides with different available modes of action,), resistance to the above insecticides could occur for pests in UK oilseed rape, thus reducing the efficacy. Neonicotinoid resistance has already been detected in the peach—potato aphid in Southern Europe. Pyrethroids and carbamates had the lowest expected efficacy against the peach—potato aphid (1–20%), which could be partly due to the insecticidal resistance occurred for this pest in UK (Table E.4). Although in 2010, a HGCA study found that pyrethroids were efficient against pollen beetle (about 90% control,), experts in our study only expressed a median 21–50% control, possibly due to growing resistance since its appearance in 2006 (Table E.4). The perceived low confidence in cabbage stem flea beetle control is worrying. Apart from neonicotinoid seed treatments, the other available chemical pyrethroids were perceived to exert lower than 50% control, which could partly due to the resistance occurred in this pest (Table E.4). Across all experts, the perceived efficacy rarely exceeded 90% from all available chemical groups used in UK oilseed rape in 2013/14. This may reflect the lack of confidence in assigning efficacy for insecticides used in oilseed rape, but also the fact that many insecticides, even when newly introduced, cannot provide 90% pest control. The perceived low and/ or uncertain efficacy in the available chemical groups and the uncertainty in pest control after the neonicotinoid restriction indicate an urgent need for robust, accountable and updated information about efficiency of insecticides for pest control for this crop in the local fields. ## Side effects of seed treatments versus sprays Experts generally perceived seed treatments to be less harmful than sprays. This is to be expected as: (i) compared with sprays, seed treatments have less direct contact with operators and non-target species; (ii) less surface runoff; and (iii) reduced concentration in the environment. However, many counter-arguments have arisen in the past few years against this application method, especially for the neonicotinoid group. Traces of neonicotinoid residues have been detected in humans, pollinators and beehives, soil and water. Negative impacts have also been found on human health, pollinators, natural enemies, earthworms and aquatic invertebrates. Possible reasons behind these side effects are the systemic characteristics of neonicotinoids (residues in treated plant tissues), low soil absorption (high leaching potentials), and great toxicity to invertebrates. On the other hand, sprays can also cause side effects for human health and the wider environment. For example, although relatively few human poisonings from pyrethroids have been reported despite their extensive use worldwide, sub-lethal reactions have been found, including paresthesia and nausea. Negative impacts have also been detected for pollinators, natural enemies, soil and water systems. Since the survey finished in June 2015, more knowledge has been accumulated on the side effects of insecticides toward non-target species and the wider environment. Take the impacts of neonicotinoid seed treatments on pollinators as an example: since the research gaps were identified by (a literature review up to June 2015), many studies have investigated other active ingredients besides imidacloprid, pollinator species besides honey bees (*Apis*), and/or impacts on the colony besides individual development. Although more work needs to be done to improve the evidence base on this issue, respondents’ opinions could be changed by the new evidence. However, with the currently available evidence, it is still difficult to compare the overall side effects of the two insecticidal application methods, especially between neonicotinoids and pyrethroids, since little research has done so (but see on earthworms). ## Hazard Quotient ratios from sprays By using the in-field Hazard Quotient method, a temporal comparison of hazard levels on non-target arthropods among available foliar active ingredients for UK oilseed rape was presented. However, limitations exist in interpreting the results because this method is based on the laboratory acute toxicity tests, one being that these tests are difficult to account for the influence from the environment. The high toxicity of dimethoate to human health and the wider environment has been widely recognized. Although its use in UK oilseed rape was stopped a decade ago, it is still approved for use in wheat. The high threat potential of lambda- cyhalothrin towards non-target arthropods has also been reflected in previous studies. Although thiacloprid has been shown to have negative effects on arthropod natural enemies, it has been found to have limited side effects on bees by some studies. According to, pymetrozine has less of an effect than pirimicarb on aphid natural enemies. Indoxacarb has been shown to be less toxic than lambda-cyhalothrin to arthropod predators and parasitoids, but its potential hazard to honey bees could be high. Relatively low side effects from pirimicarb against natural enemies have also been recorded. ## Perceptions on neonicotinoid restriction The tendency for respondents from the universities, private research institutes and government to choose the ‘Neutral’ opinion on the neonicotinoid restriction debate is worth discussing. These respondents may have had an actual mid-point opinion on this issue, that they neither opposed nor favoured the restriction. It may have been because there was a lack of interest in this topic, or they considered similar overall costs and benefits for either side. Uncertainty could also have been important to this group: they may have had more recent information about the effects of neonicotinoids on crop yields, pollinators and the wider environment, but because of the complexity of this issue and gaps in the current evidence base, they could not estimate the net costs and benefits of neonicotinoids. This is also reflected in that the reasons most frequently chosen for both the *oppose* and *favour* groups were all around pollinators and the wider environment (Figs). On the other hand, they may have been less well informed about the field situations or other related risks for farmers than consultants, who have provided more advice services to farmers (Table E.3). It is also possible that some respondents chose the neutral option to avoid the cognitive costs of selecting the most appropriate opinion, even though they may lean towards one side, while others chose this midpoint as a ‘hidden don’t know’. Nevertheless, the proportion of the last reason is estimated to be small in this study, since a ‘Not sure’ was included as an option to avoid the ambiguity, and online surveys potentially have less of this issue since people are more free to express their true opinions. ## Expected time, money and yield comparisons between 2013/14 and 2014/15 To our knowledge, little publicly available information has been made available on the time and money spent before and after the neonicotinoid restrictions, and limited research has been done regarding the impact of neonicotinoids on crop yield, but see. According to the average response on this issue, compared with 2013/14 season, agronomists spent more time on inspecting oilseed rape crops for pest damage, farmers spent more money for insecticide purchases, and crop yield would be reduced in 2014/15. Changes in the time spent on pest control activities should be taken into account when considering the pros and cons of neonicotinoid seed treatments, because it represents a hidden benefit if agronomists/ farmers spend less time on pest control, but more time on other activities. As for insecticide purchases, indicates a negative relationships between neonicotinoid seed treatments and foliar sprays for UK oilseed rape, which could potentially lead to more insecticide costs to farmers after the restriction. Although most respondents expected lower oilseed rape yield in 2014/2015 than 2013/2014, this does not reflect the actual average yields (3.9 and 3.6 t/ ha respectively,). It is difficult to assess the impact of neonicotinoid restriction on the yield change based on one year data, since the increased yield may well be due to nicer weather and lower pest pressure during the year. ## Alternative methods of pest control in oilseed rape The results from this study suggest a clear preference towards using new insecticides if these become available. This partly reflected a lack of confidence in the old chemical groups, but also an acknowledgment of the importance of insecticides for crop protection. Developing new insecticides (especially with new modes of action) would help current insecticide resistance problems in oilseed rape. However, insecticide discovery has been a challenge, with the shrinking number of agri-chemical companies involved in the research, and the expensive and time-consuming development process. Respondents also advised growers to use new oilseed rape varieties if available. Indeed, crop breeding in the UK has contributed to yield protection by improving crop resistance to pests and diseases. A new oilseed rape variety ‘Amalie’ has been recommended for use against TuYV in the 2016/17 growing season, and a recently completed pre-breeding project has further explored the potential to develop commercial oilseed rape varieties to tolerate this virus infection. With those who advised farmers to grow small areas of oilseed rape in the future, this concern has also been expressed through a farmer survey during 2014/15 by the Farm Business Survey (FBS) team, where the most important reasons for a future reduction in area were crop rotations, reduced crop price, and cabbage stem flea beetle damage. When comparing the 2014/15 with 2013/14, the total area has decreased by about 3% (22,000 ha). In line with the experts’ suggestions, the importance of IPM has also been emphasized by EU and UK policymakers, and numerous UK organizations (e.g., LEAF -Linking Environment and Farming, Natural England). In order to develop further IPM for oilseed rape production, one of the most crucial aspects is to understand insecticide efficacy on pest control, and its changes over time due to resistance: this will be important for developing action thresholds to use chemicals strategically. Another crucial aspect of IPM, as expressed by experts in this study, is natural pest control. Many studies have been carried out to evaluate the impact of natural enemies on pest suppression in oilseed rape, and to seek methods of conserving them. However, knowledge gaps still exist in this area. A big challenge will be to combine these two aspects when developing IPM strategies, so that the side effects of insecticides on natural enemies could be reduced to a minimum. Indeed, by conducting research among eight EU countries, have found consistent negative effects of insecticides on biological control potential. Farmers would adopt new strategies only if they work better than current practices. Profit is one fundamental aspect in the judgement. However, to our knowledge, little literature is available which estimates the influence of natural enemies on crop yield or net profits, especially for large scale field crops by conventional or IPM farmers, and none has focused on oilseed rape, partly because of the difficulty of conducting field experiments. More research is needed to estimate the economic value of this important service provided by natural enemies. # Conclusion Insecticides used in UK oilseed rape production have been designed to be more efficient in controlling pests, and less harmful to non-target species and the wider environment. However, their efficacy levels are not fully understood, and may not be sufficient in the long-term, due to the limited publicly available studies and fast development of insecticide resistance in pest species. Similarly, it is difficult to assess their side effects, partly because little research has comprehensively compared the impacts of different insecticides in a standardised manner. The type and extent of benefits for farmers are also fundamental when assessing insecticides. For these reasons, the decision as to whether further to restrict neonicotinoid seed treatments in oilseed rape needs careful evaluation. It is a challenge to take into account the multi-faceted aspects when assessing an insecticide; one way forward could be to translate each aspect into economic values, and then apply cost benefit analysis. In order to do so, more research is needed regarding the influence of a chemical on crop protection, farm profit, the environment and related ecosystem services. This study provided an insight into these aspects, but limitations exist due to a relatively small sample of expert opinions. Integrated pest management presents an important potential future strategy for oilseed rape production, and the importance of insecticides and natural pest control should be better recognised and incorporated. Economic valuation of pest control services by natural enemies for oilseed rape needs to be quantified, coupled with improved communication and knowledge exchange between government, researchers, consultants and growers. # Supporting Information We sincerely thank all the experts who participated, distributed, and commented on this survey. We sincerely thank that the Fera Science Ltd. shared the insecticide data for this study. [^1]: The authors have declared that no competing interests exist. DG is an employee of Fera Science Ltd. This does not alter our adherence to PLOS ONE policies on sharing data and materials. [^2]: **Conceptualization:** HZ TB AB SGP. **Data curation:** HZ. **Formal analysis:** HZ. **Investigation:** HZ. **Methodology:** HZ TB AB DG RH SGP. **Resources:** HZ TB AB DG RH SGP. **Software:** HZ. **Supervision:** TB AB SGP. **Validation:** HZ TB AB DG RH SGP. **Visualization:** HZ. **Writing – original draft:** HZ. **Writing – review & editing:** HZ TB AB DG RH SGP.

# Introduction Epithelial-stromal tumours of the serous histopathological subtype represent the largest group of epithelial ovarian cancers (EOC) and account for significant morbidity and mortality. Ovarian serous tumours may present as benign, low malignant potential (LMP) or malignant disease. Benign tumours account for up to 60% of ovarian serous tumours, present bilaterally in 20% of cases, and are cured through surgical removal of the disease. LMP tumours account for up to 15% of ovarian serous tumours and present bilaterally in 30% of cases. Although about 75% of LMP tumours are stage I at diagnosis, where survival rates exceed 90%, patients with advanced stage disease may die from complications due to extragonadal spread throughout the pelvic cavity. Approximately 15% of LMP tumours may recur up to 20 or more years after the initial diagnosis, and these cases usually have a poor outcome. About 30% of all ovarian serous tumours are malignant and 60% of these cases are bilateral. Serous tumours make up more than 50% of all malignant EOC. Although various grading methods have been used, it appears that the vast majority of malignant serous tumours are high grade ovarian serous carcinomas (HGOSC), with only about 10% presenting as low grade carcinomas (LGOSC). Treatments for both include surgery and chemotherapy, but most cases are diagnosed at advanced stages where the overall 5-year survival rate is less than 30%. Although patients with LGOSCs have a longer survival than those with HGOSCs, they respond poorly to conventional platinum and taxane-based chemotherapy, suggesting that the molecular pathways involved in the etiology of the diseases may differ. Although approximately 10% of EOC, particularly tumours of the serous subtype, occur in women harbouring germline mutations of the cancer susceptibility genes *BRCA1* or *BRCA2*, the etiology of the remainder of ovarian serous neoplasms remains unknown. Karyotyping and array comparative genomic hybridization (aCGH) studies of benign, LMP and malignant serous tumours indicate an increasing frequency of chromosomal abnormalities, with the most extensive aneuploidy and structural abnormalities occurring in malignant tumours –. Genetic analyses of *TP53* have identified rare somatic mutations in benign, LMP tumours and LGOSCs, and a very high frequency in HGOSCs. Mutually exclusive somatic mutations in either *KRAS* or *BRAF* are often reported in LMP tumours and LGOSCs (30–50%), but rarely in HGOSCs (\<12%). This mutation spectrum has been used as an argument that favours at least two distinct, but not mutually exclusive, pathways for the development and progression of ovarian serous tumours. One pathway involves a continuum of development involving benign, LMP tumours and LGOSCs, originating from surface epithelial cells of the ovary. The other pathway describes the *de novo* development of HGOSCs originating from either ovarian surface epithelial cells, or epithelial cells of the fallopian tube fimbriae. Defining the genes involved in the etiology of ovarian serous neoplasms would provide a means to further stratify patients for optimal treatment regimens, as well as identify new molecular pathways to explore in the development of biomarkers. This is particularly prescient for LMP cases given that the majority of patients do not succumb to the disease, although most cases are usually subjected to aggressive management. Although studies of DNA ploidy in LMP tumours have been used to stratify patients for aggressive treatment, the overall impact on survival is not clear. Karyotype studies have implicated chromosome 3p genes in EOC, and loss of heterozygosity (LOH) analyses have suggested that 3p genes may function as tumour suppressors. We have previously reported LOH of 3p14-pcen in benign, LMP tumours, LGOSCs and HGOSCs. Although the studies were limited by sample size, it is tempting to speculate that gene(s) residing in this genomic region may be involved in the tumourigenesis of ovarian serous neoplasms. This notion is supported by functional complementation studies involving the transfer of ‘normal’ 3p fragments, including the 3p12-pcen region, which rendered an aggressive EOC cell line harbouring LOH of the 3p arm, non-tumourigenic. LOH of 3p25-ptel was also reported in benign, LGOSCs, and HGOSCs, suggesting more than one tumour suppressor gene may be involved in the etiology of ovarian serous neoplasms. Whole genome expression analyses and targeted analyses of 3p25-ptel and 3p14-pcen genes also have identified promising candidates for further molecular analyses. In this study we have performed an extensive genetic analysis of benign and LMP ovarian serous tumours to further characterize somatic genetic events associated with the most indolent form of ovarian disease. We performed a targeted LOH analysis of the 3p12-pcen locus of interest generated from our previous analyses of benign, LMP and malignant ovarian carcinomas, in benign ovarian serous tumours to determine the extent of loss of 3p alleles in this disease. To further characterize genomic anomalies, we applied high-density genome-wide genotyping bead array technology to benign and LMP ovarian serous tumour samples. Genome-wide genotyping array studies have already shown the occurrence of specific anomalies, such as 3p loss, attesting to earlier findings that genomic aberrations are not necessarily random in malignant EOC (reviewed in Gorringe *et al.*, 2009). However, genotyping array analyses have largely focused on HGOSCs, and previous genome-wide studies of benign, LMP tumours and LGOSCs were limited by the density of genetic markers or by sample size,. We relate our results to the mutational spectra derived from *TP53*, *KRAS* and *BRAF* genetic analyses, as these genes are mutated in ovarian tumours with varying frequencies depending on the pathology of the disease. In some cases, we were also able to investigate synchronous bilateral ovarian tumours. We also analyzed a set of LGOSC, as these cancer samples have rarely been genetically characterized due to their paucity relative to HGOSC cases. This study represents the largest sample of ovarian serous tumours examined to date using high density genotyping technologies. The integration of targeted genetic analyses with global genomic effects may contribute to our understanding of the etiology of benign and LMP ovarian serous tumour samples. The results of our targeted genetic and genomic analyses support the hypothesis that LGOSC could arise from serous benign and LMP tumours, but do not exclude the possibility that HGOSC may also be derived from LMP tumours. # Results ## Genetic analysis of chromosome 3p LOH of 3p has been reported in up to 20% of benign ovarian serous tumours. As previous studies were limited by sample size, we used polymorphic microsatellite repeat markers to investigate LOH of regions on 3p in 50 benign ovarian tumour samples. We focused our analysis on the 3p26.2, 3p21.31, 3p12.3, 3p12.2, and 3p11.2 regions shown to exhibit LOH in serous benign and/or malignant tumours. Although the analysis was informative for at least one marker per region examined in 78–90% of the samples, no evidence of LOH was observed in any of the samples analyzed. To increase the resolution of markers in order to detect LOH events in tumour samples, we applied Illumina's HumanHap300-Duo Genotyping BeadChip, which assays approximately 317,500 SNPs across the human genome, to three benign ovarian tumour samples. As proof of principle, we investigated sample 1781T, a benign ovarian serous tumour that has been shown to exhibit LOH of 3p14-pcen. Samples BOV-1329GT and BOV-2564DT, which did not exhibit evidence of LOH in the present study, were also examined. BeadChip analysis identified a 9.1 Mb run of homozygosity (ROH) at 3p12-p11 in 1781T that did not display a corresponding decrease in the Log R ratio, which would have been consistent with a deletion occurring in this region. No 3p anomalies were inferred from the BeadChip analyses of samples BOV-1329GT and BOV-2564DT (data not shown). Genetic analysis of DNA from normal tissues from case sample 1781T using seven polymorphic microsatellite repeat markers suggested that the 3p ROH also occurred in constitutional DNA (data not shown). A higher density array, the Human610-Quad Genotyping BeadChip (610K), which contains over 600,000 markers, was used to genotype DNA extracted from two portions of the 1781T tumour specimen. Interestingly, 1781T-A exhibited consistent allelic imbalance across the entire lengths of chr3 and chr9, in contrast to normal genotypes observed on all chromosomes in the 1781T-B DNA preparation.The 3p12-p11 ROH is present in both preparations, affirming earlier findings that this ROH is likely present in constitutional DNA. These results are interesting in light of our recent studies that suggest the presence of an ovarian cancer tumour suppressor gene located in the 3p12-pcen region. ## High-density genome-wide genotyping of benign tumours, LMP tumours and LGOSCs To investigate the possibility that LOH analyses underestimated the frequency of 3p abnormalities in benign and LMP serous tumours, we applied the 610K BeadChip technology to an additional 21 benign ovarian serous cases (32 tumours) and 53 LMP ovarian serous cases (58 tumours), of which 10 benign and 5 LMP cases included samples taken from both the left and the right ovaries. We also included 11 LGOSC cases (12 tumours), for which both bilateral tumours of one patient were arrayed. HGOSCs have already been shown to demonstrate LOH and abnormalities of 3p using genotyping arrays. Using the Genome Viewer module of the BeadStudio software, we visually assessed the data, which was aligned according to genomic position. The B allele frequency and Log R ratio were examined in order to infer allelic imbalance of whole chromosomes or chromosomal arms and intrachromosomal breaks (**;** **,** **,**). As summarized in, allelic imbalance of 3p was observed in only two LMP samples, TOV-1068T and TOV-3922GT, both of which also harboured allelic imbalances of other chromosomes. Breaks involving the 3p arm were observed in two LMP tumours, TOV-942T and TOV-1685T. Chromosome 3p breaks were more frequently observed in LGOSCs (4 of 11 cases); however, intrachromosomal breaks were also observed on other chromosomes in all of these cases. Overall, chromosomal aberrations were more commonly observed in LMP cases (30 of 53 cases or 56.6%) than in benign cases (3 of 22 cases, 13.6%) ( **and**). The most commonly affected chromosomal arms in LMP cases were 12p (12/53), 12q (9/53), 8p (7/53), 8q (7/53), 1p (6/53), and 22q (6/53). Allelic imbalance was more frequently observed on chr12 and chr8, whereas intrachromosomal breaks were observed more often on 1p and 22q. Chromosomal abnormalities were observed in all but two of the LGOSC samples (TOV-682T and TOV1284T). Bilateral tumours from 16 samples in this study were genotyped. None of the 10 paired bilateral benign tumours exhibited any evidence of genomic anomalies. Of the five paired LMP samples examined, one or both tumour samples exhibited evidence of chromosomal abnormalities. Some of these cases exhibited identical (cases TOV-1775 and TOV-920) or similar (case TOV-4054) abnormalities, suggesting the possibility of common clonal origins in these cases, as has been proposed for malignant ovarian cancers. An identical spectrum of chromosomal abnormalities was observed in the one case of paired LGOSC samples (case TOV-854). Homozygous deletions may be inferred by identifying markers associated with a downward deviation of the Log R ratio and the absence of allele frequency scores. Null alleles resulting from somatic homozygous deletions are of particular interest, as they may affect the function of tumour suppressor genes. Furthermore, breaks occurring adjacent to cancer-associated genes may affect their regulation. provides the coordinates where both alleles are likely to be deleted, along with the affected genes. Some of the intervals are known to harbour germline copy number variants (CNV) as reported in the Database of Genomic Variants (projects.tcag.ca/variation). Thirty-six homozygous deletions were found to be unique to a single case. Some of these deletions may possibly affect the function of genes located within or adjacent to the deleted intervals. Homozygous deletions were observed in benign, LMP tumours and LGOSCs, although relative to the number of samples in each group, homozygous deletions were observed more often in LGOSCs. This is likely to be an overestimate, as LGOSC sample TOV-490T harboured several chromosomes with reduced copy number. On these chromosomes the Log R ratio is decreased, resulting in the coincidental appearance of three adjacent markers with a Log R ratio of ≥−2. Given the large ROH overlapping the 3p12-p11 region in the benign tumour sample 1781T, we investigated whether ROHs of this interval were also observed in other samples. This analysis was restricted to the benign and LMP samples, as they exhibited low levels of generalized genomic instability. We examined ROHs larger than 5 Mb, as previous studies have shown that smaller ROHs, particularly those less than 1.5 Mb, may be common occurrences.There were no significant differences in the occurrence of at least one ROH \>5 Mb per sample studied: 4/22 (18.1%) benign cases and 11/53 (20.4%) LMP cases contained at least one. Notable is the large number of ROHs (n = 14) observed in the benign bilateral tumour samples BOV-1588DT and BOV-1588GT, as compared with benign and LMP cases exhibiting no (n = 61), one (n = 8), two (n = 5) or four (n = 1) ROHs \>5 Mb. Both the left and the right ovarian tumours exhibited the same pattern of ROHs, accounting for about 7% of the genome. Genotyping of peripheral blood DNA from the same patient suggested that the ROHs occur in the germline and were not somatically acquired during the development of these tumours (data not shown). Interestingly, more ROHs were observed on chr3 than on any other chromosome. Two LMP samples (TOV-1694DT and TOV-933DT) exhibited ROHs overlapping the 3p12-p11 ROH observed in the benign sample 1781T. Additionally, two benign and/or LMP tumour samples displayed overlapping ROHs located at 2.6–5.1 Mb and 190.2–196.4 Mb on chr3. ## Genetic analysis of *TP53*, *BRAF* and *KRAS* and association with genomic anomalies Mutations of *KRAS*, *BRAF* and *TP53* were only detected in LMP tumours and LGOSCs. As reported in independent studies, samples with mutations in *KRAS* or *BRAF* were mutually exclusive. Concordant mutation results were observed in all but one of the bilateral tumour samples (LMP case TOV-1010DT/GT). There were significantly more *KRAS* and *BRAF* mutations (26 of 53, 49.1%) and fewer *TP53* mutations (1 of 53, 1.9%) in LMP cases as compared with *KRAS* and *BRAF* mutations (3 of 11, 18.1%) and *TP53* mutations (5 of 11, 45.5%) in LGOSCs ( **and**) (p = 0.00049). In general, the LMP and LGOSC cases with somatic *TP53* mutations harboured disorganized genomes, particularly large numbers of intrachromosomal breaks ( **and**). The LMP sample with a *TP53* mutation (TOV-1685GT) has 30 of 41 chromosomal arms harbouring an aberration, similar to the average number (33.4) of chromosomal arms harbouring an aberration in the *TP53* mutation positive LGOSCs. LMP cases with *KRAS* mutations contained an average of 5.3 chromosomal arms harbouring an aberration, whereas cases with *BRAF* mutations had an average of 1 chromosomal arm with an aberration. LMP mutation-negative tumours had an average of 1.5 chromosomal arms with an aberration. In the LMP tumours, there were significantly more *KRAS* mutation-positive cases that were associated with a gain of 12p (8 of 12, 66.7%) than there were in *KRAS* mutation-negative tumours (4 of 41, 9.8%) (p = 0. 0.0002). This is an interesting observation, as *KRAS* is located at 12p12.1. Moreover, the only other LMP sample to exhibit overt disorganization of its genome, sample TOV-942GT, harboured a high-level 1.59 Mb amplification containing 12 genes, including the *KRAS* locus. A gynecologic pathologist independently reviewed the LMP and LGOSC samples that were found to harbor *TP53* mutations in a blinded manner to confirm their histopathological classification. All LMP samples retained their classification status. Interestingly, none of the LGOSC samples harbouring *TP53* mutations maintained their designation. TOV-553EPT and TOV-490T were reclassified as high grade carcinomas; TOV-812EPT was reclassified as a metastatic serous carcinoma, grade not determined; TOV-947DT was reclassified as a possible LMP; and TOV-81DT was reclassified as a non-invasive implant ( **and**). ## Global analysis of copy number aberrations of benign, LMP and LGOSCs Genotyping data were analyzed by GenoCNA to evaluate various states of copy number variations that include allelic content occurring within each group of benign and LMP samples. The LGOSCs were not analyzed, given the small number of cases within the group and the fact that a number of cases were later designated by histopathology as not LGOSC. As noted in, very few chromosomal abnormalities were observed within the group of benign tumours, which was reflected in the GenoCNA analyses. Discrete gains and losses occurred throughout the genome at low frequencies (\<20%), representing CNVs. The most common regions of gains are adjacent to the centromeres on many of the chromosomes, likely indicating repetitive regions. Frequent regions of loss include the HLA region of chr6 (80%), and other common homozygous deletions (as catalogued). Discrete CNVs and somatic gains and losses of whole chromosomes and chromosomal arms are reflected in the GenoCNA analyses of the LMP tumours. As expected, chr1 shows loss of the p arm and gain of the q arm in 10–15% of samples, whereas chr12 and chr8 show gains of the entire chromosome. Losses of chr4, chr5, chr6p, chr9p and chr13 are apparent, as are gains of chr7 and chr20 (5–15%). ## Characterization of chromosome 3p12-pcen interval The ROH in the 3p12-p11 interval, along with the allelic imbalance of chr3 observed in the benign tumour sample 1781T, is interesting in light of recent research in our group suggesting the possibility of tumour suppressor gene(s) in this interval. To investigate this further, we performed mutation analysis in 1781T of protein coding regions and intron/exon splice junction sites of the top 3p12-pcen tumour suppressor gene candidates, *ROBO1*, *GBE1* and *VGLL3*. Several variants, but no apparent deleterious mutations, were observed. BeadChip analysis of 1781T demonstrated extensive allelic imbalance of chr3 and chr9. Chromosome 3 harbours *RASSF1A* (at 3p21.31) and *MLH1* (at 3p22.2), and chr9 harbours *CDKN2A* (9p21.3). These genes have been shown to exhibit tumour suppressor activity, which are often silenced by promoter methylation. Although the frequency of these events appears to be low in ovarian cancer, we tested the possibility of promoter methylation silencing in the benign tumour case 1781T and our well-characterized EOC cell lines. There was no evidence of promoter methylation of these genes in the analysis of either 1781T-A and 1781T-B, in contrast to evidence of methylated *RASSF1A* alleles in OV-90, TOV-112D, TOV-21G, and TOV-2223G, methylated *CDKN2A* alleles in TOV-112D, and methylated *hMLH1* alleles in TOV-21G (data not shown). ## Characterization of putative homozygous deletion affecting gene function The inferred 242.5 kb homozygous deletion observed at 6q22.1 in LMP tumour TOV-4054DT stood out in part because it is much larger than the size of the average homozygous deletion (28.3 kb) observed in the present study (**,**). The deletion is predicted to affect the function of *ROS1*, *DCBLD1* and *GOPC*, with breakpoints occurring in all three genes. A literature review of these genes reported that *ROS1* and *GOPC* are partners in an oncogenic fusion gene found in the glioblastoma cell line U118MG, created by a 240 kb intrachromosomal deletion. In U118MG, the fusion gene is transcribed from the 5′ end of *GOPC* and contains the first 7 *GOPC* exons and the last 9 *ROS1* exons. Log R ratios indicate that the breakpoints of the 6q22.1 deletion in TOV-4054DT occurred in genomic regions that could possibly result in the creation of an identical fusion gene. To investigate this possibility, we designed an RT-PCR assay to detect the presence of a fusion transcript in cDNA prepared from TOV-4054DT. As shown in, TOV-4054DT harbours an aberrant transcript not present in the well- characterized ovarian cancer cell line, OV-90neo^r, which does not harbour a 6q22.1 anomaly (data not shown). However, a faint band corresponding in size to the aberrant 6q22.1 transcript was also visible in the RT-PCR analysis of the contralateral LMP tumour TOV-4054GT, suggesting a clonal origin of cells that contain this anomaly. This is consistent with observation that both LMP tumours harbor allelic imbalance of the chr6q arm which include the *ROS1*, *DCBLD1*, and *GOPC* loci. Sequence analysis of the aberrant transcript revealed that it was comprised of an in-frame fusion between exon 7 of *GOPC* and exon 35 of *ROS1*. We attempted to detect the fusion protein by Western blot, but the only tissue available for protein extraction was embedded in OCT medium, which was not amenable to further experiments. Interestingly, the fusion transcript is identical to that reported in the U118MG glioblastoma cell line, and to one of the fusion genes identified in a set of cholangiocarcinomas. A review of genotyping data from a minimum of an additional 200 ovarian cancer samples and cell lines of various grades and histopathologies from our laboratory suggest that this chromosomal anomaly is unique to case TOV-4054 (data not shown). # Discussion Although LOH of 3p has been reported in benign serous tumours at frequencies of up to 20%, loss of 3p alleles were not observed in the analysis of 50 new cases. Interestingly, the sample 1781T, which exhibited 3p14-pcen LOH in our previous LOH study, was shown to exhibit allelic imbalance of chr3 and chr9. Although methylation of *RASSF1A* (3p21.31) and *CDKN2A* (9p21.3) has previously been reported in benign serous tumours at a low frequency, we observed no evidence of alteration of promoter CpG methylation in sample 1781T. We also did not detect evidence of promoter methylation of *MLH1* (3p22.2); however, this alteration is more commonly observed in low grade ovarian carcinomas of the endometrioid histopathological subtype. LOH analyses of 3p loci are consistent with SNP analyses suggesting that 3p anomalies are rare occurrences in benign serous tumours, as are anomalies associated with other chromosomes. Although independent LOH analyses have shown low frequencies of loss of chromosomes 6, 7, 9 and 10, only a limited number of loci were examined. Array CGH studies have identified both gains and losses of chr6, and losses of 1p, 4q and 5q,. The absence of *KRAS* and *BRAF* mutations in our set of benign tumours is consistent with the paucity of somatic events observed in independent reports. It has been proposed that the acquisition of a *KRAS* or *BRAF* mutation in a benign tumour might initiate the progression to an LMP tumour. The underlying molecular genetic events associated with the development of benign ovarian serous cancer samples remains elusive. It is possible that an excess of contaminating stromal cells may have obscured chromosomal anomalies in a subset of the samples analyzed. Previous studies using LOH analysis or CGH have observed chromosomal abnormalities without enriching for tumour cells, as chromosomal anomalies present in even 40% of cells can be detected by SNP array analyses. Microdissection of tumour tissues would have necessitated a round of whole genome amplification (WGA), which is discouraged by Illumina. The Illumina Infinium protocol includes a WGA step, and an additional round of WGA has been shown to reduce the call rate and may introduce allelic bias. Hence, it is possible that chromosomal anomalies are underreported in this study. While it was not possible to array constitutional DNA from every patient, a subset of abnormalities observed could be germline CNVs. It is interesting that the 9.1 Mb ROH at 3p12 observed in sample 1781T overlaps a tumour suppressor region identified by our group using a functional complementation study involving the transfer of chr3 fragments into an EOC cell line, and using comparative transcriptome analysis of ovarian serous cancer and normal samples. Although no mutations were identified in 1781T in the targeted analysis of tumour suppressor gene candidates *ROBO1*, *GBE1* and *VGLL3*, miRNAs or other noncoding RNAs (ncRNAs), either located within this region or acting upon expression of genes in the region, may play a role in the development of these tumours. Several ncRNAs, predicted to contain miRNA target sites, have been identified in the 3p12.3-pcen interval and shown to be differentially expressed in cancers compared with normal tissues. It is notable that the 3p12 interval was the region of the genome most commonly present in ROHs longer than 5 Mb, and that chr3 harboured both the most number and the longest ROHs (up to 56.6 Mb) of any chromosome within this study. The significance of this observation is unknown but could be influenced by founder effects, as the majority of samples analyzed in our study were from the French Canadian population of Quebec known for its unique genetic demography. A recent genome-wide SNP array analysis of 140 French Canadians from different geographic locations within Quebec reported that subpopulations varied in their genomic structure and degrees of relatedness, and contained significantly more ROHs than samples from European populations. Case BOV-1588 exhibited the most extensive ROHs, as approximately 212 Mb of the genome (7.1%) occurred in ROHs longer than 5 Mb. These ROHs were confirmed to be germline in this patient. As the offspring of first cousins are expected to have about 6.25% genomic autozygosity, it is possible that the extensive ROHs observed in BOV-1588 were the consequence of a consanguineous mating. Upon further review of the medical history of this case, it was revealed that the patient has schizophrenia, a condition that has recently been associated with ROHs. As ROHs may play a role in the etiology of genetic diseases, including cancer, further studies are required to determine the significance of these regions in benign ovarian serous tumours. The chromosomal abnormalities observed in 58 LMP samples from 53 cases mirror those previously reported in the literature, where 1p and 22q are subject to losses, and chr12 and chr8 display increases in copy number, ,. As expected, *KRAS* and *BRAF* mutations were observed in a mutually exclusive manner. Interestingly, gain of chr12 was significantly associated with the presence of *KRAS* mutations, a finding that has been previously observed. This association was observed in both non-small cell lung cancer (NSCLC) and lung adenocarcinomas, although not in colorectal cancers. Increased *KRAS* expression was observed in NSCLCs harbouring modest increases of copy number of chr12. Another study indicated that NSCLC patients with both a *KRAS* mutation and gain of chr12 had a worse prognosis than those harbouring only one of these aberrations. It would be interesting to investigate this association in LMP cases, but this may be difficult with the low frequency (\<15%) and the long average time (\>15 years) of recurrences for this disease. To date, only one LMP case, TOV-942GT, has died of cancer, which occurred within a year of the LMP tumour diagnosis; however, the cause of death was pancreatic carcinoma. TOV-942GT harboured an amplification of the *KRAS* locus, and while pancreatic carcinomas have been shown to have a high *KRAS* mutation rate, the pathology review excluded the possibility of metastasis in this case. The low frequency of *TP53* mutations in LMP samples is also consistent with independent reports. The *TP53* mutation positive case (TOV-1685GT) was identified in a young patient (age 26), who has remained cancer-free for the follow-up period of 6.5 years. Interestingly, both TOV-1685GT and TOV-942GT harboured extensive evidence of chromosomal instability (CIN) by SNP array analyses. However, low levels of CIN were also observed in a number of *BRAF* and *KRAS* mutation negative cases. Although the relationship between somatic mutations in these genes and genomic anomalies is unknown, the high frequency of CIN in the context of *TP53* mutations combined with the role of p53 in DNA damage response has been proposed in numerous studies (reviewed in Negrini et al., 2010). Collectively, our results indicate that ovarian serous LMPs are a heterogeneous group, composed of tumours displaying a range of genetic and chromosomal anomalies. It remains to be determined what effects the various anomalies observed in this study have on the clinical presentation of the disease. The genetic spectrum of abnormalities observed in our small set of LGOSC cases is also consistent with independent reports, particularly when factoring in an independent review of the histopathology of cases. All five LGOSC cases that harboured a somatic *TP53* mutation exhibited extensive CIN and were later reclassified. The overlap in the genetic spectrum of anomalies observed in LGOSC samples with those observed in LMP samples supports the notion that they may share a common molecular genetic etiology. However, the rare instances of *TP53* mutation positive LMP samples (including the LGOSC reclassified as a LMP case) would also support the notion that some LMP samples share common origins with HGOSC as they often exhibit somatic *TP53* mutations and extensive CIN. Regardless of the putative origins of EOC, our results suggest that a combination of *TP53* mutation testing and SNP array analyses may facilitate the classification of malignant serous cases. Identifying methods to improve histopathological classification of serous EOC cases may prove useful as improvements in patient management emerge for treating LGOSC cases. Few unique homozygous deletions were inferred in the samples analyzed, and none overlapped regions containing known tumour suppressor genes. It is interesting that 28 genes reported as differentially expressed in transcriptome studies of LMP samples are located directly adjacent to or within homozygous deletions identified in our SNP analyses of LMP samples. Furthermore, pairs of differentially expressed genes directly flank six of the observed homozygous deletions. Given the presence of contaminating stromal cells in the samples analyzed, it is likely that many of the homozygous deletions represent germline CNVs, even those found to be unique to a specific case. As CNVs may contain regulatory elements, it is possible that these germline homozygous deletions may affect the expression of adjacent genes, thus contributing to tumour risk or progression (reviewed by Henrichsen et al.). It is also possible that the presence of homozygous deletions may affect chromatin folding, affecting the expression of multiple genes in the region. A 242.5 kb homozygous deletion at 6q22.1 was observed in the LMP tumour sample TOV-4054DT. Molecular genetic characterization suggests that this resulted in the creation of a transcriptionally active *GOPC-ROS1* fusion gene. To the best of our knowledge, this is the first fusion gene reported in an ovarian LMP context. An identical fusion gene has been described in the glioblastoma cell line U118MG, as well as in a cholangiocarcinoma tumour. Both groups have demonstrated that the GOPC-ROS1 fusion protein is capable of transforming non- malignant cells. This variant protein retains tyrosine kinase activity and is targeted to the Golgi membrane. While it does have oncogenic activity, its aggressivity was augmented when expressed in mice with a disrupted *p16Ink4a* and *p19Arf* locus. Another *GOPC-ROS1* fusion gene was observed in a different cholangiocarcinoma tumour, which resulted in a smaller open reading frame, different cellular location and more potent transforming ability. Although targeted mutation analyses of *ROS1* or *GOPC* have not been performed in cancer samples, the Sanger Wellcome Trust COSMIC database (<http://www.sanger.ac.uk/genetics/CGP/cosmic/>) reported low frequencies of *ROS1* sequence variations in ovarian (1/84), lung (8/131), breast (2/201), stomach (2/60), colorectal (1/133) and CNS tumours (3/477). However, a recent large scale exomic genome sequencing analysis of 316 HGOSCs by The Cancer Genome Atlas Research Network identified 5 cases with verified sequence variants. Of the 22 sequence variations observed in either the ovarian TCGA study or in multiple tumour types in the Sanger Wellcome Trust COSMIC database, 17 are missense mutations, with 4 occurring in the tyrosine kinase domain. In total, six mutations have been observed and validated in ovarian tumours, including four missense mutations and two silent mutations. Likewise, one mutation has been observed in *GOPC*; a missense mutation in an ovarian clear cell tumour. The fusion gene occurred in a *TP53*, *KRAS* and *BRAF* mutation-negative context, with evidence of a modest level of CIN in the case sample. The LMP case was bilateral, and although the anomaly was more evident in the right tumour, molecular genetic analysis suggested that both harboured the fusion gene. The clinical and biological significance of this genetic abnormality is not clear. To date, the patient has been cancer free for 1.5 years. However, there is no evidence from a review of SNP array data that it is a common event in LMP, benign or LGOSC samples. Our group is currently investigating SNP array results from HGOSCs and EOC cell lines, and no evidence of a homozygous deletion affecting this region in these aggressive EOC tumours and cell lines were observed (data not shown). It would be interesting to test the effect of the GOPC-ROS1 fusion protein in the context of LMP tumours, but this awaits the development of a suitable cell line model system for this variant of ovarian cancer. Thus we can only speculate based on the effect the identical fusion protein has on the transforming ability in transfected cells, and propose that it may have played a role in the pathology of this LMP tumour. Our results support the hypothesis that LGOSCs are derived from LMP ovarian serous tumours. Interestingly, chromosomal aberrations, but not genetic mutations, were observed in benign serous tumours. It is possible that acquisition of a mutation, such as *KRAS* or *BRAF*, represents the moment of transition from a benign tumour to an LMP. A number of LMP tumours lacking *KRAS* or *BRAF* mutations harboured genomic aberrations, indicating that different initiating events may be present in these tumours. Indeed, a fusion gene known to be oncogenic in other tumour types was found in a single LMP case. While it is unlikely that this fusion gene is a frequent event in the development of LMP tumours, its presence indicates that other initiating, growth-promoting events may be found. The data from this study also indicates that at least some HGOSCs may be derived from LMP tumours. This study also illustrates that there is potential for high-density genotyping arrays in combination with targeted mutation screening to become useful in classifying ovarian serous tumours, and could thus have important implications in management of patients where therapy is targeted based on histopathological subtype. # Materials and Methods ## Clinical Specimens Tumour samples and peripheral blood lymphocytes were collected with informed consent from participants undergoing surgeries performed at the Centre hospitalier de l'Université de Montréal-Hôpital Notre-Dame or from surgeries performed at the McGill University Health Centre – Montreal General Hospital. The study is in compliance with the Helsinki declaration, and has been granted ethical approval by the respective Research Ethics Boards of Centre hospitalier de l'Université de Montréal-Hôpital Notre-Dame and The McGill University Health Centre. Clinical features such as disease stage, and tumour characteristics such as grade and histopathological subtype, were assigned by a gynecologist- oncologist and gynecologic-pathologist, respectively, according to the criteria established by the International Federation of Gynecology and Obstetrics. ## EOC cell lines EOC cell lines were derived from a stage IIIc/low grade papillary serous adenocarcinoma (TOV-81D), a stage III/high grade clear cell carcinoma (TOV-21G), a stage IIIc/high grade endometrioid carcinoma (TOV-112D), the ascites fluid of a stage IIIc/high grade adenocarcinoma (OV-90), a stage IIIc/high grade serous carcinoma (TOV-2223G), and both the tumour and the ascites fluid of a stage IIIc/high grade serous tumour (TOV-1946 and OV-1946), all from chemotherapy- naïve patients, as described. OV-90neo^r is a pSV2NEO-transfected clone of OV-90, which confers resistance to Geneticin®.Cells were cultured in OSE Medium supplemented with 2.5 µg/mL amphotericin B, 50 µg/mL gentamicin and 10% FBS as described previously. ## Nucleic acid extraction DNA was extracted from EOC cell lines, fresh frozen tumour specimens and peripheral blood lymphocytes as described previously. For case sample 1781T, non-tumour DNA was extracted from a paraffin-embedded lymph node sample using a previously described method. Total RNA was extracted with TRIzol™ reagent (Invitrogen Canada Inc., Burlington, ON) from the OV-90neo^r cell line grown to 80% confluency in 100 mm Petri dishes, or from fresh frozen TOV-4054DT/GT tumours as described previously. RNA quality was assessed by gel electrophoresis or 2100 Bioanalyzer analysis using the RNA 6000 Nano LabChip kit (Agilent Technologies, Mississauga, ON). ## LOH analysis LOH analysis was performed using polymorphic microsatellite repeat markers representing various 3p loci: *D3S1304* and *D3S1515* at 3p26.2; *D3S1581* and *D3S3640* at 3p21.31; *D3S1274* and *D3S1542* at 3p12.3; *D3S1538* and *D3S2388* at 3p12.2; and *D3S2386* and *D3S2318* at 3p11.2. Genetic analysis of the 3p12 locus in the tumour sample 1781T was determined using seven polymorphic microsatellite markers: *D3S3507*, *D3S1274*, *D3S3049*, *D3S3508*, *D3S3633*, *D3S3679*, *and D3S2318*. The genomic location of the markers was based on February 2009 GRCh37/hg19 assembly of the human reference sequence. LOH analysis was performed using a previously described PCR-based assay, with the primers sets for each marker described in the UniSTS Database (<http://www.ncbi.nlm.nih.gov/unists>). LOH or allelic imbalance was scored based on the absence or difference in the relative intensity of alleles in tumour DNA as compared with the DNA from patient-matched peripheral lymphocytes or, in the case of 1781T, DNA from paraffin-embedded lymph node. ## Gene sequencing analysis Mutation analysis of tumour DNA samples was designed to detect variants in the protein coding exons 2 to 11 of *TP53*, as well as the common mutations in exon 2 of *KRAS* and exons 11 and 15 of *BRAF*. Peripheral blood lymphocyte DNA from case sample TOV-1685GT was also examined for *TP53* mutations in exon 10. Mutation analyses of case sample 1781T were also performed to identify variants in protein coding regions of the chr3 genes *ROBO1*, *GBE1* and *VGLL3*. Mutation analysis was performed using PCR-based assays followed by sequencing of both genomic strands using the 3730XL DNA Analyzer system platform from Applied Biosystems at the McGill University and Genome Quebec Innovation Center ([www.genomequebecplatforms.com](http://www.genomequebecplatforms.com)) as previously described. Primer sequences for each assay were reported previously, with alternate primers used for some reactions. Primers were designed using Primer3 software based on the genomic structures available from the February 2009 GRCh37/hg19 assembly of the human reference genome. Sequence chromatograms, reviewed by at least two observers, were compared with NCBI reference sequence (RefSeq) reported in GenBank: NM_133631.3 (*ROBO1*), NM_000158.3 (*GBE1*), NM_016206.2 (*VGLL3*), NM_000546.4 (*TP53*), NM_004985.3 (*KRAS*) and NM_004333.4 (*BRAF*). Sequence variants were compared with those reported in the SNP Database ([www.ncbi.nlm.nih.gov/SNP](http://www.ncbi.nlm.nih.gov/SNP)). In addition, *TP53* variants were evaluated based on information in the International Agency for Research on Cancer (IARC) TP53 Database ([www-p53.iarc.fr](http://www-p53.iarc.fr)). ## Promoter methylation analysis Promoter hypermethylation of *MLH1*, *RASSF1A* and *CDKN2A* was examined using methylation-specific PCR assays following bisulfite conversion of cytosine residues. The bisulfite conversion reactions were performed using the Imprint™ DNA Modification Kit (Sigma) with 200 ng of DNA from EOC cell lines or tumour tissue. Primer sequences for each assay have been published previously. ## High-density genotyping Genome-wide chromosomal anomalies in three benign ovarian tumours were inferred using the Infinium™ genotyping technology with Illumina's HumanHap300-Duo Genotyping BeadChip (Illumina, San Diego, CA, USA), which assays \>317,500 SNPs. Genotyping of 32 benign ovarian serous tumours (including the 3 tumours assayed on the 300K BeadChip), 58 serous LMP tumours and 12 LGOSCs was performed using Illumina's Human610-Quad Genotyping BeadChip (Illumina, San Diego, CA, USA). This BeadChip assays 620,901 markers, where over 560,000 are SNPs with an average spacing of 4.7 kb per marker (median spacing is 2.7 kb). Both genotyping, using 750 ng of DNA from frozen tumours, and scanning, using the BeadArray™ Reader, were performed at the McGill University and Genome Quebec Innovation Centre (<http://gqinnovationcenter.com/index.aspx>). All samples had call rates (the percentage of valid genotype calls) within the range of 0.914 and 0.999 (average 0.992). Genotyping results are available at Array Express (in progress). Genotyping analysis was performed using the Genome Viewer module in BeadStudio Data Analysis software v2.2.22 (Illumina, San Diego, CA, USA.). The software aligns genotyping data for each marker with genomic map coordinates based on March 2006 NCBI36/hg18 (Build 36.1) assembly of the human reference sequence (genome.ucsc.edu/cgi-bin/hgGateway). An image file was created for inferring genomic rearrangements based on the allele frequency and copy number (Log R ratios) for each marker assayed. LOH was inferred by B allele frequency, where values that deviate from 0.5 (less than 0.4 and greater than 0.6) indicate allelic imbalance when reviewed for a series of adjacently mapped markers. Breakpoints were inferred based on deviation of allele frequencies relative to those of adjacently mapped markers. Log R ratios deviating from 0 suggest copy gain or loss. Homozygous deletions were inferred based on Log R ratios ≤−2 for at least three adjacently mapped markers, and sizes were estimated based on the location of nearest flanking markers with Log R ratios above −2. Regions suggesting extensive homozygosity (or runs of homozygosity; ROH), spanning intervals \>5 Mb were inferred from heterozygous SNP markers. ROHs were required to have an average frequency of 1 SNP per 10 kb, and a heterozygous call for a marker was allowed if it was flanked by at least 100 SNP markers with homozygous scores. The distribution of mutations in *KRAS*, *BRAF* and *TP53* between the LMP and LGOSC cases was compared using the Fisher Exact test (Statistical Product and Service Solution Package, SPSS, Chicago, IL). Normalized SNP intensity files were also analyzed by GenoCNA. This software uses a hidden Markov model containing 9 different tumour states, encompassing loss of 1 or 2 copies, copy number neutral LOH, and 5 different gain states allowing for different patterns of allele retention. This model explicitly allows for normal tissue contamination in the samples. Graphs show the percentage of the samples with gains or losses based on the GenoCNA inference, where the percentage is calculated in expectation, using the average of the probabilities of relevant states at each marker. ## Gene expression analysis Expression of the *GOPC-ROS1* fusion gene was assayed by RT-PCR in TOV-4054DT/GT and OV-90 neo^r (negative control) using cDNA synthesized as previously described. Approximately 200 ng of a 1∶10 dilution of the reverse transcribed cDNAs were used in PCR assays. Primers were designed using Primer3 software based on the genomic structures of *GOPC* and *ROS1* and on mRNA sequences available from the February 2009 GRCh37/hg19 assembly of the human reference genome. # Supporting Information We thank Manon de Ladurantaye for her helpful expertise. Research was conducted at The Research Institute of the McGill University Health Centre which receives support from the Fonds de la recherche du Québec - Santé (FRQS). Clinical specimens were provided by the Banque de tissus et de données of the Réseau de recherche sur le cancer of the FRQS affiliated with the Canadian Tumour Repository Network (CRTNet). [^1]: Conceived and designed the experiments: PNT. Performed the experiments: AHB SLA KKO. Analyzed the data: AHB SLA KKO CMG A-MM-M PNT. Contributed reagents/materials/analysis tools: KR AKW DP CMG A-MM-M PNT. Wrote the paper: AHB PNT. Pathology review: KR AKW. [^2]: The authors have declared that no competing interests exist.

# Introduction Research has shown that 60 minutes moderate-to-vigorous physical activity (MVPA) per day is the minimum amount to benefit health in youth. MVPA is defined as activity above three age-adjusted metabolic equivalents (METs), which are for example jogging, swimming or playing soccer. Active involvement in physical exercise not only promotes physical health, but it also may enhance neurocognitive functioning (e.g. memory, information processing speed, inhibition). Yet most children and adolescents have not reached a healthy standard of exercise. According to research, 21% of the children up to 11 years of age, and 13% of the adolescents between 12 and 17 years of age, meet the recommended 60 minutes of daily moderate-to-vigorous physical activity (MVPA). Potential mechanisms underlying the beneficial effects of exercise on neurocognitive functioning include enhanced cerebral blood flow, growth factor release, neurogenesis, and angiogenesis. Acute (immediate) beneficial effects of physical exercise in youth have been firmly established in a laboratory setting, but it remains largely unknown whether structural daily life physical activities are also positively associated with neurocognitive functioning. Previous studies have shown that organized sports, active transport, physical education (PE) and outdoor play contribute significantly to the total MVPA in school-aged children, among which organized sports is most strongly associated with MVPA.\[–\] Therefore, the first aim of the present study is to investigate the association between a range of physical activities including sports, active transport, PE and outdoor play, and neurocognitive functioning (including short term memory, working memory, inhibition, attention and information processing speed) in preadolescent children. Such insights are not only of considerable scientific interest, but are also important for designing specific intervention programs to promote physical activity in youth. In this study we’ve focused on comparing three groups of preadolescent children varying in the amount of physical activity and sedentary behaviour: children who do not participate in any organized sports, children regularly participating in sports, and children very frequently participating in sports. For the latter two groups, we have chosen to include soccer players, as soccer is the most popular sports in the Netherlands, with a club member rate of 39% among boys between 5 and 18 years of age. Considering the above mentioned positive relationship between physical activity and neurocognitive functioning in young people, it is quite alarming that an increase in sedentary behaviour among youth has been demonstrated during the last decade. Sedentary behaviour are activities such as TV-watching, playing computer games, and driving to school. Sedentary behaviour is largely independent of physical activity, indicating that physical activities are not always replaced by sedentary behaviour. A negative association has been suggested between sedentary behaviour and neurocognitive functioning. The precise mechanisms are unknown, but sedentary behaviour may result in higher inflammation risks, disturbed insulin regulation, and high levels of triglycerides may play a role. As far as we know, no study to date has investigated the relationship between sedentary behaviour and neurocognitive functioning in children. Therefore, the second aim of the present study was to investigate the relationship between several sedentary behaviour including TV-watching, gaming and computer time, and neurocognitive functioning in preadolescent children. # Methods ## Participants A total of 168 preadolescent children, aged 8–12 years, were recruited from primary schools, an amateur soccer club, and a professional soccer club. The sample included 51 boys not involved in any organized sports (non-athletes), 48 boys who regularly participated in sports (non-elite soccer players), and 69 boys who participated very frequently in sports (elite soccer players). Non- athletes were not involved in any organized sports activities; they were neither member of a sports club, nor participating in an extracurricular sports program at school. Non-athletes were recruited at elementary schools. The non-elite players were recruited from an amateur soccer club. The elite players were recruited from the youth academy of a Dutch professional soccer club and were following the talent development program of the youth academy. Details on this group are provided in Verburgh et al. Participants were free of known behavioural, learning and medical conditions that might impact neurocognitive functioning. To fully understand the tasks that were administered in this study, participants were included when they had an IQ\>70 as measured by a short version of the Wechsler Intelligence Scale for Children III. The study was approved by the Institutional Review Board (IRB) of the Vrije Universiteit Amsterdam (Faculty of Behavioural and Human Movement Sciences). All participants and parents and/or legal guardians were informed about the procedures of the study before giving their written informed consent prior to participation. ## Materials ### Neurocognitive tasks The Stop Signal task was used to measure motor inhibition and involved two types of stimuli: Go stimuli and stop stimuli. Go stimuli were left- or right- pointing airplanes requiring a left or right button response, respectively. In a semi-randomly selected 25% of the trials, go stimuli were followed by a visual stop signal (traffic stop sign) superimposed on the go stimulus, requiring the participants to withhold their response. The delay between the go and stop signal (SSD) varied trial by trial using a tracking algorithm that increased or decreased the delay by 50 ms, depending upon whether or not the previous stop trial resulted in successful inhibition. This procedure yielded 50% successful inhibitions and 50% failed inhibitions. The dependent variable that reflects the latency of the inhibitory process is stop signal reaction time (SSRT). SSRT was calculated by subtracting average SSD from mean reaction time (MRT) calculated for correct responses on go trials. Shorter SSRTs reflect a faster and more efficient inhibitory process. Two aspects of memory were assessed: short term memory and working memory. Verbal short term memory was assessed using Digit Span Forwards of the Digit Span task of the WISC III. Participants had to verbally reproduce dictated series of digits in a forward order, increasing in length after every two trials. Participants received one point for each correct response. Visuospatial short term memory was assessed using an adapted version of the task developed by Bergman-Nutley and colleagues, in which participants were required to repeat sequences of yellow circles presented in a four by four grid (using the computer mouse) in a forward order (VSTM Forwards). Difficulty level was increased during the course of the task by manipulating position of the stimuli and increasing span. For both tasks, the total number of correct responses multiplied by highest difficulty level passed was calculated and the composite score of the verbal short term memory and visuospatial short term memory task (the average of both z-scores) was included as dependent variable in the analyses. Working memory was examined using the composite score (the average of both z-scores) of Digit Span Backwards of the Digit Span task of the WISC III and the backward condition (VSTM Backwards) of the adapted version of the task developed by Bergman-Nutley and colleagues. These conditions were similar to the forward conditions, but participants had to repeat the stimuli in reversed order, which appeals to working memory because information must be manipulated. Three aspects of attention were assessed: alerting, orienting, and executive attention. A modified version of the Attention Network Test (ANT) was used to measure alerting and orienting attention and a modified version of the Flanker task was used to assess executive attention. Alerting attention was measured by the relative change in MRT between alerting trials and neutral trials and orienting attention was measured by the relative change in MRT between orienting trials and alerting trials. The relative change in MRT between incongruent trials and congruent trials from the Flanker task was used as a measure of executive attention. For a detailed description of these measures, see Verburgh and colleagues. Consistency in information processing speed was examined using individual response time distributions derived from correct go trials of the Stop Signal task. Each participant executed 145 go trials, which allows reliable analyses of processing speed. The ex-Gaussian distribution model combines a normal distribution shape of individual reaction times with an exponential component on the right side of the distribution. With this model, Mu is calculated to determine the average speed of processing corrected for extreme slow responses. Furthermore, Sigma (fluctuations in speed of processing) and Tau (proportion of extreme slow responses, measuring lapses of attention) were calculated. ## Estimated full-scale IQ IQ was estimated by the Wechsler Intelligence Scale for Children III. Two subtests (Vocabulary and Block Design) were administered, correlating highly (*r* \>.90) with full-scale IQ. ## Body Mass Index Participants’ height was measured using a stadiometer, with the child standing against a wall without shoes. Weight was measured for each participant to the nearest 0.1 kg by a weighing scale (Soehnle White Sense). Body Mass Index (BMI) was calculated from weight (in kg) / height × height in meters. ## Physical activity and sedentary behaviour Involvement in physical activities and sedentary behaviour were assessed using a questionnaire consisting of 13 questions on physical activity (e.g., ‘How many days a week are you going to school walking or cycling?’) and six on sedentary behaviour (e.g., ‘How many days a week do you watch television?’). Participants were required to indicate how many days per week and how many minutes per day they participated in each of the activities listed. Included dependent variables were total minutes spent in: Organized sports, active transport, PE, outdoor play, TV-watching, computer use, as well as active gaming (e.g., Wii Sports). Adequate reliability and validity have been reported for this questionnaire. ## Procedure Data of the non-athletes were collected at elementary schools during regular school hours or immediately after school. Data of the elite soccer players and non-elite soccer players were collected at the soccer club during the competitive soccer season. Participants from both soccer player groups were tested prior to soccer training, the non-athletes were tested on a day without PE class. All participants were tested in a quiet room by trained assessors using standardized instructions. There were two sessions with a duration of approximately one hour for each individual participant. First, body height and body weight were measured, followed by the WISC III and the neurocognitive tasks in fixed order. ## Statistical analyses MATLAB was used to subtract Mu, Sigma, and Tau from the individual reaction times of the Stop Signal task. SPSS version 22.0 was used for all statistical analyses (SPSS IBM, New York, U.S.A). Five participants were removed from all analyses due to not attending the second test session, technical difficulties or not speaking fluent Dutch (N = 3 non-elite soccer players, N = 2 elite soccer player). For the 163 remaining cases, total missing data of demographic variables was less than 5% (N = 6 for IQ, N = 8 for the questionnaire on physical activity and sedentary behaviour in elite soccer players). Missing data were missing completely at random and were replaced by expectation maximization. Standardized scores were used in all analyses and for the VSTM backwards and Digit Span Backwards, van der Waerden transformations were applied as they were not normally distributed. Exploratory analyses on the nine neurocognitive measures showed only significant (and if so, small) correlations between some of the neurocognitive measures. This showed that each of the neurocognitive measures assesses unique and largely independent aspects of neurocognitive functioning. Possible group differences in physical activity, sedentary behaviour, age, BMI and IQ were tested using univariate analyses of variance (ANOVA) and were further explored with Tukey post hoc comparisons. Pearson correlations were performed to determine the possible relationship between the neurocognitive measures and age, BMI and IQ. If necessary, age, BMI and IQ were entered as a covariate in subsequent analyses. Next, following Field a series of multiple backward regression analyses were conducted to investigate the relationship between the measures of physical activity and sedentary behaviour: organized sports, active transport, PE, outdoor play, TV- watching, computer use, as well as active gaming, and the neurocognitive measures: inhibition (SSRT), short term memory (Digit Span Forwards, VSWM Forwards), working memory (Digit Span Backwards, VSWM Backwards), attention (Alerting, Orienting, Executive Networks), and information processing speed (Mu, Sigma, Tau). The alpha-level was Bonferroni-adjusted for the number of predictors (α =.05/7), resulting in p-values smaller than 0.007 (two-tailed) considered statistically significant. Results were expressed in terms of R-squared (R²) and standardized regression coefficients (β) with values of 0.10, 0.30 and 0.50, referring to small, medium and large effects, respectively. # Results ## Preliminary analyses Group characteristics and data on all neurocognitive measures are shown in Tables and. Total minutes of physical activity and sedentary behaviour per week were not significantly correlated (*r* =. -08, *p* =.31). Age was associated with all assessed neurocognitive functions (0.26 \> r’s \< 0.53.001 \> p’s \<.01) and was therefore included as covariate in all subsequent analyses. There were no significant relationships between both BMI and IQ and the neurocognitive measures, with two exceptions: IQ significantly correlated with both short term memory (r =.30, p \<.001) and working memory (r =.32, p \<.001). Because group differences were found for IQ, analyses were performed with and without IQ as covariate in regression analyses with short term memory and working memory as dependent variables. ## Regression analyses Collinearity statistics of the predictors yielded tolerance values between 0.72 and 0.95, with variance inflation factors between 1.1 and 1.9, indicating that the validity of the regression models was not threatened by multicollinearity. The multiple (backward) regression analyses revealed that time spent *in organized sports* was positively associated with inhibition (β = 0.25, p =.003, 95% CI 0.09–0.42, R² =.084), short term memory (β = 0.18, p =.006, 95% CI 0.05–0.31, R² =.082), working memory (β = 0.25, p \<.001, 95% CI 0.12–0.38, R² =.10), and lapses of attention (tau) (β = 0.32, p \<.001, 95% CI 0.15–0.50, R² =.11). Additional analyses with IQ as covariate for short term memory and working memory revealed the same significant results. The only difference was that for short term memory the Beta and effect size became larger, which could be explained by the somewhat lower IQ of the elite soccer player group, while the elite soccer players showed the best short term-memory scores. As we found elite soccer players to outperform both the non- elite soccer players and the non-athletes on inhibition and the high amount of sports the elite players participated in, we reran the regression analysis with SSRT as dependent variable while eliminating the group of elite soccer players. This analysis allows us to examine the possibility that the found relationship between organized sports and inhibition was due to superior performance of the elite soccer players. Results showed that time spent in organized sports was not associated with inhibition (β = -0.16, p =.38, CI. -0.52–0.19) when the elite soccer players were excluded, indicating that the results of the regression analysis with all participants were largely driven by the scores on SSRT of the elite soccer players. This interpretation was further supported by the non- significant Pearson correlations between time spent in sports and inhibition within the non-elite and non-athlete group (r =.03, p =.78) and within the elite soccer player group (r =.07, p =.60). As was also reported in, there were group differences on short term memory, working memory and lapses of attention. Therefore, we repeated the method described above for inhibition (SSRT) for those three measures. By excluding the elite soccer players, sports was still associated with better performance on short term memory (β = 0.09, p =.02, 95% CI 0.014–0.17, R² =.05) and working memory (β = 0.39, p \<.001, 95% CI 0.009–0.25, R² =.15). Sports was not associated with lapses of attention when the elite soccer players were excluded (β = 0.08, p =.15, 95% CI -0.003–0.018). *Outdoor play* was positively associated with working memory in the analyses with and without IQ as covariate (β = 0.19, p =.002, 95% CI 0.07–0.30, R² =.028, and β = 0.37, p = \<.001, 95% CI 0.24–0.49, R² =.10, respectively). Last, regarding sedentary behaviour, *computer use* was negatively related to inhibition (β = -0.27, p =.001, 95% CI -0.42–-0.12, R² =.024). The other measures of physical activity (PE and active transport) and sedentary behaviour (TV-watching and active gaming) were not associated with scores on any of the measured neurocognitive functions. # Discussion The present study addressed the relationship between both physical activity and sedentary behaviour and neurocognitive functioning in preadolescent boys. Results showed that time spent in organized sports was positively associated with short term memory, working memory, and lapses of attention. Moreover, time spent playing outdoors was also associated with working memory. In contrast, time spent at the computer or gaming was negatively associated with inhibition. ## Benefits of physical activity The observed positive relationships between organized sports and the neurocognitive measures: short term memory, working memory, and lapses of attention, as well as the positive relationship between outdoor play and working memory are highly important, as these neurocognitive functions are the key to daily life functioning, including academic achievement, autonomous behaviour, and quality of life. For example, working memory is a major mediator in academic achievement: It has been shown that working memory is highly predictive for reading and spelling achievement in school-aged children. Our findings receive support from previous studies that observed enhanced working memory after single bouts and regular sessions of physical exercise in healthy preadolescent children, suggesting that even short sessions of active outdoor play during school time may be beneficial for working memory. Concerning lapses of attention (Tau), it has been shown that short losses of attention may lead to educational problems. Interestingly, integrity in several important white matter tracts is positively associated with Tau. Moreover, there is some evidence showing that physical exercise may lead to improved white matter integrity in these brain areas. Our findings raise the intriguing possibility that physical activity might be a promising method to enhance short term memory and working memory. When we excluded the group of elite soccer players from the analyses because they showed better performance on the inhibition, short term memory, working memory and attentional tasks, there still was a significant positive relationship between short term memory and working memory, and participation in soccer. In contrast, a recent study showed that soccer players outperformed non-athletic children on a psychomotor vigilance task (PVT) and had better cardiovascular fitness. Interestingly, no relationship between cardiovascular fitness and the PVT was found, which may be due to the possibility of the PVT appealing less to executive functions, or methodological limitations (i.e. small sample size) of the study. Another possible explanation of the findings of Ballester and colleagues is the ‘cognitive component skills theory’ suggesting innate excellent cognitive skills in elite athletes. Indeed, several studies show already differences between elite and sub-elite soccer players at a very young age. However it is still topic of debate whether the excellent cognitive skills of elite youth athletes result from training (e.g. a result of many training hours, high quality training facilities and coaches) or are innate. Our study adds to this debate by showing that especially on motor inhibition and lapses of attention, the elite soccer players showed superior performance. All in all, we believe that the current study together with the findings of our meta-analysis and other recent studies, provides enough support for the recommendation to include more sports in the school curriculum, for instance by intensifying PE classes and by encouraging active play during recess at school. Nevertheless, we emphasize the need for further research on the relationship between sports and neurocognitive functioning and moderators such as MVPA and, cardiovascular fitness. Other possible moderating factors such as improved motor skills (needed in many neurocognitive tasks) or motivational aspects, may play a role as well in the relationship between sports participation and neurocognitive functioning. ## Costs of sedentary behaviour Results indicated that more computer use including using the internet and gaming, may lead to poorer motor inhibitory skills. However, direction of the finding and causality are unknown. It may also be that poor inhibitory control leads to more computer use as was found in a study of Little and colleagues, who reported that diminished inhibitory control might underlie excessive computer use and even game addictions. Either way, our findings are relevant to health issues, because of the worldwide increase of sedentary behaviour among youth and the importance of inhibitory control for daily life. ## Limitations and future research One limitation is that measures of physical activity and sedentary behaviour were based on self-report. While we acknowledge the limitations of this method, self-report measures have been shown to provide valid measures of time spent in sedentary behaviour, and unlike other measures (such as pedometers or accelerometers) are able to provide insight into the types of activities (e.g., gaming, doing homework or watching TV). For future research, it is recommended to use objective measurements next to self-reports, such as accelerometers to objectify measurements of physical activities and sedentary behaviour. In addition, as technology and the daily use of new devices grow rapidly, there is a high need for validated but up-to-date questionnaires for measuring sedentary behaviour in children. For instance, the questionnaire we used was developed in 2007, in a time in where tablets, smartphones or online shopping nearly existed. Second, the cross-sectional design used in the present study prevents us from drawing conclusions about the causality underlying the findings. Future research should focus on high-quality RCT’s to draw conclusions about causality and the optimal duration, frequency and intensity of physical activity in preadolescent children to enhance neurocognitive functioning. More specifically, there is no consensus in the literature on beneficial effects of chronic exercise, MVPA and/or cardiovascular fitness in children on neurocognition and executive functioning in particular, which is complicated by the use of a large variety of tasks including tasks partly appealing to executive functions. Therefore, thoroughly designed studies are required to draw firm conclusions about the effects of exercise in youth. Last, as the present study only included boys, our results may not generalize to girls. In addition, many studies have shown that girls become increasingly less active during adolescence, which in turn has negative effects on health. ## Conclusion The present research complements and extends previous research on benefits of physical activity, and costs of sedentary behaviour, on neurocognitive functioning. The gains of physical activity include key aspects of cognition that are likely to be relevant in any situation that requires basic levels of short-term memory, working memory, and attention (e.g., traffic, conversation, group tasks). The costs of sedentary behaviour seem to include inhibition. Scientifically, the findings may give direction to experimental research needed to unravel cause and effect. Societally, although only part of the puzzle between physical activity, sedentary behaviour, and neurocognition is addressed, we recommend interventions to promote of physical activity and outdoor play to enhance both health *and* neurocognitive functioning in future generations. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** JO ES LV PvL. **Formal analysis:** LV JO. **Funding acquisition:** JO. **Investigation:** LV. **Methodology:** JO LV. **Project administration:** LV. **Resources:** LV JO ES. **Software:** LV. **Supervision:** JO. **Validation:** LV JO PvL. **Visualization:** LV. **Writing – original draft:** LV. **Writing – review & editing:** LV JO ES PvL.

# Introduction HIV vaccines and microbicides hold promise for preventing the acquisition of HIV-1 infection, but successful design of such agents requires a clear understanding of the mechanisms of HIV-1 transmission at the initial site of infection. Most HIV-1 infections occur during heterosexual intercourse, and women are more likely to become infected than men. Initial exposure to HIV-1 during sexual transmission occurs in the genital tract; however, little is known about HIV-1-specific immune responses at this site, as well as the effect of HIV-1 on mucosal immunity. Human leukocyte antigen (HLA)-G is a non-classical major histocompatibility class I protein, characterised by limited polymorphism and tissue-restricted distribution. HLA-G is expressed as membrane-bound (HLA-G1, -G2, -G3 and -G4) and soluble (HLA-G5, -G6, -G7) isoforms as a result of alternative splicing. The major isoforms present in the plasma are soluble HLA-G (sHLA-G)-1 and -G5 which are generated by shedding or proteolytic cleavage of membrane-bound HLA-G1 isoform and by secretion of a soluble form, respectively. Under physiological conditions, sHLA-G levels correlate with gender and HLA-G genetic polymorphisms. The level of sHLA-G is higher in women than in men. Healthy individuals carrying the HLA-G\*01:01:03 and HLA-G\*0105N alleles have lower plasma sHLA-G levels than subjects carrying the more frequent HLA-G\*01:01:01 allele. In addition, individuals with the latter allele have lower plasma sHLA-G levels than those with the HLA-G\*01:04 allele. Polymorphisms in the 3′-untranslated region (3′UTR) can also affect the production of HLA-G molecules. The presence of a 14-bp sequence insertion in HLA-G 3′UTR has been associated with lower levels of sHLA-G in serum of healthy subjects. HLA-G expression can be induced during pregnancy, antiretroviral (ART) therapy, and in pathological conditions such as autoimmune diseases, cancers, transplantations, and viral infections. HLA-G molecules inhibit the activity and mediate apoptosis of natural killer (NK) cells and cytotoxic CD8⁺ T cells, as well as CD4⁺ T cell proliferation and induce tolerogenic dendritic cells (DC) and regulatory T cells. The immunosuppressive properties of HLA-G might contribute to the susceptibility to HIV-1 infection. Recent studies have shown that HLA-G polymorphisms are associated with altered risks of heterosexual acquisition – and vertical transmission, of HIV-1. Plasma sHLA-G expression, at the protein level, was recently associated with increased risk of HIV-1 infection and more rapid disease progression. However, initial exposure to HIV-1 during sexual transmission occurs in the female genital tract and no data are available on the possible association between genital HLA-G expression and susceptibility to HIV-1 infection. We have therefore measured the genital levels of sHLA-G in HIV-1-infected and HIV-1-uninfected female commercial sex workers (CSWs), as well as HIV-1-uninfected non-CSW women at low risk for exposure to investigate whether sHLA-G expression is associated with HIV-1 infection. # Methods ## Study population Female CSWs were enrolled through a dedicated sex worker clinic in Cotonou, Benin and were divided into two groups: HIV-1-uninfected CSWs (n = 52) and ART- naïve HIV-1-infected CSWs (n = 44). The HIV-1-uninfected non-CSW control subjects at low risk for exposure (n = 71) were enrolled from a general health clinic in Cotonou. Women were invited to participate in the study as they attended clinics. Women were excluded from the study if \<18 years old, menstruating, or pregnant. At enrolment, participants were asked to answer a questionnaire about demographic information, sexual behaviour, duration of prostitution, number of sex partners, condom use, vaginal douching practices, and reproductive history. Each participant underwent a genital examination by a physician. Vaginal specimens were obtained for diagnosis of candidiasis and bacterial vaginosis by microscopic examination. Endocervical swabs were obtained to test for *Neisseria gonorrhoeae* and *Chlamydia trachomatis* infection using BD ProbeTec ET system (Strand Displacement Assay, Becton Dickinson, Heidelberg, Germany). Peripheral blood was taken for HIV, HLA-G and CCR5 genotype analyses. Plasma and serum were kept frozen at - 80°C until use. HIV-1 positivity was defined by the presence of HIV-1 antibodies tested with Vironostika HIV Uni-Form II Ag/Ab (Organon Teknika, Boxtel, The Netherlands). Non-reactive samples were considered HIV-seronegative, whereas reactive samples were tested with Genie II HIV-1/HIV-2 (Bio-Rad, Hercules, CA). Genie II dually reactive samples (to HIV-1 and HIV-2) and discordant samples (Vironostika reactive/Genie II non-reactive) were further tested by INNO-LIA HIV I/II Score (Innogenetics NV, Technologiepark 6, Gent, Belgium). Viral loads were determined in the plasma of all HIV-1 infected CSWs using VERSANT HIV-1 RNA 3.0 Assay (bDNA) (Siemens Medical Solutions Diagnostics, Tarrytown NY). DNA samples were genotyped for the CCR5 32-bp deletion allele and all women were found to be homozygous for the wild- type allele. ## Mucosal sample collection and preparation Cervicovaginal lavage (CVL) samples were obtained from all study participants by a physician, using a 10-ml syringe filled with sterile phosphate-buffered solution and aimed directly into the cervical os. CVL fluids were then collected, transferred immediately into 20 ml of RPMI-1640, kept on ice, and processed within 1 hour. CVL samples were centrifuged at 1500 r.p.m. for 10 min to remove cells and debris, and supernatants were stored at −80°C until shipped on dry ice to Montréal, Canada. CVL samples were concentrated with Amicon Ultra-15 5 kDa (Millipore, Billerica MA) prior to sHLA-G measurement. ## Soluble HLA-G measurements and HLA-G genotyping sHLA-G CVL levels were measured using the Human sHLA-G Immunoassay kit (Alexis Biochemicals, San Diego, CA, USA), which allows simultaneous detection of HLA-G1 and -G5 soluble proteins without discrimination. The final concentration of sHLA-G in the CVL sample was determined as follows: concentration obtained with the sHLA-G Elisa assay (units per ml)/(CVL concentration factor)×total CVL volume prior to concentration. HLA-G alleles were determined by direct DNA sequencing analysis of the nucleotide regions encompassing HLA-G exons 2–4 and using purified DNA from blood samples as described previously. HLA-G 3-UTR polymorphisms were determined according to the protocol previously described by. ## Statistical analysis Statistical analysis was performed using the GraphPad PRISM 5.0 for Windows (GraphPad Software, San Diego, CA). One-way analysis of variance and Chi-square tests were used to assess the significance of the associations between continuous and categorical variables across all study groups. Comparisons of continuous and categorical variables between two groups were assessed by the Mann-Whitney *U* and Chi-square or Fisher exact tests, respectively. Spearman's rank test was used to determine correlations between continuous variables. Multiple logistic regression analysis was used to define independent predictors identified as significant in the crude analysis. Odds ratio (OR) and 95% confidence interval (CI) were calculated with the exact method. Differences were considered significant at P≤0.05 or P≤0.015 when comparing two or three groups, respectively. ## Ethics statement Written informed consent was obtained from all subjects who participated in the study and the investigation reported in this paper was approved by the Comité National Provisoire d'Éthique de la Recherche en Santé in Cotonou and the CHUM Research Ethics Committee. # Results Sociodemographic and clinical characteristics of the study population are described in. These data were collected to address the issue of confounding variables for risk of HIV-1 infection. The three groups were similar with respect to age, days from last menses, vaginal douching, and the presence of vaginal candidiasis. The HIV-1-infected CSWs were more likely to have a bacterial vaginosis (P = 0.003) than the HIV-1-uninfected non-CSWs. The HIV-1-unifected non-CSWs, were less likely to have *Chlamydia trachomatis* or *Neisseria gonorrhoeae* genital infections than the HIV-1-uninfected (P = 0.027) and HIV-1-infected (P = 0.022) CSW groups. The average number of clients was higher in HIV-1-uninfected CSWs than in HIV-1-infected CSWs (P = 0.044), whereas the duration of sex work, and condom use were equivalent between the two CSW groups. HIV-1-infected CSWs had significantly higher levels of sHLA-G in their CVL samples (94±145 units/ml) than did the HIV-1-uninfected CSWs (35±53 units/ml; P = 0.009) and the HIV-1-uninfected non-CSW women (26±53 units/ml; P = 0.0006). There was no significant correlation between the HIV-1 plasma viral load and the sHLA-G level in the CVLs of the HIV-1-infected CSWs (r² = −0.162, P = 0.344). Since sHLA-G expression has been associated with HLA-G polymorphism, we looked at the distribution of sHLA-G levels, either between study groups or in the total population, according to the HLA-G genetic variants. The HLA-G\*01:01:02 genotype, in the heterozygous or homozygous states, was associated with increased expression of genital sHLA-G in HIV-1-infected CSWs compared with those in both the HIV-1-uninfected CSW (P = 0.051) and non-CSW (P = 0.002) groups. In the overall population, women carrying the HLA-G\*01:04:04 heterozygous or homozygous genotypes expressed the highest levels of genital sHLA-G molecules when compared with those expressed by women harbouring other genotypes (P = 0.038). However, there was no significant association between HLA-G alleles and sHLA-G levels within the three groups taken separately. Because HLA-G polymorphism can also be associated with HIV-1 infection, we looked at the distribution of the HLA-G genetic variants among the study groups and found no significant association between HLA-G alleles and HIV-1 infection (data not shown). The presence of bacterial vaginosis could potentially affect the genital level of sHLA-G molecules and since the rate of bacterial vaginosis was significantly higher in the HIV-1-infected CSWs, we investigated the possible correlation between sHLA-G levels and the presence of bacterial vaginosis. We found that the expression of sHLA-G in genital samples was significantly associated with bacterial vaginosis among the HIV-1-infected CSWs (P = 0.035). When adjustment was made for all significant variables found in the crude analysis (HIV-1 infection, bacterial vaginosis, HLA-G\*01:01:02 and HLA-G\*01:04:04 genotypes), the expression of sHLA-G in the genital mucosa remained significantly associated with both HIV-1 infection (OR: 3.0, 95% CI = 1.17–7.53, P = 0.02) and bacterial vaginosis (OR 3.4, 95% CI = 1.10–10.5, P = 0.03). # Discussion High level of sHLA-G in the genital mucosa is associated with HIV-1 infection in Beninese CSWs. In the present study, we have carefully controlled for potential confounding factors that could influence HLA-G expression such as gender, pregnancy, ART therapy, and HLA-G polymorphism. All study participants were ART- naïve nonpregnant women. The HLA-G\*01:01:02 and HLA-G\*01:04:04 genotypes were significantly associated with sHLA-G expression in the crude analysis but these associations disappeared after adjustment was done for HIV-1 infection. In contrast to previous studies, HLA-G polymorphism was not associated with risk of HIV-1 infection among the Beninese CSWs. The relatively small number of subjects analysed in each groups have limited the power of the present study to reproduce previous findings. We have previously measured the level of sHLA-G in the blood of these women and found that HIV-infected CSWs had lower plasma levels when compared to HIV- uninfected CSWs and non-CSWs. This is in sharp contrast with that found in the genital mucosa of these women. The discordance in the production of sHLA-G between the two compartments may depend on local factors such as immune cells, micro-organisms and derived products that could affect sHLA-G expression. sHLA-G plays a crucial role in the regulation of both innate and adaptive immunity by modulating the function of DC, NK and T lymphocytes. These effects depend on interactions of HLA-G molecules with inhibitory receptors expressed on myeloid cells (immunoglobuline-like transcripts (ILT)-4), on myeloid and lymphoid cells (ILT-2) and on NK cells (killing inhibitory receptor (KIR)-2DL4). The outcome of the immune response may therefore vary according to the specific interactions of sHLA-G with the different types of cells and receptors. Interaction of sHLA-G with ILT-2 receptor on DC and NK cells decreased the release of interferon (INF)-gamma and increased the production of interleukin (IL)-10 and transforming growth factor (TGF)-beta. IL-10 has been shown to induce HLA-G expression and HLA-G can also stimulate IL-10 expression in peripheral blood monocytes. Triggering ILT-4 by sHLA-G induces tolerogenic DC and T regulatory cells. On the other hand, interaction of sHLA-G with KIR2DL4 receptor on peripheral blood monocytes and NK cells promotes the production of pro-inflammatory cytokines and chemokines,. We have previously measured the cytokine and chemokine expression patterns in the genital samples of our study subjects and found that HIV-1-infected CSWs had significantly higher levels of IFN-gamma tumor necrosis factor (TNF)-alpha, monocyte chemotactic protein (MCP-3/CCL7) and monokine induced by IFN-gamma (MIG/CXCL9) compared with those in both the HIV-1-uninfected CSW and non-CSW groups. The same observations were made for IL-1 beta and IL-8 (data unpublished). High level of IL-1 beta and TNF-alpha in the female genital tract has been associated with enhanced HIV-1 shedding at this site. The inflammatory response observed in the genital mucosa of HIV-1-infected women may promote the recruitment, differentiation and activation of immune cells, which act as targets favouring viral replication and viral dissemination at the initial site of infection. As to whether sHLA-G is directly involved in the induction of such mucosal inflammation via its interaction with KIRD2L4 on monocytes and NK cells in the female genital tract remains to be confirmed. Although the genital mucosa levels of sHLA-G correlate significantly with those of the cytokines and chemokines in the HIV-1-uninfected groups, these correlations were not significant in the HIV-1-infected CSW group. Thus, in the absence of HIV-1, genital levels of the immunosuppressive sHLA-G molecules and pro-inflammatory cytokines and chemokines are low and correlate to maintain mucosal homeostasis. Conversely, in the presence of HIV-1, there is an aberrant and independent production of both factors in the female genital tract that may reflect a viral strategy of immune piracy, allowing for the simultaneous production of chemokines/cytokines to recruit and activate HIV-1 target cells and sHLA-G to induce immune tolerance towards HIV-1. Interestingly, the increased level of sHLA-G in genital samples was also significantly associated with the presence of bacterial vaginosis. Although HIV-1-infected CSWs had higher levels of sHLA-G and were more likely to have a bacterial vaginosis than the HIV-1-uninfected non-CSWs, the association between sHLA-G levels and bacterial vaginosis remained significant after adjusting for HIV-infection. This suggests that genital sHLA-G level is independently associated with both bacterial vaginosis and HIV-1 infection. Bacterial vaginosis is an established risk factor for HIV infection. It has been suggested that bacterial vaginosis increases risk of HIV infection by inducing a clinical or subclinical mucosal inflammatory response, recruiting target cells and breaching of intact cervico-vaginal mucosa. Indeed, bacterial vaginosis has been associated with increased levels of IL-1 beta, IL-6, IL-8, IL-10 and TNF-alpha, RANTES (CCL5), macrophage inflammatory protein (MIP-1 alpha/CCL3) and MIP-1 beta (CCL4) in genital samples. However, bacterial vaginosis was not associated with the production of these cytokines and chemokines in the genital tract of the Beninese women ( and). Altogether, these results suggest that in the context of HIV-1 infection, sHLA-G expression in the female genital tract is a complex process modulated by many factors such as HIV-1, bacterial vaginosis HLA-G genotypes, and cytokine/chemokine expression patterns, which may all contribute to an immunological environment promoting viral replication and escape from the mucosal immune response. # Supporting Information We are indebted to N. Geraldo, A. Gabin, C. Assogba and C. Agossa-Gbenafa for their clinical expertise, to M. Massinga-Loembe, G. Ahotin, L. Djossou, and E. Goma for their technical assistance and to G. Batona and other field workers who helped with recruitment of commercial sex workers. [^1]: Conceived and designed the experiments: VT JL JP MR. Performed the experiments: VT JL. Analyzed the data: VT MR. Contributed reagents/materials/analysis tools: A-CL MDZ KRF JP MA. Wrote the paper: VT MR. Lead investigator of this study: MR. [^2]: The authors have declared that no competing interests exist.

# Introduction Leukemia, lymphoma and multiple myeloma are the three main types of highly heterogeneous hematological malignancies that are derived from myeloid and lymphoid cell lineages. Acute myeloid leukemia (AML) is characterized by abnormal expansion of immature myeloid cells and their accumulation in the bone marrow and blood, interfering with normal cellular growth. AML is a highly aggressive cancer with poor prognosis. It is also the most common type of acute leukemia in adults. Treatment strategies and success rates vary depending on many factors, including the subtype of AML, prognostic factors, age and general health status of the patient. Standard treatment regimens based on patient stratification include the combination of chemotherapeutics such as Cytarabine, Daunorubicin and Etoposide with or without radiotherapy. However, high heterogeneity of clinical outcomes in AML patients suggests that current classifications fail to distinguish patient subgroups sufficiently. A not so well studied protein network in the context of AML is the Renin- Angiotensin System (RAS). RAS is composed of several gene products which play a critical role in regulating blood pressure, renal vascular resistance and the fluid/electrolyte balance. The idea of a local RAS operating independent of the circulating RAS was brought into light by demonstrating localized RAS elements in organs other than liver (angiotensinogen), kidney (renin) and lung (ACE). Localized RAS elements were found in many organs such as the brain, blood vessels and heart. It is predicted that locally produced angiotensins have important homeostatic functions and may contribute to local tissue dysfunction and diseases. The presence of local RAS specific to the hematopoietic bone marrow microenvironment was reported for the first time in 1996. Major RAS molecules have been identified in the bone marrow microenvironment, such as renin, angiotensinogen, angiotensin receptors and angiotensin converting enzymes (ACEs). Locally active bone marrow RAS affects important stages of physiological and pathological blood cell production through autocrine, paracrine and intracrine pathways. Local bone marrow RAS peptides control the development of hematopoietic niche, myelopoiesis, erythropoiesis, thrombopoiesis and other cellular lineages. Local RAS is also active in the primitive embryonic hematopoiesis phase. The presence of renin, ACE, angiotensin II (Ang-II) and angiotensinogen in leukemic blast cells has been demonstrated, and local bone marrow RAS has been shown to play a role in the development of neoplastic malignant blood cells. Establishing a role for genes involved in the development and biology of cancers, as prognostic and chemotherapeutic markers, is one of the most effective and successful approach used in the classification of malignancies. Thus, here we aimed to define AML subgroups based on expression of RAS genes. We also aimed to test if the resulting tumor subtypes differ in their responses to drugs and to demonstrate distinct prognostic profiles. # Materials and methods ## *In silico*  ### Datasets Cancer Genome Project (CGP) gene expression data (E-MTAB-783) was downloaded from ArrayExpress website (<https://www.ebi.ac.uk/arrayexpress/>), and drug screening data was downloaded from the CGP database. Microarray dataset GSE12417, corresponding to AML patients, was downloaded from the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) database. The training cohort of 163 patients in GSE12417 was used for our analyses since this was the same microarray platform as the one used for CGP. ### Data normalization and variance analysis CGP gene expression data was normalized by the RMA method using the BRB-Array Tools software. In order to choose the RAS genes that would be used in real-time PCR for validation studies, we first aimed to choose the most variable genes that would likely give detectable fold differences *in vitro* by PCR. Analysis of variance was performed using all 39 probesets corresponding to the 25 genes in the RAS and genes with at least 0.8 of variance. Above these thresholds, the mean expression was at least 5.5 and the log fold difference between min to max was above 3 for all probesets. Thus, nine probesets corresponding to eight genes (*CTSG*, *CPA3*, *AGT*, *ANPEP*, *IGF2R (two probesets)*, *RNPEP*, *ATP6AP2 and CTSA*) were selected to be used in further analyses. ### IC50 calculation methods In order to calculate drug response parameters such as IC50, EC50, activity area and Amax, the growth rate of the cells were depicted as a function of drug concentration by being modeled with non-linear logistic regression as explained in De Lean *et*. *al*, which is also reported in NIH/NCGC assay guidelines. While the non-linear logistic regression function used to model data is used widely for cytotoxicity calculations, here for the first time we used six different versions of this function and selected the one with the lowest standard error rate among all for the calculation of cytotoxicity values. We name this approach the 6-model (6M). Thus, six different models were derived from the following non-linear logistic regression function: $$Y = \left( a - d/\left( 1 + (X/c)^{b} \right) + d \right)$$ where *Y* is the percent growth of the cells, *X* is the arithmetic drug concentration, *a* is the percent growth of the cells when the cells are not treated with the drug, *d* is the percent growth of the cells for infinite dose, i.e. a dose for which there is no additional effect when increased, c is the dose corresponding to percent growth exactly between *a* and *d*, and *b* is the Hill slope factor that is used to define the steepness of the curve fitted. The following are the conditions required for the generation of 6-models: 1. **3-Parameter model:** Curves were fitted without using Hill slope factor *b*. 2. **3-Parameter Top 100 model:** Curves were fitted without using Hill slope factor *b* and with *a* = 100. 3. **3-Parameter Bottom 0 model:** Curves were fitted without using Hill slope factor *b* and with *d* = 0. 4. **4-Parameter model:** Formula is used as it is. 5. **4-Parameter Top 100 model:** Curves were fitted with *a* = 100. 6. **4-Parameter Bottom 0 model:** Curves were fitted with *d* = 0. Six different drug response parameters are calculated out of the fitted curves as follows: - **IC50:** Value of *X* when *Ŷ* = 50% - **IC90:** Value of *X* when *Ŷ* = 90% - **IC95:** Value of *X* when *Ŷ* = 95% - **EC50:** Value of *X* when *Ŷ* = *a*+*d* - **Amax:** *a* − *d* - **Activity Area:** Σ*ŶX*, (sum of *Ŷ*s for each 0.01 increment of *X* fitted), where *Ŷ* is the predicted value of *Y* by the curve fitted. With the 6M approach we recalculated IC50 values that were also included in the raw CGP data for the 17 AML cell lines treated with four drugs (ATRA, Cytarabine, Etoposide and Doxorubicin) using an in-house R script “*SixModelIC50 V3*.*r*” (<https://github.com/muratisbilen/6-Model_IC50_CalculationV3.git>). These drugs were selected as we obtained AML chemotherapy treatment protocols from the Department of Hematology, Hacettepe University and compiled a list for all drugs in these protocols. Among these only ATRA, Cytarabine, Etoposide and Doxorubicin were present in the CGP database. We referred to the recalculated IC50 data as 6M IC50 and performed a Pearson r correlation analysis between CGP IC50s and recalculated 6M IC50s to test the compatibility. In addition, IC50 values were calculated using the 6M approach on the data obtained from in vitro analysis in which nine AML cell lines were treated with Doxorubicin and Etoposide. ### Linear regression analyses We performed correlation analysis between expression values of the eight genes and drug data (both CGP IC50 data and 6M IC50 data) individually. To identify if multiple genes can be used to better identify the relationship between gene expression and drug sensitivity data, linear regression analyses were performed using the Minitab 17 software (<https://www.minitab.com>). Seventeen AML cell lines from the CGP database were either randomly divided into two groups, the discovery group (12 cell lines) and the test group (five cell lines), or chosen manually so that the sensitivity range of cells in both groups spanned as large variance as possible. To generate a linear regression model for each drug (ATRA, Cytarabine, Etoposide, Doxorubicin), IC50s of the discovery cell line group obtained either from CGP or recalculated as 6M IC50, and expression of the eight RAS genes which were selected from variance analysis, were used as predictors. As a measure of the response variable variation explained by each linear regression model, we used the adjusted (adj.) R² values. To test consistency of the linear regression models generated with the eight genes, we replicated the random divison of groups ten times and reported the average of the adjusted R². Furthermore, to identify a minimal gene list for the prediction of chemosensitivity, the discovery group was used to fit a model explaining the drug response using “best subsets” function of the software, which runs all possible regression models with one variable, two variables and so on, based on a list of predictors, enabling the user to choose a smaller set of predictors that can explain the response. The subset with the highest R² (adj.) was selected as the best model. Regression formula of the best models (*y* = ±*a* + \[*n*1 × *x*1\] ± \[*n*2 × *x*2\] ± \[*n*3 × *x*3\] ± \[*n*4 × *x*4\]…) were applied for the test group of each drug. In the regression formula *y* (predicted IC50 values) were calculated where *a* and *n* are the constant values, *x*: gene expression values of the 12 cell lines in the discovery group. Also, the goodness of fit measure Sy.x were computed by Graphpad. Sy.x is a standard deviation of the residuals that here has been used to describe the difference in standard deviations of CGP IC50 and 6M IC50 versus predicted IC50s. It is a goodness-of-fit measure used to show how well our predicted IC50s fit with CGP and 6M IC50 values. All the correlations were calculated with Graphpad software as Pearson’s r and p values. ### Hierarchical clustering analysis Cluster 3.0 (<http://bonsai.hgc.jp/~mdehoon/software/cluster/software.htm>) and Java Treeview (<http://jtreeview.sourceforge.net/>) software were used for hierarchical clustering analysis with mean standardized gene expression values for each dataset. Hierarchical clustering was performed by clustering both genes and arrays using Euclidian distance as similarity metric and complete linkage as clustering method. ### Gene sets enrichment analysis—GSEA Gene set enrichment analysis was performed using the GSEA guideline (<https://www.gsea-msigdb.org/gsea/index.jsp>). Briefly, dataset E-MTAB-783 has 22277 probesets IDs and these were collapsed into 13321 genes. For genes with more than one probeset, one with the highest expression was selected. “C5_all Gene ontology v6.1 database” was used for the analysis which has gene sets that contain genes annotated by the same GO term. We used default filtering criteria in GSEA for gene set sizes, which includes genesets with sizes between 15–500. After applying this filter, analysis was performed for 4081 gene sets. ### Mutation analyses Mutation data of AML cell lines was downloaded from Genomics of Drug sensitivity in Cancer database (<https://www.cancerrxgene.org/downloads/bulk_download>). 14 out of the 17 AML cell lines used in our analyses were available. Seven genes which were mutated in at least three AML cell lines were analyzed further. ## In vitro ### Cell lines and cytotoxicity experiments HEL92.1.7 (2111706), and QIMR-WIL (86030601) cell lines were purchased from Sigma Aldrich (St. Louis, Mo., USA), KASUMI-3 (CRL-2725), GDM-1 (CRL-2627) and CESS (TIB-190) cell lines were purchased from ATCC (Virginia, USA) and P31/FUJ (JCRB0091), NOMO-1 (IFO50474), KASUMI-1 (JCRB1003) and SKM-1 (JCRB0118) cell lines were purchased from JCRB Cell Bank (Osaka, Japan). Cell lines were authenticated by manufacturers, all cell lines were morphologically checked by microscope and routine mycoplasma testing was performed by PCR. HEL92.1.7, GDM-1, CESS, P31/FUJ and NOMO-1 were cultured and maintained in RPMI-1640 medium (Sigma-Aldrich, R0883 (St. Louis, Mo., USA)) supplemented with 10% fetal bovine serum (FBS) (Sigma-Aldrich, F6178 (St. Louis, Mo., USA)), 1% penicillin- streptomycin (Sigma-Aldrich, 11074440001 (St. Louis, Mo., USA)), and 1% 200 mM L-glutamine (Sigma-Aldrich, G7513 (St. Louis, Mo., USA)). KASUMI-1, SKM-1 and KASUMI-3were cultured in RPMI-1640 medium but with 20% FBS. QIMR-WIL was cultured in DMEM medium (Sigma-Aldrich, D6546 (St. Louis, Mo., USA)) but with 10% FBS, 1% penicillin-streptomycin, and 1% 200 mM L-glutamine. All cell lines were cultured at 5% CO² and 37 °C in a humidified incubator. Doxorubicin (D1515) and Etoposide (E1383) were purchased from Sigma-Aldrich (St. Louis, Mo., USA) and were dissolved in DMSO. Cell viability was measured using CellTiter-Glo reagent (G7572, Promega, Fitchburg, Wisconsin, USA). 7000 cells/well in 90 μl medium were plated in each well of a 96-well plate. Cells were treated with six different concentrations of Doxorubicin or Etoposide separately (20, 10, 2, 1, 0.2, 0.1 μM). After 72 hours of drug treatment, cells were treated with CellTiter-Glo reagent and the luminescence signal was then recorded with a microplate luminometer (Turner Designs, CA, USA). All drug treatment experiments were repeated three times. Growth percentages were calculated for each drug and cell line, and cytotoxicity values were calculated using the 6M approach. ### qRT-PCR *AGT*, *ANPEP*, *ATP6AP2*, *CPA3*, *CTSA and IGF2R* genes’ expression was quantified using SYBR ^™ Green master mix (Bio-Rad, \#1725150, (USA)). PCR reactions were run under cycling conditions according to manufacturer’s instructions. GAPDH was used as a reference gene in all reactions. qRT-PCR relative gene expression data was calculated using ddCT method. Using qRT-PCR relative gene expression data, predicted IC50 values were calculated with the formulas generated by linear regression analyses of *in silico* data using qRT-PCR based expression values as predictors. Primers used in this study are shown in. GAPDH was used as endogenous control. ## Clinical data validation ### Log Rank with Multiple Cutoffs (LRMC) and survival analysis In regression analysis, four formulas were generated for Doxorubicin and Etoposide using both CGP and 6M IC50 data. *IGF2R*, *CTSA*, *ATP6AP2* are common in three of the four formulas except for 6M IC50 data for Etoposide. Therefore, these genes were chosen to test relationships with clinical outcome. Clinical data were obtained from the training cohort of the GSE12417 dataset (AML Cooperating Group 1999). In the AMLCG 1999 cohort, patients were treated with TAD: Thioguanine, Cytarabine and Daunorubicin, or HAM: Cytarabine, Mitoxantrone protocols followed by the TAD protocol. We used an in-house R script (<https://github.com/muratisbilen/LRMC.git>) (Log Rank Multiple Cutoff, LRMC) by which log-rank test-based p-values associated with hazard ratio (HR) could be obtained using all possible cutoff values representing each sample in a given dataset and best cutoff is selected as in. Using this approach, we selected best cutoffs for *IGF2R*, *ATP6AP2 and CTSA* genes to be used for clinical correlation studies and Kaplan-Meier plots. Patients with gene expression lower than cutoff, for each gene individually, were labeled as ‘Low’ (low expression) and higher than cutoff were labelled as ‘High’ (high expression). Univariate cox regression analyses were performed and Kaplan-Meier graphs were drawn using SPSS Statistics 19 (IBM, Chicago, IL, USA). Additionally, the expression of all these three genes (*IGF2R*, *CTSA* and *ATP6AP2)* was evaluated together as good and bad prognostic groups. Patients were grouped as “Good” if they have high expression levels of IGF2R and CTSA and low expression levels of ATP6AP2 defined by expression value cutoffs in previous analysis. Rest of the patients were grouped as “Bad”. Then Kaplan Meier survival analysis was performed for these groups. # Results ## Discovery of RAS drug sensitivity biomarker genes The RAS consists of the 25 genes, corresponding to 39 probesets in Affymetrix HG-U133A, a microarray platforms used in the Cancer Genome Project (CGP). For the 17 AML cell lines, both drug cytotoxicity and gene expression data are available in the CGP database. We focused only on genes which showed high variation in expression for further validation and therefore, selected nine probesets (eight genes) as described in the methods section. We recalculated IC50 values using the 6M approach applied to raw CGP cytotoxicity data. Using Pearson correlation we observed strong correlations between CGP IC50 and 6M IC50 for drugs widely used in AML (Cytarabine, Etoposide, Doxorubicin) but not for ATRA. To identify biomarkers of chemosensitivity, we calculated Pearson correlation between gene expression and IC50 values obtained from CGP and generated by 6M approach. We thus identified six gene/drug cytotoxicity correlations which were significant with CGP IC50 values, and seven significant correlations with 6M IC50 values. Four gene/drug associations were common to both analyses. Linear regression analysis was then performed to test whether the combined expression analyses of genes could correlate better with drug sensitivity data or not. Thus, we generated discovery and test groups. Each group include a wide range of cell line IC50 values as possible. Linear regression models for drug sensitivity prediction were generated for the discovery group (12 cell lines) using expression data of highly variant eight genes and IC50 values obtained from CGP and 6M IC50 of four drugs in Minitab 17. Then, obtained results tested with the validation group (five cell lines). The models generated with combined expression analyses of the eight genes resulted in high R² (adj) values for Etoposide, Doxorubicin and Cytarabine but no model could be generated for ATRA. As independent datasets with drug sensitivity data for these compounds do not exist, we utilized a cross- validation method to test the robustness of the proposed models by generating the discovery and test groups 10 times, with 12 and five cell lines, respectively. The average of 10 R² values generated from discovery groups was calculated for both CGP and 6M IC50s. Our results showed that the 10 random models of sensitivity to Doxorubicin had an average R² above 85% for both CGP and 6M IC50s, but R² decreased slightly for models of sensitivity to Etoposide while R² values highly decreased for models of sensitivity to Cytarabine when compared to those generated for cell lines that were manually selected. We therefore, focused on Doxorubicin and Etoposide for further analyses. We then aimed to identify the minimal number of genes that needed to be included in combinations into the models that would give the highest correlation using the ‘best subsets function’ of Minitab. The software selected three genes/probesets for Doxorubicin when either CGP and 6M IC50 values were used and, four and five genes/probeset combinations for Etoposide using CGP and 6M IC50 values, respectively; all together corresponding to a total of six genes (*AGT*, *ANPEP*, *ATP6AP2*, *CPA3*, *CTSA and IGF2R (two probesets)*), when the analysis was performed with the discovery group. Applying the resulting models to the test group showed the reliability of all models. As shown in, the goodness of fit measures (R sq. and Sy.x) were 0.9 and 0.21 for Doxorubicin as modeled using 6M IC50 data and 0.89 and 0.34 when we used CGP IC50 values. Similarly, for Etoposide, these two measures were 0.78 and 0.34 for 6M IC50 and 0.77 and 0.57 for CGP IC50 values. ## *In vitro* validation of biomarker genes We next asked if the linear regression models generated *in silico* could predict *in vitro* cytotoxicity. For this purpose, we determined gene expression values by qRT-PCR for the six RAS genes (*AGT*, *ANPEP*, *ATP6AP2*, *CPA3*, *CTSA and IGF2R (two probesets)*) and used these to predict *in vitro* IC50 values obtained for Etoposide and Doxorubicin calculated with 6M approach for nine AML cell lines. Correlation analysis showed that *in silico* and *in vitro* gene expression data were highly concordant except for CTSA (r: \>0.7 and p-value \<0.05). We applied normalized gene expression values obtained in vitro to the *in silico* generated linear regression models (using four regression formulas). Thus, utilized linear regression formulas with qRT-PCR gene expression data showed a good correlation with *in silico* data for Doxorubicin but not Etoposide. ## Biological features of drug sensitive and resistant cells Cell lines sensitive to Etoposide and Doxorubicin were almost identical. To determine molecular mechanisms underlying differential response to Etoposide and Doxorubicin, we performed gene set enrichment analyses (GSEA) with sensitive and resistant subgroups for Gene Ontology (GO) gene sets. Several gene sets were significantly enriched among sensitive and resistant cell lines (FDR q-value\<0.25). Gene sets enriched in sensitive cells with a FDR q-value of lower than 0.25 included TNF-receptor interacting, and response to type I IFN stimulus; while gene sets such as regulation of TGF-beta production and FN- binding were enriched in resistant cells suggesting a mesenchymal phenotype. To determine if the differentially expressed genes could be reflecting Epithelial-Mesenchymal Transition (EMT), we compared E-Cadherin (epithelial marker) and Vimentin (mesenchymal marker) expression using t-test, between sensitive and resistant cell groups. EMT is the process that epithelial cells lose the apical-basal polarity and cell adhesion, and transform to invasive mesenchymal cells. It is known to play an important role in biological and pathological processes such as cancer progression, metastasis and drug resistance. In our analysis, E-cadherin and Vimentin expression were not significantly different between sensitive and resistant groups defined in (p\>0.1). Then, in order to understand whether the mutational profile is involved in sensitivity to Doxorubicin, we analyzed mutational data of sensitive, intermediate and resistant groups of AML cell lines. Although we have a small sample size, especially in the resistant group (n = 2), we observed that both of the resistant cell lines are NRAS and P53 mutant, whereas all of the sensitive cell lines (n = 5) were wild type for these genes. However, these results need to be validated in larger sample sizes to be conclusive. ## RAS genes are prognostic biomarkers for AML We then asked if the RAS gene expression could help prognosticate AML patients. For this purpose, we utilized the training set within the GSE12417 dataset, including 163 samples of bone marrow or peripheral blood mononuclear cells from adult patients with untreated acute myeloid leukemia. Patients in this cohort were also-treated with TAD protocol which contains Daunorubicin, which is also used as the starting material for semi-synthetic manufacturing of Doxorubicin. We found that high expression of genes *IGF2R* and *CTSA* were both associated with better overall survival, while the opposite was true for *ATP6AP2* when patients were classified in either “High” or “Low” groups based upon LRMC cutoffs for each gene separately. We then stratified patients into “Good” and “Bad” prognosis groups using the best cutoff values obtained for these three genes as explained in the method section. As shown in, it was revealed that there was a striking difference in overall survival in the groups that were predicted as "Good" and "Bad". The "Good" group showed better survival than the "Bad" group. Since the patients were all treated with Daunorubicin, these data suggest that the expression pattern of these genes was able to identify patients which are responders of this therapy. # Discussion RAS’ local presence in the marrow affects the most important stages of physiological and pathological blood cell proliferation, and also has important roles in the development of blood cancers. It has been shown that RAS plays important roles in drug resistance to chemotherapeutic agent in addition to angiogenesis, invasion and proliferation. Inevitably, most of these processes are interdependent. Most of the increased metastasis and invasion occurs due to an active RAS results in angiogenesis. AT1R upregulation in ovarian cancer and increased expression of AT1R and ACE in prostate cancer, and AGTR1 in breast cancer; localized RAS presence in gastric cancer and its correlation with tumor spread and progression; demonstrate strong associations of RAS with various cancers. Irregularity of RAS components in cancer is strongly associated with increased angiogenesis and metastasis, and these parameters are associated with poor prognosis. Gene expression profiling has revealed various AML subtypes related to diagnosis, therapy response and prognosis. Although gene expression profiling has not yet been integrated into clinical practice, this is expected to happen in near future. In our study, we focused on RAS genes and identified their association with Doxorubicin and Etoposide sensitivity. We also show that RAS genes can be used to stratify AML patients into groups with distinct prognoses. Similar to our findings, low expression of IGF2R in non-small cell liver cancer has been associated with poor prognosis and high expression in bladder cancer has been associated with good prognosis. Although high CTSA expression was associated with poorer outcome in breast ductal carcinoma *in situ*, it was also found to suppress invasion and metastasis of colorectal cancer, suggesting tissue-specific differential roles. Recent studies linked *ATP6AP2* up- regulation to the progression of glioma and colorectal cancer, due to its roles in aberrant activation of the Wnt/beta-catenin signalling pathway. *ATP6AP2* was also shown to be a key component of the pro-angiogenic/proliferative arm of the RAS, which plays a role in the growth and spread of endometrial cancer. Compared to the presence in the lysosome, it is found more in the cell membrane. Thus, it is clear that in this way it induces TGF-beta pathway activation. *IGF2R* is located in the membrane of organelles and is responsible for the transport to lysosome, and its intracellular functions have not yet been clearly identified. *CTSA* is a protease found in the lysosome. The fact that these three genes function together in the lysosome suggests that lysosomal functions can contribute to cell sensitivity. *ATP6AP2* gene was found to cause disruption of V-ATPase formation and defects in the lysosomal glycosylation and autophagy. Supportively, Doxorubicin has been reported to cause autophagy induced cell death in AML cells. GSEA revealed that sensitive cells were correlated with TNF-receptor interacting and response to type I IFN gene sets and resistant cells were correlated with regulation of TGF-beta production and FN-binding gene sets in AML, suggesting a mesenchymal phenotype. A good and reliable subgrouping which can predict Doxorubicin sensitivity in AML was performed with the *ATP6AP2*, *IGF2R*, and *CTSA* gene combination. For those analyses, we utilized a Daunorubicin treated cohort, which is used as the starting material for semi-synthetic manufacturing of Doxorubicin. Therefore, the combination of these genes which can predict the sensitivity of Doxorubicin in AML patients may, therefore, be confirmed *ex vivo*. The mutational analyses performed in this study had a small sample size with only two resistant cells. Therefore more conclusive results would be reached when this type of analysis is performed with larger sample sizes, or when mutational profiling is performed in patients treated with Doxorubicin, which may shed light on future studies. # Conclusions As a result, we identified *IGF2R*, *CTSA* and *ATP6AP2* gene biomarkers, which can subgroup AML patients into distinct good and bad prognostic groups. *ATP6AP2* was associated to resistance and *IGF2R* and *CTSA* were associated to sensitivity for Doxorubicin *in silico* and *in vitro*. In future studies, it is important to investigate whether these genes can be used for personalized treatment and improve the effectiveness of treatments. # Supporting information 10.1371/journal.pone.0242497.r001 Decision Letter 0 Bertolini Francesco Academic Editor 2020 Francesco Bertolini This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 7 May 2020 PONE-D-20-09590 Renin Angiotensin System Genes are Biomarkers for Personalized Treatment of Acute Myeloid Leukemia with Doxorubicin or Etoposide PLOS ONE Dear Dr. Turk, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process by both Reviewers. We would appreciate receiving your revised manuscript by Jun 21 2020 11:59PM. When you are ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. To enhance the reproducibility of your results, we recommend that if applicable you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission- guidelines#loc-laboratory-protocols> Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). This letter should be uploaded as separate file and labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. This file should be uploaded as separate file and labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. This file should be uploaded as separate file and labeled 'Manuscript'. Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out. We look forward to receiving your revised manuscript. Kind regards, Francesco Bertolini, MD, PhD Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Please provide additional information about each of the cell lines used in this work, including any quality control testing procedures (authentication, characterisation, and mycoplasma testing). For more information, please see [" ext-link-type="uri" xlink:type="simple"\>http://journals.plos.org/plosone/s/submission- guidelines#loc-cell-lines."](http://journals.plos.org/plosone/s/submission- guidelines#loc-cell-lines.) 3\. Please provide the source, product number and any lot numbers of the doxorubicin and etoposide obtained for your study.” 4\. Please note that PLOS does not permit references to “data not shown.” Authors should provide the relevant data within the manuscript, the Supporting Information files, or in a public repository. If the data are not a core part of the research study being presented, we ask that authors remove any references to these data. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Turk and colleagues in their research article entitled “Renin Angiotensin System Genes are Biomarkers for Personalized Treatment of Acute Myeloid Leukemia with Doxorubicin or Etoposide” perform a series of analyses with the aim to to verify if RAS genes can be good predictors of the sensitivity of two chemoterapeutics. Their bioinformatic approach identifies a series of genes that have been, in this research, tested with in vitro experiment. Additionally, applying again a computational approach, the authors stratify a cohort of patients previously sequenced on the basis of the previously mentioned genes. Although this research article is a good piece of work, I think that, in its current state, it is not suitable for publication but it can be potentially interesting if some modifications will be done to the analyses and to the manuscript. The main concern here is about the methodology implied in the first computational section. I please recommend to specifically indicate, particularly in the method section, if the workflow-analyses performed have been either applied in previous researches or are reported here for the first time. One example is in the “IC50 Calculation Methods” section: the six different models seem introduced by the authors for the first time while, in the result section (line 207) it is referred to them as the “NCBI proposed 6-model approach”, is it the same? Can the author add a reference to this? Finally I suggest to be more consistent and clear with the numbers/genes along the text. The major points are listed here. In the “Data normalization and variance analysis” is there a reason why “the genes whose variance was above 0.8 SD of the mean” were chosen? Additionally this number is not the same of the results in which is reported “which showed high variation in expression and therefore, selected 9 probesets (8 genes) with standard deviation values above 0.9” (line 205), I would suggest to add a reference or better explain this method. I was wondering why the author did not consider to calculate and consider adjusted p-value for the genes selected. In “linear regression analyses” section the authors need to better clarify the steps they followed during this methodology, I suggest either to insert some references or clarify the steps. Please also clarify if in this case all the genes or only the 8 were used. Moreover, I wonder if the Pearson’s correlation was always applied on normal distribution of data, if this is not the case I would suggest a Spearman correlation test. In the results section the authors refer to 6M data which have not been explained before in the method section, these data likely are deriving from the raw CGP after applying the six model approach, I would suggest to the authors to add this information in the method section. I suggest to replicate the random division of the groups and test if the results are consistent with the one obtained here. Moreover in the method section there is no mention of such a random division, please, add it. If there is a reason why the division was not replicated, please, mention it. In vitro experiment the genes used are six, and the primers reported in the table 1 are for seven genes. Please clarify this and explain the reason why the authors did not consider all the eight genes from the in silico workflow. Along the text it is not clear if the sub-groups of genes belong to the initial eight. Please refromulate the text in order to give a better explanation of these numbers and other numbers of genes. My suggestion is to either reorganize the figures or change the captions: in Fig.3 there is no explanation of the three panels (A, B and C) and neither of the colors. Moreover, beside the Kaplan-Meier curves there are other 3 plots which are not explained. The same for Fig. 4. Minor points: \- Line 79, when E-MTAB-783 is indicated, please cite the two research articles that contributed to produce these data: 1\) Garnett MJ, Edelman EJ, Heidorn SJ, Greenman CD, Dastur A, Lau KW, Greninger P, Thompson IR, Luo X, Soares J, Liu Q. Systematic identification of genomic markers of drug sensitivity in cancer cells. Nature. 2012 Mar;483(7391):570-5. 2\) Venkova L, Aliper A, Suntsova M, Kholodenko R, Shepelin D, Borisov N, Malakhova G, Vasilov R, Roumiantsev S, Zhavoronkov A, Buzdin A. Combinatorial high-throughput experimental and bioinformatic approach identifies molecular pathways linked with the sensitivity to anticancer target drugs. Oncotarget. 2015 Sep 29;6(29):27227. The 1) is already present in the manuscript as ref number 32. -Line 89, please insert the article “the” when referring to the 17 AML and to the 25 genes that are taken by CGP and are indicated in the results section. Moreover, consider to add this info also on the methods. -Lines 121, please insert the website of Minitab 17 -Line123-124 if the authors are referring to the same eight genes that have been mentioned in the MM Normalization section I would suggest to point it out. -Lines138 Please cite the reference or website for Cluster3.0 and Java Treeview software. \- The link at line 144 does not work, please indicate the number of pathways and the number of genes that were present in the C5_all Gene ontology database, and which version of the database was used. -Line 188 if the script is available provide it as supplementary information or in a github repository -Line 217 the authors are referring to 4 drugs and 8 genes, are these numbers and data the same that have been identified in the previously mentioned analyses? Why did the authors perform linear regression analyses at this step? Please report this information in this section of the manuscript and if the genes/drugs are the ones already mentioned add the definite article “the”. -Lines 245-247 please refer to which correlation analysis the authors are referring to. Moreover, PLOS does not accept references to “data not shown.” \- Line 248 please indicate why only four formulas were applied and change 4 in four. -Lines291-292 when the authors refer to “We then stratified patients using the best cut-off values obtained for these 3 genes” please add, “as explained in the method section”. -Line 292 please substitute 3 with three -Figure 1 A) and B) are not indicated. Define the Sy.x parameter, is it the value for the residuals? Please add this information also in the methods. -Line562 please reformulate “ve resistant” -Uniform the numbers, below 10 the number should be indicated as word. Reviewer \#2: Seyhan Turk and co-worker in their work demonstrate that expression of Renin-Angiotensin System (RAS)-related genes predict drug responses (Doxorubicin and Etoposide) in AML patients. Moreover, authors show that identified RAS genes expression stratify AML patients into different subtypes with distinct prognosis. Overall, presented data support use of RAS gene expression analysis as novel tool for AML drug-sensitivity and disease prognostication. Unfortunatley, altough an elegant set of in-silico approaches have been employed, the lack of experimental analyses with appropriate functional in-vitro and in-vivo represents the main drawback of the entire work. In detail: Major points • 17 AML cell lines included in CGP database have been chosen for in-silico analysis. In parallel, 9 AML cell lines have been testd for in vitro studies. Are those the same cells included in short list for in-silico analysis? Furthermore, did you see any differences based on their specific genetic background (mutational analysis)? • Importantly, GEP analysis have been performed on genes, among those of RAS system, with higher expression variability. Why did you reject those with less variation for your analysis? • Could you please detail the NCBI proposed 6-model used approach? • To make in vitro date more consostent, could be useful including gene expression analysis as well as IC50 values for all tested AML cell lines. • Data showed in Table 2 are not clear. Could you please describe it better? • The prognostic role of 3-gene expression signature need deeper analysis. Why did you analyze only 3 genes? What about other RAS genes? Did you perform a cumulative analysis of RAS-related genes? • As per Authors own admission, the major study limitation is lack of mutational data analysis. Indeed, it’s worth to be investigated AML patiens subclasses including those carryng poor prognostic mutations such as FLT3. To this aim a detailed description of used AML cell lines could help (i.e. carryng FLT3-ITD or WT, NPM1 etc.) Minor • Please pospone figures legend at the end of manuscript right after refernces • The gene set enrichment analysis revealed findinds that are not supported by experimental data. Overall these data are somehow confusing because are not conclusive at all. In my opinion it’s better including these data as supplementary results to make conclusion more focused • In the Materials and methods the first sentence of paraghraph is quite misleading. Additionally, please include reference for CGP database. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: Yes: Michele Cea \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0242497.r002 Author response to Decision Letter 0 21 Jun 2020 Dear Editor, We would like to thank the Editorial Board and the Referees for all of the important contributions, which will improve the paper. We have carefully reviewed the comments and revised the manuscript accordingly. Below please find the answers to the Editor’s and Reviewer’s comments. Yours faithfully, Journal Requirements: Comment 1. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> Response 1. PLOS ONE's style requirements fully checked and revisions were done. Comment 2. 2\. Please provide additional information about each of the cell lines used in this work, including any quality control testing procedures (authentication, characterization, and mycoplasma testing). For more information, please see <http://journals.plos.org/plosone/s/submission-guidelines#loc-cell-lines>." Response 2. Additional information is added for the used cell lines (see In vitro section in Materials and Methods). Comment 3. 3\. Please provide the source, product number and any lot numbers of the doxorubicin and etoposide obtained for your study.” Response 3. Drug information has been added to the manuscript. Comment 4. 4\. Please note that PLOS does not permit references to “data not shown.” Authors should provide the relevant data within the manuscript, the Supporting Information files, or in a public repository. If the data are not a core part of the research study being presented, we ask that authors remove any references to these data. Response 4. The “data not shown” was removed and the data was added as “S8 Table”. Comments to the Author Reviewer \#1: Comment 1. The main concern here is about the methodology implied in the first computational section. I please recommend to specifically indicate, particularly in the method section, if the workflow-analyses performed have been either applied in previous researches or are reported here for the first time. One example is in the “IC50 Calculation Methods” section: the six different models seem introduced by the authors for the first time while, in the result section (line 207) it is referred to them as the “NCBI proposed 6-model approach”, is it the same? Can the author add a reference to this? Finally, I suggest to be more consistent and clearer with the numbers/genes along the text. Response 1. Here, we report the 6-model (6M) approach for the first time which depends on a non-linear logistic regression function explained in NIH/NCGC assay guidelines. We derived six different versions of this function and select the one with the lowest error rate among all for the calculation of cytotoxicity values. References were added and the “IC50 Calculation Methods” section has been re- written. No inconsistency could be found with the numbers/genes given throughout the entire paper. The major points are listed here. Comment 2. In the “Data normalization and variance analysis” is there a reason why “the genes whose variance was above 0.8 SD of the mean” were chosen? Additionally, this number is not the same of the results in which is reported “which showed high variation in expression and therefore, selected 9 probesets (8 genes) with standard deviation values above 0.9” (line 205), I would suggest to add a reference or better explain this method. I was wondering why the author did not consider to calculate and consider adjusted p-value for the genes selected. Response 2. In order to choose the genes which will be used in Real-time PCR for validations, we aimed to choose the most variable genes which could give detectable fold differences in vitro. The variance value of 0.8 was chosen arbitrarily. For these analyses, variance and standard deviation are the same values; we decided to use “variance”. The “Data normalization and variance analysis” section has been expanded and detailed. Since, we are here trying to choose the most variant genes, we did not calculate and consider adjusted p-value for the genes selected. Comment 3. In “linear regression analyses” section the authors need to better clarify the steps they followed during this methodology, I suggest either to insert some references or clarify the steps. Please also clarify if in this case all the genes or only the 8 were used. Moreover, I wonder if the Pearson’s correlation was always applied on normal distribution of data, if this is not the case I would suggest a Spearman correlation test. Response 3. The linear regression analyses section has been re-written. Linear regression analysis was performed with the highly variant eight RAS genes and minimal gene combinations which are now explained more clearly. We performed Pearson r correlation analysis throughout the paper as we consistently obtained better p values with it, as compared to Spearman’s test. Comment 4. In the results section the authors refer to 6M data which have not been explained before in the method section, these data likely are deriving from the raw CGP after applying the six model approach, I would suggest to the authors to add this information in the method section. Response 4. With 6M approach we recalculated IC50 values from raw CGP data for 17 AML cell lines treated with four drugs (ATRA, Cytarabine, Etoposide and Doxorubicin) using an in-house R script. We refer to this data as 6M IC50s. Additionally, we treated 9 AML cell lines with Doxorubicin and Etoposide in vitro and their IC50 values were calculated using 6M approach, as well. We refer to this data as in vitro IC50s. These are explained in the methods section. Comment 5. I suggest to replicate the random division of the groups and test if the results are consistent with the one obtained here. Moreover, in the method section there is no mention of such a random division, please, add it. If there is a reason why the division was not replicated, please, mention it. Response 5. In response to this comment we divided 17 cells randomly 10 times and generated 10 different discovery and test groups consisting of 12 cell lines and 5 cell lines, respectively. Linear regression models were generated in the 10 discovery groups separately. The 10 random models of sensitivity to Doxorubicin still have an average R2 above 85% for both CGP and 6M IC50, but R2 decreased slightly for models of sensitivity to Etoposide. Average R2 values now given in “S5 Table”. We added the 10 times random division to the method section and also mentioned it in the results section. However, the reason we performed our analyses without random divisions was because we wanted both the discovery and test cohorts to contain cells that spanned as large a sensitivity interval as possible. We therefore, present in this revised version results from both analyses. Comment 6. In vitro experiment the genes used are six, and the primers reported in the table 1 are for seven genes. Please clarify this and explain the reason why the authors did not consider all the eight genes from the in silico workflow. Along the text it is not clear if the sub-groups of genes belong to the initial eight. Please reformulate the text in order to give a better explanation of these numbers and other numbers of genes. Response 6. GAPDH was used as endogenous reference control. That’s why the primers are seven in the Table 1. We used six genes because in our linear regression analyses, highest correlation is observed in IGF2R/ATP6AP2/CTSA combination with Doxorubicin (both CGP and 6-model) and highest correlation is observed in IGF2R/ATP6AP2/CTSA/CPA3 combination with Etoposide (CGP) and ANPEP/ATP6AP2/CTSA/CPA3/AGT combination with Etoposide (6-model). Prediction model contains totally six genes for Doxorubicin and Etoposide. The section has been re-written for the sake of clarity. Comment 7. My suggestion is to either reorganize the figures or change the captions: in Fig.3 there is no explanation of the three panels (A, B and C) and neither of the colors. Moreover, beside the Kaplan-Meier curves there are other 3 plots which are not explained. The same for Fig. 4. Response 7. Figures and their legends were re-organized. Minor points: Comment 8. \- Line 79, when E-MTAB-783 is indicated, please cite the two research articles that contributed to produce these data: 1\) Garnett MJ, Edelman EJ, Heidorn SJ, Greenman CD, Dastur A, Lau KW, Greninger P, Thompson IR, Luo X, Soares J, Liu Q. Systematic identification of genomic markers of drug sensitivity in cancer cells. Nature. 2012 Mar;483(7391):570-5. 2\) Venkova L, Aliper A, Suntsova M, Kholodenko R, Shepelin D, Borisov N, Malakhova G, Vasilov R, Roumiantsev S, Zhavoronkov A, Buzdin A. Combinatorial high-throughput experimental and bioinformatic approach identifies molecular pathways linked with the sensitivity to anticancer target drugs. Oncotarget. 2015 Sep 29;6(29):27227. The 1) is already present in the manuscript as ref number 32. Response 8. Citations were added and the text was reorganized. Comment 9. -Line 89, please insert the article “the” when referring to the 17 AML and to the 25 genes that are taken by CGP and are indicated in the results section. Moreover, consider to add this info also on the methods. Response 9. In the method and results sections article "the" were inserted. Source of AML cell lines, CGP, was indicated. Comment 10. -Lines 121, please insert the website of Minitab 17 Response 10. The website was added. Comment 11. -Line123-124 if the authors are referring to the same eight genes that have been mentioned in the MM Normalization section, I would suggest to point it out. Response 11. These are the same genes, which is now indicated. Comment 12. -Lines138 Please cite the reference or website for Cluster3.0 and Java Treeview software. Response 12. The websites for software were added. Comment 13. \- The link at line 144 does not work, please indicate the number of pathways and the number of genes that were present in the C5_all Gene ontology database, and which version of the database was used. Response 13. The reference was added. Dataset E-MTAB-783 has 22277 probesets IDs and these were collapsed into 13321 genes. For genes with more than one probeset, one with the highest expression was selected. “C5_all Gene ontology v6.1 database” was used for the analysis which has gene sets that contain genes annotated by the same GO term. Default filtering criteria in GSEA for gene sets is that it should have minimally 15 genes and maximally 500 genes. After applying this filter, analysis was performed for 4081 gene sets. This information was added to the method section, as well. Comment 14. -Line 188 if the script is available provide it as supplementary information or in a github repository. Response 14. The R script was provided as supplementary. Comment 15. -Line 217 the authors are referring to 4 drugs and 8 genes, are these numbers and data the same that have been identified in the previously mentioned analyses? Why did the authors perform linear regression analyses at this step? Please report this information in this section of the manuscript and if the genes/drugs are the ones already mentioned add the definite article “the”. Response 15. Yes, four drugs and eight genes are the same in the previously mentioned analyses. To examine the correlation of drug sensitivity with combined expression profile of the eight genes, linear regression analyses were performed. “Linear regression analyses” section was reviewed. Article “the” was added when needed. Comment 16. -Lines 245-247 please refer to which correlation analysis the authors are referring to. Moreover, PLOS does not accept references to “data not shown.” Response 16. This correlation analysis is referring to test the compatibility of in silico and in vitro gene expression data with each other. “data not shown” removed from the text and data now given in “S8 Table”. Comment 17. \- Line 248 please indicate why only four formulas were applied and change 4 in four. Response 17. We applied four formulas since there are four linear regression models for Doxorubicin (CGP and 6M IC50s) and Etoposide (CGP and 6M IC50s) with minimal gene lists, we used these four formulas to predict IC50 values for both drugs in test groups. “4” was changed to “four”. Comment 18. -Lines291-292 when the authors refer to “We then stratified patients using the best cut-off values obtained for these 3 genes” please add, “as explained in the method section”. Response 18. The text was rewritten. Comment 19. -Line 292 please substitute 3 with three Response 19. The number was indicated as “three”. Comment 20. -Figure 1 A) and B) are not indicated. Define the Sy.x parameter, is it the value for the residuals? Please add this information also in the methods. Response 20. A), B), C) and D) were indicated in the Figure 1. Sy.x is a standard deviation of the residuals. In our linear regression analyses the residual standard deviation used to describe the difference in standard deviations of CGP and 6M IC50s versus predicted IC50s. It is a goodness-of-fit measure used to show how well our predicted IC50s fit with the CGP and 6M IC50s. It is also mentioned in the methods section. Comment 21. -Line562 please reformulate “ve resistant” Response 21. The text was reformulated. Comment 22. -Uniform the numbers, below 10 the number should be indicated as word. Response 22. Below 10 the number are indicated as word. Reviewer \#2: Major points Comment 1. • 17 AML cell lines included in CGP database have been chosen for in-silico analysis. In parallel, 9 AML cell lines have been testd for in vitro studies. Are those the same cells included in short list for in-silico analysis? Furthermore, did you see any differences based on their specific genetic background (mutational analysis)? Response 1. Yes, they are same cell lines. 17 cell lines used in CGP data are listed in S1 Table. These are CTV-1, HL-60, GDM-1, HEL92.1.7, KASUMI-1, KMOE-2, K052, ML-2, MONO-MAC-6, NKM-1, NOMO-1, P31-FUJ, THP-1, QIMR-WIL, CMK, CESS, OCI-AML2. Cell lines studied in vitro experiments are shown in the method section “Cell lines and cytotoxicity experiments”. These are HEL92.1.7, KASUMI-3, GDM-1, QIMR- WIL, CESS, P31/FUJ, NOMO-1, KASUMI-1 and SKM-1. Seven cell lines used for in silico analyses were also used for in vitro experiments (HEL92.1.7, KASUMI-1, GDM-1, QIMR-WIL, CESS, P31/FUJ, NOMO-1), as stated in the manuscript. We performed a comprehensive mutational analysis in order to see if any mutational pattern overlaps with the sensitivity profile, and added the results in the manuscript. Also, nine AML cell lines were WT for NPM1 and FLT mutations. Comment 2. • Importantly, GEP analysis have been performed on genes, among those of RAS system, with higher expression variability. Why did you reject those with less variation for your analysis? Response 2. In order to choose the genes that will be used further in Real-time PCR for validations, we first aimed to choose the most variable genes which are also highly likely to give detectable fold differences via PCR. Comment 3. • Could you please detail the NCBI proposed 6-model used approach? Response 3. It is detailed in the “IC50 Calculation methods” section. NCBI methodology or "NIH/NCGC-proposed methodology" suggests a calculation methodology similar to 6-model approach. The function used to model the data is widely being used for cytotoxicity calculations. We derived different versions of this function, which is partly suggested by NIH/NCGC, and select the one with the lowest error rate among all for the calculation of cytotoxicity values. Comment 4. • To make in vitro date more consistent, could be useful including gene expression analysis as well as IC50 values for all tested AML cell lines. Response 4. To make more consistent in vitro cytotoxicity IC50s, predicted IC50 and QPCR relative gene expression values are now given in “S7 Table” and “S9 Table”. Comment 5. • Data showed in Table 2 are not clear. Could you please describe it better? Response 5. The data has been described in more detail. Comment 6. • The prognostic role of 3-gene expression signature need deeper analysis. Why did you analyze only 3 genes? What about other RAS genes? Did you perform a cumulative analysis of RAS-related genes? Response 6. First, we started analyzing the 25 RAS genes all together. After applying a variance cut-off, this number decreased to eight genes. After linear regression analysis, these eight genes decreased to six genes (four regression formulas contain totally six genes for Etoposide and Doxorubicin and for two different IC50s (CGP and 6M)). But three genes (IGF2R, CTSA, ATP6AP2) were common for all regression formulas except for one “Etoposide 6M”. Therefore, we analyzed that three genes’ biomarker potential. Even for Doxorubicin, only these three genes come out in the regression formulas with both CGP and 6M IC50s. Only one additional gene (CPA3) comes out in the regression formula for Etoposide CGP IC50s. Analyses of six genes was necessary to produce predicted IC50s from regression formulas. Comment 7. • As per Authors own admission, the major study limitation is lack of mutational data analysis. Indeed, it’s worth to be investigated AML patiens subclasses including those carryng poor prognostic mutations such as FLT3. To this aim a detailed description of used AML cell lines could help (i.e. carryng FLT3-ITD or WT, NPM1 etc.) Response 7. Since there is no mutational data in the patient dataset we could not perform mutational analyses with clinical data. However, we evaluated the mutational profile of 14 out of 17 AML cell lines used in in silico analyses. Seven genes (TP53, RBMX, NRAS, ANKRD36C, TNS1, TTN and ASXL1) which are mutated in at least three cell lines were included in mutational analysis and our results are given in “S4 Fig”. For FLT3 and NPM1 gene, none of the used in vitro AML cell lines were mutants. Minor points Comment 8. • Please pospone figures legend at the end of manuscript right after refernces Response 8. PLOS Journal requirements state the following: Place figure captions in the manuscript text in read order, immediately following the paragraph where the figure is first cited. Comment 9. • The gene set enrichment analysis revealed findinds that are not supported by experimental data. Overall these data are somehow confusing because are not conclusive at all. In my opinion it’s better including these data as supplementary results to make conclusion more focused. Response 9. The gene set enrichment analysis results have been changed as supplementary “S2 Fig” and “S10 Table”. Comment 10. • In the Materials and methods the first sentence of paraghraph is quite misleading. Additionally, please include reference for CGP database. Response 10. Paragraph was re-written. The CGP database references were added. 10.1371/journal.pone.0242497.r003 Decision Letter 1 Bertolini Francesco Academic Editor 2020 Francesco Bertolini This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 14 Jul 2020 PONE-D-20-09590R1 Renin angiotensin system genes are biomarkers for personalized treatment of acute myeloid leukemia with doxorubicin as well as etoposide PLOS ONE Dear Dr. Turk, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process, particularly by Reviewer \#1. Please submit your revised manuscript by Aug 28 2020 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols> We look forward to receiving your revised manuscript. Kind regards, Francesco Bertolini, MD, PhD Academic Editor PLOS ONE \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: (No Response) Reviewer \#2: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Partly Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: I appreciate the revision of the methods and the extension of the non-linear regression model section, now the methods are sufficiently explained and more clear. However I have noticed that the new version of the manuscript has some parts that need to be reformulated, many typos and inconsistencies along the text. Additionally it is missing the explanation of the statistics applied in some parts of the manuscript and my concern here is whether or not these tests were rigorously applied. Finally, I suggest a revision of the Supplementary Materials provided and that the R scripts file will be provided. For these reasons I am still not considering the manuscript suitable for publication in its current form but I suggest the following points to be addressed to be taken in consideration for publication. Line 82 the authors should insert the references in the correct location, if the reference number 30 is referring to the GSE12417 dataset it should be inserted right after it. As this, there are other similar cases along the text, therefore, I please invite the authors to check this in the manuscript. Line 91, please remove “were chosen arbitrarily”, if any other study used this parameter please cite it. Line 99-100 “explained in De Lean et. al, which is also explained” there is a repetition of the word explain, please substitute with “reported”. Line129 please, insert here that the IC50 values were included in the CGP. If I am not wrong the authors mentioned this already at line 243. Line 130-131 Please provide the R script and refer to it as supplementary material. Line 131 “We refer to this data as 6M IC50”, please be consistent with this nomenclature along the text, sometimes it is called “6M IC50” others “IC50 6M” other only “6M” Line 135 “9 AML cell lines were treated with” the number 9 needs to be written in words. Additionally I suggest to reformulate or clarify the meaning of the sentence “IC50 values were calculated using the 6M approach in vitro.”, did the authors mean that the values were calculated using the 6M approach on the data obtained from the in vitro analysis? Line 138 8 needs to be written as a word. The correlations to which the authors are referring here is the one shown in S3 Table, how was this correlation calculated? If it was with graphpad I would suggest inserting a sentence at the end of this section saying that all the correlations were calculated with Graphpad software. Moreover, Pearson correlation should be used when both the variables are normally distributed; in the response to the reviewers the authors mentioned that they got better results with this method but I was wondering if the two variables were tested for normal distribution or not. Line 151 please delete “for 10 models” at the end of the sentence, it is redundant. Line 161 please substitute “that is used to describe” with “that here has been used to describe”. Line 173-174 please correct “Gene set enrichment analyses was” with “Gene set enrichment analysis was” Line 223 please correct “Clinical data was” with “Clinical data were” Line 226 please, when the R script is mentioned in the text, refer to the supplementary material in which it is contained. Lines 228-229 These lines need a reformulation. First, you should put as references the two studies (PMID: 31949498, PMID: 28607584), second, please change “ 'Low' = low expression” with “‘Low’ (low expression)” and “‘High = high expression’” with “‘High’ (high expression)” and finally, substitute “our previous studies” with “as in refX and refY”. Line 242 I ask the authors to rewrite the citations when CGP database is mentioned. Line248 the authors mention: “We observed strong correlations between IC50 values obtained from CGP”, as said above, this correlations need to be clarified in the MM section. Line254 The sentence “We then asked whether combined expression analyses of genes could correlate better with drug sensitivity data.” should be linked with the next paragraph. Line257-262 please re-arrange these sentences because they are not clear. The explanation of the workflow used has some typos and english grammar errors. Moreover, some parts are already mentioned in the MM section and should be removed. Line 264 please, remove “RAS genes” it is a redundancy; if it is not, I please ask the authors to explain why it is mentioned here. TableS4 and S5 should be merged into the same file. The name of the columns should be revised, precisely: on the top of the column referring to CGP please indicate CGP and on the top of the column referring to 6M IC50 indicate 6M IC50. I also suggest to name the sheets of the.xlsx table according to the table. The two above mentioned indications are applicable also to the other S Tables. Line277 6M IC50 needs to be indicated accordingly along the text. TableS6, the columns referring to each group need to be marked. Figure 1: The names in the title need to be consistent with the content of the text, therefore 6M needs to be substituted with 6M IC50. Line278 I would suggest to report also here the name of the six genes that have been selected. Line294 a comma between “ANPEP ATP6AP2” is missing. I suggest the authors either merge table S7 and S8 or put them in different sheets of the same file. S1 Fig: I suggest that the legend of this figure will also include the meaning of the colors. Maybe a scale of colors should be provided in the figure. Line320-322 EMT acronym is not explained along the text and there is no mention of the statistical test involved when the authors say: “there were no significant differences between the two groups”. The authors need to mention the test performed (Wilcoxon or t-test according to the distribution of the data) and/or report it in the mm section. S4 Fig needs to be included as a table or a different figure should be provided. Line 334-335 I was wondering how the authors investigated the up-regulation of the three genes. Additionally, as previously mentioned from the reviewer 2, it is still not clear to me the choice of these three genes; if it is related to the fact that these genes were the one common for all regression formulas I suggest to mention it in the text and/or mm section. Line332 I recommend to add again the reference for GSE12417 when it is mentioned. Line334 Doxorubicin needs to be indicated with the first letter in uppercase. Line 337 please add in the parenthesis together with Fig2 “see Materials and Methods section”. Line 339-341 these lines need to be reformulated. Line352 the word “cut-off” needs to be consistent along the text and the numbers need to be rounded. Since many S tables are really small I suggest to insert them in a unique pdf file and leave as excel only those that do not fit on a pdf page. The script file is missing, it should be provided in a.zip file with all the codes. The section Acknowledgments is blank. Reviewer \#2: (No Response) \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: **Yes: **Michele Cea \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0242497.r004 Author response to Decision Letter 1 27 Aug 2020 Dear Editor, We would like to thank the Editorial Board and the Reviewer for all the contributions. You can find all the responses and the necessary revisions based on the reviewer’s comments below. Yours faithfully, Dr. Seyhan TURK, PhD Review Comments to the Author Reviewer \#1: I appreciate the revision of the methods and the extension of the non-linear regression model section, now the methods are sufficiently explained and more clear. However I have noticed that the new version of the manuscript has some parts that need to be reformulated, many typos and inconsistencies along the text. Additionally it is missing the explanation of the statistics applied in some parts of the manuscript and my concern here is whether or not these tests were rigorously applied. Finally, I suggest a revision of the Supplementary Materials provided and that the R scripts file will be provided. For these reasons I am still not considering the manuscript suitable for publication in its current form but I suggest the following points to be addressed to be taken in consideration for publication. Response: Accordingly, all sections were reformulated. All typos and inconsistencies have been corrected along with the text. All necessary explanations of all applied statistics were added to the manuscript. All tests were applied rigorously. All Supplementary Materials were reviewed and R script file was provided as a new Supplementary (S2 Table). Comment 1: Line 82 the authors should insert the references in the correct location, if the reference number 30 is referring to the GSE12417 dataset it should be inserted right after it. As this, there are other similar cases along the text, therefore, I please invite the authors to check this in the manuscript. Response 1: The locations of all references have been checked and necessary insertions were done. Comment 2: Line 91, please remove “were chosen arbitrarily”, if any other study used this parameter please cite it. Response 2: The “were chosen arbitrarily” were removed from the text. Comment 3: Line 99-100 “explained in De Lean et. al, which is also explained” there is a repetition of the word explain, please substitute with “reported”. Response 3: The “explained” was substituted with “reported”. Comment 4: Line129 please, insert here that the IC50 values were included in the CGP. If I am not wrong the authors mentioned this already at line 243. Response 4: Text was reviewed. The “IC50 values that were also included in the raw CGP data” and “We used two versions of CGP data, one original CGP data, second is recalculated 6M IC50 data by R script” were added to the text. Comment 5: Line 130-131 Please provide the R script and refer to it as supplementary material. Response 5: The R script was provided in the text and it was referred as a new “S2 Table” Comment 6: Line 131 “We refer to this data as 6M IC50”, please be consistent with this nomenclature along the text, sometimes it is called “6M IC50” others “IC50 6M” other only “6M” Response 6: “6M IC50” it was corrected in all necessary locations in the manuscript. Comment 7: Line 135 “9 AML cell lines were treated with” the number 9 needs to be written in words. Additionally I suggest to reformulate or clarify the meaning of the sentence “IC50 values were calculated using the 6M approach in vitro.”, did the authors mean that the values were calculated using the 6M approach on the data obtained from the in vitro analysis? Response 7: The number 9 was written in words. The sentence reformulated as “In addition, IC50 values were calculated using the 6M IC50 approach on the data obtained from in vitro analysis in which nine AML cell lines were treated with Doxorubicin and Etoposide” to clarify the meaning. Comment 8: Line 138 8 needs to be written as a word. The correlations to which the authors are referring here is the one shown in S3 Table, how was this correlation calculated? If it was with graphpad I would suggest inserting a sentence at the end of this section saying that all the correlations were calculated with Graphpad software. Moreover, Pearson correlation should be used when both the variables are normally distributed; in the response to the reviewers the authors mentioned that they got better results with this method but I was wondering if the two variables were tested for normal distribution or not. Response 8: The number 8 was written in words. At the end of the section “all the correlations were calculated with Graphpad software” were added. Pearson's correlation analysis was applied only on normally distributed data. We observed that there was no evidence to reject normality in any variables (p\>0.05) except for CPA3 gene (marked in red). The Supplementary data was updated so that an Excel file for the normality test results was added as "Correlation Analyses Results - Kolmogorov Smirnov (KS). Comment 9: Line 151 please delete “for 10 models” at the end of the sentence, it is redundant. Response 9: The “for 10 models” was removed from the text. Comment 10: Line 161 please substitute “that is used to describe” with “that here has been used to describe”. Response 10: The “that is used to describe” was substituted with “that here has been used to describe”. Comment 11: Line 173-174 please correct “Gene set enrichment analyses was” with “Gene set enrichment analysis was” Response 11: The “Gene set enrichment analyses was” was corrected with “Gene set enrichment analysis was” Comment 12: Line 223 please correct “Clinical data was” with “Clinical data were” Response 12: The “Clinical data was” was corrected with “Clinical data were” Comment 13: Line 226 please, when the R script is mentioned in the text, refer to the supplementary material in which it is contained. Response 13: It was referred to the “S2 Table” when the R script is mentioned in the text. Comment 14: Lines 228-229 These lines need a reformulation. First, you should put as references the two studies (PMID: 31949498, PMID: 28607584), second, please change “ 'Low' = low expression” with “‘Low’ (low expression)” and “‘High = high expression’” with “‘High’ (high expression)” and finally, substitute “our previous studies” with “as in refX and refY”. Response 14: The necessary references were added (for PMID: 31949498, PMID: 28607584). The “'Low' = low expression” was changed to “‘Low’ (low expression)” and “‘High = high expression’” was changed to “‘High’ (high expression)” And the “our previous studies” was changed to “as in \[38,39\].” Comment 15: Line 242 I ask the authors to rewrite the citations when CGP database is mentioned. Response 15: The citations were rewritten. Comment 16: Line248 the authors mention: “We observed strong correlations between IC50 values obtained from CGP”, as said above, this correlations need to be clarified in the MM section. Response 16: The section were clarified in the MM section. The “We performed a Pearson r correlation analysis between CGP IC50s and 6M IC50s to test the compatibility and observed strong correlations between them for drugs widely used in AML (Cytarabine, Etoposide, Doxorubicin) but not for ATRA (S3 Table)” was added. Comment 17: Line254 The sentence “We then asked whether combined expression analyses of genes could correlate better with drug sensitivity data.” should be linked with the next paragraph. Response 17: The sentences were linked together. Comment 18: Line257-262 please re-arrange these sentences because they are not clear. The explanation of the workflow used has some typos and english grammar errors. Moreover, some parts are already mentioned in the MM section and should be removed. Response 18: The section was rearranged. Comment 19: Line 264 please, remove “RAS genes” it is a redundancy; if it is not, I please ask the authors to explain why it is mentioned here. Response 19: The “RAS genes” were removed. Comment 20: TableS4 and S5 should be merged into the same file. The name of the columns should be revised, precisely: on the top of the column referring to CGP please indicate CGP and on the top of the column referring to 6M IC50 indicate 6M IC50. I also suggest to name the sheets of the.xlsx table according to the table. The two above mentioned indications are applicable also to the other S Tables. Response 20: The S4 Table and S5 Table were merged into the same file as a new “S5 Table”. The name of the columns was revised. CGP and 6M IC50 were indicated on the top of the columns. The sheets of the.xlsx tables were named according to the tables. Other S tables were also reviewed. Comment 21: Line277 6M IC50 needs to be indicated accordingly along the text. Response 21: “6M IC50” was indicated accordingly along the manuscript. Comment 22: TableS6, the columns referring to each group need to be marked. Response 22: S6 Table was revised accordingly. Comment 23: Figure 1: The names in the title need to be consistent with the content of the text, therefore 6M needs to be substituted with 6M IC50. Response 23: The names in the titles of the Fig1 were renewed. Comment 24: Line278 I would suggest to report also here the name of the six genes that have been selected. Response 24: The gene names were also reported in the text. Comment 25: Line294 a comma between “ANPEP ATP6AP2” is missing. Response 25: The comma was added. Comment 26: I suggest the authors either merge table S7 and S8 or put them in different sheets of the same file. Response 26: Accordingly, Table S7 and Table S8 were merged as a new “S7 Table”. Comment 27: S1 Fig: I suggest that the legend of this figure will also include the meaning of the colors. Maybe a scale of colors should be provided in the figure. Response 27: A color scale was added to the S1 Fig. And the meanings of the colors were added to the legend. Comment 28: Line320-322 EMT acronym is not explained along the text and there is no mention of the statistical test involved when the authors say: “there were no significant differences between the two groups”. The authors need to mention the test performed (Wilcoxon or t-test according to the distribution of the data) and/or report it in the mm section. Response 28: The EMT acronym was explained in the text. The citations were added. We used t-test for statistical test. Accordingly, t-test was mentioned in the text. Comment 29: S4 Fig needs to be included as a table or a different figure should be provided. Response 29: S4 Fig was included as a new “S10 Table”. Comment 30: Line 334-335 I was wondering how the authors investigated the up-regulation of the three genes. Additionally, as previously mentioned from the reviewer 2, it is still not clear to me the choice of these three genes; if it is related to the fact that these genes were the one common for all regression formulas I suggest to mention it in the text and/or mm section. Response 30: Accordingly, the sentence was corrected as: “We found that high expression of genes IGF2R and CTSA were both associated with better overall survival, while the opposite was true for ATP6AP2 when patients were classified in either “High” or “Low” groups based upon LRMC cutoffs for each gene separately (see Materials and Methods section) (Fig 2).” The High expression and the opposite were determined using Log-rank with multiple cut-offs (LRMCs) algorithm as described in methods sections under the heading of clinical data validation. LRMC generates all possible cutoffs with their respective p values associated with Hazard ratios (Fig 2A). So for each gene, these cutoffs were generated and one cutoff with the smallest p-value was selected. Patients with expression values above this cutoff were labeled high and the rest were labeled as low. And for these groups, Kaplan Meier graphs were generated (Fig 2B). This explanation is also given in Fig 2 legend along with cutoff values as well. The reason why these three genes are selected, as explained previously, is because they are common for both Doxorubicin and Etoposide except for Etoposide 6M IC50. As suggested by the reviewer the required explanation has been added to the “Clinical data validation-Log rank with multiple cutoffs (LRMC) and survival analysis” in the MM section. Comment 31: Line332 I recommend to add again the reference for GSE12417 when it is mentioned. Response 31: The reference for GSE12417 was added when needed along with the manuscript. Comment 32: Line334 Doxorubicin needs to be indicated with the first letter in uppercase. Response 32: The correction was done. Comment 33: Line 337 please add in the parenthesis together with Fig2 “see Materials and Methods section”. Response 33: The “see Materials and Methods section” was added. Comment 34: Line 339-341 these lines need to be reformulated. Response 34: The lines were reformulated. Comment 35: Line352 the word “cut-off” needs to be consistent along the text and the numbers need to be rounded. Response 35: All “cutoff” terms were made consistent and the numbers were rounded. Comment 36: Since many S tables are really small I suggest to insert them in a unique pdf file and leave as excel only those that do not fit on a pdf page. Response 36: Except for S1 Table, all Supplementary Tables were converted to PDF files. Comment 37: The script file is missing, it should be provided in a.zip file with all the codes. Response 37: The R Script file has only “one code” and it was provided in the new “S2 Table”. Comment 38: The section Acknowledgments is blank. Response 38: Accordingly, the title was deleted because this section is empty. 10.1371/journal.pone.0242497.r005 Decision Letter 2 Bertolini Francesco Academic Editor 2020 Francesco Bertolini This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 11 Sep 2020 PONE-D-20-09590R2 Renin angiotensin system genes are biomarkers for personalized treatment of acute myeloid leukemia with doxorubicin as well as etoposide PLOS ONE Dear Dr. Turk, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process by Reviewer \#1. Please submit your revised manuscript by Oct 26 2020 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols> We look forward to receiving your revised manuscript. Kind regards, Francesco Bertolini, MD, PhD Academic Editor PLOS ONE \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: (No Response) Reviewer \#2: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Partly Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Turks and colleagues addressed most of the points raised during the last round of revision but one of the most important points, together with some typos/inconsistencies have not been corrected and/or introduced. In order to be accepted I recommend precisely covering all the points raised and to check for possible mistakes newly introduced. Main point: the R script/scripts are still not included in the current version of the manuscript therefore I kindly ask to provide them in one of the following manners: either as file.R /.Rmd in a compressed file (.zip, gzip, tar.gz, ecc ecc) or as a link to a public repository. Currently the only file provided is a 1 x 1 table with the name of the script “SixModelIC50 V3.r” which does not include any code line. Minor points to be addressed: Line 86-87 It is not clear which dataset has been used for data normalization and variance analysis, the name “CGP microarray” combines the “CGP gene expression data” and the “microarray dataset GSE12417”. I please suggest, if the authors intend the “CGP gene expression data”, to use this name. Line 103 in the new version of the manuscript it comes out that both the models and data have the same name “6M IC50”. This intent was not clear from the previous version, since there was a little bit of confusion in the names. Therefore, I suggest to use two different names for model and data (maybe using lowercase letters in one of the two or only 6M when referring to the model while 6M IC50 when referring to the data). Plase, be consistent along the text with this nomenclature when referring to one or to the other. line 131-133: “. We referred to this data as 6M IC50. We used two versions of CGP data, one original CGP data, second is recalculated 6M IC50 data by this script.” Please, reformulate this sentence, it does not seem in the right place and it is not clear. I ask the authors to be consistent with the nomenclature and terms used in the other section. 138-139 the sentence “and observed strong correlations between them for drugs widely used in AML (Cytarabine, Etoposide, Doxorubicin) but not for ATRA (S3 Table).” needs to be moved in the result section. Plus there is a conflict on what it is mentioned in lines 254-256 of the results: “ Using t-test we observed strong correlations between CGP IC50 and 6M IC50, for drugs widely used in AML (Cytarabine, Etoposide, Doxorubicin) but not for ATRA (S3 Table).” The authors need to clarify if they used a t-test or a correlation Pearson test. The table is referring to the Pearson correlation. Line 144: this is the first time the authors refer to the “CGP 6M IC50 data”, please be consistent with the nomenclature as mentioned in the previous paragraph and revision. If the name was only 6M IC50 it needs to be like this, otherwise if a new name is introduced, it needs to be specified before and the authors should explain it. Line 232 It is quoted another R script that is not the same as the one used to calculate the 6M model but it is referred to as the same. I please ask the authors to correct this, and if it is available, to also provide this script together with the previous one. They can be put together in a compressed (.zip, gzip, tar.gz, ecc ecc) file. Line 227-228 these new inserted lines need to be reformulated. “IGF2R, CTSA, ATP6AP2 were selected for clinical correlation studies is because they are common for all regression formulas except for Etoposide 6M IC50.” Likely “is because” is a typo, and this is the first time that the authors indicate “Etoposide 6M IC50”, what are they referring to? Line 302 I kindly ask the authors to revise the use of the article “the” in this location. Have these cells been previously indicated in the text of the results? I suggest to remove the “the” and add (see Materials and Methods section) if the authors agree. Lines 324-326 the authors should clarify how they “ tested E-Cadherin (epithelial marker) and Vimentin (mesenchymal marker) expression in silico”, if this analysis is referring to the S3 Fig, I please the authors to add at the end of this paragraph (S3 Fig). Moreover, the following paragraph (Lines 332-334) should not be separated if the authors are agreed. Finally the t-test is not shown in the S3 Fig and therefore the quote “(S3 Fig)” should be removed from line 334. Fig 2 I please ask the authors to explain also the meaning of the red circle in the figure legend. Moreover, I think that the panel A legend needs to be more clear: I find it difficult to read it and it is not explicative of the figure. Reviewer \#2: Authors have addressed all my concerns thus making manuscript suitable for publication in its present form \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0242497.r006 Author response to Decision Letter 2 12 Oct 2020 Review Comments to the Author Reviewer \#1: Turks and colleagues addressed most of the points raised during the last round of revision but one of the most important points, together with some typos/inconsistencies have not been corrected and/or introduced. In order to be accepted I recommend precisely covering all the points raised and to check for possible mistakes newly introduced. Main point: the R script/scripts are still not included in the current version of the manuscript therefore I kindly ask to provide them in one of the following manners: either as file.R /.Rmd in a compressed file (.zip, gzip, tar.gz, ecc ecc) or as a link to a public repository. Currently the only file provided is a 1 x 1 table with the name of the script “SixModelIC50 V3.r” which does not include any code line. Response: R Scripts were included in the manuscript as a link to a public Github Repository. “S2 Table was removed” Minor points to be addressed: Comment 1. Line 86-87 It is not clear which dataset has been used for data normalization and variance analysis, the name “CGP microarray” combines the “CGP gene expression data” and the “microarray dataset GSE12417”. I please suggest, if the authors intend the “CGP gene expression data”, to use this name. Response 1. “CGP microarray” was changed to “CGP gene expression data”. Comment 2. Line 103 in the new version of the manuscript it comes out that both the models and data have the same name “6M IC50”. This intent was not clear from the previous version, since there was a little bit of confusion in the names. Therefore, I suggest to use two different names for model and data (maybe using lowercase letters in one of the two or only 6M when referring to the model while 6M IC50 when referring to the data). Plase, be consistent along the text with this nomenclature when referring to one or to the other. Response 2. As suggested by the reviewer, we used “6M” when referring “model” and we used “6M IC50” when referring “data”. Comment 3. line 131-133: “. We referred to this data as 6M IC50. We used two versions of CGP data, one original CGP data, second is recalculated 6M IC50 data by this script.” Please, reformulate this sentence, it does not seem in the right place and it is not clear. I ask the authors to be consistent with the nomenclature and terms used in the other section. Response 3. We reformulated and relocated the line. Comment 4. 138-139 the sentence “and observed strong correlations between them for drugs widely used in AML (Cytarabine, Etoposide, Doxorubicin) but not for ATRA (S3 Table).” needs to be moved in the result section. Plus there is a conflict on what it is mentioned in lines 254-256 of the results: “ Using t-test we observed strong correlations between CGP IC50 and 6M IC50, for drugs widely used in AML (Cytarabine, Etoposide, Doxorubicin) but not for ATRA (S3 Table).” The authors need to clarify if they used a t-test or a correlation Pearson test. The table is referring to the Pearson correlation. Response 4. We removed the “and observed strong correlations between them for drugs widely used in AML (Cytarabine, Etoposide, Doxorubicin) but not for ATRA (S3 Table)” from the Materials and Methods section. “T-test” was changed to “Pearson correlation” in the Results section. Comment 5. Line 144: this is the first time the authors refer to the “CGP 6M IC50 data”, please be consistent with the nomenclature as mentioned in the previous paragraph and revision. If the name was only 6M IC50 it needs to be like this, otherwise if a new name is introduced, it needs to be specified before and the authors should explain it. Response 5. “CGP 6M IC50 data” was changed to “6M IC50 data”. Comment 6. Line 232 It is quoted another R script that is not the same as the one used to calculate the 6M model but it is referred to as the same. I please ask the authors to correct this, and if it is available, to also provide this script together with the previous one. They can be put together in a compressed (.zip, gzip, tar.gz, ecc ecc) file. Comment 6. R Scripts were included in the manuscript as a link to a public Github Repository. “S2 Table was removed” Comment 7. Line 227-228 these new inserted lines need to be reformulated. “IGF2R, CTSA, ATP6AP2 were selected for clinical correlation studies is because they are common for all regression formulas except for Etoposide 6M IC50.” Likely “is because” is a typo, and this is the first time that the authors indicate “Etoposide 6M IC50”, what are they referring to? Response 7. We reformulated the line. Comment 8. Line 302 I kindly ask the authors to revise the use of the article “the” in this location. Have these cells been previously indicated in the text of the results? I suggest to remove the “the” and add (see Materials and Methods section) if the authors agree. Response 8. We removed the “the” and added “(see Materials and Methods section)”. Comment 9. Lines 324-326 the authors should clarify how they “ tested E-Cadherin (epithelial marker) and Vimentin (mesenchymal marker) expression in silico”, if this analysis is referring to the S3 Fig, I please the authors to add at the end of this paragraph (S3 Fig). Moreover, the following paragraph (Lines 332-334) should not be separated if the authors are agreed. Finally the t-test is not shown in the S3 Fig and therefore the quote “(S3 Fig)” should be removed from line 334. Response 9. We compared E-Cadherin (epithelial marker) and Vimentin (mesenchymal marker) expression using t-test. We clarified it and showed t-test p value on the S3 Fig. We added “(S3 Fig)” at the end of paragraph and combined with the next paragraph. Comment 10. Fig 2 I please ask the authors to explain also the meaning of the red circle in the figure legend. Moreover, I think that the panel A legend needs to be more clear: I find it difficult to read it and it is not explicative of the figure. Response 10. We simplified the legend, since we already have explanations for this method in methods section and we cited two previous studies. We also explained the meaning of red circle in the legend. 10.1371/journal.pone.0242497.r007 Decision Letter 3 Bertolini Francesco Academic Editor 2020 Francesco Bertolini This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 4 Nov 2020 Renin angiotensin system genes are biomarkers for personalized treatment of acute myeloid leukemia with doxorubicin as well as etoposide PONE-D-20-09590R3 Dear Dr. Turk, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Francesco Bertolini, MD, PhD Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#2: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: (No Response) Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: (No Response) Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: (No Response) Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: (No Response) Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Turk and colleagues substantially improved the work, and I found that the manuscript in its current state can be suitable for publication in PLOSONE journal. I only recommend that just a few very minor typos should be corrected before the final acceptance or the publication of this manuscript: Line 179: the number "2227" needs to be indicate as 2,227. Lines 268-270: please reformulate the sentence. Line 648: the sentence is starting with “and”, I please the authors to correct this. I please the authors be consistent with the name “CGP I50” when referring to it also in the Supp Tables (i.e: S3, S7 Tables). Reviewer \#2: The additional editing performed by Authors have made manuscript fully suitable for publication in PlosOne journal \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No 10.1371/journal.pone.0242497.r008 Acceptance letter Bertolini Francesco Academic Editor 2020 Francesco Bertolini This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 6 Nov 2020 PONE-D-20-09590R3 Renin angiotensin system genes are biomarkers for personalized treatment of acute myeloid leukemia with doxorubicin as well as etoposide Dear Dr. Turk: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Francesco Bertolini Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction The secretory pathway traditionally contains a number of biochemically distinct inter-related membrane organelles that continuously communicate with each other and exchange materials through membrane trafficking. The classical secretory proteins are often extended at their N-terminus by a ‘leader’ or ‘signal’ sequence of 13–30 hydrophobic amino acids. This directs the nascent protein to co-translate and vectorially transfer across the membrane of the endoplasmic reticulum (ER), and is often cleaved before completion of the transmembrane transport of the protein. Secretory proteins are then transported to the Golgi apparatus and trans-Golgi network where they undergo further glycosylation, and sorting and being packaged into vesicles, respectively. Finally the secretory vesicles are delivered to and fuse with the plasma membrane, resulting in releasing their contents into the extracellular space. However, numerous secretory proteins with normal extracellular functions have been shown to be devoid of functional signal sequences and do not appear substrates for the ER membrane translocation machinery. In addition, the secretion of these proteins is not affected by the presence of brefeldin A, a drug that blocks ER/Golgi-dependent secretory transport. These observations suggest that alternative secretory mechanisms that are independent of ER/Golgi secretory pathway exist in eukaryotic cells. Secretion of proteins without an N-terminal signal sequence is currently known as the unconventional/non- classical secretory pathway or leaderless secretion. Up to date, several unconventional secretory pathways have been reported for a few biomedically important factors, including proangiogenic mediators such as fibroblast growth factors 2 and inflammatory cytokines such as interleukin 1α and 1β in mammalian cells. Plant secretome revealed that more than half of the total identified proteins were leaderless secretory proteins, which is distinctly higher than in human and yeast secretomes, implying that this unconventional secretory mechanism is common to all eukaryotes and it is more largely used than in other eukaryotes. Furthermore, plants exposed to biotic and abiotic stresses usually significantly contained more leaderless secretory proteins in the extracelluar space than non-stressed plants, suggesting that environmental component might be involved in release of leaderless secretory proteins into the extracelluar space. However, until now, only one leaderless secretory protein, mannitol dehydrogenase (MTD) in celery, has been shown to bypass the ER-Golgi-plasma membrane exocytic pathway for its delivery to the extracellular space by molecular biology and biochemistry approaches. Synaptotagmins (SYTs) constitute a family of membrane-trafficking proteins that are characterized by an N-terminal transmembrane region, a linker of variable size, and two C-terminal C2 domains in tandem. SYTs are reported to play a vital role in neurotransmitter release and insulin exocytosis in mammalian cells.The synaptotagmin family in *Arabidopsis* has five members. SYT1, the only one characterized so far, is ubiquitously expressed and predominantly localized to the plasma membrane. Disruption of *SYT1* function in *Arabidopsis* leads to abiotic stresses hypersensitivity due to a reduced integrity of the plasma membrane. However, the subcellular localization and the functions of other SYTs remain unknown. Hygromycin B is an aminoglycoside antibiotic produced by *Streptomyces hygroscopicus* that is active against both prokaryotic and eukaryotic cells by inhibiting protein synthesis. It has been reported that hygromycin B acts by interfering with translocation and causes mistranslation. An *Escherichia coli* gene has been identified that confers resistance in transgenic plants against hygromycin B. The resistance gene codes for hygromycin B phosphotransferase (HYG^R, E.C. 2.7.1.119) that adds phosphate to position 7 of the destomic acid ring of hygromycin B, which results in complete loss of biological activity both *in vitro* and *in vivo*. Although HYG^R has been mainly used as a positive selective marker for transgenic cells, few studies have examined the subcellular localization and trafficking of HYG^R and the putative elements that regulate the tolerance of *HYG^R*-expressing cells to hygromycin B. Here, we provided several lines of evidence about localization of *Arabidopsis* synaptotagmin SYT2. More importantly, we found that HYG^R is present both in the cytoplasm and the extracelluar space in *HYG^R*-*GFP*-transgenic plants. The loss of SYT2 caused inhibition of HYG^R-GFP trafficking. Based on the fact that HYG^R-GFP lacks a signal sequence and its secretion is not sensitive to brefeldin A treatment, we propose that HYG^R-GFP is not secreted via the conventional secretory pathway and SYT2 plays an important role in regulating the unconventional protein trafficking from the cytosol to the extracelluar matrix in plant cells. # Results ## Characterization of *Arabidopsis* SYT2 Protein The synaptotagmin 2 gene (*SYT2*, *At1g20080*) is one of five putative *SYTs* in *Arabidopsis*. It comprises 12 exons and 11 introns, based on information available in the *Arabidopsis* Information Resource database (TAIR; <http://www.arabidopsis.org/>). Homology analysis using amino acid sequence data showed that SYT1 is the closest relative of SYT2 in *Arabidopsis*, with about 66% amino acid identity between them. Compared to SYT1, all amino acid residues thought to play crucial roles in coordinating calcium (Ca) ions are conserved in the C2A domain of SYT2. Unlike SYT1, however, only four putative amino acids of the SYT2 C2B domain are involved in Ca binding (lacking the fourth putative residue). According to *SYT1* expression profiles based on microarray expression data obtained from Geneinvestigator (<http://www.genevestigator.ethz.ch>), *SYT2* is highly expressed in pollen grains, whereas expression in other organs, such as roots or leaves, is detectable but low. A secretory signal peptide was predicted in the SYT2 amino acid sequence but neither a chloroplast transit peptide nor a mitochondrial targeting peptide was identified using the TargetP 1.1 server (<http://www.cbs.dtu.dk/services/TargetP/>). ## SYT2 does not Colocalize with BFA Compartment To investigate the subcellular localization of SYT2, we fused the gene encoding green fluorescent protein (GFP) to the *SYT2* gene (C-terminal end of the encoded protein) under the control of the 35S promoter of the cauliflower mosaic virus (*CaMV35S*) and used these constructs to transiently or stably transform *Nicotiana tabacum* and *Arabidopsis*. The resulting fusion protein (SYT2-GFP) was primarily detected in mobile punctate structures in leaf cells of transiently transformed *N. tabacum* and *Arabidopsis*. Plant lines stably expressing SYT2-GFP were also analyzed by laser scanning confocal microscope (LSCM) to localize the fusion protein. The fluorescence signals appeared as punctate structures with a dim cytosolic background in root hairs, root meristem cells and elongation cells. To investigate whether the SYT2-positive structures were of endosomal origin, transgenic SYT2-GFP *Arabidopsis* seedlings were incubated with FM4-64, a fluorescent marker internalized by a clathrin-dependent process and sequentially labels early endosomal, late endosomal, and vacuolar compartments. As shown in, the internalized FM4-64 dye rarely co-localized with SYT2-GFP-containing compartments in root cells even after 2 h of incubation, during which time FM4-64 was detected in vacuolar membranes. Co-localization studies were also performed using seedlings expressing VHA-a1-GFP, ARA6-GFP, and ARA7-GFP, all of which have been reported to reside on endosomes and regulate endosomal fusion. Co-localization of FM4-64 with large amounts of VHA-a1-GFP and ARA6-GFP and lesser amounts of ARA7-GFP, was detected after 30 min, demonstrating that SYT2-GFP is targeted to a compartment independent of endosomal membranes. The endosomes in *Arabidopsis* root tips are the main target of the fungal toxin brefeldin A (BFA). This drug inhibits certain ADP ribosylation factor/guanine nucleotide exchange factors (ARF-GEFs) and causes the endocytic tracer FM4-64 to rapidly aggregate throughout vesicle agglomerations known as BFA compartments. To investigate whether SYT2-GFP was associated with BFA-sensitive endosomes, the transgenic plants were treated with 25 µM BFA. After BFA treatment, the punctate SYT2-GFP structures were almost intact and did not accumulate in BFA compartments ( to), while VHA-a1-GFP (early endosome marker) and ARA6-GFP (late endosome marker) perfectly overlapped with and was located at the periphery of the BFA compartments, respectively ( to),. To further demonstrate that SYT2-GFP-containing structures are excluded from the late endosomes, we analyzed the effect of wortmannin, which inhibits the biosynthesis of phosphatidylinositol 3- and 4-phosphates as well as phospholipids in plant cells. Exogenous application of wortmannin causes the late endosomes to dilate or form ring-shaped structures, but has no effect on the Golgi apparatus and early endosomes. The morphology of SYT2-GFP structures was not altered and the two markers were almost separated ( to). As previously reported for ARA6-GFP and ARA7-GFP, both of which localize on the late endosomes, wortmannin caused the formation of ring-shaped structures ( to). Furthermore, SYT2 protein did not colocalize with ARA7-GFP by immunofluorescent labelingusing anti-SYT2 and anti-GFP antibodies ( to). Taken together, these data demonstrate that SYT2-GFP does not localize on the late endosomes. ## SYT2 is Localized on the Golgi Apparatus The punctate structures labeled by SYT2-GFP were insensitive to BFA and wortmannin treatment and did not become labeled with FM4-64, reminiscent of the Golgi apparatus. In addition to its punctate appearance, the Golgi apparatus did not become co-localized with the endocytic tracer FM4-64 or with BFA compartments in *Arabidopsis* root-tip cells. To determine if SYT2-containing structures were associated with Golgi apparatus, we performed an immunofluorescent study on wild-type and SYT2-GFP-overexpressing plants. Antibodies were generated against the cytoplasmic region (300 aa–535 aa) of SYT2 (anti-SYT2). Affinity purified anti-SYT2 antibodies was analyzed by western blotting of wild type and SYT2-GFP transgenic *Arabidopsis* seedlings, recognizing proteins of 61 KD and 87 kD corresponding to both native SYT2 and recombinant SYT2-GFP respectively. In order to determine whether the generated antibodies were specific for SYT2 and did not recognize other SYT family members, including the closely related SYT1, a mutant in *SYT2* from the SALK collection (SALK_135307, *syt2-1*) with the T-DNA located in the 9th exon of *At1g20080* was isolated and further analyzed. No mRNA transcripts and proteins were detected in the homozygous *syt2-1* line, despite the fact that *SYT1* was expressed at wild-type levels in the mutant. In view of the lower *SYT2* transcripts in vegetative tissues analyzed by microarray ananlysis, it is suggested that SYT2 protein production is probably regulated at the level of translation. As shown in, SYT2 became immunolocalized into punctate structures were similar to those observed in plants expressing SYT2-GFP. Furthermore, most of the co- localization between SYT2 and SYT2-GFP occurred in transgenic plants over- expressing SYT2-GFP, as confirmed using anti-SYT2 and anti-GFP antibodies. SYT2 localization was further analyzed by double-immunofluorescent labeling with anti-SYT2 and anti-GFP antibodies in plants expressing ST-YFP, a well described Golgi marker. SYT2 likewise co-localized with ST-YFP. Immuno-labeling of ultra- thin sections of *Arabidopsis* root cells using anti-SYT2 antibodies showed that gold particles mainly deposited on the Golgi apparatus. ## Co-expression of *SYT2* and *HYG^R* Leads to Hypersensitivity to Hygromycin B Hygromycin B is an aminoglycoside antibiotic produced by *Streptomyces hygroscopicus* that is active against both prokaryotic and eukaryotic cells. The hygromycin B phosphotransferase (HYG^R) phosphorylates and inactivates hygromycin B, and has been widely used as a selectable marker in the generation of transgenic plants. To investigate the subcellular localization of SYT2 in plant cells as described above, the binary vector (pCambia1301-SYT2/HYG^R), which was generated by pCambia1301 from T-DNA containing a SYT2-GFP expression cassette and a hygromycin B-selectable marker, was introduced into *Arabidopsis* seedlings by *Agrobacterium*-mediated transformation. However, we noticed that the positively transgenic plants (named *SYT2/HYG^R*) grew weakly when selected on 20 µg/mL hygromycin B-containing medium. These plants had low viability or showed slow growth and the apparent loss of apical dominance (inhibition of the primary inflorescence) following the development of two symmetrical axillary buds after their transfer to soil ( to). T2 and T3 seedlings exhibited wild-type growth in the absence of hygromycin B. To observe whether SYT2 was associated with uptake of hygromycin B, wild-type and *syt2-1* plants were incubated on ½ Murashige and Skoog (MS) medium containing different concentrations of hygromycin B. As shown in, the phenotype of *syt2-1* is not obviously different from that of wild type under hygromycin B treatment, indicating that SYT2 is not probably related with the uptake of hygromycin B in *Arabidopsis*. Therefore, it is presumed that SYT2 contributes to the detoxification of *HYG^R* in *HYG^R*-containing plants. To investigate the effect of SYT2 on hygromycin B tolerance, wild-type and *syt2-1* plants were transformed with a *35S-HYG^R* construct and the transgenic plants were named as *HYG^R* and *syt2-1/HYG^R*, respectively. We analyzed the expression of the *HYG^R* gene in these transgenic lines using semi-quantitative RT-PCR. The expression level of *HYG^R* gene was similar in the selected lines. We further investigated the growth of these lines on ½ MS medium agar plates with different concentrations of hygromycin B. The growth of the primary roots and hypocotyls of *SYT2/HYG^R* seedlings was greatly inhibited even on the medium containing as low as 5 µg/mL hygromycin B. The roots of *SYT2/HYG^R* seedlings were 30.1%, 9.3% and 7.1% of that of *HYG^R* seedlings in the presence of 5, 10 and 20 µg/mL of hygromycin B, respectively. Microscopic observation also revealed that the root hairs and roots of *SYT2/HYG^R* seedlings were greatly shortened ( to), suggesting that a reduced function of HYG^R caused by the over-expression of *SYT2*. ## HYG^R-GFP is Exported via an Unconventional Secretory Pathway To obtain the clues as to why co-expression of SYT2-GFP and HYG^R caused hypersensitivity to hygromycin B in *Arabidopsis*, we first analyzed the subcellular localization of a translational fusion between HYG^R and GFP under the control of the constitutive promoter (*CaMV35S*) in stable transgenic lines. The fluorescent signals from HYG^R-GFP were present on the cell surface of root cells and interestingly HYG^R-GFP was preferentially expressed in leaf-tip zones. In the plasmolyzed cells, HYG^R-GFP was found in the cytoplasm as well as in the cell walls. As it has been mentioned that the HYG^R protein could be secreted in plants, we first investigated the secretory property of the HYG^R-GFP protein by protein gel blot using mesophyll protoplasts of *HYG^R-GFP* plants. Unexpectedly, in protoplast lysates, the expected band of full length HYG^R-GFP (about 65 kD) was not detected when anti-HYG^R antibody was applied, but a band with a molecular weight between 34–43 kD (about the molecular weight of HYG^R protein) was detected both in the medium and in the protoplast lysates, suggesting that HYG^R protein was secreted into extracellular space. When anti-GFP antibody was applied, the HYG^R-GFP protein appeared as a single band of approximately 30 kD, slightly bigger than GFP, in both the medium and the protoplast lysates of *HYG^R-GFP*-expressing plants. Using tubulin as an intracellular marker, it was found that contamination of the medium with intracellular proteins was below the level of detection. These results suggest that HYG^R-GFP had been efficiently truncated at carboxyl terminus of HYG^R shortly after it was synthesized in *HYG^R-GFP*-expressing plants. To assess whether HYG^R-GFP was secreted via the conventional secretory pathway, protoplasts were treated with BFA. Although the Golgi apparatus in *Arabidopsis* root tissues appears to be BFA-resistant, it turns out that BFA indeed exerts a marked effect on the Golgi apparatus in non-root tissues of *Arabidopsis*. Confocal microscopy revealed that the classic re-absorption of Golgi membranes back into the ER in BFA-treated *Arabidopsis* leaves. Furthermore, the secretion of acid phosphatase inhibited by BFA treatment was also reported in mesophyll protoplasts of tobacco, indicating that BFA inhibits the conventional secretory pathway in leaf cells. In order to analyze whether BFA affected the secretion of HYG^R-GFP by immunoblot analysis, the total protein of protoplast lysates and the medium was harvested after 5-h BFA treatment, respectively. As shown in, HYG^R was detected in similar amounts in the absence and presence of BFA in the protoplast lysates or in the medium, respectively. The effectiveness of BFA on the conventional ER/Golgi pathway was verified by measuring the activity of acid phosphatase (AcPase) at hourly intervals in the medium and protoplast lysates. As expected, an obvious inhibition of AcPase secretion upon BFA treatment was observed at each individual measurement time as previously reported. The facts that BFA inhibited AcPase secretion but did not inhibit secretion of HYG^R-GFP suggested that HYG^R-GFP secretion indeed follows an alternative secretory pathway. ## SYT2 is Required for the Unconventional Secretion of HYG^R To further examine whether SYT2 was involved in the unconventional secretory process of HYG^R-GFP in *Arabidopsis*, the HYG^R-GFP was introduced into *syt2-1* plants by *Agrobacterium*-mediated transformation and the resultant transgenic plants (*syt2-1/HYG^R-GFP*) had similar phenotype to *syt2-1/HYG^R* plants under hygromycin B treatments. As shown in, expression of HYG^R-GFP resulted in an increase in GFP fluorescence owing to intracellular accumulation of HYG^R-GFP. HYG^R-GFP accumulated in whole leaf cells besides in leaf-tip zones. When we examined the tissues at higher resolution by LSCM, the fluorescence signals were found on punctate structures in cytoplasm. After being plasmolyzed, *syt2-1/HYG^R-GFP* plants showed that fluorescence signals of *HYG^R-GFP* were primarily localized on intracellular punctate structures and vacuoles. To characterize the fluorescent proteins in *syt2-1/HYG^R-GFP* plants, total proteins extracted from protoplasts and medium were analyzed by protein gel blot using anti-GFP antibody. As shown in, the total protein in the medium contained no detectable tubulin, indicating that the medium was not obviously contaminated by protoplast proteins. HYG^R-GFP in the medium extracts of *syt2-1/HYG^R-GFP* protoplasts similarly exhibited a single band that co-migrated with the GFP- fusion protein in the extracts from *HYG^R-GFP* protoplasts and medium. However, the protoplast extracts had three forms of GFP fusion protein in *syt2-1/HYG^R-GFP* plants. The greatest band migrated with an apparent molecular weight of about 55 kD which was lower than the expected full- length of HYG^R-GFP. Apart from this upper GFP fusion protein, two less intense bands with the molecular weight of about 43 kD and 30 kD were recognized below it, implying that HYG^R-GFP had undergone partial truncation at its amino terminus with different extents in the *syt2-1/HYG^R-GFP* plants. Immunogold-labeled ultrathin sections for electron microscopy showed the gold particles situated on the cell wall both in concentrated and dispersed manner in the root cells of *HYG^R-GFP*-expressing plants. Little or no gold particles were detected on the cell wall in the root cells of the *syt2-1* plants expressing *HYG^R-GFP*. However, several gold particles well deposited close to, or in the vacuoles in these cells. No obvious signals were found in the vacuoles of the *HYG^R-GFP* transgenic plants or in the whole cells of non- transformed plants. It was of interest to note that *syt2-1/HYG^R* seedlings also exhibited the inhibition of root elongation under higher hygromycin B treatments (\>5 µg/ml). To confirm that the sensitivity of *syt2-1/HYG^R* to hygromycin B in root tip growth is caused by the deficiency of SYT2, the binary construct containing *SYT2-GFP* and *HYG^R* was introduced into *syt2-1* mutants by *Agrobacterium*-mediated transformation. T3 progeny were subjected to hygromycin B and it was found that the *SYT2/HYG^R* plants restored the *syt2-1* phenotype to the *HYG^R transgenic* plants with respect to the root elongation and root morphology. These results confirmed that the deficiency of SYT2 in *syt2-1* resulted in the increased sensitivity to higher concentrations of hygromycin B during *Arabidopsis* seedling growth. ## *SYT1* Expression is Up-regulated in Hygromycin B-treated *syt2-1* Mutant The detoxifying ability of HYG^R in the *syt2-1* mutant was reduced under higher concentrations of hygromycin B when compared with that in the wild- type plants, but was much higher than that in SYT2-overexpressing plants. We hypothesized that the other members of SYT family in *Arabidopsis*, especially the SYT1, which has the highest homology with SYT2, might contribute to the decreased resistance of *syt2-1* to hygromycin B. To address this possibility, the expression level of *SYT1* in hygromycin-treated *syt2-1* plants was examined by semi-quantitative RT-PCR. Under normal condition, *SYT1* is expressed at similar level in *syt2-1* to that in wild-type plants. However, the expression of *SYT1* was greatly enhanced in *syt2-1* plants under hygromycin B treatment for 3 h and 15 h. To further examine whether the up-regulated expression level of *SYT1* in *syt2-1* has a role of enhancing the sensitivity to hygromycin B, we investigated the phenotype of *syt1-2* (*SYT1* knock-out mutant, Schapire et al., 2008) and *SYT1*-overexpresing plants both which contain *HYG^R* gene. From, it is evident that co-expression of *SYT1* and *HYG^R* led to hypersensitivity to hygromycin B in *Arabidopsis*. # Discussion Like most proteins involved in vesicular trafficking, the localization of synaptotagmins provides important information about the biological functions of these proteins. Thus, we firstly investigated the subcellular distribution of SYT2 using different approaches. In all cases, SYT2 was detected in punctate structures, raising the possibility that it was targeted to the membrane trafficking pathway. This was further supported by the results obtained from immunolocalization studies using anti-SYT2 and anti-GFP antibodies. These studies showed that SYT2 was broadly distributed on the Golgi apparatus. This result is analogous to mammalian Syt 4, which localizes to the Golgi apparatus in undifferentiated neuroendocrine PC12 cells. However, the localization of SYT2 is in contrast to the plasma membrane localization of SYT1 in *Arabidopsis*. It has been long appreciated that the Golgi apparatus forms the heart of the secretory pathway and it is where secretory materials are posttranslationally modified before being sorted for delivery to their final destination, such as the plasma membrane or extracellular space,. Indeed, in plant cells, some Golgi-localized proteins have been shown to be involved in the transport of cargo from the Golgi apparatus to the cell surface. Therefore, the localization of SYT2 on the Golgi apparatus in *Arabidopsis* suggests a role in the secretory pathway. HYG^R has been shown to be effective in selection with various plant species, including dicots, monocots and gymnosperms. However, the data available at present appear insufficient to provide complete knowledge of mechanism of HYG^R secretion. In the present experiment, HYG^R-GFP was present in both intracellular and extracelluar space, suggesting that HYG^R-GFP may be excreted from cytosol into the extracellular space. Furthermore, anti-GFP antibodies recognized a band with a molecular weight of about 30 kD in the protoplast lysates from *HYG^R-GFP* plants, which was slightly greater than the full-length GFP, implying that *HYG^R-GFP* was truncated at the carboxyl terminus of HYG^R shortly after its synthesis and HYG^R-GFP was secreted in its truncated form. Interestingly, co-expression of *HYG^R* and *SYT2* in *Arabidopsis* caused hypersensitivity to hygromycin B, suggesting that SYT2 may have a role in regulating the detoxification of HYG^R for hygromycin B. To confirm whether SYT2 is involved in the trafficking of HYG^R, we examined the existing form of HYG^R-GFP in *syt2-1* mutant. We found that the loss of SYT2 partially inhibited the truncation of HYG^R-GFP at the carboxyl terminus of HYG^R, which subsequently accumulated in intracellular punctate structures and vacuoles in several truncating forms, suggesting SYT2 has a vital role in regulating the trafficking of HYG^R-GFP for its secretion in plant cells. Proteins can be secreted in plant cells via either the conventional or the unconventional secretory pathway. The unconventional secretory proteins not only lack of canonical signal sequence, but are also resistant to the export processes affected by BFA, an inhibitor of ER/Golgi-dependent protein secretion in both animals and plants. Interestingly, no conventional signal peptide sequence was found in HYG^R predicted by SignalP 3.0 Server (<http://www.cbs.dtu.dk/services/SignalP/>) or TargetP 1.1 Server (<http://www.cbs.dtu.dk/services/TargetP/>). Therefore, it is possible that HYG^R-GFP was synthesized on the free ribosomes in cytoplasm and exported by a signal peptide-independent secretory process. This unconventional secretion has been thoroughly demonstrated in mammalian and yeast cells. We further analyzed the secretion of HYG^R-GFP protein in *HYG^R*-*GFP*-expressing plants in response to BFA treatment by protein immunoblot. As expected, upon BFA treatment, HYG^R-GFP secretion in the transgenic *Arabidopsis* was not perturbed. Thus HYG^R-GFP secretion displays the features of leaderless or unconventional protein secretion, including the absence of a canonical signal peptide in the protein and the insensitive export in the presence of brefeldin A, Therefore, we can safely conclude that an unconventional secretion is involved in *HYG^R*-transgenic *Arabidopsis* plants. It is unexpected that the detoxifying ability of HYG^R in the loss-of- function *syt2-1* mutant was also destroyed, causing the plants to grow slowly and weakly under higher concentrations of hygromycin B, although these plants showed stronger resistance compared with *SYT2/HYG^R* ones. The most probable explanation for this phenomenon is that the trafficking of HYG^R-GFP in *syt2-1* plants is inhibited and the protein is partly transported into vacuoles, as revealed by the localization of HYG^R-GFP in this mutant, and may then be degradated. Co-expression of *HYG^R* and *SYT1*, the latter, the most similar member to *SYT2* in *Arabidopsis SYT* family, also led to hypersensitivity to hygromycin B. In any case, this result provided direct evidence that the contributor to the weakened resistance of *syt2-1* to hygromycin B might be SYT1. We further found that the transcriptional expression of *SYT1* in *syt2-1* plants remained unchanged under normal conditions, but obviously enhanced under hygromycin B treatment. The SYT1 knock-out mutants also showed sensitivity to hygromycin B even at lower concentration (5 µg/mL), although they have much stronger tolerance than *SYT1/HYG^R* plants. Considering the similar responses of SYT2 and SYT1 to hygromycin B, we conclude that SYT1 may contribute to the resistance of *HYG^R*-harboring plants via a different secretory route. Unconventional secretion can be classified into non-vesicular and vesicular mechanisms. Non-vesicular mechanisms are based on direct translocation of cytoplasmic proteins across the plasma membrane via a specific plasma membrane ATP-binding cassette transporter or some lipids, such as phosphatidylinositol 4,5 bisphosphate \[PI(4,5)P₂\] in the inner leaflet of the plasma membrane.Vesicular mechanisms of unconventional secretion involved multi- vesicular bodies and exosomes that need to fuse with plasma membranes to release cargo into the extracellular space. In our study, SYT2 was not presented on the multivesicular bodies (PVC in plant cells), indicating that SYT2 protein may regulate the unconventional secretory pathway by a distinct manner from the multivesicular body-mediated secretion of exosome in mammalian cell. However, SYT2 is not the only Golgi-localized protein that regulates unconventional secretion. Golgi-localized protein GRASP (Golgi reassembly stacking protein) in *Dictyostelium discoideum*, is also required for Golgi-independent cell-surface transport of a non-signal-peptide-containing protein, acyl-CoA binding protein (AcbA), which triggers terminal differentiation of spore cells. In *Drosophila melanogaster*, GRASP modulates Golgi-independent cell surface transport of α intergrin. In a *D. melanogaster grasp* mutant, the α integrin subunits are not properly deposited at the plasma membrane and instead retained intracellularly. From sequence comparison of all the available genomes, it was revealed that plants lack a bona fide GRASP homolog. Very recently, an *Arabidopsis* protein Exo70E2 was found to be present in some double membrane structures (named EXPO) and did not colocalize with the Golgi apparatus, the TGN or PVC. Exo70E2 served to release a leaderless protein (SAMS2) into the extracelluar space, indicating that there may be diverse proteins which localize on different organelles and modulate the release process of unconventional proteins in plant cells. Therefore, SYT2 is the first protein, to our knowledge, that resides on Golgi apparatus and regulates unconventional protein secretion in plants. # Methods ## Plant Material and Growth Conditions Seeds expressing ARA6-GFP and ARA7-GFP were kindly provided by Takashi Ueda and Thierry Gaude. Construct of ST-YFP was kindly made available by Jingbo Jin. *Arabidopsis* mutant *syt2-1* (SALK_135307) was obtained from the *Arabidopsis* Biological Resource Center at Ohio State University. Other transgenic plants were generated based on the protocol in. *Arabidopsis* seeds were pretreated in 70% ethanol for 5 min, surface-sterilized in 50% bleach for 1 min, and washed with sterile distillated water at least five times. Seeds were planted on 1% agar containing ½ MS salts with or without the indicated concentrations of hygromycin B, allowed to imbibe for 3 days at 4°C, and germinated in a vertical orientation. Seedlings were grown at 22±3°C under a 16-h light/8-h dark regime. Experiments were performed using 3- to 4-day-old seedlings for microscopic observation, or 7- to 10-day-old seedlings for measurement of root and shoot lengths. ## Antibody Preparation and Protein Gel Blot Analysis For protein gel blot analysis, a polyclonal antibody was raised against a truncated form of SYT2. The C-terminal region of SYT2 (235 amino acid residues) and the full length of HYG^R were expressed in *E. coli* respectively as recombinant proteins using the expression vector pET28b (Invitrogen, Carlsbad, CA). The recombinant proteins were expressed and purified according to the manufacturer's protocol, and the purified proteins were injected into a rabbit to raise antibody according to a published protocol. The polyclonal antibody was purified according to Park et al.. Monoclonal anti-GFP antibodies were purchased from Sigma-Aldrich (St. Louis, MO). Total protein extracts were obtained by grinding 100 mg of wild-type, *syt2-1*, or SYT2-GFP-overexpressing plants in protein extraction buffer \[20 mM Tris-HCl, pH 7.5, 5 mM ethylenediaminetetraacetic acid (EDTA), 5 mM ethylene glycol tetraacetic acid (EGTA), 10 mM dithiothreitol (DTT), 0.05% sodium dodecyl sulfate (SDS), and 1 mM phenylmethylsulfonyl fluoride (PMSF)\]. The extracts were spun for 10 min at 4°C, and the resulting supernatant loaded on a SDS-PAGE gel with loading buffer. For HYG^R protein hybridization, mesophyll protoplasts were prepared from the leaf tissues of 3- to 4-week-old *Arabidopsis* plants which stably expressed HYG^R-GFP protein. After being washed five times with W5 solution (154 mM NaCl, 125 mM CaCl₂, 5 mM KCl, 2 mM 4-Morpholineethanesulfonic acid, pH 5.7), the protoplasts were incubated with 25 µM BFA for 5 hours. At the end of the incubation, the medium and the protoplasts were collected respectively. The protoplast proteins were extracted as described by Wu et al. The medium proteins were precipitated by trichloroacetic acid method and resolved in the SDS-PAGE loading buffer. The samples were boiled for 10 min and loaded on polyacrylamide gel. After electrophoresis, the separated proteins were transferred to a nitrocellulose membrane for 2 h. The nitrocellulose membrane was then incubated in a 1∶800 dilution of anti-SYT2, 1∶500 anti-HYG^R, 1∶1000 anti- tubulin or 1∶4000 anti-GFP antibodies in phosphate-buffered saline (PBS) buffer (pH 6.9). Horseradish-peroxidase-conjugated secondary antibody (Sigma-Aldrich) was used at a 1∶5000 dilution, and the results were interpreted using an enhanced chemiluminescence detection system, with visualization by enhanced chemiluminescence detection reagents (Applygen Technologies Inc., Beijing, China) according to the manufacturer's recommendations. ## Fluorescent Dye and Treatments with BFA and Wortmannin To visualize putative endosomes, seedlings were mounted in ½ MS liquid with 3 µM FM4-64 \[Invitrogen; T13320; diluted from a 3 mM stock solution in dimethyl sulfoxide (DMSO)\] on slides for a specified time. For BFA treatment, seedlings were incubated in ½ MS liquid containing 25 µM BFA diluted from a 50 mM stock solution in DMSO and then mounted on slides in the presence of BFA. For the wortmannin treatment, seedlings were incubated in ½ MS liquid containing 20 µM wortmannin diluted from a 20 mM stock solution in DMSO for 1 h before observation. ## Immunofluorescent Labeling Four-day-old seedlings were fixed in 4% paraformaldehyde in PEM buffer (50 mM PIPES, 5 mM EGTA, and 5 mM MgSO₄, pH 6.9) for 1 h at room temperature, followed by washing with 0.1 M glycine in PEM buffer. Fixed cells were partially digested with 2% (w/v) driselase (Sigma-Aldrich) for 30 min at 37°C. The plasma membrane was permeabilized with 0.3% Triton X-100 and 10% DMSO in PBS for 1 h at room temperature. Seedlings were incubated in blocking solution for 1 h at room temperature and then incubated with primary antibodies of anti-SYT2 (1∶50) or anti-GFP (1∶200; Sigma-Aldrich) again for 1 h at room temperature. Primary antibodies were washed out with blocking solution three times for 5 min and the seedlings then incubated with fluorochrome-conjugated secondary antibodies in the dark at 37°C for 3 h. Secondary antibodies (purchased from Sigma-Aldrich) were used at the following concentrations: fluorescein isothiocyanate-conjugated anti-rat IgG, 1∶100; fluorescein isothiocyanate-conjugated anti-rabbit IgG, 1∶100; rhodamine (TRITC)-conjugated anti-rabbit IgG, 1∶100. ## Fluorescence Microscopy Fluorescence microscopy was performed using a TCS SP5 confocal laser-scanning microscope (Leica, Oberkochen, Germany). All LSCM images were obtained using the Leica Confocal software and a 63× water-immersion objective. GFP or GFP/FM4-64 was excited at 488 nm and emission was detected between 500 and 530 nm for GFP and between 620 and 680 nm for FM4-64. To visualize GFP/RFP, GFP and RFP were excited at 488 nm and 543 nm, respectively, and emission detected at 500 and 530 nm for GFP and between 565 and 600 nm for RFP. Images were edited using the LAS AF Lite image browser (Leica) and Adobe Photoshop CS3 (Adobe Systems, San Jose, CA). ## Immunoelectron Microscopy For immunogold labeling of SYT2 and HYG^R, roots of *Arabidopsis* were fixed with 4% paraformaldehyde and 1% glutaraldehyde for 4 h and then were embedded in LR White resin (Sigma) and polymerized by heat. Ultrathin sections were obtained and transferred to nickel grids that were then blocked with 5% BSA and incubated subsequently with the primary anti-body (anti-SYT2, 1∶200; anti- HYG^R, 1∶200) at 37°C for 1 h. After five washes with PBS for 20 min, the sections were treated with the secondary antibody (goat anti-rabbit IgG coupled to 10-nm gold particles, Sigma, 1∶50) at 37°C for 1 h. Finally, the sections were stained with 2% uranyl acetate for 10 min and observed under JEM-1230 TEM (JEOL). ## Acid Phosphatase Assay The activity of the AcPase was measured according to Pfeiffer by measuring the release of *p*-nitrophenol (*p*NP) from *p*-nitrophenyl phosphate (*p*NPP). Samples of 200 µl were incubated with 200 µl of reaction buffer containing 40 mM MES-Tris, pH 5.5, 5 mM *p*NPP, and 10 mM MgCl₂, for 45 min at 30°C. The reaction was stopped by the addition of 5 ml of 40 mM NaOH, and the concentration of *p*NP was determined at 405 nm wavelength. All assays were performed as triplicate. # Supporting Information [^1]: Conceived and designed the experiments: HZ LZ LJ JL. Performed the experiments: HZ LZ BG HF. Analyzed the data: HZ LZ BG JJ. Contributed reagents/materials/analysis tools: JJ MB LJ. Wrote the paper: HZ LZ JL. [^2]: The authors have declared that no competing interests exist.

# Introduction Cistanches Herba (Rou Cong Rong), known as “Ginseng of the desert”, originates from dried succulent stems of *Cistanche deserticola* Y.C. Ma and *Cistanche tubulosa* (Schrenk) Wig according to the Chinese Pharmacopoeia (2010 edition), and is popular for its tonifying the kidney-yin, benefiting life essence and relaxing bowel. Currently, Cistanches Herba is mainly distributed in arid and warm deserts in northwest China, particularly in Xinjiang and Inner Mongolia provinces. However, the two origins of Cistanches Herba differ in terms of their pharmacological activity and chemical components. Tu et al. investigated the decoction of three *Cistanche* species (*C. deserticola*, *C. tubulosa*, *Cistanche salsa*) and found that *C. tubulosa* showed the lowest effect in the Yang-deficiency mouse model. Zhang et al. compared pharmacological activity between *C. deserticola*, *C. tubulosa* and *C. salsa*, and found that these species had medicinal functions such as anti-fatigue and hypoxia tolerance, but not on the same extent. Previous research reported the chemical component, and indicated the difference of chemical component and content for plant origins of Cistanches Herba. As for the clinical application and market circulation,as a tonic,*C. tubulosa* has been traditionally used as a blood circulation-promoting agent and in the treatment of impotence, sterility, lumbago, body weakness in Japan. Consequently, it is of great significance to discriminate two origins of Cistanches Herba for the quality control and clinical application. However, there is no research focus on discrimination of two origins of Cistanches Herba. Many researched methods, including microscopy, ultraviolet and infrared detection, inter-simple sequence repeats method have been used to identify the genus of Cistanches, but not only for two origins specially. Here, we conjunctively utilized chemical and molecular techniques to distinguish two origins of Cistanches Herba, UPLC-QTOF/MS (ultra-performance liquid chromatography coupled with quadrupole time-of-flight mass spectrometry) and DNA barcoding. UPLC-QTOF/MS provides information more rapidly and efficiently compared with other techniques. The high selectivity and sensitivity of UPLC- QTOF/MS have resulted in its application for both quantitative and qualitative analyses, as well as in metabolite analysis and identification of complex compounds in Traditional Chinese Medicine. Principal component analysis (PCA) and orthogonal projection to latent structure-discriminant analysis (OPLSDA) are also developed to identify potential marker compounds. DNA barcoding, an easier and more universal molecular marker technology, uses a DNA fragment to identify species or genera. It is objective, more accurate, and easier to perform than traditional identification methods and other molecular marker technologies. Moreover, DNA barcoding has successfully been applied to identify animal and plant, including medicinal plants. The purpose of this research is to establish a scientific method system, combined UPLC-QTOF/MS and DNA barcoding, for discrimination of two plant origins of Cistanches Herba. # Materials and Methods ## Ethics statement We confirm that the field studies did not involve endangered or protected species. GPS coordinates have included in the sample information, please see." ## Plant materials and reagents Succulent stems of Cistanches Herba were collected from wild desert region in Inner Mongolia, Qinghai Provinces, Xinjiang Uygur Autonomous Region, People's Republic of China in May 2012. The samples of the research were all collected in wild desert region, not in private land, where no specific permissions were required. The botanical identities of the stems were confirmed by Dr. Linfang Huang. Voucher specimens were deposited at The Institute of Medicinal Plant Development. High-performance liquid chromatography (HPLC)-grade acetonitrile (Merck KGaA, Darmstadt, Germany) and formic acid (Tedia, USA) were utilized for UPLC analysis. Deionized water was purified using a Milli-Q system (Millipore, Bedford, MA, USA). All other chemicals were of analytical grade. ## Sample preparation Cistanches Herba samples (1.0 g, 65-mesh) were transferred into a 50-mL conical flask, and 50 mL of 70% methanol was added. After soaking for 30 min, ultrasonication (35 kHz) was performed at room temperature for 30 min. After centrifugation at 10,000 rev/min for 10 min, the supernatant was stored at 4°C and filtered through a 0.22-μm membrane before injection into the UPLC-QTOF/MS system for analysis. ## UPLC-QTOF/MS For UPLC analysis, the following systems/parameters were used: Waters Acquity system (Waters) equipped with a binary solvent delivery pump, auto-sampler and PDA detector connected to a Waters Empower 2 data station; ultrasonication (250 W, 50 kHz, Kunshan Ultrasonic Instrument Co., Zhejiang, China); and an electronic analytical balance model AB135-2 (Mettler-Toledo., Greifensee, Zurich, Switzerland). A Waters Acquity UPLC BEH C₁₈ column (1.7 µm, 2.1×100 mm, Waters) and a Waters C₁₈ guard column (same material, waters) were used and maintained at 30°C. The mobile phase was 0.1% formic acid aqueous solution (A) and acetonitrile (B) with a gradient program as follows: 0–3 min, 10–22% B; 3–4 min, 22–23% B; 4–6 min, 23–35% B; 6–8 min, 35–37% B; 8–11 min, 37–42% B; 11–12 min, 42–48% B; 12–15 min, 48%–50% B at a flow rate of 0.3 mL/min. The injection volume was 5 µL. The UPLC/MS analysis was performed on a QTOF Synapt G2 HDMS system (Waters, Manchester, UK) equipped with an electrospray ionization (ESI) source operated in the negative-ion mode. N₂ was used as the desolvation gas. The desolvation temperature was set at 450°C at a flow rate of 800 L/h, and the source temperature was set at 120°C. The capillary and cone voltages were set to 2500 and 40 V, respectively. Data were collected between 50–1200 Da with a 0.1-s scan time and a 0.01-s interscan delay over a 15-min analysis time. Argon was used as the collision gas at a pressure of 7.06661023 Pa. All MS data were collected using the LockSpray system to ensure mass accuracy and reproducibility. The \[M-H\]^- ion of leucine-enkephalin at m/z 554.2615 was used as the lock mass in negative ESI mode. ## Data analysis UPLC-QTOF/MS data for Cistanches Herba samples were analyzed to identify potential discriminant variables. Peak finding, alignment and filtering of ES raw data were carried out using the Marker Lynx applications manager, version 4.1 (Waters, Manchester, UK). The parameters used were as follows: retention time (t_R) of 0–15 min, mass of 50–1200 Da, retention time tolerance of 0.02 min, and mass tolerance of 0.02 Da. Three replicate samples collected from each geographic location were used (n = 3). A total of 6, 339 variables were used to create the model. ## DNA barcoding: DNA extraction, PCR amplification and sequencing Samples taken from dried fleshy stems of *C. deserticola* and *C. tubulosa* (30 mg) were rubbed for 2 minutes at a frequency of 30 r/s. DNA was extracted according to the manufacturer's instructions (Tiangen). Specifically, the protocol was modified such that chloroform was replaced with a mixture of chloroform: isoamyl alcohol (24∶1 in the same volume), and buffer solution GP2 with isopropanol (same volume). The rubbed powder was put into 1.5 ml eppendorf tubes, added 700 µL 65°C preheated GP1 and 1 µL β-mercaptoethanol to mix using vortex for 10–20 s, and incubated for 60 minutes at 65°C; Adding 700 µL mixture of chloroform: isoamyl alcohol (24∶1), centrifuge for 5 minutes at 12000 rpm(∼13400×g); Pipette supernatant to a new tube, adding 700 µL isopropanol, blending for 15–20 minutes; Piping all the mixture into spin column CB3 and centrifuge for 40 s at 12000 rpm; Discarding the filtrate and adding 500 µL GD(adding quantitative anhydrous ethanol before use), centrifuge at 12000 rpm for 40 s; discarding the filtrate and adding 700 µL PW(adding quantitative anhydrous ethanol before use) to wash the membrane, centrifuge for 40 s at 12000 rpm; Discarding the filtrate and adding 500 µL PW, centrifuge for 40 s at 12000 rpm; Discarding the filtrate and centrifuge for 2 minutes at 12000 rpm to remove residual wash buffer PW; Transferring the spin column CB3 into a clean 1.5 ml eppendorf tube, and drying at room temperature for 3–5 minutes; Centrifuge for 2 minutes at 12000 rpm to obtain the total DNA. Primers for polymerase chain reaction (PCR) were based on sequences reported previously. PCR reaction mixtures contained 2-μL DNA template, 8.5-μL ddH₂O, 12.5-μL 2× Taq PCR Master Mix (Beijing TransGen Biotech Co., China), 1/1-μL forward/reverse (F/R) primers (2.5 µM), in a final volume of 25 µL. PCR amplification was conducted as described by Kress et al.. The primer of PCR reaction were fwd PA: GTTATGCATGAACGTAATGCTC (5′-3′) and rev TH: CGCGCATGGTGGATTCACAATCC (5′-3′). PCR products were separated and detected by 1% agarose gel electrophoresis. PCR products were purified following the manufacturer's protocol and directly subjected to sequencing. ## Sequence alignment and analysis ITS and ITS2 sequences were collected from the GenBank database. Sequences from sequencing of the samples were submitted to GenBank database (Accession numbers were listed), assembled with CodonCode Aligner 3.7.1 (CodonCode Co., USA) and aligned using ClustalW. Kimura 2-Parameter (K2P) distances, GC content of base and Neighbor-joining (N-J) trees were calculated and constructed using the MEGA 5.05 with the Bootstrap method (1000 resampling) and K2P model. Barcoding gap (spacer region that was formed between intra- and inter-specific genetic variations) and identification efficiency (the ability of identification for comparing different barcodes) were drawn and calculated based on the method reported by Meyer and Paulay. # Results ## Tentative peak assignment by UPLC-QTOF/MS Representative chromatograms of *C. deserticola* and *C. tubulosa* from different producing areas are shown in. The fingerprint chromatogram indicated similarities among Cistanches Herba samples. A total of 23 qualified mass peaks were detected and 16 peaks were identified by matching the retention times and mass spectra with those reported previously. Peaks 2, 3, 5, 6, 7, 8, 9, 11, 12, 15, 16, 17, 20, 21, 22, and 23 were tentatively identified as cistanoside F, mussaenoside acide, cistanbuloside C1/C2, campneoside II, isomer of campneoside II, echinacoside, cistanoside A, acteoside, isoacteoside, syringalide A-3′-α-L- rhamnopyranoside, cistanoside C, 2′-acetylacteosid, osmanthuside B, cistanoside D, tubuloside B, and cistancinenside A, respectively. Chemical constituents were determined to be primarily phenylethanoid glycosides (PhGs), while one compound, mussaenoside acide, was an iridoid polysaccharide. PhGs are the main active compounds in terms of treatment of kidney deficiency, and antioxidant and neuroprotective effects. ## *PCA of* C. deserticola *and* C. tubulosa PCA was employed to distinguish samples of different plant species. PCA is an unsupervised multivariate data analysis method that aims to visualize the similarities and/or differences within multivariate data of secondary metabolite composition. The two-component PCA model cumulatively accounted for 46.04% of the variation (PC1, 36.43%; PC2, 9.61%). shows that 24 samples were clustered into two groups in the PCA scores plotted according to species origin, indicating that the chemical composition of *C. deserticola* and *C. tubulosa* differed significantly. ## OPLS-DA and marker identification To identify potential chemical markers for discrimination of the two species, the S-plot of OPLS-DA was generated. In the S-plot, each point represents one t_R–m/z ion pair. The X and Y axes represent the contribution and confidence of the ion, respectively; the farther the distance the ion *t*_R–*m*/*z* pair points from zero, the larger the contribution/confidence of this ion is to the difference between the two groups. Thus, the t_R–m/z ion pointing to the two ends of the ‘S’ represent the characteristic markers with the highest confidence in each group. The OPLS-DA results showed that UPLC-QTOF/MS could be used to distinguish *C. deserticola* from *C. tubulosa*. A total of six credible and significant markers were determined to facilitate discrimination of these groups. The identities of three potential markers were tentatively assigned. The components correlated with these three ions were tentatively identified as isomers of campneoside II, cistanoside C and cistanoside A. The marker compounds **a**, **b** and **c** could be used to distinguish the two plant species, as the ion intensities of **a** and **b** in *C. deserticola* was higher than in *C. tubulosa*, and marker **c** could be detected in *C. tubulosa*, but not in *C. deserticola*. ## DNA barcoding: sequence information and identification efficiency Sequence information was shown in. The average genetic distance of *psb*A-*trn*H (0.1732) was larger than other two regions (0.0740, 0.1197) significantly. The average GC content of *psb*A-*trn*H (20.64%) was smaller than other two regions (55.00%, 55.00%). Though the success rate of ITS and ITS2 was not obtained in this study, the *psb*A-*trn*H region performed well in PCR amplification and sequencing (100%, 87.23%). Identification efficiency was achieved by BLAST1 analysis and the nearest-distance method, and mainly reflected the success rate of the barcodes. The *psb*A-*trn*H region was clearly higher than the other two barcodes in identification efficiency based on two methods. The shortage of sequences is most likely the reason that ITS region exhibited 100% identification efficiency based on BLAST1 method, and 0 based on the nearest- distance method. ## Analysis of genetic divergence using six parameters Six parameters were used to analyze intra-specific variation and inter-specific divergence using three barcodes. The significant difference between inter- and intra-specific variations was indicative of the utility of the DNA barcodes. Here, the minimum interspecific distance of three barcodes was all higher than the maximum intraspecific distance. Moreover, *psb*A-*trn*H region had larger maximum intraspecific distance and average interspecific distance than the other two barcodes, indicated that *psb*A-*trn*H region performed well in discrimination of two origins of Cistanches Herba. ## Analysis of barcoding gap to identify C. deserticola and C. tubulosa The barcoding gap presents the remarkable variation of inter- and intra-species, and demonstrates that separate, non-overlapping distributions between intra- and inter-specific samples. In this study , the distance range was set to 0–0.45, because the greatest K2P distance of *psb*A-*trn*H between *C. deserticola* and *C. tubulosa* was close to 0.45. The three barcodes exhibited distinct gaps in the distributions of intra- and inter-specific variation. Furthermore, the gap of *psb*A-*trn*H was significantly larger than other two barcodes. Therefore, *psb*A-*trn*H region could be an ideal barcode for discriminating two origins of Cistanches Herba. ## Neighbor-joining (NJ) tree An NJ tree illustrates the relationship among species and facilitates determination of their clustering. In this study, NJ tree of three barcodes were built based on K2P model. The results demonstrated that two origins of Cistanches Herba clustered into two clades separately. Thus, the NJ tree clearly distinguished between *C. deserticola* and *C. tubulosa*. # Discussion and Conclusions Cistanches Herba is an important medicinal material commonly used to nourish in the Asian community. However, the two origins of Cistanches Herba, *C. deserticola* and *C. tubulosa*, have different chemical compositions and pharmacological activities respectively. Concurrently, the two origins differ in clinical application and commodity market. The classification of *Cistanche* is confused and massive substitute and adulterants flood the market due to the shortage of resources and special growing environment for Cistanches Herba. Genus of *Cistanche* is accepted to include four species and one variant: *C. deserticola*, *C. tubulosa*, *C. sinensis*, *C. salsa*, and *C. salsa var. albiflora*. Researchers in Japan considered the origin of Cistanches Herba as *C. salsa*, while it was identified as *C. deserticola* by Tu. Therefore, it is confused in classification of *Cistanche*, and it is hard to discriminate the two origins of Cistanches Herba. Traditional methods for quality control of Cistanches Herba are morphological identification, microscopic identification and TLC (Thin-Layer Chromatography), FTIR (Fourier Transform Infrared Spectroscopy), HPLC (High Performance Liquid Chromatography). Morphological and microscopic method can easily differentiate species from different genera or families that possess big difference in morphological and microscopic characteristics, while it is hard to distinguish sibling species. TLC and FTIR can clearly discriminate species that possess different kind of chemical compositions, whereas it is difficult to determine the chemical component and content. HPLC is mainly used for differentiating species with different chemical elemente contents, nevertheless, the time of analysis is longer and the sensitivity is relatively lower compared to UPLC. Correspondingly, UPLC-QTOF/MS technology was faster and more accurate in determining chemical composition than other chemical methods. Molecular identification methods exhibit well in discrimination based on the genetic variation, such as SDS-PAGE (Sodium Dodecyl Sulfate-Polyacrylamide Gel Electrophoresis), AFLP (Amplified Fragment Length Polymorphism). However, these molecular methods are not easy to operate and are not universal. Correspondingly, DNA barcoding could discriminate species more universally, quickly and accurately than other molecular methods. For the species from same genus and close genetic relationship, those methods alone may not perform well in identification. Here, we combined UPLC-QTOF/MS and DNA barcoding in identifying *C. deserticola* and *C. tubulosa*, and evaluated the chemical and molecular markers that would allow them to be discriminated. 23 qualified mass peaks were detected and 16 were identified by using UPLC–QTOF/MS, and three potential marker compounds were firstly found to facilitate the discrimination of two origins by PCA and OPLS-DA analysis. Furthermore,four indicators were assessed by DNA barcoding technology in terms of their ability to differentiate two origins: Identification efficiency, genetic efficiency, barcoding gap, and NJ tree analysis. The *psb*A-*trn*H region was supported as a suitable DNA barcode for discriminating *C. deserticola* and *C. tubulosa*. In conclusion, we firstly established a new molecular and chemical analysis- combined method for discriminating and quality control in two origins of Cistanches Herba. DNA barcoding can discriminate two origins in genetic variation and authenticate species universally and accurately; UPLC-QTOF/MS technology can analyze chemical composition to evaluate the quality of medicinal materials rapidly and accurately. The combined method of DNA barcoding and UPLC- QTOF/MS technology guarantee the identification in multiple sources of medicinal materials more accurately and scientifically, and may serve as method for identifying other confusing species or genus in classification. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: LFH SHZ LBW. Performed the experiments: SHZ LBW. Analyzed the data: SHZ LBW. Contributed reagents/materials/analysis tools: LFH. Wrote the paper: SHZ XJ LBW. Check the manuscript: ZHW.

# Introduction Electronic cigarettes, or vaping devices, produce an inhalable aerosol that usually contains nicotine, flavorings and other chemicals. While the name “vaping” conjures the idea of harmless water vaper, the aerosol can expose users to heavy metals, volatile organic compounds and other harmful ingredients know to have adverse health effects. In August 2019, Wisconsin reported the first cluster of lung injuries to the Centers for Disease Control and Prevention (CDC). The CDC worked with federal and state partners to address a multistate outbreak of e-cigarette, or vaping, product use–associated lung injury (EVALI). A total of 2807 cases of EVALI and 68 deaths had been reported to the CDC as of February, 2020. All patients with EVALI have reported using vaping products. Nationally, most of these patients reported using *Tetrahydrocannabinol* (THC)-containing products, but a minority reported exclusive use of nicotine- containing products. Beginning in September 2019, news surrounding a vaping illness simultaneously prompted restrictions and bans by local governments because of the questions about the safety of electronic nicotine delivery systems, *and* a defense of e-cigarettes by industry representatives claiming that it was bootleg, cannabis- based cartridges that were leading to the injuries, and not the nicotine-based cartridges that were sold legally. Extensive coverage of EVALI was likely effective in spotlighting the issue of e-cigarettes safety. However, conflicting information about the cause, and uncertainty about the cause when the news first broke, may have limited the ability of the news to perform one of its functions and accurately inform people about the actual and potential dangers. The U.S. Food and Drug Administration (FDA) acknowledges that in comparison to combustible cigarettes, which kill up to half of life-long smokers, e-cigarettes lie on the lower end of a continuum of risk. Yet, many smokers in the U.S. believe e-cigarettes are at least as harmful to health as combustible cigarettes. This may dissuade them from switching to e-cigarettes and, thus, have a detrimental impact on population health. Widespread news coverage of the EVALI outbreak may have increased confusion about the relative harms of these products. Media scholars posit that mass media serves four core functions: (*a*) information distribution, reporting events to the public; (*b*) interpretation, providing the context for and meaning of issues and events; (*c*) socialization, cultivating community values, beliefs, and norms; and (*d*) entertainment, providing diversion and escape from everyday life. Though news media can affect policy and behavior change through providing a source of health and science information, the effect of that information is not consistent. Thus, we sought to examine whether the *initial* widespread news coverage of EVALI changed perceptions and beliefs about e-cigarettes, especially beliefs about the harms and risks of e-cigarettes among users. # Methods ## Overview This paper combines the results from two studies to provide data on our research questions. Study one was conducted to assess awareness, harm perceptions and beliefs about three different tobacco products: e-cigarettes, snus, and heat- not-burn tobacco, that have the potential to be authorized as modified risk tobacco products (MRTPs). This study was completed before EVALI news coverage. Within a week of concluding data gathering for study one, news coverage of EVALI increased substantially. A month before our first study (July 28th to Aug 28th), a ProQuest search in the recent news database using the terms “vaping illness” OR “e-cigarette illness” OR “mysterious lung disease” OR “lung illness” yields 0 news stories. From Aug 29^th- Sept 29^th, the same search yields 297 entries, including a front-page New York Times article. Given this opportunity for a natural experiment, we went back into the field with the same measures in order to assess how perceptions and beliefs about e-cigarettes had changed after the news coverage (study two). Like study one, study two survey also assessed awareness, harm perceptions, and beliefs about e-cigarettes. We added items to study two survey that assessed awareness of the news story. Study two was conducted before the “mysterious vaping illness” was given the name EVALI, and before THC or vitamin E acetate was widely accepted as the cause. ## Participants ### Study one We recruited 865 adult current and former smokers to complete an online survey about MRTP beliefs through Dynata from August 27–28, 2019. Using established definitions of smoking status, participants were considered current smokers if they had smoked at least 100 cigarettes in their lifetime and currently smoke every day or some days, and former smokers if they had smoked at least 100 cigarettes in their lifetime, and currently did not smoke at all. Additionally, participants could not have participated in more than two online surveys about cigarette smoking or other tobacco products in the last three months. Participants included 450 men and 414 women, with a mean age of 47.5 years. A little more than half of the participants were current smokers or ever-users of e-cigarettes. Participants had similar demographics to a national sample of smokers, as they were diverse in race, education, and income, though our sample was slightly more educated than the population of US smokers. The purpose of study one was to understand what people believe about MRTPs, and how information they get from brands might influences those beliefs. The sample size for study one allows for comparisons between the three products within subgroups of smoker and non-smoker, and between groups with differing experiences with MTRPs. ### Study two Using Dynata and the same inclusion and exclusion criteria, we recruited an independent sample of 344 adult current and former smokers from September 27–29, 2019. Participants included 163 men and 181 women, with a mean age of 46.7 years. Participant characteristics were statistically similar between the two studies. The sample size for study was chosen to allow for meaningful comparisons between subgroups of e-cigarette users and between time periods. ## Procedures ### Study one Eligible participants answered demographic survey questions and information about their current smoking behavior. Study one’s purpose was to understand what people believe about MRTPs, and how information they get from brands might influences those beliefs. Participants read a paragraph describing the potential for the FDA to authorize MRTPs and a brief description of MRTPs. Participants were randomly assigned to either the control condition, in which they read a generic description about e-cigarettes, snus, and heat-not-burn tobacco, presented in random order; or to the corporate social responsibility condition, in which they read the generic description, plus a corporate responsibility statement crafted using press releases and text from IQOS, General Snus, or JUUL’s website respectively. There were no differences between the conditions, so the conditions were collapsed in subsequent analyses. After reading descriptions of MRTPs, and a description of the products, participants answered questions about awareness, harm perceptions, use, and susceptibility to each of the products. Then the survey assessed 15 different beliefs for the three products. The belief questions were worded using three different variations and participants were randomly assigned one wording variation per product using a Latin-square design. Participants read a consent form that provided the approximate time it would take to complete the survey, emphasized that their participation was voluntary, and supplied the contact information for the study PIs and the Institutional Review Board. After reading through the consent form, they were instructed to click through to the next page. The consent form advised those who did not wished to participate to close their internet browsers. The University of Pennsylvania institutional review board approved the consent process, study procedures, and study materials. ### Study two Study two was designed to take advantage of the fact that study one had occurred just prior to a large national news story about one of the products. Participants in study two completed a very similar survey. The survey was identical up until the Latin-square design assessment of the 15 beliefs. Study two only assessed beliefs about e-cigarettes. Measures assessing awareness of health news stories from the previous two months were added after the belief assessment in study two. ## Measures in study one and study two ### E-cigarette use Participants who answered that they had heard of e-cigarettes before the time of the survey indicated whether they had ever used an e-cigarette or vaping device, even one or two times. Participants who answered yes responded with how many of the last 30 days they had used the device(s). Participants who were unaware of e-cigarettes, or who had never used an e-cigarette were considered never users. Participants who answered that they had used an e-cigarette but reported that they had not used one in the last 30 days were considered former users. Those who had used an e-cigarette at least once in the last 30 days were considered current users of e-cigarettes. Because the participants were all current or former smokers, the former and current users have experience with both cigarettes and e-cigarettes. ### Harm perceptions Participants indicated how harmful they thought e-cigarettes were to their health on a scale of not at all harmful (1), to extremely harmful (4). This measure was adapted from the Population Assessment of Tobacco and Health survey. ### E-cigarette beliefs The surveys assessed 15 different beliefs about e-cigarettes: being risky, having long term health benefits, causing lung damage, tasting good, feeling harsh, being odorless, being easy to use, looking cool, making second hand smoke, not being addictive, containing nicotine, helping smokers quit, untrustworthy science about the product, appealing to kids, and being expensive (Appendix A). The decision to focus on these beliefs was based on prior qualitative research, surveys on salient beliefs, and examining the MRTP applications and materials made publicly available. The 5-point response scale ranged from strongly disagree (1) to strongly agree (5). The survey assessed these beliefs using three wording variations in study one: a comparison to cigarettes, a self-referent, and an absolute statement. Spearman ranked correlations of the beliefs between the absolute wording belief ranking and the self-referent wording ranking was very high (ρ (rho) =.98, p \<.001) indicating that participants did not meaningfully differentiate the two wording variations. Consequently, in study two, only the comparison and self-referent wording variations were used (ρ (rho) =.63, p \<.05). ## Measure in study two ### Awareness of EVALI story The survey presented the participants with five one-sentence descriptions of news stories about health that had been run in the past two months. These five stories were presented in a random order and included “Several patients around the country have died because of a mysterious lung ailment tied to vaping.” Other story options were about cardiovascular health of dog owners, allergic reactions because of tattoos, pharmaceutical companies facing fines because of their ties to opioids, and the percent of American youth who had tried vaping. Participants indicated whether they had heard of, read, or seen the story (coded as 1) or not (coded as 0), or if they didn’t know (coded as 0). ## Statistical analysis The research questions motivating the analysis were: 1. Did harm perceptions and beliefs about e-cigarettes change between study one and study two, particularly for beliefs about health harms or benefits? 2. Did changes between study one and study two differ by e-cigarette user status? We conducted t-tests adjusted for unequal sample sizes and *X*² in STATA 14.0 to examine differences in harm perceptions between time points. To prevent multiple comparisons from increasing the false positive rates, we conducted a MANOVA for beliefs, followed by post hoc analysis when the MANOVA indicated significant differences between study one and study two. The University of Pennsylvania institutional review board approved the procedures. # Results ## News coverage and perceived harm Between study one and study two, during which news coverage of EVALI was intense relative to baseline, the information circulating in the public information environment was associated with an increase in e-cigarette harm perceptions (mean = 2.67 (sd =.90) to m = 2.90 (.97), p \<.001). This change was largely driven by ever-users of e-cigarettes, specifically former-users (m = 2.61 (.81) to m = 2.99(.94,) p \<.01). Never users’ perceptions trended in the same with a smaller and non-significant difference between time 1 and 2 (m = 2.98(.88) to m = 3.14(.92), p =.051;). ## Fifteen beliefs about e-cigarettes Between study one and study two some important beliefs did not change significantly after intense news coverage of EVALI; the risk of e-cigarette use, lung damage and long-term health benefits. Some social beliefs about e-cigarettes being easy to use, cool, and appealing to kids decreased between study one and two by.2 to.4 points (p \<.05;). When *the comparison to cigarettes* was invoked, the belief that e-cigarettes were riskier, more likely to cause lung damage, and cooler increased by.3 to.4 points between study one and two (p \<.05). Increased information in the public information environment about e-cigs is affecting beliefs about cigarettes, but only when a comparison to cigarettes is invoked. A MANOVA did not indicate significant differences between study one and study two for the set of 15 beliefs for current users, or when a comparison to cigarettes was invoked. A MANOVA indicated significant differences between study one and study two for the set of 15 beliefs for never and former users, therefore we conducted post hoc t-tests to examine which of the 15 beliefs changed significantly between study one and study two. *For current users*, this indicates that there was no significant difference in the endorsement of beliefs about e-cigarettes being risky, having long term health benefits, or causing lung damage between study one and study two. In comparison, *never user’s* beliefs about e-cigarettes causing lung damage, being risky to use, and feeling harsh, increased after EVALI news coverage (p \<.05;). The group most at-risk (current users) did not accept the lung damage risk assessment, while the group at lowest risk (never users) did. Information in the public information environment about e-cigarettes is not affecting the at-risk current user group. ## Awareness of coverage Seventy eight percent of participants in study two were aware of the EVALI story, the highest of any of the stories we asked about. In comparison, 22% of people had heard about allergic reactions because of tattoos, 36% about the cardiovascular health of dog owners, 56% about the percent of American youth who had tried vaping, and 57% about pharmaceutical companies facing fines because of their ties to opioids. Those aware of the EVALI story were more likely to endorse the belief that e-cigarettes were risky compared to cigarettes (m_unaware = 3.05, m_aware = 3.43, p =.03); a similar pattern occurred for the parallel question focused on one’s own risk (m_unaware = 3.70, m_aware = 4.05, p =.07;). Awareness was not related to the belief that e-cigarettes would cause lung damage. Sub-groups of aware-unaware by user status are small and unstable but potentially instructive about the impact of the information circulating in the public environment. Due to the small numbers, we compared current users to non- users (both former users and never users) in sub-group analysis. For non-users, awareness of the story was associated with endorsement of the belief that using e-cigarettes was risky (m_unaware = 3.95, m_aware = 4.4, p =.06, Hedges’ g effect size =.59;). There was no difference in endorsement between aware and unaware current users (m_unaware = 3.33, m_aware = 3.30, ns, effect size =.03). For these same subgroups, when a *comparison to cigarettes is primed*, unaware non-users exhibited less acceptance of e-cigarette risk compared to cigarettes than aware never users (m_unaware = 2.96, m_aware = 3.61, p \<.01; Hedges’ g effect size =.69;). For long-term health benefits of e-cigarettes compared to cigarettes, lower endorsement occurred among the aware compared to the unaware (m_unaware = 3.00, m_aware = 2.27, p \<.01, Hedges’ g effect size =.55). Story awareness among current users of e-cigarettes exhibited no differences or trends approaching significance for beliefs that e-cigarettes are riskier, cause lung damage, or have long-term health benefits *compared to regular cigarettes*. # Discussion Study two took place before EVALI had been officially named, and before confirmation of THC and vitamin E acetate as the likely causes of EVALI. Some news stories mention these as potential causes in the coverage of the “mysterious vaping illness.” In a public information environment that was working perfectly in service to public health at the time of study two, we would expect EVALI news to cause an increase in harm perceptions, an increase in the belief that e-cigarettes were risky and caused lung damage, and a decrease in the beliefs that e-cigarettes had long-term benefits. Our study did not find such neat and tidy results, leading us to examine why the information environment led to some of the expected and hoped for outcomes, but not all. This study did find that attitudes and beliefs about e-cigarettes changed after EVALI news became a well-known story. But rather than perceptions and beliefs moving systematically within the population, there were differences in how much the beliefs changed over time between e-cigarette never, former, and current users and when evaluating e-cigarettes on their own, or in comparison to cigarettes. There were no beliefs that changed significantly for all user groups *and* when referencing cigarettes and not. From a public health perspective, we would hope that the EVALI news stories in the information environment would reach and move the most at-risk population, in this case, the current users of e-cigarettes. While harm perceptions about e-cigarettes increased overall after the EVALI coverage, the change among never users was small. Importantly, never users already had higher perceptions of harm, and the perceptions of harm among current users after EVALI are still lower than never users and former users before the news coverage. While overall harm perceptions increased for current users, their beliefs about lung damage, risk, and long-term health benefits of e-cigarettes did not change. The fact that there are differences between never, former, and current users is consistent with research demonstrating the role of involvement on message processing and attitude change. Because current users have more of a personal interest in the safety of e-cigarettes than never users, they are likely more attentive to the information, and as a result are affected in different ways. These data suggest that contrasting information processing motivations between groups with different experiences using e-cigarettes moderated the effects of a prominent news event on beliefs about the risks and benefits of vaping. That current users’ harm perceptions increased following the EVALI news, suggests that safety concerns are likely a key difference between the groups. For never users, the news of EVALI validated their previous behavior, whereas it forced current users to reconcile their behavior with information that it is harmful. It is instructive to note that different beliefs changed when the *comparison to regular cigarettes* was primed. While it is expected that public information environment after EVALI would increase the beliefs about lung damage and risk, it is perplexing that those beliefs only increased when in comparison to regular cigarettes. This may be a result of the larger news narrative around e-cigarettes. The EVALI story and news concerning e-cigarettes in general has focused extensively on their popularity with young people, while e-cigarette industry marketing has focused on the benefits of e-cigarettes in comparison to cigarettes. Thus, smokers are evaluating EVALI in light of the larger news context. EVALI may make e-cigarettes less appealing to use and have seemingly fewer benefits and more risks compared to cigarettes. Changes in participant’s perceptions of e-cigarettes were associated with awareness of the coverage of this prominent news story, underscoring the importance of working to ensure that coverage is a scientifically accurate as possible. Changing beliefs and perceptions after news coverage is consistent with other news topics. It is particularly important to use these news events that capture the public’s attention to provide accurate and not misleading information. A morning consult poll conducted in mid-September indicated that 34% of adults believed that the lung disease deaths were related to marijuana and THC-containing vapes, while 58% percent said nicotine e-cigarettes such as Juul were to blame. Given the evolving nature of the story in the time leading to study two, which news reports were seen by the participants may have been important in how beliefs and harm perceptions were changed. However individual news exposure and awareness of specific news stories were not measured in this study. One of the limitations of this study is that it was not originally designed to capture changes in the beliefs about e-cigarettes in reaction to the public news environment. Therefore, we do not have data in study one about news consumption, exposure, or awareness, nor do we have information about whether users in either study used THC in their e-cigarettes in addition to tobacco. Cases of EVALI varied by state, and while awareness of the EVALI story was very high compared to other news stories, some of the differences in awareness may be due to location of the participants, which we did not measure. Because the first study’s aim was to capture beliefs about MRTPs among former and current smokers, we do not have e-cigarette users who have never used cigarettes in our sample. We did not anticipate conducting study two while we were planning and conducting study one. As such, we had not designed study one to allow for follow up with participants, the cross-sectional nature of these studies does not allow us to infer causality. We believe that even with these limitations, the data provide insight into how beliefs about tobacco products can change when there are major news events. We used mainstream news coverage about the mysterious vaping illness as an indicator of the public information environment. This was not intended to be a content analysis of mainstream or social media news sources, which could tell us which stories were circulating and influential. Instead, we can describe what beliefs and perceptions changed and did not change at the height of EVALI new coverage. When the public information environment is dominated by news stories about the health and safety of tobacco products, it presents an opportunity to change beliefs that are frequently targeted by paid health campaigns. The EVALI story, like any other, appears subject to selective perception. Our experiences, attitudes, and existing beliefs shape how we view and interpret news stories. A news story about a novel and complex issue like the safety of e-cigarettes is particularly likely to evoke motivated reasoning processes, particularly among e-cigarette users who have both a physical and emotional interest in the issue. Current users may have read the stories more carefully or followed the nuances of the evolving story whereas never users may not have followed the news as closely. However, as current users have a personal interest in whether e-cigarettes are safe, they are motivated to process information in such a way that allows them to maintain that belief rather than change their behavior. # Supporting information [^1]: The authors have no conflicts of interest to disclose.

# Introduction ## 1 On Modeling and Optimization Signaling pathways are of utmost importance for understanding cellular function and predicting cellular response to perturbations. Recent advancements in text mining and the construction of Protein-Protein Interaction (PPI) networks have led to large databases of signaling pathways, showing how proteins interact with each other,. However, compilation and visualization of protein connectivity in signaling networks is just the first step towards understanding the cell's signaling mechanisms. The modeling and analysis of these networks either at the connectivity level or down at the level of signal transduction mechanics between nodes is a crucial next step towards the construction of functional models, predictive of the cell's biology. A variety of methods have been proposed for this task, each adopting a different perspective on the nature of the included reactions, and focusing on different properties of the signaling network. Two wide classes of network analysis can be distinguished: i) *Topological analysis* of the signaling network, that extracts insight into the cells' function by investigating the structural characteristics of the signaling network (e.g., feedback loops, strongly connected components). ii) *Network identification*, which identifies the network structure (i.e. connectivity of signaling species), or reaction parameters that define the mechanics of signal transduction from one node to the next. Typically a mathematical formalism is adopted to model how signal transduction takes place and an executable model is constructed by combining this formalism with a prior knowledge network (PKN) that serves as a scaffold. By simulating the model under different node and reaction parameters, conclusions can be drawn for the importance of each node and reaction on the propagation of the signal. Amongst the most widely used formalisms are the various forms of logic modeling, and ordinary differential equations (ODEs). In certain cases, the initial model is trained to signaling data via an optimization approach, to compute the values of model parameters that better fit the data at hand, or a sensitivity analysis approach is used , to compute the influence of model parameters to the overall response of the model. The incorporation of signaling data allows the construction of cell-specific, tissue-specific, or disease-specific pathways. The selection of the modeling approach, and subsequently of the optimization procedure, is very close related to the availability of data and biological question at hand. For example, if time course data are available and the dynamics of signaling reactions are of interest, then an ODE-based approach may be suitable, especially if the interrogated signaling network is small in size. To this end significant work has been published on parameter estimation of ODE- based models using a wide spectrum of methods including general purpose optimization methods (gradient based algorithms, stochastic search algorithms, branch and bound strategies, geometric programming, Dynamic Flux estimation and others). However, large scale signaling networks cannot easily be addressed within an ODE framework because of excessive CPU times and lack of proper constrain of the association-dissociation constants. If data are available for large pathways but on a single time point, then logic based modeling (Boolean or fuzzy logic, simulated at a ‘pseudo steady-state’) can be used to identify the structure of the signaling pathway. ## 2 Boolean Modeling In Boolean modeling, signal transduction is modeled using the rules of Boolean logic. Protein nodes assume only binary values {0,1}, denoting the activation (or not) of the corresponding signaling molecule, and signal is propagated from the receptor level to downstream nodes using a combination of OR and AND gates. In an approach was introduced to compress a protein network and convert it into Boolean models that are trained against signaling data. In the approach, implemented in the tool CellNOpt, reactions that appear to contradict the data are removed from the PKN, and thus measurement-prediction mismatch is minimized. In CellNOpt a Genetic Algorithm (GA) was used to prune the pathway by identifying and removing the contradicting reactions. The GA offered a robust and flexible optimization framework and managed to uncover structural differences between normal and cancer liver cell types. In a more recent study, the optimization problem was formulated as an Integer Linear Program (ILP), and was solved through CPLEX ([ILOG CPLEX 9.0,](http://www.ilog.com/products/cplex/))and GUROBI (Gurobi Optimization, Inc., <http://www.gurobi.com/>)viaGAMS(<http://www.gams.com/>). In contrast to GA, the ILP formulation guaranteed global optimality and required a fraction of the CPU time needed by the GA. The computational efficiency of the ILP formulation allows the rapid optimization of large scale signaling networks, as illustrated in a study, numbering around 120 nodes and 230 reactions (3 times bigger than the ones interrogated previously) offering a systems wide view of the signaling network in primary human hepatocytes. ## 3 Constrained fuzzy logic Even though Boolean modeling successfully addresses proteins' connectivity and directionality within the signaling pathway, it offers merely a qualitative view of signal transduction. In reality protein activities assume a continuous rather than a 0/1 pattern in signal transduction, making Boolean logic a rough approximation of how signal transduction really takes place. Constrained fuzzy logic (cFL) was introduced to offer a more detailed view of the cell's signaling mechanisms and implemented in the package CellNOpt-cFL. In cFL, a quantitative, yet static view of the signaling network is adopted. Proteins assume real values and a transfer function (TF) is introduced to propagate the signal from one protein to the next. A set of parameters in the TF defines its behavior and allows the calibration of the model to signaling data, in similar fashion to the pruning of the pathway in Boolean modeling. In a two- step method was proposed, wherein first a GA was used to remove all reactions that appear not to be functional based on the data at hand and estimate a rough approximation of transfer function parameters and in a subsequent step, a gradient based/greedy algorithm was used to give a better estimate of the parameters. The cFL approach performed significantly better than Boolean modeling in terms of fitting the data but resulted in more parameters, raising concern about model over-parameterization and causing the training process to be computationally more expensive. ## 4 Proposed approach Computational efficiency and availability of data are amongst the main limiting factors in modeling via cFL. In the present work we introduce two new approaches for more efficient optimization of signaling pathways in a fuzzy logic framework. Firstly, we formulate the signaling activities as a regular optimization problem (i.e., a nonlinear program (NLP)), solved through IPOPT (Interior Point OPTimizer, <https://projects.coin-or.org/Ipopt>) under GAMS. Secondly, we introduce an aggressive compartmentalization scheme similar to the equivalent classes concept published in, to simplify the model at hand so it can be constrained with small datasets. In contrast to previous compression methods, the new compartmentalization procedure is capable of addressing complex connectivity patterns and feedback loops, decreasing in a more efficient manner network size, CPU time, and over-parameterization/non-identifiability caused by the lack of data. As a result, the proposed NLP formulation allows for fast optimization of medium-scale topologies, and can also address the quantitative modeling of large scale signaling pathways. As a case study, we tackle the construction of cell type specific pathways in normal and transformed hepatocytes, to prove that our approach works for pathways as large as 15 receptors wide, numbering around 120 nodes and 230 reactions. # Results Our approach is based on the utilization of a transfer function (TF) to model how signal propagates between nodes of the signaling network. Briefly, we implemented and tested the following transfer functions: (i) Unity function, (ii) linear function and (iii) normalized Hill function. The normalized Hill function was chosen for being continuous, differentiable, monotonic, and fitting the expected qualitative trends of signaling reactions (sigmoid curve). The normalized Hill function was used in modeling signal transduction in. Reactions with multiple inputs are supported via AND and OR gates. In the case of an AND gate, all of the upstream nodes must be activated for the signal to propagate downstream, while in the case of an OR gate, one of the upstream nodes is enough to activate the downstream node (See Methods section §1). Normalized Hill function, AND and OR gates are shown in. In this work, we implement an NLP formulation to optimize the value of reaction parameters (*a, p* and *n* for every reaction), minimizing the difference between model predictions and measured data, resulting in a cell-type specific model of the signaling pathway. We then investigate if all reactions were necessary to fit the data by examining the parameters of the reactions and testing to determine if their removal significantly affect model fit. ## 1 Optimization of a Toy Model To illustrate how the proposed formulation fits parameters *a, p* and *n* to signaling data, we used the 10-node toy model shown in consisting of two stimuli (green nodes); two inhibitors (red nodes); 5 measured signals (gray nodes);4 OR gates (e.g.,TNFα OR PI3K→JNK); 4 AND gates (e.g.,TGFα AND NOT MEK1/2i →MEK1/2); and 4 NOT gates (total number of parameters = 20). *In-silico* data are shown in and consist of 3 stimuli (green nodes); the activation levels of 5 signals (gray nodes); and 2 inhibitors (red nodes) (total number of data points = 45). The red background color in the data represents the initial and after-optimization measurement-prediction mismatch of the model. For example, MEK1/2 signal under TNFα, without any inhibitor being present, was initially misfitted by the PKN. i.e. The data showed no activation, while in the PKN, MEK1/2 was clearly activated by TNFα. After the optimization procedure the red background was removed, implying that, in the optimized model, TNFα did not activate MEK1/2. The goal of the NLP formulation is to minimize the fitness error by searching for optimum values of the parameters *a, p* and *n* within predefined bounds. For the toy problem the bounds were:, and while the exponent was held constant n = 4. The upper and lower bounds for *p* were defined in such a manner that p = 0.3 corresponded to an over-responsive transfer function and p = 0.7 corresponded to an under-responsive transfer function, while p = 0.5 was the initial guess for the *p* parameter. Parameter *a* acts as a scaling factor and serves to limit the activity of those reactions that appear not to be functional based on the data at hand. Although the initial selection of upper and lower bounds for the *p* parameters together with the value of *n* was done arbitrarily, in case of high remaining fitness error these values can be updated and the algorithm rerun to guarantee the best possible solution (see also Material and Methods section 5.2 – Definition of search space). present the optimization results of the toy model. In, the activity of each reaction is visualized using arrows in gray scale; reactions with larger *a* parameters effectively transmit more signal downstream (are more active) and have a more solid color. The transfer functions themselves are illustrated in. The efficiency of our approach is validated by the eradication of most of the fitness error as shown in (red background). The optimization eliminated the PI3K to JNK, PI3K to P38, and PI3K AND NOT MEK1/2i to MEK1/2 reactions (bottom right panels in). Manual inspection of the data and the initial topology can confirm this decision: JNK and P38 were activated upon TNFα stimulation alone; therefore reactions from the TGFα pathway to JNK and P38 were not active. On the other hand, TNFα stimulation induced AKT activation but did not affect MEK1/2 or ERK1/2, implying that the PI3K to MEK1/2 reaction was not active. To validate that reactions i) PI3K to MEK1/2, ii) PI3K to JNK and iii) PI3K to P38 were not active in the optimized model; we manually removed them from the initial model and run the NLP algorithm once again. No significant differences were observed between the two optimized models, indicating that these three reactions were not vital to fit the data (data not shown). ## 2 Optimization of a medium-scale signal transduction pathway ### 2.1 Background Next, we tested the proposed NLP approach to the medium-scale signaling pathway used in, which numbers a total of 52 reactions and 37 species (total number of model parameters = 104). The training dataset was constructed using the xMAP technology on transformed human hepatocytes (HEPG2 cells) and numbers a total of 728 datapoints. The initial topology and the experimental dataset are illustrated in. The proposed approach was implemented in 3 steps: (i) definition of search space for the *p* parameters of each reaction, (ii) generation of a family of solutions and (iii) exhaustive removal of reactions from the PKN to address over-parameterization (see Methods section §2, 3, 4). ### 2.2 Optimization results contains an “average” pathway for 500 solutions. The solid lines are the minimum set of reactions needed to fit the experimental dataset and the opacity of each of these edges corresponds to the maximum activity (*z_i^k*) of the respective reactions. The dashed lines are reactions that were present in the family of models but could be removed for being redundant based on the analysis in Methods section, §4. presents the signaling dataset together with the measurement prediction mismatch for the optimized model (red background). The average CPU time of each run was 10 minutes. Several interesting features can be uncovered from the proteomic-driven optimization of the generic pathway: LPS pathway was deactivated altogether since it only partially affected the AKT signal. IGF1 and TGFα signaled through PI3K and activated AKT, GSK3 and P70. Moreover, TGFα activated MEK1/2, P90, CREB, IRS1S and HISTH3 via RAS. TNFα and IL1α also had partially overlapping pathways signaling through the MAP3Ks. IL1α signaled through TRAF6 to MAP3K7 and then to JNK, CJUN, P38, HSP27 and IKB. IL1α also activated MEK1/2 via TRAF6 and then P70S6, P90RSK, CREB, IRS1S and HISTH3. TNFα, on the other hand, signaled through MAP3K7 but had clear effects only on IKB, while partially activated a number of signals such as CJUN and P53. Moreover, TNFα partially activated P70S6, CREB, IRS1S and MEK via PI3K. As shown in, most of the measurement-prediction mismatch has essentially been removed by the optimization procedure. The remaining fitness error is below 8% (mean fitness error). Residual errors appear either in areas of the pathway where the a priori knowledge was poor, or where erroneous measurements in the experimental dataset conflicted each other. The latter is shown in the JNK signal under IL1α and JNKi. Even though JNKi was supposed to have inhibited JNK activation upon IL1α stimulation, the data shows that JNK remained active. In such cases the NLP algorithm is not able to reproduce the respective datapoint. Similar case consisted the misfitting of i) CJUN under IL1α and JNKi, ii) MEK1/2 under IL1α, IL6, TGFα and MEK1/2i, iii) P38 under IL1α and P38i, iv) GSK3 under IGF1, TGFα and GSK3i, and so forth. Those residual errors appeared in almost all optimization procedures. In conclusion, despite the residual error, the optimized model successfully captured the patterns underlying the signaling dataset. ### 2.3 Cross-validation For the optimization of the PKN, the signaling dataset in its entirety is used. Herein, however, to better evaluate the performance of the proposed formulation, we performed a cross validation study where random portions of the dataset, of increasing size, were left out of the training process, model predictions corresponding to this data were computed and then compared to the measured data evaluating the measurement prediction mismatch. illustrates the fitness error corresponding to all measured data (total fitness error), in blue, together with the error corresponding to the excluded data (in red). Interestingly, up to 40% of the dataset could be removed before the fitness error started increasing significantly, implying the proposed formulation is robust against missing data. Moreover, the algorithm performed relatively well even with 80% of the dataset missing. After this point, a steep increase in the overall error was observed, since key pathways were removed and the fitness error quickly reached that of the null solution. ## 3 Optimization of a large-scale signal transduction pathway ### 3.1 Background In order to evaluate the performance of our optimization procedure, we asked whether we could apply the procedure to larger pathways. Here, we focused on pathways that are experimentally identifiable using ELISA type of assays and thus are limited in well-known signal transduction mechanisms. The resultant PKN accounts for dozens of stimuli and their downstream nodes. The pathway contains 228 reactions and 117 species (total number of model parameters = 456). The corresponding data were measured using the xMAP technology on primary human hepatocytes and consist of a total of 120 multi-combinatorial experiments. Cells were perturbed with combinations of 15 stimuli and 3 inhibitors (including the No-inhibitor treatment), while 14 key phosphoproteins were measured (total number of data points = 1680). Before the optimization procedure, the pathway was compartmentalized to reduce the parameters space (the compartmentalized pathway numbers 44 species and 69 reactions, total number of model parameters = 138), while a family of solutions was obtained to guarantee that the algorithm is not trapped in a local minimum (see Methods section §3, 5). ### 3.2 Optimization Results In the optimized, compartmentalized version of the large-scale pathway is shown, together with the measurement-prediction mismatch. To demonstrate how the compartmentalization scheme works, we first examined the pathways downstream of EGF, TGFα, BTC, NRG1 and IL6. ERBB3 was placed in a group alone (C11) since it was the only node activated by NRG1; ERBB4 was also placed alone (C12) for having been activated by BTC and NRG1. ERBB2 and SHC were grouped together since they were both activated by EGF, TGFα, BTC and NRG1. Moving further downstream, INPP5D, JAK1, JAK2, INPPL1, GRB2, GAB2, GAB1, SOS, RAS, CRK, CRKL, DOCK1, BRAF, RAC1 and the MAP3Ks were grouped into C2 since all of them were activated by EGF, TGFα, BTC, NRG1 and IL6. This example demonstrates how the proposed compartmentalization scheme is based on the experimental treatments present in the dataset. If for example, another ligand was introduced activating via a different pathway RAC1, then the extensive compartment C2 would be broken into 2 smaller ones. First, INPP5D, JAK1, JAK2, INPPL1, GRB2, GAB1, GAB2, SOS, RAS, CRK, CRKL and BRAF, activated by EGF, TGFα, BTC, NRG1 and IL6; and second, RAC1 and the MAP3Ks (MAP3K2,3,4,6,9,10,11,12,13,15), activated by EGF, TGFα, BTC, NRG1, IL6 and the new ligand. With the proposed compartmentalization scheme, the interrogated pathway is never larger than what can be constrained by the data at hand. In, the optimized pathway of was mapped back to the PKN. Reactions within the same compartment were plotted in blue and were not involved in the optimization procedure. The rest of the reactions were plotted in black and their thickness corresponds to the maximum activity of each reaction in the optimized model. The resulting pathway reveals well known characteristics of signaling cascades (See ): EGF, TGFα, BTC and NRG1, all signaled through the EGFR and then through the cluster of SHC, GRB2, GAB1, SOS, RAS to either activate MAP2K1, ERK, RPS6KA1, GSK3 and STAT3, or go through PI3K to AKT and subsequently to RPS6KB1 and IRS1S. On the other hand IL1β, FLAGELLIN and IL1α signaled through TRAF6 and mainly activated IKB, JNK, MAPK14 and HSP27. CD40LG and TNF activated the same signals but went through TRAF5, TRAF2 and MAP3K7. The solution obtained herein, when compared to the Boolean solution in was able to decrease the remaining fitness error up to 75% (mean fitness error). The algorithm completed within 20 minutes. Even though the two solutions share the same basic connectivity patterns, the constrained fuzzy logic approach handles conflicts in the data more efficiently, since it allows partial activation of the signaling species. For instance, GSK3 was removed from the Boolean solution for having been activated in an inconsistent manner (it was activated under very few combinatorial treatments and remained unaffected by either PI3Ki or MEKi). Under the constrained fuzzy logic approach, however, GSK3 was activated by RPS6KA1. By fitting the *p* and *a* parameters of this and the upstream reactions the model predictions for GSK3 matched the data and the fitness error was reduced. Similarly, IRS1S and RPS6KB1 were activated under constrained fuzzy logic, in contrast to the Boolean approach. # Discussion In this paper we introduced a Non Linear Programming (NLP) formulation for the *quantitative* modeling of signal transduction pathways, based on signaling data. We employed a fuzzy logic approach to model signal transduction mechanisms and coupled it to an NLP optimization formulation. The proposed method allowed for fast optimization of signaling pathways to high throughput signaling data in a quantitative framework. As case study, three pathways of different scale were interrogated, a small, medium and a large-scale one. For the latter two, i) the systematic definition of the search space, ii) the generation of a family of solutions, iii) and the identifiability/over-parameterization of the pathway were addressed to ensure the best possible performance of the proposed formulation. The systematic definition of search space guaranteed that a representative set of solutions was obtained while at the same time minimized the required CPU time. The collection of a family of near optimal solutions decreases the probability of biologically relevant solutions remaining unreported. The proper size for the family of solutions was also addressed (see Supporting Information 1). By addressing over-parameterization either by exhaustively removing reactions from the PKN, or via the proposed compartmentalization scheme, we decreased CPU time and guaranteed that only reactions vital for fitting the data were included in the solution. Finally, results on both the medium and the large scale signaling pathways were compared with the ones obtained by alternative approaches. Our NLP formulation presents several advantages and limitations in pathway optimization. On the negative side, it is clear that verification of the presence or absence of each reaction in the generic topology, or unique identification of its parameters is not possible given the relatively small dataset at hand. shows the un-compartmentalized version of the initial pathway where 116 out of 228 reactions (∼50%) could not be identified if they are present or not given the data at hand (blue lines). This implies that the optimization problem incorporates more parameters than what it is possible to constrain. However, the exhaustive removal of reactions from the PKN, in the case of the medium scale topology, and the adoption of the equivalent classes concept (introduced) as a compartmentalization scheme, in the case of the large- scale topology, limited the number of redundant/non-identifiable reactions left in the model. Another inherent limitation of the proposed approach is our restriction to connectivity present in the PKN. The formulation we use, by optimizing the values of model parameters (*a* and *p*), minimizes measurement prediction mismatch. Essentially reactions can be removed by setting the gain parameter of the respective reactions to zero, however, there is no support for adding new connections. Thus, the connectivity of proteins in the solution is a subset of the connectivity in the PKN. If the data dictates connectivity that is not supported by the PKN, there will be remaining fitness error in the solution. Even though methods have been developed to address this based on the inference of physical interactions of proteins from the signaling data, adding new connectivity in the PKN can lead to poorly confined solutions and further research is needed to tackle this issue. Another limitation is the single time point measurement of the signaling activity. All the incorporated signaling data from HepG2 cells were obtained from the same time-point (30 min). Consequently, any activity that takes place earlier or later on will not be accounted for. To alleviate this limitation an average “early” time point was employed in the phosphoprotein activity of primary hepatocytes that incorporates the average activity of 5 and 25 minutes.The single time point measurements also prevent us from capturing the dynamics of the signaling reactions. Even though a dynamic representation is closer to reality, and can be potentially handled within a logic framework, both the experimental cost and the number of parameters required, make it difficult to model large topologies. On the positive side, our approach is a significant advancement of the Boolean Logic that successfully addresses both the protein connectivity and the activity/intensity of reactions in large signaling pathways that –as shown- number ∼120 species and ∼230 reactions. When compared to Boolean modeling, the proposed approach provides a quantitative view of the signaling pathway, supporting continuous values for the activation of the included species. Moreover, each reaction is modeled via a sigmoid curve (normalized hill function) that more closely replicates its actual mechanics. As a result, the proposed approach gives lower fitness error than the Boolean counterpart. When compared to other fuzzy models, the proposed algorithm performed equally good to previous approaches interrogating the optimization of the medium scale pathway to signaling data. Even though the two procedures follow different workflows, the topology of the solutions is very similar and the goodness of fit is of the same level, whereas CPU times favors the NLP approach (∼60 minutes per run for CellNOpt-cFLagainst ∼15 mins for NLP). The computational efficiency of the NLP approach allowed the interrogation of large-scale pathways, namely the one introduced in. It performed significantly better than the Boolean approach in terms of goodness of fit, decreasing the fitness error up to 75% (mean fitness error). Although the CPU time was increased, the solution remained computationally feasible. Overall, the proposed approach addressed successfully the optimization of medium and large-scale signal transduction networks. It allowed the fast optimization of signaling topologies by combining the versatile nature of logic modeling with state of the art optimization algorithms # Methods ## 1 NLP formulation The proposed NLP formulation is built based on a pre-existing ILP (Integer Linear Programming) formulation first published in and thus uses the same nomenclature, repeated here for consistency. ### 1.1 Definitions A pathway is defined as a set of reactions *i = 1*, …,*n_r*; and species *j = 1*, …,*n_s*. Each reaction has three corresponding index sets. Namely the index set of signaling molecules (or reactants) *R_i*, inhibitors *I_i*, and “products” *P_i*. These sets are all subsets of the species index set ; Typically, these subsets have very small cardinality (few species), e.g., \| *R_i* \| = 0,1,2 ; \| *I_i* \| = 0,1 ; \| *P_i* \| = 1,2 ; \| *R_i* \|+\| *I_i* \| = 1,2. A set of in-silico experiments is performed mimicking the conditions of each actual experiment. The experiments are indexed by the superscript *k = 1*, …,*n_e*. In each experiment a subset of species is introduced to the system and another subset is excluded from the system, in similar fashion to the “actual” experiments where a combination of stimuli and inhibitors are introduced to the cells. The predicted activation value of species *j* in experiment *k* is represented by the constant. If available, the corresponding measured value is represented by. The last group of variables introduced, represent the activity of reaction *i* in experiment *k.* ### 1.2 Objective Function The objective function to be minimized is and represents the weighted measurement-prediction mismatch; are user- set weights that may favor the fit of specific nodes in the pathway. In the present study, all nodes are considered equally important (have equal weights a_j^k). ### 1.3 Single reactant – single product reactions Reactions with a single reactant and a single product are modeled using the following transfer function (TF): ) represents a normalized Hill function. Parameter *p* defines the midpoint of the curve (i.e. the value of *x* for which *f(x)* equals to 0.5), *n* is the Hill coefficient and defines the steepness of the curve whereas *a* is a scaling factor. The activity of reaction *i* in experiment *k* equals to:, where. The activation value of the downstream node equals to:, where. In case species *j* is inhibitory we use:, where. ### 1.4 Multiple reactants – single product reactions (AND gates) In case more than one reactants are needed to propagate the signal to the downstream species, the activity of reaction *i* is modeled as a function of the bilinear product of the reacting species: The activation value of the downstream node equals to:. The bilinear product is chosen for satisfying key properties, such as continuity, differentiability and for reproducing the Boolean AND gate for 0 and 1 values of the reacting species. ### 1.5 Multiple reactions leading to same product (OR gates) In case more than one reactions lead to the same product, the activation value of the downstream species is given by the following formulation:where, is the set of all reactions that have species *j* as their product. Let *i₁, i₂*, …,*i_\|Tj\|* denote the elements of. Then, is calculated recursively as: ### 1.6 Implementation The goal of the NLP formulation, described above, is the identification of optimal values for *a, p* and *n* parameters of each reaction to minimize the difference between model predictions and experimental data, as captured by the objective function in (1). The NLP was solved through IPOPT under GAMS. Additionally, an interface was developed in BASH scripting language to preprocess the PKN and generate the input files for the NLP algorithm. The DataRail toolbox was employed in MATLAB to handle and plot the dataset. The optimization was run on Dual Quad Core Intel® Xeon® Processors E5530 2.4 GHz, 12 GB, DDR3 RDIMM Memory, 1066 MHz. All results presented in this MS were computed using a single cor. ## 2 Definition of the search space A systematic definition of the search space is vital for obtaining the best possible solutions within reasonable CPU time. A wider search space accounts for a bigger number of feasible solutions, possibly including some that minimize the objective function, but often increases the CPU time. The model parameters to be estimated are: *a, p* and *n; a* serves as a scaling factor to limit protein activity in case the reaction appears not to be functional based on the data at hand, and is defined in ; *p* defines the midpoint of the curve (i.e. when *x_j^k* equals to 0.5) and can be any real number; *n* can be any positive integer, but here is fixed to 4, since the remaining parameters suffice to fit the data. In the toy model *p* was arbitrarily defined in. For the medium and large-scale topologies, we test a number of different upper-lower bound pairs, ranging from 0.1 to 2.0, to determine the one for which the algorithm performs best, in terms of goodness of fit, as well as decrease the required CPU time, facilitating the generation of a family of solutions. Goodness of fit is quantified by the mean absolute error (MAE) as calculated by the following formula Results for the medium-scale topology are shown in. The x-axis (0.1→2.0) corresponds to the lower bound of *p* range; y-axis (0.1→2.0) corresponds to the upper bound; while the z-axis corresponds to the MAE of the solution. shows that the quality of the solution mostly depends on the lower bound and less on the upper bound of *p.* In the corresponding CPU time is shown. As expected widening the range of *p* drastically increases the CPU time, since the search space becomes bigger. Based on these graphs the bounds of choice for *p* is 0.1 → 0.4375, since they provide both an excellent fit and low CPU time. ## 3 Generation of a family of solutions Instead of collecting a single solution that minimizes the objective function in (1), we collect a family of 500 near optimal solutions to account for slightly suboptimal pathways that may bare strong biological significance, and avoid as much as possible terminating with a significantly suboptimal local minimum. The proposed NLP approach optimizes the values of *a* and *p* to minimize the measurement – prediction mismatch as shown). However, as long as the optimizer used is local, there is no guarantee that the obtained solution is a global minimum of (1). Moreover, there might be more than one solution (with different values for *a* and *p*), scoring the same (optimal) goodness of fit, which should be taken in consideration when biological insight about the interrogated system is to be extracted. Therefore, a large number of runs is performed each one starting from different (random) initial guesses, to obtain a family of near optimal solutions. shows the MAE of 500 solutions, obtained from equal runs of the proposed NLP approach each one starting from a different initial guess for the parameters *a* and *p.* Most of the runs resulted in solutions with very similar (±3%) MAEs. This indicates that although the IPOPT optimizer, used herein, is not global, it furnishes near-optimal solution points independently on the initial guess. In, an “average” pathway for these 500 runs is illustrated. The opacity of each of these edges corresponds to the average activity of the respective reactions over the 500 runs. For a discussion on the optimum size of the family of solutions see Supporting Information 1. ## 4 Removing conflicting and redundant reactions from the PKN Optimization of the PKN to the data at hand results in a set of values for the model parameters (*a* and *p*) that minimize the measurement prediction mismatch, as defined). Subsequently, we iteratively remove reactions from the PKN (every time a reaction is removed we re-optimize the PKN) while monitoring the fitness error to identify all reactions that are not vital in fitting the signaling dataset, either because they directly contradict the data, or because they are non-identifiable. Non-identifiable reactions are those whose presence in the model cannot be validated nor disproven based on the data at hand. This may occur when signal transduction from a cytokine to a measured protein can be achieved by a number of different pathways, and there is no definite way to identify which one is really functional. Consequently, removing a non- identifiable reaction from the PKN has no effect on the fitness error. In an attempt to remove conflicting reactions and tackle over-parameterization, we gradually remove reactions from the PKN until the fitness error starts increasing (i.e., the algorithm can no longer fit the dataset at hand). At that point there are no more conflicting or non-identifiable reactions left in the model, but all of the remaining ones are vital for fitting the data. At every iteration, the reaction with the lowest activity is removed (variable *z_i^k* in the formulation). The activity of each reaction mostly depends on the parameter *a* (gain) of the reaction and directly correlates to the “amount of signal” propagating downstream. In this manner, the least significant reaction is removed at every iteration. Even though the sequence reactions are removed by will affect the obtained solution (i.e., the solution is not unique), it is guaranteed to be optimal since only conflicting/non-identifiable reactions are removed and key property of these reactions is that their removal does not affect the fitness error of the solution. Results are illustrated in. shows how the algorithm performs when reactions of the PKN are removed in order of increasing significance. The x-axis corresponds to the number of removed reactions, while the y-axis corresponds to the MAE of the solution. As illustrated in, up to 10 reactions can be removed (20% of the initial topology) without affecting the goodness of fit of the solution. More than that, vital reactions are missing and the MAE increases significantly. Small fluctuations in the figure are attributed to variations of the fitness error of the solutions (±3%). shows the solution after removing conflicting and non-identifiable/redundant reactions.The above-mentioned procedure results in the identification of one of possibly many optimal and identifiable solutions, the superposition of which is the family of solutions as defined in paragraph 6.3. ## 5 Compartmentalization of the large-scale topology Before optimizing the large-scale model, the PKN is compartmentalized by grouping together nodes that share identical response under all experimental conditions, to reduce the parameter space. In similar fashion to the medium-scale model in, the large-scale pathway in also includes a number of non-identifiable reactions, in the sense that signal transduction from a cytokine to a measured protein can be achieved by a number of different pathways and there is no definite way to identify which one is truly functional. In pathways of this size, however, is not efficient to exhaustively remove reactions until the optimizer can no longer fit the data at hand. Instead we propose an alternative method for reducing the parameter space. We propose a compartmentalization scheme, based on the “equivalent classes” concept introduced in, for “grouping” nodes that share identical responses under all experimental conditions; thus resulting in an equivalent (compartmentalized) model where nodes have been replaced with their respective compartments, and reactions between nodes are now reactions between compartments. In more detail, we define a *compartment (C) as* every set of *non-measured* species, such that for every. Where \- *k = 1*, …,*n_e*, is the set of experiments. \- *x_j^k* is the predicted value of species *j* in experiment *k*. In this case study, we simulate the pathway running the NLP formulation under all experimental conditions present in the signaling dataset with nominal values for all parameters; subsequently, we format the simulation results in a 2d matrix, rows corresponding to the nodes in the pathway and columns corresponding to the different experimental conditions; we identify the nodes that share the same response under all conditions (i.e., identify replicate lines) and group them together in compartments; we replace every node in the PKN with its corresponding compartment and remove replicate reactions. This procedure is implemented using BASH. Since the nodes in a compartment share identical responses under all experimental conditions, their connectivity inside the compartment cannot, in principle, be interrogated based on the data at hand. Thus, it is purposeful to group these nodes together and update the PKN replacing nodes with the compartments they belong into. By doing so, we drastically decrease the parameters space. ## Application of the compartmentalization scheme to an illustrative example To better illustrate how the proposed compartmentalization scheme works to simplify the interrogated model, we construct the example model of. Node “A” serves as input to the pathway (stimuli), and activates nodes B1, B2; these interact with each other and finally activate node “C” that serves as a readout (signal). The proposed scheme groups B1-B2 into “Cmp” and simplifies the model as illustrated in. If data dictates: *A = 1;C = 1*, then reactions A→Cmp and Cmp→C are conserved. Else if *A = 1;C = 0*, then at least one of the above mentioned reactions have to be removed. demonstrate how the compartmentalization scheme can be too restrictive and may decrease the quality of the solution. In input nodes A1, A2 are connected to latent nodes B1 and B2; B1 activates C1 and B2 activates C2. After the compartmentalization procedure, B1 and B2 are replaced with compartment “Cmp” that activates C1 and C2. In the case where C1 is activated by A1, and C2 by A2; then either C1, or C2 will be misfitted in the compartmentalized model, since differential activation of C1 and C2 is possible only if either CMP→C1, or CMP→C2 are removed from the pathway. However, if either one of the two reactions are removed, then the respective signal (C1 or C2) will remain inactive under all conditions, thus misfitting the data. If no compartmentalization is performed, then the pathway can be optimized by removing (or decreasing the activity) of A1→B2 and A2→B1. This increase in fitness error caused by the compartmentalization procedure implies that grouping nodes B1 and B2 in the compartment Cmp should not have taken place if data were to fit perfectly. Cases like this may arise when limited experimental conditions are available, since it is more likely for nodes to be grouped together. E.g. If only one condition is available, then all nodes will be grouped in a single compartment. In such cases compartmentalization of the PKN is not recommended. In all cases the solution should be manually inspected to ensure that the remaining fitness error is not caused by the aggression of the compartmentalization scheme. # Supporting Information We would like to acknowledge the contribution of (i) Aikaterini D. Chairakaki for performing the proteomic experiments used in the optimization of the large- scale signaling pathway; (ii) Thomas Weiss for providing the primary hepatocytes used in the proteomic experiments; (iv) The Computational Infrastructure for Operations Research (COIN-OR) project for providing the Interior Point Optimizer (IPOPT) free of charge. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived the Non Linear Programming formulation: AM MKM INM. Implemented the Non Linear Programming formulation: AM. Edited the manuscript: MKM JSR AM DAL. Analyzed the data: INM. Contributed reagents/materials/analysis tools: AM JSR INM. Wrote the paper: INM LGA.

# Introduction Cochlear implant (CI) users experience greater difficulty than normal-hearing (NH) listeners to understand speech when background noise is present. In addition to this general problem, speech-in-noise performance also varies considerably across CI users. Some CI users show speech understanding that is comparable to that of moderately hearing impaired listeners, whereas in others a speech reception threshold (SRT) in background noise cannot be specified, because 50% speech understanding cannot be reached even in quiet. Many individual factors of CI users may influence their speech-in-noise performance. One factor that is widely discussed in the literature is the limited spectral resolution available to the CI user compared to NH listeners. Spectral resolution in CI users can be assessed in different ways. Objective (physical) measures include electrical field imaging (EFI,,) and electric compound action potentials (ECAPs,), which offer electrode-specific and thus frequency-specific measures of the electrical field spatial spread in the cochlea. Subjective (perceptual) measures include place pitch discrimination, spatial tuning curves, and electrode discrimination. These subjective measures also characterize spectral resolution frequency- specifically, whereas other subjective measures such as spectral ripple discrimination or detection and spectral modulation thresholds usually employ broadband stimuli with variable spectral contrast, which are more similar to speech. Direct strong relations between spectral resolution and speech intelligibility in these studies have so far remained elusive. There are to our knowledge currently no links investigated between speech intelligibility and spatial spread assessed using EFI. Spatial spread assessed using ECAPs was not found to correlate significantly to speech-in-noise performance. Subjective, frequency- specific measures show modest correlation to speech performance, such as for tuning curves inferred from gap detection, pitch ranking, or electrode discrimination, but other studies also show no correlations to speech performance. Correlations between speech performance to subjective spectral resolution measures with broadband stimuli show mixed results with some studies claiming strong correlations using, e.g., spectral modulation thresholds, but also studies which did not find such correlations,. Possible reasons for these mixed results may be other individual factors involved in determining speech-in-noise performance, which limit the predictive power of the single factor spectral resolution. Individual factors independent of spectral resolution that influence speech-in- noise performance are numerable. The most important investigated so far are age, duration of deafness, duration of hearing impairment, etiology, hearing aid usage, socioeconomic status, and a general cognitive ‘ability’, which can be measured using cognitive tests. The predictive power of these factors for speech performance either alone or combined is, however, relatively low, explaining less than typically 25% of the variance in speech tasks. Computer model studies not involving human subjects allow systematic investigations of individual factors on the predicted speech in-noise- performance. Without comparison to actual CI users, however, these studies remain theoretic predictions. The human subject in these studies is replaced by a pattern recognizer that labels the processed acoustic signals (restricted by the factors investigated) according to its training and thus “recognizes” the speech items. The recognizer can either work with restricted training, for example in the form of a “frozen speech approach”, which means that exactly the same speech recording (and only one recording per item) is used for training and testing,, or with statistical speech models based on several recordings per speech item,. Fredelake and Hohmann showed that wider electric field spatial spread functions that are uniform across electrodes resulted in higher SRTs and thus poorer speech-in-noise performance using restricted training. A similar trend was observed in their study when the cognitive ability was modelled by adjusting internal noise applied on the speech features. Stadler and Leijon showed with a statistical speech recognition backend that an incorporation of a measure of spectral resolution has some predictive power for individually modelled SRTs. However, their work also shows how difficult it is to estimate spatial spread reliably and that such a reliable estimation is crucial for SRT-predictions, with large intra-individual differences across test-retest. The current study aims at systematically analyzing the separate and combined effect of electrical field spatial spread and internal noise standard deviation on predicted speech-in-noise performance in a computer model for electric stimulation of the auditory system in combination with a statistical model of speech, by employing an automatic speech recognition system. Furthermore, it is investigated if an incorporation of one or the other factor, as newly collected in a group of individual CI users using Cochlear devices, improves the goodness of prediction of individual CI users’ speech-in-noise performance. Such a computer model approach allows to go beyond linear contributions of each of these factors to speech-in-noise performance, because both factors electrical field spatial spread and internal noise will nonlinearly interact within the model. The manuscript is organized as follows: After a systematic evaluation about the effect of electrical field spatial spread and internal noise standard deviation in isolation on SRTs predicted by a physiologically-inspired computer model, the measurement data of individual CI users is investigated in terms of predictive power for measured SRTs using linear tools, such as correlation coefficients and a generalized linear model. The physiologically-inspired computer model is then individualized systematically to different degrees, based on measurement data: individualization based on spatial spread alone, internal noise alone, and combined individualizations are realized. Predicted and measured SRTs are compared, and the goodness of prediction is quantified. # Methods ## Model structure ### Model front end This study uses the model front end of Fredelake and Hohmann, which is based on the dissertation of Hamacher. The model is used here essentially as previously reported in, therefore, the model description will be kept brief. A sketch of the model structure is shown in. The speech and noise mixture (at a given SNR) is first processed by the advanced combinational encoder (ACE) CI speech coding strategy (cf.,) giving an electrical pulse stimulation pattern on 22 electrodes. In agreement with, the electrodes were positioned centrally within a 35mm long, 1-dimensional cochlea. Subsequent to the electric stimuli, a spatial spread function on each of the 22 electrodes is used to simulate the transfer of the electric pulse onto each one of the auditory nerves, which were equally distributed along the entire length of the cochlea. In and in experiment 1 of the current study, each spatial spread function is an idealized symmetrical double-sided exponential function with width λ (i.e., the distance from the center of the double-sided exponential to 1/e of the maximum amplitude) in millimeters. However, these spatial spread functions can also be individualized according to spatial spread functions measured in actual CI listeners. This spatial spread function serves as one of the major factors investigated in this study on speech-in-noise performance. The auditory nerves (AN) are modeled as leaky integrate-and-fire neurons with stochastically variable absolute and relative refractory times, latency and jitter, as well as a neuronal membrane noise. In the current study, 1000 AN cells were modelled. Afterwards, non- overlapping groups of adjacent auditory nerve cells are formed each associated with the electrode closest to the group. The spatial limits of each group are defined as the arithmetic midpoints between the position of the associated electrode and the positions of its left and right neighbors. Beyond the most basal and apical electrodes this grouping procedure is applied with a constant group width of 0.75 mm. Spike trains within the groups are temporally integrated including a forward masking model. This results in an “internal representation” (IR), a spectrogram-like matrix of 46 rows, and columns at a frame update rate of 500 Hz. The excitation in each IR (amplitude of each time-frequency element) typically ranges between 0 and 50, in agreement with IRs shown in (their). Each element of the IR was multiplied with Gaussian noise (with a mean of 1 and a variable standard deviation, typically between 0.025 and 0.3), which is termed “internal noise”. This internal noise limits the predicted speech-in-noise performance and is used as the second major factor whose effect on individual and systematic SRTs is investigated in the current study. ### Model backend The Framework for auditory discrimination experiments (FADE) was used as speech pattern recognizer that provides a good generalization about the trained speech in the sense that it uses a statistical model generated from several speech utterances for a given word. The same framework was also used in combination with the electric model of Fredelake and Hohmann in. The details of this approach are given in and will be briefly described here: 120 sentences of the Oldenburg sentence test mixed with stationary OLnoise, each at -12 dB SNR to 21 dB SNR in 3 dB steps were processed by the model front end resulting in whole- sentence IRs. This procedure was repeated 8 times for each SNR with different temporal passages of the noise, where 7 of these repetitions served as training and 1 as test material. Whole-word models with 6 states in a standard Hidden- Markov-Model (HMM) based on the Hidden-Markov-Model Toolkit (HTK,) were trained using a Gaussian mixture model consisting of only one Gaussian distribution (with parameters mean and standard deviation). These models were used for the recognition of 600 presented words (contained within 120 sentences of 5 words each). Note that this approach does not receive separate words, but processes the entire sentence. The FADE framework automatically looks for word boundaries, because the HTK grammar was restricted to containing five subsequent words framed by a start silence model and a stop silence model. All combinations of training and testing SNRs were calculated resulting in combinations with low scores (at low SNRs) and high scores (with both training and testing having high SNRs at the same time), showing iso-score lines across different combinations. An interpolation between the two lowest testing SNRs along the 50%-iso-scoreline was then chosen as the predicted SRT. The motivation for this procedure was that also humans have acquired their speech discrimination and identification ability at a variety of different SNRs and should be able to make use of the “best- matching” training SNR to base their decision (in order to get best possible performance). ## Experiment 1: Systematic model evaluation The aim of Experiment 1 was to systematically investigate the effect of spatial spread and internal noise on model-predicted SRTs. Therefore, SRTs were predicted as a function of different electrical field spatial spreads with a constant internal noise standard deviation σ_int = 0.19. This σ_int was chosen as the average strength used also in the individualization experiment 4 (see below). Furthermore, SRTs were predicted as a function of σ_int with constant electrical field spatial spread λ equal to 9 mm. The same spatial spread function for all electrodes was chosen within a given spatial spread (in mm), giving a homogenous array, for simplicity. ## Experiment 2: Linear models of measurement results from individual CI listeners ### Participants 14 CI users aged between 34 and 85 years (median 64.5 years) participated in this study at the German Hearing Center of the Medical University Hanover. All participants were using Cochlear devices equipped with the ACE sound coding strategy and had at least 1 year of experience with their own CI. Therefore, the tested group of listeners was controlled for having the same device and signal processing strategy. For bilateral CI users only the side obtaining the best speech performance was tested. If a CI was worn on the other side, it was switched off during the measurements. Demographic information about the participants is shown in. The study protocol was approved by the institutional medical ethics committee of the Medical University of Hanover. All CI users gave their informed written consent to participate in the study. ### Electrical field measurements The electrical potential distribution in the perilymph was measured using the Nucleus Interface Communicator (NIC; Cochlear Corp., Sydney, Australia) to stimulate and record from the electrodes of each CI user. It is known that the potential distribution depends on individual factors such as the geometry of the cochlea and the electrode positions. Each electrode was stimulated in monopolar mode using biphasic pulses with amplitude 106.50 μA, a phase width of 25 μs, and an inter-phase gap of 8 μs. The voltage was recorded on the same and on all the other electrodes, and normalized by the current amplitude of the stimulating biphasic pulses, resulting in an intra-cochlear potential map. Note that the physical unit of this normalized voltage is given in Ω. More details about the measurement procedure can be found in. ### Text-reception-threshold test An adjusted version of the Text-reception threshold (TRT) test, in detail described in, was used to assess the performance of the listener in visually combining fragments of words to a full sentence. This test displays sentences of the Oldenburg sentence test (e.g., “Peter kauft drei nasse Schuhe”, engl. “Peter buys three wet shoes”,) on a computer screen and masks them with random bars, mimicking the masking effect of a fluctuating noise with speech-like modulation. The random bars masker was chosen, because this masker has shown highest correlations to SRTs in stationary noise (out of three tested masking patterns,). The participant is asked to repeat the words that he/she can read. The percentage of sentence coverage with bars is adaptively adjusted during a measurement run (consisting of 20 sentences) until 50% of the words are correctly repeated. This coverage serves then as a non-audiological estimate about the ability of the participant to combine word fragments. Before the actual measurement data collection, two familiarization runs of 20 displayed sentences each were finished by each participant. ### Anamnesis assessment The participant’s anamnesis was assessed using a questionnaire, following procedures described in and. Age, year of first notice of the hearing loss, start of profound hearing loss (defined by inability to use the telephone), usage of hearing aids during the phase of profound hearing loss, year of implantation, and self-reported etiology were assessed on this questionnaire. ### Speech intelligibility measurements Speech intelligibility in noise was assessed using the Oldenburg sentence test (Wagener et al., 1999) adaptively, aiming at the SNR corresponding to 50% speech intelligibility (defined as SRT). Stationary, speech-shaped noise (OLnoise) and speech were presented using a frontal loudspeaker at 1 m distance to their own speech processor. The presentation level was set at 60 dB SPL (A). Two test lists were conducted in advance to the actual measurement to familiarize the CI user to the test. ## Extraction of parameters for model individualization ### Electrical field spatial spread The spatial spread of the electrical field in the perilymph was estimated by fitting single-sided exponential functions to each side of the off-diagonal elements of the intracochlear potential map, allowing a vertical offset to be present (i.e., exponential functions were not forced to approximate 0 for abscissa positions towards ± infinity). Separate offsets were chosen for the apical and basal ends of the curves, which allowed much better fits to the normalized voltage data than without. Within this manuscript the recordings at the stimulating electrode are disregarded, as these values are dominated by the electrode-tissue impedance and not by the anatomy. A linear interpolation was done in the region of ±0.75 mm around the stimulating electrode and an extrapolation was done to regions of the cochlea not covered by the electrode array. The linear interpolation was done in contrast to (who extended the exponential fits towards the stimulating electrode), because the steepness of some of the single-sided exponential functions would have resulted in extraordinary peaky spatial spread functions that would have dominated the signal transmission in the CI model. The procedure resulted in 22 spatial spread functions per CI user—one for each electrode. shows spatial spread functions (gray continuous lines) that were fitted to measured raw normalized voltage data for participant 08 as a typical example. For electrode 11 both the fitted spatial spread function (black continuous curve) and the 21 raw normalized voltage data points (black diamonds) are shown. The fit closely matches the measured data in the region of the cochlea covered by the electrodes. The spatial spread functions across the electrodes (gray lines) exhibit large differences in this participant. To quantify the width of each spatial spread function, full-width-half-maximum (FWHM) values were extracted from each fitted double-exponential curve as the full width halfway between the maximum and 0kΩ. shows the FWHM values of each fitted spatial spread function for each electrode (a) and each participant (b). FWHMs are highly variable across electrodes and across participants. There is a tendency to wider spatial spreads for low electrode numbers (more apical electrodes with a median of 10.3 mm for electrode 1) compared to narrower spatial spreads for high electrode numbers (more basal electrodes with a median of 5.0 mm for electrode 22). FWHMs of spatial spreads averaged across all electrodes are between 5.1 mm for participant 46 and 9.8 mm for participant 61. ### Internal noise modelling Internal noise standard deviation σ_int is adjusted in the current study using two different factors, which are the patient anamnesis and the cognitive performance of the patient. The phenomenological model of and was used to calculate the “auditory performance” (AP) from the factors assessed in the anamnesis questionnaire, which is a number (in %) that quantifies the expected detriment in speech recognition performance from the individual anamnesis data according to this phenomenological model. This factor may be interpreted as the deprivation of the auditory system preceding the implantation, which depends on duration of moderate and severe/profound hearing impairment, usage of hearing aids, age at implantation and etiology. In detail, the AP is calculated using. <img src="info:doi/10.1371/journal.pone.0193842.e001" id="pone.0193842.e001g" /> AP = Dur ( mHL ) · ( − 0.23 % / y ) \+ Dur ( sHL ) · Δ s \+ B 1 \+ B 2\. In Dur(mHL) is the duration of moderate hearing loss in years, Dur(sHL) is the duration of severe hearing loss in years, Δs is a factor that depends on the usage of hearing aids during the phase of severe hearing loss prior to implantation (-0.83%/y for no, -0.64%/y for one, and -0.45%/y for two hearing aids). These terms were taken from, who inferred these by investigating data of 2251 CI patients. Duration of moderate hearing loss is defined as the difference in years between first self-reported notice of hearing impairment to inability to use the telephone with the impaired ear. Duration of severe hearing loss is defined as the difference in years between inability to use the telephone to implantation date. B1 and B2 (both in %) are taken from, who investigated the same pool of CI patients. B1 and B2 reduce or increase the AP based on the patient’s age at implantation (B1, see) and etiology (B2, see). The TRT-test result was used to quantify the (non-audiological) cognitive performance of the participant. Three different ways were realized to determine the individual σ_int: (1) using the TRT-test result only, (2) letting TRT-test and anamnesis data contribute with equal weights and (3) using the anamnesis data only. Pilot testing with the model showed that a reasonable range of internal (multiplicative) noise standard deviations is between σ_int = 0.15 and σ_int = 0.25 (σ_int is a scalar without a unit). Therefore, the ranges of individual factors were then linearly mapped onto this range. This means that the poorest performer was assigned the highest noise standard deviation of 0.25 and the best performer was assigned the lowest σ_int (0.15). shows individual σ_int values for all participants derived either using the aforementioned three combinations of TRT- test result and patient anamnesis. A color code was chosen to visually highlight good (green), moderate (black), and poor (red) performance. Note that adjustment of the internal noise due to the results of either of those tests can only be a very coarse model of limiting human cognitive performance and is not intended to model the details of functional or dysfunctional cognitive processes in human listeners. ## Generalized linear model A generalized linear model (GLM) was used to assess the predictive power of each of the three individually extracted parameters: average FWHM of the spatial spread, total auditory performance (AP), and TRT-test result. Statistical independence and a linear combination of the three normally distributed variables were assumed. ## Experiment 3: Different degrees of model individualization Experiment 3 investigates the question if an individual incorporation of either the EFI data (assessing the electrical field spatial spread) or the internal noise (σ_int parametrized by the TRT-test data or the AP or both) into the physiological model of CI user’s speech intelligibility can improve the prediction SRTs. Therefore, a step-wise approach was taken using three sub- experiments: 1. Experiment 3a: Internal noise individualization using either only the AP, only data from TRT-test, or a combination of both AP and TRT-test with equal weights. 2. Experiment 3b: Electrical field spatial spread individualization only 3. Experiment 3c: Full individualization of electrical field spatial spread and internal noise with noise strength estimated from either only the AP, or only data from TRT-test, or data from both AP and TRT-test with equal weights. # Results Three experiments have been designed to assess the efficacy of the model to predict SRTs of CI users. Experiment 1 performs a systematic analysis of the different parameters of the physiologically-inspired CI model described in the methods section. Experiment 2 presents the individual factors measured in CI subjects which may either in isolation or combined (linearly) correlate with speech performance. Finally, experiment 3 incorporates the individual factors into the physiologically-inspired CI model and compares the model predictions with the actual speech performance measured in each CI user. ## Experiment 1: Systematic model evaluation shows SRT predictions varying the electrical field spatial spread (in the form of the parameter λ) systematically and uniformly across all electrodes. An average σ_int = 0.19 was chosen for this model variation, as this value is also used as an average for the internal noise strength for model individualization in experiment 3. Predicted SRTs increase (i.e., speech-in- noise discrimination is poorer) systematically as the electric field spatial spread of the model widens. shows SRT predictions varying the internal noise strength systematically. An average electrical field spatial spread function of λ = 9 mm was chosen also for this model variation. Predicted SRTs increase, as σ_int increases. Note that the test-retest reliability of the predicted SRTs was calculated to 0.4 dB, based on several repetitions of predicting the same SRT. ## Experiment 2: Linear models of measurement results from individual CI listeners ### Correlations of raw measurement data shows scatter plots of average FWHM of the spatial spread (panel A), auditory performance (panel B), and TRT-test result (panel C) on the ordinate against individual SRT. Each participant is denoted using her/his ID number. The range of SRTs covered by the participants is between -0.1 dB SNR and 6.2 dB SNR, which corresponds to the range of SRTs covered in the systematic model evaluation. Neither the average electrical field spatial spread ( panel a), nor the auditory performance alone ( panel b) correlated strongly with the measured SRT using Pearson’s correlation coefficient (i.e., linearly). The trend of the (non- significant) correlation even showed the opposite sign than expected beforehand (wide spatial spread tended to be related to low SRTs and high auditory performance tended to be related to high SRTs). In contrast, the TRT-test result in the form of percentage of tolerated sentence coverage ( panel c) correlated highly, r = -0.72 (p \< 0.01) with measured SRT, indicating that participants, who could well combine fragments of words in a written sentence showed also better speech-in-noise performance and vice versa. The most probable linear regression line is plotted (green dashed) in those panels with significant correlations. ### Predictions using a generalized linear model A generalized linear model (GLM) was used to assess the predictive power of each of the three parameters: average FWHM of the spatial spread, auditory performance (AP), and TRT-test result. The fitted GLM can be described by : $$SRT_{pred}\left( {dB} \right) = 11.62 - 0.0183 \cdot FWHM\left( {mm} \right) + 0.0644 \cdot AP - 0.1403 \cdot TRT(\%)$$ The GLM-predicted SRTs as a function of the measured SRTs are shown in). The fitted GLM provided a significantly better prediction than the null hypothesis of a constant model (F = 5.57, p = 0.017). In line with the correlation analyses above, only the TRT-test result provided significant predictive value for the SRT (p = 0.015). The SRTs predicted by the fitted GLM showed a highly significant correlation coefficient with measured SRTs (r = 0.79, p = 0.001), explaining 62% of the total variance. ## Experiment 3: Model individualization Three different degrees of individualization in the physiological model of CI user’s speech intelligibility were tested: One version that individualizes the electric field spatial spread only, one version that individualizes the internal noise only, and one that individualizes both factors combined. These three model versions were chosen to get a comprehensive picture about which factors are crucial in a nonlinear model mimicking speech-in-noise performance of CI listeners. shows Pearson’s correlation coefficient, the probability p that the null hypothesis of no correlation between measured and predicted SRTs needs to be rejected, RMS-error, and Bias between measured and predicted SRTs. In general, the model shows a negative bias of 2–3 dB with respect to the measured data, i.e., it underestimates the average performance of the listeners. There is only one significant correlation within the table of results: If the model’s internal noise is individualized to the TRT-test result only (taking an average spatial spread that is uniform across all electrodes), the highest correlation between measured and predicted SRTs is obtained. These SRT-predictions correlate highly significantly (p \< 0.01) with SRT-measurements (r = 0.68). No significant correlations were found when individualizing both the internal noise and the electrical field spatial spread in combination, or when individualizing the electrical field spatial spread only. shows scatter plots (predicted vs measured SRTs) with a part-individualized model version (individualizing internal noise only from TRT-test result) in panel (a) and a full individualization (internal noise also from TRT-test) in panel (b). When individualizing internal noise only, the high correlation (r = 0.68) to measured SRTs is clearly visible in panel a): predicted SRTs follow a diagonal direction with respect to measured SRTs. However, the model produces SRTs (around 4 to 7 dB SNR) that are more in line with listeners showing poorer SRTs and there remains a bias towards listeners with better SRTs. The highly significant correlation found with individualizing the model using internal noise only (panel a) is lost if additionally the electric field spatial spread is individualized (panel b). Note that the model predicted speech intelligibility scores below 50% for all SNRs tested for participant 048 in panel b. Therefore it was not possible to predict an SRT for this listener. The listener was thus excluded from the correlation coefficient calculation in this panel. The range of predicted SRTs is small (4 to 7 dB SNR) when individualizing internal noise alone and is sufficiently larger (-2 to 20 dB SNR) when individualizing both factors, which highlights the nonlinear behavior of the model with respect to these two factors. # Discussion This study systematically evaluated whether linear tools or a (nonlinear) state- of-the-art CI computer model can be used to predict individual speech performance of real CI users. The effect of individualizing different parameters of the front-end model (electrical spatial spread, cognitive noise) for predicting individual speech performance in a group of 14 CI users was investigated. In general, the results of non-individualized versions of the computational model show that the model predicts an improvement (decrease) of individual SRTs with narrower electrical field spatial spread and smaller internal noise standard deviation σ_int in agreement with the expectation. However, only an incorporation of σ_int estimated from the individual TRT-test result shows highly significant correlations to measured SRTs in CI subjects. The amount of correlation is of the same magnitude as the raw (linear) correlation between TRT-test and SRT. An additional incorporation of electrical field spatial spread renders this correlation insignificant. ## Effect of front end The two factors electrical field spatial spread and σ_int within the model front end both showed considerable influence on predicted SRTs in the systematic evaluation (experiment 1). In line with predictions by and using the same model front end, a systematic decrease in electrical field spatial spread or in σ_int improves (decreases) SRTs. In the model, the wider electrical field spatial spread functions will cause wider modelled neural excitations, resulting in spectrally smeared IRs. With regard to spectral resolution, the same trend can also be found in vocoder studies, i.e., studies with NH subjects listening through an acoustical simulation of the CI user’s signal processing. Vocoder studies indicated that the number of independent frequency channels in CI users is effectively much lower than in NH listeners without vocoder processing, which limits speech-in-noise performance. The overlapping bandwidth of the vocoder channels spectrally smears the vocoder output and thus limits the spectral resolution. This affects speech recognition with higher speech scores for narrower bandwidths, improved SRTs for steeper vocoder filter slopes, and poorer speech scores going alongside reduced spectral ripple discrimination. These vocoder studies (similar as in computer model studies) carry potentially less individual variability than studies with actual CI users due to the systematic control over the spectral resolution in the vocoder and a larger homogeneity across the NH subjects. In actual CI users, however, the literature gives a less coherent picture about the effect of spectral resolution on speech perception. Psychophysical measures of spectral resolution,, evoked potentials, and spectral shape perception, have been reported to correlate in varying degrees to speech perception. Highest correlations were found using those measures that assess spectral resolution across the whole cochlea, possibly because the stimuli used in these tests are closer to actual (broadband) speech stimuli. ## Individual predictions Computer models of CI listeners currently work well for contrasting different preprocessing algorithms and different acoustic situations with averaging over CI individuals. Also within one CI listener high correlations between predicted and measured SRTs were found by using an envelope-correlation measure based on the electrodograms generated by the individual user’s CI. In contrast to the study of, the current study focused on correlations (between predicted and measured speech performance) across individuals in one specific (standard) acoustic situation that is widely used as a clinical test. Such correlations across individual CI users in one test are very rarely reported. An exception is the study of Stadler and Leijon. In their study, a simple model as well as a physiologically detailed model of signal processing in CI users was individually adjusted due to results of a subjective spectral discrimination task. They found that both models could account for a large proportion of the speech-in-noise performance variance measured in CI users with a standard speech test. However, the measure that used to assess spectral resolution uses wide-band signals, which makes this spectral resolution task closer to a speech-in-noise task (providing potentially a higher predictive power from the raw data) than the measure of spectral resolution used in the current study. In the current study, the individualization of spatial spread is based on intra-scalar voltage distribution measures (EFIs) that are electrode-specific and completely objective, i.e., they can be measured without interaction by the CI user. The hypothesis is that wider voltage distributions in the scala tympani should lead to increased spatial spread and in turn to poorer speech intelligibility (higher predicted SRTs) both in the CI user and in the model. However, both the raw data and simulation data with individualized spatial spread by using measured EFIs on each electrode and CI user in the model were not found to correlate directly to the SRT or to reduce the RMS error in predicting the SRT in the physiologically- inspired computer model (2.9 dB for the non-individualized spatial spread and 6.3 dB for the individualized spatial spread, see). Even a decrease of correlation coefficient is found when spatial spread is individualized in addition to the internal noise. This, and the additional absence of correlation to SRTs using the electrode-averaged electric field spatial spread widths indicates that this peripheral factor (as measured in the current study) is not predictive for individual SRTs. This result should be interpreted with caution, because it does not prove that human neural resolution has a negligible effect on speech-in-noise performance of CI users. It could also mean that the normalized electrical potential distributions across CI electrodes include variations that may not correspond to human neural resolution and are thus inadequate as a measure for these. Since model results in experiment 1 and other studies such as have shown that the human neural resolution is an important factor for speech-in-noise predictions of actual CI users, it is worthwhile to pursue this research further. EFI, as measured in the present study, however, can be excluded as a technique yielding predictive value for SRT-prediction. The internal noise standard deviation inferred from the TRT-test result showed a high predictive value with correlation coefficients ranging between r = -0.72 (raw TRT-test result correlated with SRT) and r = 0.68 (with the part- individualized model). This is in line with the data reported in, who found significant correlation between TRT and SRT in a much larger sample size of 90 CI users. Their correlation coefficient was substantially lower (r = -0.27) and it is currently unclear what the reason for the difference is. One difference is that the study recruited participants with CIs from three different manufacturers and different signal processing strategies across and within one manufacturer, whereas the current study controlled for these variables. The fact that the model simulation with TRT-individualized internal noise led to a similar correlation as with the raw TRT data is not surprising, because in this model version the variance of the internal noise is artificially forced to correlate with the TRT results. However, the relatively high correlation coefficient suggests that the internal noise individualization is a meaningful way of representing some of the more central factors in the model. ## Limitations of the current study and other factors This study focuses on the assessment of several, but not all individual factors that may contribute to individual speech-in-noise performance. One important other factor not implemented in the computer model so far is the involvement of the status of the afferent spiral ganglion cells. The EFI measure can roughly correspond to neural excitation of the spiral ganglion cells only if a homogenous distribution of functional AN cells is assumed and if the distance from electrode to the nervous tissue is constant along the electrode array. However, dendrites of AN cells may have retracted, AN cell density locally or totally decreased, or even dead regions of completely missing AN cells in the cochlea may occur. Better diagnostics are needed in order to include this factor in an individualized CI model, because currently there is no reliable test to estimate the status of the afferent spiral ganglion cells non-invasively in CI listeners. A constant distance of the electrode to that part of the nervous tissue where action potentials are generated is a further hypothesis that may be reasonable, at least in the first turn of the cochlear spiral, due to the circular placement of the electrode array. Further factors that may play a role are different individual TCL and MCL values and different loudness-growth functions. The internal noise, as it is applied in the present study, can only be a very coarse model of some of the cognitive processes that are involved in speech perception of actual CI users. From a signal processing point of view the internal noise is merely a distortion of the input signal to the central stage (the FADE speech recognizer) that remains unchanged in all model versions. To mimic more realistically differences in human cognitive processes, also variations in the back-end would be needed, such as smearing the state- transition probabilities of the trained HMM or randomly deleting some HMM states. This was out of the scope of the present study and even when doing so it would be hard to prove that such artificial modifications of the backend provide a good model for variations in cognitive processes in actual human listeners. Currently it is still unknown how to exactly model human cognitive speech processing and this paper has not improved our understanding of this problem. Future enhancements of the model could include spread of excitation measures using ECAPs instead of EFI measures, because ECAPs may be a better measure of human neural resolution. However, since ECAP spread of excitation measures are produced by auditory nerve responses, they are subject to a “double-application” of the spatial spread function from the electrode to the auditory nerve. A deconvolution as proposed by would be suitable to implement these measures in the CI model. Additional possibilities to improve the modeling of individualized measures of spectral spread include combinations of psychophysics and imaging data. To refine the modeling of the individual electrode-nerve interface, more detailed 3-dimensional models based on computer tomography data might be helpful (cf.,) to use within the frontend of the model. # Conclusions This study systematically evaluated a nonlinear model of CI user’s speech-in- noise performance with respect to the model-inherent factors electric field spatial spread and internal noise. Furthermore, the hypothesis was tested if an individual assessment of these factors with incorporation into the model can result into an improvement of individual SRT prediction. The predictions were compared to predictions with linear standard tools. The following conclusions can be drawn: 1. Predicted SRTs decrease (improve) with narrower electric field spatial spread, and with smaller internal noise standard deviation. 2. Only an incorporation of internal noise standard deviation estimated from the individual TRT-test result shows highly significant correlations to measured SRTs. The amount of correlation is of the same magnitude as the linear correlation between TRT-test result and SRT. An additional incorporation of electrical field spatial spread, as measured using normalized data, renders this correlation insignificant. This may suggest that spatial spread estimates from EFI data are not sufficient to capture individual differences in neural spectral resolution and hence differences in speech-in-noise performance. As the TRT-test has shown high predictive value in this study with a highly variable group of participants in terms of age and etiologies, the TRT test is recommended as an important factor for individual speech-in-noise performance. This factor can also be measured pre-surgically with the purpose of predicting SRTs post-surgically. This study shows that it is difficult to incorporate other factors into the individual prediction, at least with the simplifying assumptions that have been taken in the current study. # Supporting information This work was supported by the DFG Cluster of Excellence EXC 1077/1 "Hearing4all". Special thanks to Nils Schreiber for performing the FADE model predictions. The authors appreciate thorough and constructive feedback by Manuel Malmierca, Arne Leijon, and one anonymous reviewer on an earlier version of the manuscript. ACE advanced combinational encoder AN auditory nerve AP auditory performance CI cochlear implant dB deciBel DTW dynamic time warping ECAPs electric compound action potentials EFI electrical field imaging FADE framework for auditory discrimination experiments FWHM full width half maximum GLM generalized linear model IR internal representation NH normal-hearing RMS root-mean-square σ_int internal noise standard deviation SNR signal-to-noise ratio SPL sound pressure level SRT speech reception threshold TRT text reception threshold [^1]: The authors have declared that no competing interests exist.

# Introduction Poliovirus surveillance is essential to the success of the Global Poliovirus Eradication Initiative (GPEI). With poliovirus eradication nearing, rapid detection of polioviruses from specimens collected through acute flaccid paralysis (AFP) and environmental surveillance systems is crucial to monitor eradication progress. Improving methods and procedures by increasing sensitivity and robustness is a major objective of the Global Polio Laboratory Network (GPLN). Molecular methods, like real-time reverse transcription PCR (rRT-PCR), can identify and distinguish wild and vaccine-like polioviruses isolated from AFP cases or environmental sources, but algorithms rely on sequencing as the gold standard to provide final verification. Intratypic differentiation (ITD) by rRT-PCR is key to the GPLN surveillance workflow, to rapidly screen poliovirus isolates of programmatic importance after cell culture isolation and before verification of a subset of isolates by sequencing. Over the years, as technology and polio eradication needs have evolved, ITD assays have changed multiple times, from version 1 to version 6, to better meet the needs of the global network. The rRT-PCR screening kit, ITD 5.0, consists of six assays (EV+Sabin quadruplex, PanPolio \[PanPV\] assay, wild poliovirus type 1 (WPV1), PV type 2 (PV2) assay, wild poliovirus type 3 (WPV3)-I and WPV3-II assay used in conjunction with a decision algorithm to identify polioviruses of programmatic importance to be referred for sequencing. Poliovirus sequences inform the molecular epidemiology of the virus to help guide vaccination campaigns. The ITD 5.0 suite of assays has been adapted and modified from the previous version, ITD 4.0; evaluations with the ABI7500 rRT-PCR system report 97.7%–99.1% specificity and 92%–100% sensitivity. Most of the 114 accredited ITD laboratories use ABI7500 real-time PCR systems, but other instrument platforms, such as the Bio-Rad CFX96, Stratagene MX3000P, and Qiagen Rotor-Gene Q systems are used as well. The ability of the additional instruments to work with the poliovirus suite of diagnostic assays (e.g., with high sensitivity and specificity) is important to provide adequate global coverage in testing poliovirus. Most real-time platforms, like ABI7500, Bio-Rad CFX96, and Stratagene MX3000P, use Peltier elements for the regulation of heating and cooling of samples, and can produce different results from the Rotor-Gene Q system, which uses a rotary mechanism in which samples are spun continuously and heated and cooled with air. Differences between the two platform types are most apparent with assays using highly degenerate primers and probes, like the PanPV assay, which utilizes 21 mixed-bases and 8 deoxyinosine residues in order to identify all polioviruses. In this study, the PanPV assay was investigated because it showed an increased background signal on the Rotor-Gene Q platform that may lead to false-negative results. Accordingly, a new poliovirus probe was needed to reduce fluorescence background and improve sensitivity. Here, we report on an evaluation of ITD performance on different real-time PCR platforms and on pilot results for the novel poliovirus probe for the updated ITD 5.1 kit tested in five GPLN laboratories (Philippines, Pakistan, Madagascar, India and Democratic Republic of the Congo). # Material and methods ## Virus isolates and RNA transcripts for real-time RT-PCR platform validations A virus panel (n = 184) encompassing all PV serotypes (N = 158), non-polio enterovirus (NPEV) (N = 15), and non-enterovirus (NEV) (N = 11) (CPE-positive cultures but enterovirus-negative by ITD), was used to measure assay performance with the six ITD assays according to previously described methods. Isolates were derived from stools of AFP surveillance cases collected between 1999–2015 from GPLN and sent to the Polio laboratory at the Centers for Disease Control and Prevention (CDC) for confirmation or processing. In addition, two plaque purified WPV1 isolates (Accession no. KY941931 and KY941934) from Human Rhabdomyosarcoma cell lines (RD cells, ATCC cat# CCL-136) were included; these strains last circulated in the African (AFR) and Eastern Mediterranean (EMR) WHO regions, respectively. A reference polio type 1 virus from the National Institute for Biological Standards and Control (NIBSC, UK) (Sabin 1, Accession no. AY184219) was used for the evaluation of the updated PanPV probe. Synthetic poliovirus RNA transcripts were used to assess the performance of the initially designed PanPV probes following a previously published method. We generated three RNA transcripts derived from capsid viral protein 1 (VP1) sequences of the Sabin 1, Sabin 2, and Sabin 3 vaccine strains. All RNA standards were stored in single-use aliquots at −80°C until needed. Each ITD PCR reaction consisted of 10 μl of qScript^™ XLT One-Step RT-qPCR ToughMix^® (Quanta Biosciences, Beverly, MA), 1 μl of primers/probe(s) mix (contained in the ITD kit; Centers for Disease Control and Prevention \[CDC\], Atlanta, GA), 8 μl RNase-free water, and 1 μl of template (virus culture supernatant or RNA). ## Real-time PCR platform evaluation Each of the six ITD assays (EV+Sabin, PanPV, WPV1, PV type 2, WPV3-I and WPV3-II assays) was tested on four real-time PCR platforms: Applied Biosystems 7500 Real Time PCR System (ABI7500, ThermoFisher Scientific, Waltham, MA); Bio-Rad CFX96 (Bio-Rad Laboratories, Hercules, CA); Stratagene MX3000P (Agilent Technologies, Santa Clara, CA), and Rotor-Gene Q (Qiagen, Hilden, Germany). The ABI7500 is the most frequently used real-time PCR platform in the GPLN (90%) and is considered the “gold standard”. In order to accommodate the deoxyinosine-containing primers and probes, in earlier versions of the ITD run method, the ramp speed between annealing and extension for the ABI7500 was reduced to achieve higher specificity and sensitivity for the PanPV assay. The reduced ramp rate is a standard run method for the GPLN procedure. The ramp speed for the Rotor-Gene would also require a slowdown but because the Rotor-Gene Q software does not have that option, an additional temperature step was added between annealing and extension. The thermocycling conditions for each PCR cycler are listed in. ## Validation of updated PanPV assay The virus panel consisting of 184 polioviruses and non-polio enteroviruses was tested with the current PanPV assay and with the best-performing updated PanPV probe on the Rotor-Gene Q at the Polio laboratory at CDC Atlanta. The same samples were then tested with the updated PanPV assay on the other PCR cyclers (ABI7500, CFX96, and MX3000P) using the run method described earlier. ## Limit of Detection (LOD) for updated PV probe The LOD and ITD reactions were run as previously described. Briefly, WEAF-B1 WPV1, SOAS WPV1, and Sabin 1 isolates were tested in triplicate serial dilutions (10⁷ to 10⁰ CCID₅₀·ml^-1). The 95% LOD of the updated PanPV assay was determined by testing 20 replicates of the last dilution step with 95% positivity in the ABI7500. Thermocycling conditions were the same as described in. ## Pilot tests for updated PV probe in five GPLN laboratories After completing the evaluation of the updated PV probe in the Polio and Picornavirus Laboratory at the CDC in Atlanta, newly developed Zen PV probes were combined with primers at the CDC. The PanPV Zen primer and probe mix was shipped to five GPLN laboratories for pilot testing of CPE positive virus isolates from AFP surveillance stools. ## Data management, statistical, and visual analysis Any sample with cycle threshold (Ct)value \< 40 was considered a positive a positive result for the assay-by-assay comparison. To analyze all real-time RT- PCR data, Ct values were recorded for each sample and target. Results were compiled and edited using R. The McNemar test was used for parallel testing analysis using the gmodels package in R. Data visualizations were made using ggplot package in R and Prism 7.0, Graph Pad Software (San Diego, CA). To analyze background fluorescence between PanPV and updated PanPV from testing 32 WPV1 isolates, the raw fluorescence data from cycle 6, where background fluorescence stabilizes, was exported and compiled in Excel and R. The non- parametric Wilcoxon signed rank test was run in R to determine any significant difference in background fluorescence between the two assays. ## Ethical considerations CDC’s internal program for Human Subjects Research Determination deemed that this study is categorized as public health non-research for the purpose of human subject regulations. # Results ## High concordance between ABI7500, Rotor-Gene Q, CFX96, and MX3000P A total of 158 poliovirus isolates from 1999–2015 were selected from the CDC database. All poliovirus serotypes were confirmed by VP1 sequence using standard methods. Isolates were re-tested with the ITD 5.0 kit using ABI7500. All serotypes were detected by corresponding assays in the ITD (e.g., PanEV+, Sabin1+, PanPV+) including mixtures. The complete set of virus isolates was tested on the CFX96 and MX3000P, resulting in a 100% match for all six ITD assays (n = 184). Five of the ITD assays had 100% concordance between ABI7500 and Rotor-Gene Q; PanPV had 6 false-negatives out of 158 (3.8%) poliovirus isolates on Rotor-Gene Q. The six false negative virus isolates (clarified supernatant) were diluted 1:10 in Minimum Essential Media(MEM) and re-tested with the PanPV assay. All 6 virus isolates were positive after diluting and the results from all 6 assays were 100% concordant among the ABI7500, CFX96, MX3000P, and Rotor-Gene Q, indicating the false negatives were due to high background signals. ## Updated PV probe with Zen^™ quencher was superior to the standard PanPV probe A Zen^™ quencher was added as a second, internal quencher in the PanPV probe at the 8^th, 9^th, or 10^th base from the 5’ reporter dye sequence, respectively. The best probe (Zen at position 8) had reduced background compared to the standard PanPV probe (7.05 ± 0.35 and 61.02 ± 3.74 respectively). The updated PanPV probe with Zen^™ at position 8 was selected because it showed the lowest background combined with the highest fluorescent signal when tested with synthetic poliovirus RNA transcripts (Sabin 1, Sabin 2 and Sabin 3;). A total of 184 virus isolates (including 32 WPV1 isolates that were of programmatic importance) were tested in parallel using both PanPV and updated PanPV assays. The WPV1 isolates showed a lower average background fluorescence in the updated PanPV probe than the standard PanPV probe (8.13 ±0.001 and 68.01 ± 0.15, respectively); the difference was statistically significant (*P*\< 0.05). Six false-negative samples previously missed on the Rotor-Gene Q by the standard PanPV assay were positive with the updated PanPV probe. All poliovirus serotypes were detected by the updated PanPV assay when tested on the ABI7500, CFX96, MX3000P, and Rotor-Gene Q. Interpretation of results was simplified because the updated PanPV assay significantly reduced background signals. Even weaker positive signals were more defined, with curves clearly separated from the background and a higher signal-to-noise ratio. The interpretation remained the same for the other PCR platforms that were evaluated. ## Comparable limit of detection for updated PanPV and standard PanPV probes The 95% LOD was determined with three plaque-purified polioviruses representing three wild type 1 poliovirus genotypes as part of the quality assessment for any new assay deployed to the GPLN to show non-inferiority. The updated Zen8PV assay maintained the same sensitivity as the PanPV assay: the LOD was 1 CCID₅₀·μl^-1 for Sabin 1 reference virus and 10 CCID₅₀·μl^-1 for WPV1-SOAS and AFRO-WEAF-B1 reference virus templates, the same LODs as previously identified. ## PanPV probe with double quencher showed comparable results to PanPV with single quencher in five GPLN laboratories in pilot tests The new PanPV assay was piloted in GPLN laboratories that have AFP surveillance samples of programmatic importance from three WHO regions. The National Institute of Biomedical Research (INRB, Democratic Republic of Congo); Institut Pasteur in Madagascar; the National Institute of Virology Mumbai (NIVMU, India); the Research Institute for Tropical Medicine (RITM, Philippines); and the National Institute of Health (NIH, Pakistan) pilot tested the Zen PV assay. In collaboration with these GPLN partners, the updated PanPV assay was validated screening 293 poliovirus and non-poliovirus isolates (n = 17) from AFP surveillance at the INRB (n = 18); Pasteur Institute (n = 50); NIVMU (n = 41); RITM (n = 75); and NIH (n = 126). Most of the virus isolates were serotyped as Sabin 1 or Sabin 3 (n = 263). Both PanPV assays, using the newly designed Zen8PV probe and the standard PanPV probe, were run on the ABI7500 concurrently on the same plates for Ct value comparison. The specificity and sensitivity of the updated PanPV assay were 100% concordant in non-Rotor-Gene Q platforms compared to the previous version of the PanPV assay. In Rotor-Gene Q platforms, the updated PanPV assay did not have false negatives, unlike the previous version of the PanPV assay. The mean Ct values of the PanPV assay and Zen8PV were 24.3 and 22.9, respectively. The Zen8PV assay was 1–2 Ct’s lower compared to the standard PanPV assay, and interpretation of results was simplified due to reduced background with higher signal-to-noise ratios. # Discussion Rapid poliovirus detection from poliovirus isolates remains a crucial component of laboratory surveillance. We evaluated four machines, Rotor-Gene Q, Stratagene MX3000P, Bio-Rad CFX96, and the ABI7500, for performance with the GPLN assays. Previous versions of the ITD polio diagnostic real-time RT-PCR assays contained a standard PanPV assay (versions 1 through 5). We validated a new updated PanPV assay with Zen^™ quencher to increase sensitivity and interpretation based on the reduced background noise for the most common real-time PCR platforms in the network. All real-time PCR platforms had concordant results with the appropriate run profiles. The updated PanPV probe improved poliovirus detection by reducing background signals and making overall analysis simpler in both the poliovirus panel and in the pilot study. The validation of various platforms and the release of an updated PanPV assay will increase the robustness of the assays used by the GPLN and will decrease time spent analyzing data. Though the majority of its 114 polio diagnostic labs use ABI7500, the GPLN also includes laboratories using alternative real-time RT-PCR platforms, such as Rotor-Gene Q and Stratagene MX3000P. The rationale for choosing different platforms includes many reasons such as compatibility with other (non-polio) assays, institutional standardization, local sales and service, or other instrument availability. The advantage of real-time systems like the BioRad CFX96 and the Rotor-Gene is that they do not necessitate bi-annual calibrations. Since they use light emitting diodes (LED) as their light source, no lamp changes are needed, unlike the halogen lamps used in the AB7500 or Mx3000. The high nucleotide sequence diversity among polioviruses presents a challenge to the design of nucleic acid-based assays. Genomic sequences that encode strong amino acid conservation can still be highly variable because of codon degeneracy. To accommodate this variability, degenerate codon positions on the template were matched by mixed-base or deoxyinosine residues on the primers and probe. The specificity of the updated PanPV assay was 100% (184 out of 184 poliovirus and non-poliovirus isolates) and the sensitivity was 1 to 100 CCID₅₀·μl^-1 (Sabin 2, and WPV1 respectively). The PanPV real-time RT-PCR assay in the ITD 5.0 has excellent diagnostic specificities for a diverse array of poliovirus genotypes (100%). However, this results in higher background signals, leading to the potential misinterpretation of results as false-negative. To replace the old version of the PV assay, we updated the PanPV assay by adding a Zen^™ quencher as a second internal quencher within the PanPV probe in the 8^th position from the 5’ end of the probe sequence. The updated PanPV assay sensitivity and specificity were assessed with 15 non-polio enteroviruses, 11 CPE-positive enterovirus-negative samples, and 158 virus isolates, including all PV serotypes and relevant genotypes circulating in the past decade. One limitation of this study was that the limited number of non- polio enterovirus isolates available due to the selective nature of the L20B cells used for virus isolation. Multi-site validation of the updated PanPV assay in the GPLN showed identical or better results compared to the previous PanPV assay. In addition, the new updated PanPV assay reduced background signals, which simplified interpretation of results. A limitation of the updated PanPV pilot testing was the sample size: only 310 virus isolates were parallel tested with the updated and standard PanPV assay; most of which were Sabin 1 and Sabin 3 viruses, because there were not many wild PV isolates and only 17 PV type 2 to include in this comparison. This was primarily due to the advanced state of the global eradication program, where wild type 1 poliovirus is found in only Afghanistan and Pakistan in the WHO EMR region. In addition, the Rotor Gene Q cycler was used without an approved run method, thus a head-to-head comparison using the same cycler was not possible. The pilot testing was performed in laboratories that had both, programmatically important AFP surveillance samples (i.e. WPV1, PV2) and an approved cycler (i.e. AB7500). In August 2020 the WHO AFR region was certified wild poliovirus-free and it is now increasingly difficult to test assays prospectively. Retrospective testing is also becoming challenging since even potentially infectious material that might contain poliovirus type 2 has been discarded in order to comply with World Health Organization GAPIII requirements. Since its development and piloting, the new PanPV assay has been deployed to the GPLN for use in 2019 to laboratories that use Rotor Gene Q cyclers among others. The work described here serves one key purpose. It establishes a uniform standard for future ITD evaluations by providing a baseline to screen alternative platforms suited to a lab’s financial, scientific, diagnostic needs, as well as considerations of the GPLN. The continuous update and validation of methods and procedures are critical for any network, whether domestic or international, to incorporate new technologies and to improve detection sensitivity. Diagnostic networks must be prepared to handle evolving issues, both human-made, like logistics and rules and regulations, and molecular evolution of the pathogen (e.g. genetic drift). The example of the GPLN illustrates the need for collaboration and necessary background work required for a functional global network, which can serve as a model for future global laboratory networks. # Supporting information The authors wish to acknowledge Naomi Dybdahl-Sissoko and Hong Pang for their assistance in growing and handling reference virus isolates. We acknowledge the contributions of Ling Wei for her expertise with plaque purification of viruses, and Elizabeth Henderson and Jane Iber for their help with the specimen and VP1 sequencing databases. We thank David R. Kilpatrick for his contributions to the early development of several of the real-time RT-PCR assays. **Disclaimer:** The findings and conclusions in this report are those of the authors and do not necessarily represent the views of the Centers for Disease Control and Prevention. The use of trade names is for identification only and does not imply endorsement by the CDC or the U.S. government. [^1]: The authors have declared that no competing interests exist.

# Introduction Cerebral ischemia induces the loss or reduction of oxygen and glucose delivery to brain region affected causing disruption in production of adenosine triphosphate, increased reactive oxygen species (ROS) production, and sparking inflammatory cascades that may culminate in the death of both neurons and astrocytes. At the core of the infarct zone, near complete death is observed through necrosis within minutes, however the region surrounding this core (known as the ischemic penumbra) is partially perfused and does not immediately experience irreversible damage. Within the ischemic penumbra, maintenance of astrocyte viability is critical since neurons are dependent on close interactions with astrocytes for survival. In fact, astrocyte survival can promote synaptic remodeling and neurite outgrowth to compensate for neurons lost through the ischemic insult. Several studies have reported that astrocytes may be particularly susceptible to ischemia. Astrocytes exposed to ischemic challenges experience loss of astrocyte marker proteins and evidence of astrocyte cell death prior to histologic evidence of neuronal death has been observed. Furthermore, prolonged astrocyte survival in areas of cerebral infarction can contribute to protecting neurons from cell death by means of astrocyte-mediated glutamate clearance, astrocyte release of metabolic intermediates such as lactate, alanaine, citrate and α-ketogluterate – and finally through scavenging of ROS, particularly through glutathione. Relaxin is a peptide hormone with many diverse actions in multiple tissues. Whilst classically thought of as a hormone of female reproduction, the fact that it is present in the male, and has actions outside of the reproductive system, indicates the dogma no longer stands. In addition to the many physiological actions of relaxin that have been reported, relaxin has been shown to protect tissues from ischemia, particularly in models of myocardial infarction and the brain –. Wilson et al. (2006) reported that intracerebral injection of relaxin directly into the cortex prior to middle cerebral artery occlusion (MCAO) reduced ischemic cerebral lesion size indicating a direct action of relaxin on cells of the brain. This group also reported that inhibition of nitric oxide synthase (NOS) blocked this response, implicating nitric oxide (NO) in this observation. These neuroprotective mechanisms may be due to local vasodilation induced by relaxin. However it is also possible that relaxin is acting directly to protect neural tissues and other neuroprotective actions may be possible; experiments from this laboratory on cultured brain slices indicated that in slices exposed to hypoxic conditions, relaxin prevented cell death. Given that these experiments were devoid of a functional circulation, the results show relaxin may have a direct, neuroprotective effect. In the current study, the direct effect of relaxin on astrocytes in an *in vitro* model of hypoxia was examined. It was hypothesized that relaxin peptides would prevent the production of ROS and thus protect astrocytes from cell death that normally arise from hypoxic conditions. Two types of relaxin, relaxin-2 and relaxin-3 as well as a relaxin chimera peptide, R3/I5, were used in these experiments. Relaxin-2 was used since other reports used this form of relaxin in MCAO or brain slice studies. In addition, relaxin-3, the most recently discovered relaxin-family peptide with nearly exclusive expression in the brain was employed to determine whether or not this peptide provided neuroprotection to astrocytes during hypoxia. Last, since relaxin-3 has been reported to act through both relaxin family peptide receptor (RXFP) 1 and RXFP3, a highly selective RXFP3 agonist, termed R3/I5, was used to elucidate whether or not RXFP3 was involved in relaxin-mediated neuroprotection. # Materials and Methods ## Primary Astrocyte Cell Culture Primary rat cortical astrocytes were obtained from Invitrogen (Carlsbad, CA, USA) and stored in liquid nitrogen until use. On the day of establishment, vials containing 1×10⁶ cells were thawed and suspended in astrocyte growth media warmed to 37°C; the astrocyte growth media consisted of DMEM 1x (containing 4500 mg/L glucose, 110 mg/L pyruvate, 584 mg/mL L-Glutamine), 15% FBS, and PenStrep (500 units penicillin, 500 µg streptomycin). Cells were plated on 25 cm² tissue culture-treated flasks at a seeding density of 2×10⁴ cells/cm². Flasks were then placed in a water- jacketed incubator at 37°C, 5% CO₂ and 90% humidity. Fresh, pre- warmed media was replaced every 4 days and cells were grown to 100% confluence. ## Hypoxia Induction Protocols Astrocytes were cultured in 96-well plates (for viability assays) or 24-well plates (for imaging assays) and grown to 100% confluence. On the day of the experimental protocol, the astrocyte growth media was aspirated and replaced with serum-free, glucose-free media (subsequently referred to as oxygen-glucose deprivation; OGD). Astrocytes were either exposed to untreated, glucose-free media or glucose-free media containing relaxin. The astrocytes were then placed in a polycarbonate hypoxia induction chamber (Stemcell Technologies); a gas mixture containing 5% CO₂ and 95% N₂ was used for 10 minutes to purge the ambient air from the chamber and to simulate an ischemic environment. To ensure that the chamber remained humidified throughout the hypoxia protocol, 20 mL of sterile water in a Petri dish was placed in the hypoxia chamber. The hypoxia chambers were sealed, then placed in a 37°C incubator for 12, 24 or 48 hours depending on the experimental protocol. All experiments were repeated nine times (n = 9). ## Assessment of Astrocyte Viability Astrocytes were incubated with serum-free glucose-free media alone or serum-free glucose-free media containing relaxin. The concentrations of relaxin used in these studies were: relaxin-2 (10, 50 ng/mL), relaxin-3 (10, 50 ng/mL) and a highly selective RXFP3 agonist, R3/I5 (10, 50 ng/mL). Astrocyte cell viability was assessed at 12, 24 and 48 hours by the reduction of 3-(4,5-dimethylthiazole-2-yl)-2,5-diphenyltetrazo-lium bromide (MTT). Briefly, stock concentrations of MTT (5 mg/mL) were diluted (1∶10) in each well of a 96-well plate. The MTT was incubated with the cells for 30 minutes at 37°C and the reduced formazen product was lysed from the cells using a 100% dimethylsulfoxide solution (DMSO). Absorbance was subsequently measured at 570 nm using a fluorescent microplate reader. ## Reactive Oxygen Species Detection Assay In order to determine whether or not relaxin peptides prevented the excessive production of ROS as a result of ischemic challenge, ROS were measured using the Image-iT LIVE Green Reactive Oxygen Species Detection Kit. This kit uses a fluorogenic marker, 5-(and-6-)-carboxy-2′,7′-dichlorodihydrofluorescein diacetate (carboxy-H₂DCFDA) to detect ROS in live cells. Astrocytes were plated on 24-well tissue culture treated plates, containing cover slips, and exposed to either serum-free glucose-free media alone or serum-free glucose- free media with relaxin-2 (10 ng/mL) or relaxin-3 (10 ng/mL) or R3/I5 (10 ng/mL). The astrocytes were then exposed to the hypoxia protocol described above for 12 or 24 hours. At the conclusion of the hypoxia protocol, astrocytes were washed once with warm PBS and then covered with 25 µM carboxy-H2DCFDA working solution; the astrocytes were incubated in this solution for 30 minutes at 37°C in the dark. After 30 minutes, the astrocytes were washed three times in warm PBS and mounted on glass slides. The live astrocytes were imaged immediately using an Olympus IX8I live-cell microscope using filters optimized for fluorescein at 20x. Images were processed using ImagePro Plus Version 5. ## Assessment of Astrocyte Mitochondrial Membrane Potential The production of ROS under oxidative stress is closely linked with disruptions in mitochondrial membrane potential (Δψ_m). Therefore, an investigation was undertaken in order to determine whether or not relaxin- treated astrocytes exhibited differences in Δψ_m in response to hypoxic challenge. The cationic dye, 5,5′,6,6′-tetrachloro-1,1′,3,3′-tetraethylbenzimidazolcarbocyanine iodide (JC-1) was used to assess Δψ_m. JC-1 is a cationic dye that exhibits a potential-dependent accumulation in mitochondria detected by a shift in fluorescence from green (∼525 nm) to red (∼590 nm). Consequently a decrease in the red fluorescence and an increase in green fluorescence indicate the presence of depolarized mitochondrial membranes. Astrocytes were plated on 24-well tissue culture treated plates, containing cover slips, and exposed to either serum-free glucose-free media alone or serum- free glucose-free media with relaxin-2 (10 ng/mL) or relaxin-3 (10 ng/mL) or R3/I5 (10 ng/mL). The astrocytes were then exposed to the hypoxia protocol described above for 12 hours. At the conclusion of the hypoxia protocol, the media was aspirated, astrocytes were washed once with PBS and incubated with PBS containing the JC-1 dye (2 µg/mL) and incubated for 30 minutes at 37°C. The dye was then removed and the astrocytes were washed three times. The astrocytes were imaged using an Olympus IX81 1ive cell microscope optimized for red and green fluorescence. Images were acquired at both wavelengths at a magnification of 20x. Images were processed using ImagePro Plus Version 5. ## Apoptosis Detection Assay The detection of astrocytes undergoing apoptosis that was induced by 24 hours of OGD was detected using a commercially available Annexin V conjugate assay that detects phosphatidylserine (PS) on the outer leaflet of the cell membrane. The assay is built upon the principle that in non-apoptotic cells, PS is located on the cytoplasmic surface of membrane; conversely, in apoptotic cells, PS externalizes to the outer leaflet of the membrane thereby allowing Annexin V to conjugate to the exposed PS. Twenty-four hours following the induction of OGD and treatment with relaxin-2, relaxin-3 R3/I5 or control, astrocytes were incubated with PBS-containing Annexin V (5 µg/mL) for 15 minutes. Astrocytes were then washed with PBS and immediately imaged using an Olympus IX8I live-cell microscope for FITC. Images were acquired at a magnification of 20x. Images were processed using ImagePro Plus Version 5. ## Mechanisms of relaxin-mediated neuroprotection In order to determine the involvement of NO and PI3K in the neuroprotective effects that were observed by relaxin application the following pharmacological blockers were applied for 15 minutes prior to the induction of OGD and throughout the 24 hour OGD (in addition to relaxin-2 or relaxin-3): one of *N*_ω-Nitro-L-arginine methyl ester hydrochloride (L-NAME, 1 mM) or 1H-, Oxadiazolo\[4,3-a\] quinoxalin-1-one (ODQ, 100 µM) to block NO and one of wortmannin (WORT, 10 µM) or LY294002 (LY, 10 µM). The concentrations of inhibitors were chosen based on dose responses experiments conducted in the lab (data not shown). Once a concentration of an inhibitor was chosen, the inhibitor was incubated with astrocytes for 24 hours and the viability was assessed with an MTT assay to ensure that it was not cytotoxic. ## Materials and Reagents Rat primary cortical astrocytes, DMEM, FBS and PenStrep, Image-IT LIVE green reactive oxygen species detection kit, Annexin-V and JC-1 cationic dye were purchased from Invitrogen (Carlsbad, CA, USA) and the MTT reagent was obtained from Sigma Aldrich (Oakville, ON, Canada). Recombinant human relaxin-2, human relaxin-3 and R3/I5 chimera relaxin peptide was purchased from Dr. John D Wade and Dr. Ross AD Bathgate, Howard Florey Institute, Melbourne, Australia. All inhibitors (L-NAME, ODQ, WORT and LY) were obtained from Sigma Aldrich, Oakville, Ontario, Canada. ## Data Analysis and Statistics Experiments were performed nine times (n = 9). Data involving the imaging of cells is representative of typical results obtained from experiments performed. Data (where applicable) are presented as mean ± SEM. Statistical significance was accepted if *P*\<0.05. Statistical analysis on the raw data was employed using Graph Pad Prism software (San Diego, CA, USA). Statistical significance was assessed using an ANOVA and a post hoc Tukey test for multiple comparisons. # Results ## Assessment of Astrocyte Viability Following Hypoxic Challenge The viability of astrocytes in response to OGD challenge was assessed by the uptake of MTT. Exposure of astrocytes to 12 hours of OGD did not cause a significant difference in astrocyte viability between astrocytes that were treated with relaxin-2 (10, 50 ng/mL), relaxin-3 (10, 50 ng/mL) and a highly selective RXFP3 agonist, R3/I5 (10, 50 ng/mL) compared with OGD-alone. Astrocytes were exposed subsequently to OGD for 24 hours and treated with media (serum-free, glucose-free)-alone or media with relaxin-2 (10, 50 ng/mL), relaxin-3 (10, 50 ng/mL) and a highly selective RXFP3 agonist, R3/I5 (10, 50 ng/mL). Astrocytes that were treated with relaxin peptides throughout the 24-hour OGD protocol all demonstrated a significant increase in cell viability compared to untreated hypoxic astrocytes. Relaxin-2 (10, 50 ng/mL), relaxin-3 (10, 50 ng/mL) and a highly selective RXFP3 agonist, R3/I5 (10, 50 ng/mL) did not protect the viability of astrocytes in response to 48 hours of exposure to OGD *in vitro*. ## Production of Reactive Oxygen Species by Astrocytes in Response to Oxygen Glucose Deprivation The production of ROS by astrocytes in response to OGD was assessed with the fluorogenic marker carboxy-H₂DCFDA. Astrocytes that were exposed to 12 hours of OGD demonstrated a greater amount of fluorescent signal compared to astrocytes that were treated with relaxin-2 (10 ng/mL), relaxin-3 (10 ng/mL) and a highly selective RXFP3 agonist, R3/I5 (10 ng/mL) indicating that relaxin may prevent the production of ROS in response to OGD over a 12 hour period. As a control, astrocytes that were not exposed to OGD were also loaded with H₂DCFDA and did not show any signal (data not shown). The production of ROS in astrocytes in response to OGD was also assessed over a 24 hour exposure to OGD. Astrocytes that were treated with media alone exhibited a marked increase in ROS production (indicated by the flourometric carboxy-H₂DCFDA signal) when compared to those astrocytes that were treated with relaxin-2 (10 ng/mL), relaxin-3 (10 ng/mL) and a highly selective RXFP3 agonist, R3/I5 (10 ng/mL) indicating that relaxin may prevent the production of ROS in response to OGD over 24 hours in astrocytes. ## Assessment of Mitochondrial Membrane Potential in Astrocytes Exposed to 12 Hours of Oxygen-glucose Deprivation Differences in Δψ_m were assessed using the cationic dye JC-1 (2 µg/mL). Staining with JC-1 allows cells to be excited at two different wavelengths in order to assess Δψ_m in cells undergoing oxidative stress or other challenges that would affect Δψ_m such as apoptosis. Red fluorescence (J-aggregate form, ∼585 nm excitation) indicates polarized mitochondrial membranes whereas green fluorescence (monomer form, ∼514 nm excitation) indicates depolarized mitochondrial membranes. Astrocytes that were treated with relaxin-2 (10 ng/mL), relaxin-3 (10 ng/mL) and a highly selective RXFP3 agonist, R3/I5 (10 ng/mL) exhibited a marked difference in JC-1 staining compared with astrocytes that were solely exposed to OGD. Those astrocytes that were treated with relaxin peptides show marked staining for the J-aggregate form of JC-1 and a limited amount of staining for the monomer form. This compared to astrocytes that were exposed to OGD alone that exhibited much more staining for the monomer form thereby indicating that the mitochondrial membranes of astrocytes that were exposed to OGD are more depolarized compared to that of astrocytes that were treated with relaxin-2 (10 ng/mL), relaxin-3 (10 ng/mL) and R3/I5 (10 ng/mL). ## Apoptosis Detection Assay The identification of astrocytes that were undergoing apoptosis as a result of 24 hour-OGD exposure was examined by loading astrocytes with Annexin V (an early marker of apoptosis through labeling translocated PS) and PI (a marker of cell death). Astrocytes that were not treated with relaxin exhibited characteristic Annexin V labeling as indicated by green fluorescence showing an increase in apoptosis in these cells. Astrocytes that were incubated with either relaxin-2, relaxin-3 or R3/I5 showed a marked decrease in apoptosis that was indicated by a much lower green fluorescence signal. ## Mechanisms of relaxin-mediated neuroprotection In order to determine whether or not NO and PI3K was involved in the protection of astrocyte viability from the effects of 24 hour-OGD, astrocytes were relaxin-2 or relaxin-3 and either L-NAME, OGD (to block NO) or WORT or LY (to block PI3K) during the 24 hour OGD protocol. The results show that the inhibition of NO or PI3K significantly reduced the relaxin-2 and relaxin-3 protective effects during the 24 hour-OGD protocol. # Discussion The objectives of this study were to determine whether or not relaxin-2 and relaxin-3 protected cultured astrocytes from cell death induced by hypoxia *in vitro*, and to investivate whether or not relaxin had an effect on some of the mediators of cellular damage that results from hypoxia, namely the production of ROS and disruption of mitochondrial function. The data presented indicate that relaxin-2, relaxin-3 and R3/I5 protect cultured astrocytes from cell death over a 24 hour period. These data also show that cultured astrocytes that were treated with relaxin-2, relaxin-3 and R3/I5 show a marked decrease in ROS production over the course of 12 and 24 hour exposures to hypoxia. Further, hypoxic conditions that result in mitochondrial depolarization and apoptosis appear to be reduced when astrocytes are incubated with relaxin-2, relaxin-3 and R3/I5 during hypoxic challenge. Finally, we show that the inhibition of NO and PI3K prevents the relaxin-2 and relaxin-3-mediated protection of astrocytes during OGD. In a number of previous studies, relaxin has been demonstrated to confer protective effects to tissues that are undergoing ischemic stress. Work from Bani and colleagues, have provided ample evidence that relaxin protects the heart from ischemia in a number of studies investigating myocardial infarction in guinea pigs and rats. Also in these studies, these researchers have reported on relaxin’s protective effect as a consequence of activation of NOS and the downstream production of NO. Our study also shows that NOS is involved in the protection of astrocytes from the effects of OGD. Relaxin has also been reported to reduce the lesion development as a result of cerebral ischemia. These results show that pre-treatment with relaxin-2 administered intracerebrally reduced the lesion size compared to untreated animals following MCAO. Further work from this laboratory has indicated that brain slices exposed to hypoxia *in vitro* exhibit higher viability when co-incubated with relaxin-2 compared to untreated cells. The current study provides some possible insight into some of the mechanisms by which relaxin may protect neural tissues from cerebral ischemia. The findings show that relaxin-2, relaxin-3 and R3/I5 protect astrocytes from cell death in an *in vitro* model of hypoxia. Neurons require a close interaction with astrocytes for survival as astrocytes provide a multitude of functions that ensure proper neuronal function (e.g. structural scaffolding, extracellular ion regulation, control of pH, neurotransmitter clearance etc.) and the death of astrocytes would directly impact the fate of neurons. Therefore the protection of astrocytes from cell death may not only directly protect astrocytes but positively influence (protect) the cell death in neurons as well. In addition to the demonstration that relaxin-2 affords protection to astrocytes that were exposed to a hypoxic challenge, data presented are the first to report that relaxin-3 also has this action. Given that relaxin-3 interacts with both relaxin receptors (RXFP1 and RXFP3) it was important to confirm an involvement of RXFP3 by using a highly selective RXFP3 agonist, R3/I5. In doing so these data show that activation of both RXFP1 and RXFP3 provide protection from hypoxia. Astrocytes, like other cells and tissues experiencing hypoxia, are vulnerable to the production of ROS that result in cellular damage, impaired cellular function and potentially cell death. Reactive oxygen species is a phrase used to describe molecules and free radicals (chemical species with one unpaired electron) that are derived from oxygen. Superoxide anion is one of the most common free radical precursors of most ROS produced in cells and is a byproduct of a number of enzymatic complexes of the electron transport chain of the mitochondria,. In the brain, it has been reported that complex I is the primary source of O₂⁻•. Under basal conditions free radicals are usually scavenged or converted to non-reactive species usually resulting in the formation of water; SODs are the most common means of scavenging O₂⁻•. Cells that are experiencing oxidative stress produce an imbalance of ROS that overwhelms the endogenous ROS scavenging systems and further oxidative damage resulting in the disruption of cellular proteins, lipids, polysaccharides and DNA. These data presented in the current paper show that astrocytes treated with relaxin-2, relaxin-3 and R3/I5 show a marked reduction in ROS production over a 12 and 24 hour period. Furthermore, over a 12 hour period, relaxin-2 and relaxin-3 prevented the collapse in the Δψ_m when compared to those astrocytes that were exposed to OGD alone. We show that relaxin may be working to protect astrocytes from hypoxic challenges by preventing the production of ROS and affecting (positively) the mitochondrial integrity. We have also shown that two signalling pathways (NO and PI3K) are implicated in the protection of astrocytes during exposure to OGD. A possible mechanism by which relaxin may be protecting the cell from hypoxia include inducing the expression of hypoxia inducible factor-1 alpha (HIF-1α) through a nuclear factor kappa B (NF-κB) mechanism. Relaxin has been reported to increase the expression of NF-κB, which could lead to the production of HIF-1α which helps cells resist damage as a result of hypoxia. Furthermore, induction of NF-κB by relaxin may also affect the regulation and expression of mitochondrial SOD; for example, increased levels of SOD within the mitochondrial matrix could eliminate O₂⁻• that is formed during hypoxia,. Finally, it has been reported that an increase in the availability of adenosine diphosphate (ADP) within the cytosol directly affects ROS production resulting in a decrease in the Δψ_m. Through the numerous phosphorylation events that arise from relaxin-RXFP signaling, ADP levels within the cytosol may increase – and relaxin-induced availability of ADP may directly affect the mitochondrial function of the cells and therefore act in part to prevent apoptosis. Further study is also warranted into the possibility that relaxin peptides mediate protection during reperfusion injury. Other investigations into the ability for relaxin to protect tissues during reperfusion injury have looked at the heart and observed a protective effect. Therefore, a more clear picture would emerge as to the neuroprotective effect of relaxin if ischemia/reperfusion was investigated. Further study on an ischemia/reperfusion *in vivo* model is warranted to investigate this possibility. Taken together these results presented here provide evidence that relaxin-2, relaxin-3 and an RXFP3-specific agonist, R3/I5 protect astrocytes from cell death induced by OGD. These data also indicate that astrocytes that have been exposed to relaxin peptides exhibited a reduction in the hypoxia-induced production of ROS over a 12 and 24 hour period and more viable mitochondria as shown by a maintenance of Δψ_m. These data provide insight into the mechanisms by which relaxin may act on astrocytes to provide neuroprotection and presents a possible therapeutic potential to treat this cerebral ischemia and stroke. We wish to thank Esther Semple for assistance with performing the *in vitro* experiments. We also thank Dr. Lindsay C Bergeron for assistance and support throughout the performance of these studies. Finally we thank Gregg Johns for help with the preparation of the figures. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JMW AJSS. Performed the experiments: JMW. Analyzed the data: JMW AJSS. Contributed reagents/materials/analysis tools: JMW. Wrote the paper: JMW AJSS.

# Introduction *Bartonella henselae* (*Bhe*) is a worldwide distributed, zoonotic pathogen. In its feline reservoir host, it causes an asymptomatic, intraerythrocytic bacteraemia. Accidental transmission of *Bhe* from cats to humans can manifest in a variety of clinical symptoms, ranging from the so-called cat-scratch disease in immuno-competent patients to bacillary angiomatosis or peliosis in immuno-compromised persons, respectively. *Bhe* expresses a VirB/VirD4 type IV secretion system (T4SS) that mediates translocation of the *Bartonella* effector proteins (Beps) BepA to BepG into the host cell cytosol. The Bep effectors share a common basal architecture, consisting of an N-terminal effector domain and a bi-partite translocation signal composed of at least one BID domain (*<u>B</u>artonella* <u>i</u>ntracellular <u>d</u>elivery) and a positively charged C-terminus. Effectors BepA, BepB and BepC all contain a single FIC domain in proximity to their respective N-terminus, while BepD, BepE and BepF display tyrosine/proline- rich repeats in their N-terminal portion. Interestingly, effectors BepE, BepF and BepG all contain multiple BID domains while BepG consists exclusively of four BID domains flanked by short linker regions. Bep translocation into the host cell promotes a variety of distinct phenotypes that include: (i) inhibition of apoptosis, (ii) activation of the pro- inflammatory response, (iii) capillary-like sprout formation of endothelial cell aggregates and (iv) host cell invasion by a cellular structure named the invasome. *Bhe* internalization via the invasome route is a well controlled multi-step process, consisting of *Bhe* adherence to the cell surface, *Bhe* aggregation, *Bhe* engulfment by plasma-membrane-derived membrane protrusions and eventually *Bhe* internalization. Invasome formation can be triggered in a redundant manner, either by BepG alone or by the combined action of effectors BepC and BepF. Various pathogenic bacteria translocate effector proteins into their respective host cells that interfere with Rho GTPase signaling events,. Rho GTPases interact in their GTP-bound form with multiple downstream proteins, thereby transmitting incoming signals to basal levels. In contrast, GDP-bound GTPases are not able to bind to and activate their interaction partners. GTPase signaling is in general controlled by GAP, GEF and GDI proteins. While GAPs (GTPase-activating proteins) stimulate the turn-over of the GTP to GDP, GEF (guanine nucleotide exchange factor) increase the exchange rate of GDP with GTP. GDI (guanine nucleotide dissociation inhibitor) bind to the C-terminal lipid groups of GTPases, thereby preventing membrane binding and stabilizing them in the inactive state in the cytosol. Pathogenic bacteria translocate various GAPs or GEFs into the host cell in order to subvert Rho GTPase signaling: In example, *Salmonella enterica* effector SptP or *Yersinia enterocolitica* effector YopE act as GAPs of Rho GTPases, while the *S. enterica* protein, SopE as well as *Escherichia coli* effector MAP posses GEF functionality on Rho-family GTPases. Recently, a new family of bacterial effector proteins sharing a common Trp-xxx- Glu motif (WxxxE motif) was shown to interfere with Rho GTPase signaling. These WxxxE-family proteins, later shown to be Rho GEFs, include SifA and SifB from *Salmonella*, MAP and EspM/M2 from *E.coli* as well as IpgB2 and IpgB1 from *Shigella*. The WxxxE motif was demonstrated to be essential for GEF function although it is not directly involved in establishing contact with the target Rho GTPases. Alternatively to exhibit GAP or GEF functions, bacterial effector proteins were shown to directly interfere with Rho GTPase signaling by promoting chemical modifications of GTPases (ADP-rybosylation, glucosylation, AMPylation), , or indirectly by interacting with Rho GTPase regulators such as Dock180, Crk or ELMO. In this study, we investigate the function of the *Bartonella* effector protein BepF. We show that the isolated BID-F1 or BID-F2 domains - together with BepC - are sufficient to trigger invasome establishment. Further, we demonstrate that constitutive-active Cdc42 or Rac1 can substitute for BepF in the BepC/BepF- dependent invasome formation pathway, suggesting a regulatory role of BepF on the small Rho GTPases during the process of invasome formation. # Materials and Methods ## Bacterial Strains, Growth Conditions, Conjugations *Bhe* strains were cultured as previously described on solid agar plates (Columbia base agar supplemented with 5% sheep blood and appropriate antibiotics). *E. coli* strains were grown on solid agar plates (Luria Bertani broth) supplemented with appropriate antibiotics. Triparental matings between *E. coli* and *Bhe* strains were performed as described. lists all bacteria strains used in this study. ## Plasmid Construction DNA manipulations were carried out following standard protocols. Vectors pCD353, pMS007, pPG100 and derivatives, pRS79, pMT563 and pTR1769 as well as peGFP- Cdc42, peGFP-Cdc42, pRK5mycL61-Cdc42, pRK5mycL61-Rac1 have been described before (see for plasmid origins). eGFP-Bep fusion plasmids pMT560, pMT562, pMT567, pMT591, pMT592, pMT593, pMT597. pMT612, pMT613 and pMT614 were obtained by PCR amplification of the respective insert with the corresponding primers, cutting the purified PCR products with XmaI and XbaI and their ligation into pWAY21 (eGFP, Molecular Motion, Montana Labs) cut accordingly. pMT001, pMT004, pMT005, pMT030, pMT031 and pMT52 were generated by PCR amplification of the respective insert with the corresponding primers, cutting the purified PCR products with NdeI and their ligation into NdeI-digested pPG100. All constructs were sequence confirmed. and list all plasmids and primers constructed or used in this study. ## Cell Lines and Cell Culture HeLa Kyoto β cells and NIH3T3 cells were cultured in DMEM (Gifco, invitrogen) supplemented with 10% fetal calf serum (FCS). ## Transfection and Infection Assays Transfection and infection of HeLa cells was performed as described. In brief, 4500 cells were seeded into a well of a 96-well plate, and after over-night incubation transfected with DNA using Lipofectamine2000 (invitogen), following manufacturer's instructions. Cells were washed once with phosphate-buffered saline (PBS) and supplemented with fresh DMEM/10%FCS medium 6–8 h post transfection. Cells were further incubated for 24 h at 35°C, 5% CO₂ before continuing with the respective assays. HeLa infections were carried out as described. In brief, HeLa cells were infected with *Bhe* at a multiplicity of infection (MOI) = 500 per strain in 100 µl medium M199/10%FCS supplemented with 500 µM IPTG (Promega). Following 48 h incubation cells were fixed with para-formaldehyde (PFA). Transfection of NIH 3T3 cells was performed following manufacturer's instructions. Briefly, cells were seeded out at a density of 30000/well of a 24 well plate and incubated over night. The next day, 200 µl optimem was mixed with 2 µg of plasmid DNA and 6 µl of lipofectamine2000 and incubated for 30 min. Afterwards, 100 µl of the transfection mix was added to the cells together with 400 µl of fresh DMEM/10%FCS and incubated for 4 h. Then, medium was exchanged with 500 µl fresh DMEM/10%FCS and cells were incubated for 48 h at 35°C, 5% CO₂. ## Immunoprecipitation (IP) and Immunoblot analysis IP was performed as described elsewhere. Expression of novel N-terminal FLAG- tagged and NLS-Cre-Bep fusion proteins was verified by analysis of total *Bhe* lysates obtained from *Bhe* grown on CBA plates containing 500 µM IPTG. Proteins were run on a SDS-PAGE gel for separation and transferred onto nitrocellulose membranes (Hybond, Amersham Biosciences) and probed against the FLAG epitope using mouse monoclonal anti-FLAG antibody M2 (Sigma, 1∶1000). Novel eGFP-Bep fusion proteins were assessed for their stability by analysis of total cell lysates obtained from HeLa cells transfected with plasmids encoding the respective constructs and incubated for 24 h. After protein separation by SDS- PAGE and transfer onto nitrocellulose, membranes were examined for the presence of eGFP using rabbit monoclonal anti-GFP antibody (Molecular Probes, 1∶5000). In all experiments, secondary horseradish peroxidase-conjugated antibody (Amersham, 1∶10000) was visualized by enhanced chemiluminescence (PerkinElmer). ## Immunofluorescent (IF) labeling Indirect IF labeling was performed as described. Standard 96-well plate assays were stained with TRITC-phalloidin (Sigma, 100 µg/ml stock solution, final concentration 1∶400), and DAPI (Roche, 0.1 mg/ml) using a Tecan Eoware freedom pipeting robot. Glasslides for confocal microscopy were stained with Cy5-phalloidine (Sigma, 100 µg/ml stock solution, final concentration 1∶100), and DAPI. ## Semi-automatic image analysis, invasome quantification and microscopy Image analysis and invasome quantification was performed as described. In brief, cells were automatically imaged in up to three different wavelengths depending on the applied cell staining. The number of cells per image was determined by MetaExpress in-build analysis modules (CountNuclei) and invasomes on the very same images were defined and counted by eye. In every experiment, at least 500 cells were analyzed per condition. ## Epi-fluorescence and Confocal Laser Scanning Microscopy Epi-fluorescence and confocal Laser Scanning was performed exactly as described earlier. In brief, 96-well plates were imaged with an ImagXpress Micro (IXM) automated microscope (Molecular devices). For confocal laser microscopy, specimens were visulaized using an IQ iXON spinning disc system (Andor) in combination with an IX2-UCB microscope (Olympus). Images were exported and finalized using Metamorph, ImageJ and Adobe Photoshop. ## Scanning electron microscopy (SEM) SEM analysis was performed exactly as described before. In brief, cells were seeded onto glass slides and treated as described above (infection and transfection assays). Following incubation, probes were washed and fixed with 250 µl of 2.5% glutaraldehyde for 30 min at RT. Afterwards, cells were washed twice and the samples were subsequent dehydrated with an ethanol step gradient (30%, 50%, 70%, 90%, 100%; 15 min each) at 4°C. Thereafter, samples were critical point-dried and sputter-coated with a 3 nm thick Platin layer. Images were taken on a Hitachi S-4800 field emission scanning electron microscope, using an acceleration voltage of 2 kV. # Results ## BepF tyrosine phosphorylation is not required for invasome formation In previous work, we have shown that BepC together with BepF can trigger invasome formation. However, the molecular details of the function of either of the two proteins remained to be determined. *In silico* analysis of the sequence of BepF revealed that BepF contains a tyrosine-rich repeat motif close to its N-terminus, which is linked to three BID domains. The first and the second BID domain, BID-F1 and BID-F2, are fused together while the third BID domain, BID-F3, is linked via a short spacer sequence to BID-F2. Web-based sequence analysis of BepF using Scansite (<http://scansite.mit.edu/>) and NetPhos (<http://www.cbs.dtu.dk/services/NetPhos/>) yielded in high probability predictions of multiple tyrosine phosphorylations of the tyrosine-rich motifs \[E/T\]PLYAT. Furthermore, previous work demonstrated that short, synthesized peptide fragments containing the \[E/T\]PLYAT motif of BepF are *in vitro* phosphorylated and interact with Crk, RasGAP and Grb2. To check whether the tyrosine-rich repeats of BepF are indeed phosphorylated upon host cell entry and contribute to invasome formation, we generated two BepF mutants, one having all seven tyrosine replaced with phenylalanine (further referred to as BepF-YF) and one mutant consisting only of the three BIDF domains and the positively charged C-tail (further referred to as BID-F1-3). HeLa cells were thereafter co-infected with the effector-deficient *Bhe* strain *ΔbepA-G* expressing FLAG-tagged BepC and *Bhe ΔbepA-G* strains expressing BepF or BepF mutant constructs BepF-YF, BID-F1-3 with an MOI = 500 per strain for 48 h. The stability of FLAG-tagged mutant constructs of BepF was verified by Western blotting. Following immunoprecipitation using anti-FLAG agarose beads, tyrosine phosphorylation was analyzed by Western blotting. The results clearly showed that wild-type BepF is tyrosine phosphorylated in the host cell, while neither of the two mutant constructs displayed any detectable tyrosine phosphorylation signal, indicating that the N-terminal tyrosine-containing repeat motifs are indeed phosphorylated in the host cell. Next, we investigated if the tyrosine-rich repeat is required for BepF to contribute to invasome-mediated *Bhe* internalization. Therefore, we infected HeLa cells with *Bhe* wild-type, *Bhe ΔbepA-G* or combinations of *Bhe ΔbepA-G*/p*BepC* and *Bhe ΔbepA-G*/p*bepF*, *ΔbepA-G*/p*bepF-YF* or *ΔbepA-G*/p*BID-F1-3*. Quantification of invasome formation of fixed, stained and microscopically imaged cells demonstrated that BID-F1-3 was sufficient to trigger invasome formation together with BepC to the same level as wild-type BepF or BepF-YF. To further strengthen that point, we generated eGFP-tagged fusion proteins containing either only the N-terminal part of BepF (NterF) or the BID-F1-3 region. HeLa cells were transfected with plasmids encoding for eGFP, eGFP-BepF, eGFP-NterF and eGFP-BID-F1-3 and, after 24 h incubation, infected with *Bhe ΔbepA-G*/p*bepC* at an MOI = 500 for another 48 h. Stable expression of the eGFP-fusion was verified by Western blotting. The obtained data were in line with our previous finding: HeLa cells ectopically expressing either eGFP-BepF or eGFP-BID-F1-3 and infected with *Bhe ΔbepA-G*/p*bepC* showed invasome formation at a frequency of about 10%, while HeLa cells expressing GFP- NterF and infected with the same strain did not show any invasomes. Taken together, we show that the BID domains BID-F1-3 are sufficient to trigger invasome formation together with BepC. Further, we show that, although tyrosine- phosphorylated in the host cell, the N-terminal tyrosine-containing repeat motif of BepF does not contribute to BepC/BepF-dependent invasome formation. ## The BID domains BID-F1 and BID-F2 of BepF together with BepC are sufficient to promote invasome formation In a next step, we tested whether individual BID domains of BepF could contribute to invasome formation in combination with BepC. Therefore, we first cloned FLAG-tagged BepF mutant constructs that consist of BID-F2-3 or BID-F3 and transformed the plasmids into *Bhe ΔbepA-G*. Fusion construct expression and stability was tested by Western blotting. *Bhe* strains *ΔbepA-G*/p*BID-F2-3 and ΔbepA-G*/p*BID-F3* were tested in co-infection experiments with *Bhe ΔbepA-G*/p*BepC* according to the standard protocol. Quantification of invasome formation on fixed, stained and imaged cells indicated that the removal of the first BID domain (BID-F1) reduced invasome formation by about 70% compared to BID-F1-3, while the removal of both BID-F1 and BID-F2 together lead to a complete abolishment of invasome formation. To investigate the capacity of BID-F1 and BID-F2 to contribute to invasome formation in more details, we generated plasmids encoding for eGFP-tagged constructs eGFP-BID-F1, eGFP-BID-F2, eGFP-BID-F3 and eGFP-BID-F1-2. Fusion protein stability was verified by Western blotting. Following transfection of HeLa cells with the indicated constructs, cells were infected with *Bhe ΔbepA-G*/p*bepC* for 48 h. The results showed that both BID-F1 and BID-F2 together with BepC are able to promote invasome formation while it was absent from cells expressing eGFP-BID-F3 and infected with *Bhe ΔbepA-G*/p*bepC*. Interestingly, eGFP-BID-F2 was significantly more potent than eGFP-BID-F1 to promote invasome establishment and eGFP-BID-F1-F2 was promoting invasome formation to the same extent than BID-F1-3, each in combination with BepC. In summary, our results show that BID-F1 and BID-F2, but not BID-F3 domains are individually sufficient to mediate invasome formation in combination with BepC. ## Disruption of the WxxxE motif in BID-F1 interferes with BID-F1 function In 2008, Alto *et al* proposed a family of bacterial effector proteins containing a WxxxE motif to be mimics of host cell GTPases. This statement was later revised and it was shown for multiple instances that translocated bacterial proteins containing the WxxxE motif act as GEFs for Rho family GTPases. Sequence analysis of the BIDF domains showed that BID-F1 contains a WxxxE motif as well, while BID-F2 and BID-F3 harbor a closely related motif at the same position, WxxxN. However, amino acid sequence alignments of BID-F1, BID-F2 and BID-F3 with known WxxxE-family GEFs showed low sequence conservation besides the motif itself. Nevertheless, we decided to further focus on BID-F1, since it contains an intact WxxxE motif, and mutated tryptophan-362 into alanine in various BepF-related constructs to disrupt the WxxxE motif (AxxxE). Thereafter, we co-infected HeLa cells according to the standard protocol with *Bhe ΔbepA-G*/p*bepC* and *ΔbepA-G*/p*bepF* W362A, *ΔbepA-G*/p*BID-F1-3* W362A or *ΔbepA-G*/p*BID-F2-3* and checked for invasome formation. Mutant protein stability was tested by Western blotting. The obtained results demonstrate that, upon changing the WxxxE motif to AxxxE, the capacity of BepF as well as BID-F1-3 to contribute to invasome formation decreased to the level obtained for co- infections with *ΔbepA-G*/p*bepC* and *ΔbepA-G*/p*BID-F2-3*, thus basically eradicating the contribution of BID-F1 to the process of invasome formation. Next, we introduced the mutation into our eGFP-fusion constructs and quantified invasome formation on HeLa cells ectopically expressing eGFP-fusion proteins and infected with *Bhe ΔbepA-G*/p*BepC* following standard protocols. GFP-fusion protein stability was tested by Western blotting. These results were in line with our previous findings: the introduced W362A mutation in eGFP-BID-F1-2 decreased invasome formation down to the level found for eGFP-BID-F2 alone in combination with BepC. Furthermore, mutating the WxxxE motif in eGFP-BID-F1 significantly decreased invasome formation compared to wild-type eGFP-BID-F1. Comparing the amino acid sequences of BIDF domains with characterized WxxxE- family GEF proteins, we identified a conserved serine residue located six amino acids downstream of the glutamic acid of the WxxxE motif. This serine was present in all WxxxE-family proteins except for SifA, while being present in BID-F2 but absent in BID-F3. To test whether this serine residue may play a role in BID-F1 and BDF2 functionality during invasome formation, we constructed mutant constructs encoding for GFP-BID-F1 S372A, GFP-BID-F1 W362A/S372A and GFP- BID-F2 S508A. The constructs were tested in standard transfection-infection assays and invasome formation was quantified after 48 h of infection with *Bhe ΔbepA-G*/p*BepC*. The results showed that mutation of serines 372 and 508 did not affect invasome formation, implying that the indicated residue is not critical to maintain BID-F1 and BID-F2 domain function and structure. Concluding, our results indicate that the WxxxE motif found in BID-F1 is essential for the function of the BID-F1 domain and that the conserved serine residue downstream of the WxxxE motif is not critical to maintain BID-F1 and BID-F2 functionality. ## BepF can be substituted by expression of constitutive active Cdc42 or Rac1 during BepC/BepF-dependent invasome formation Several bacterial effectors containing the WxxxE motif were shown to act as GEFs for the small GTPases RhoA, Rac1 and Cdc42. Previous work on *Bhe*-triggered invasome formation has further demonstrated that Cdc42 and Rac1, but not RhoA, are required for invasome formation. To test whether BepF interferes with Rac1- or CdC42-mediated signaling, we transfected HeLa cells with plasmids encoding for myc-tagged constitutive active Cdc42 (L61-Cdc42) or Rac1 (L61-Rac1). After 24 h of incubation, cells were infected with *Bhe ΔbepA-G*/p*BepC* at an MOI = 500 and incubated for another 48 h. Following fixation and staining, invasome formation was quantified. Our results showed that *Bhe ΔbepA-G*/p*BepC* could indeed promote invasome formation on HeLa cells expressing either L61-Rac1 or L61-Cdc42. Further, we also observed a more than 50% increase in invasome frequency on HeLa cells expressing either constitutive active GTPase and infected with *Bhe ΔbepA-G*/p*BepC* and *ΔbepA-G*/p*bepF* compared to empty vector transfected cells. Interestingly, invasome formation on HeLa cells expressing L61-Cdc42 or L61-Rac1 and infected with *Bhe* wild-type decreased compared to the empty vector control, thereby confirming previous published results. The fact that substitution of BepF with L61-Cdc42 or L61-Rac1 leads to significantly less invasome formation as the combined action of BepC/BepF indicates that the activity of Cdc42 and Rac1 is essential for certain steps of invasome establishment but may act rather inhibitory on other aspects of the entire process. ## BepF triggers the formation of filopodia-like extensions and membrane protrusions on HeLa and NIH 3T3 cells Although BepF has been shown to infrequently trigger the formation of small actin foci, the function of BepF has mainly been investigated in the context of invasome formation. Based on the finding that the constitutive active GTPases L61-Cdc42 and L61-Rac1 can substitute for BepF function we tested for a BepF- specific phenotype on the F-actin cytoskeleton level that is related to the action of L61-Cdc42 or L61-Rac1. To this end, we infected HeLa cells with various *Bhe* strains at a high MOI (1000) for 48 h to trigger maximal phenotypic penetrance. As previously reported host cell viability was unaffected under these infection conditions. After cell fixation, we analyzed the cells by scanning electron microscopy (SEM). Uninfected as well as *Bhe ΔbepA-G*, *ΔbepA-G*/p*BepC* or *ΔbepA-G*/p*bepG* infected HeLa cells showed low levels of filapodia-like structures or membrane protrusions.. In contrast, HeLa cells infected with *Bhe* wild-type or *Bhe ΔbepA-G*/p*BepF* displayed drastically changed cell morphology and showed massive formation of filopodia-like structures as well as membrane protrusions that frequently contacted neighboring cells. The previously reported small actin foci promoted by BepF on HUVECs were completely absent on HeLa cells. In a next step, we tested our eGFP-BIDF fusion constructs in the same TEM-based assay. We found that BID-F1 as well as BID-F2, but not BID-F3 or BID-F1 AxxxE induced the formation of filopodial extensions and membrane protrusions. To strengthen our findings, we repeated the experiments with the eGFP-fusion constructs in NIH 3T3 cells, a cellline well known for a highly responsive actin cytoskeleton that is often used to study stress fibers, lamelipodia and filopodia formation upon system perturbation. To this end, we transfected NIH 3T3 cells with indicated plasmids encoding for eGFP-fusion constructs as well as proper controls. After fixation and staining, cells we analyzed the actin cytoskeleton phenotype of GFP-positive cells. The results showed that eGFP- tagged full-length BepF, BID-F1 and BID-F2 induced a change in actin cytoskeleton morphology that is phenotypically comparable to the expression of L61-Rac1 or L61-Cdc42 in these cells while neither eGFP control, eGFP-tagged BID-F3 nor eGFP-tagged BID-F1 AxxxE fusion proteins affected the F-actin organization of NIH 3T3 cells. In summary, our data suggests that the BepF domains BID-F1 and BID-F2 are involved in the regulation, in particular the activation of Rac1 and Cdc42. # Discussion The *Bartonella henselae* effector protein BepF has previously been implicated in triggering invasome formation together with BepC in a cofilin1-dependent manner. Here, we show that the individual BID domains BID-F1 and BID-F2, but not BID-F3 are sufficient to promote invasome formation together with BepC. Sequence analysis of the three BepF BID domains implies that BID-F2 and BID-F3 are more homologue to each other than to BID-F1; however, the general level of sequence homology is low. Thus, from sequence comparison it is not evident why BID-F1 and BID-F2 can contribute to invasome formation while BID-F3 cannot. Besides the three BID domains, BepF contains a tyrosine-rich repeat motif that is phosphorylated in the host cell upon effector translocation. Interestingly, the replacement of all tyrosine residues as well as the complete removal of that protein portion did not interfere with BepC/BepF-mediated invasome formation, nor with BepF triggered formation of filopodial cell extensions. It is tempting to assume that BepF may interact with multiple SH2-domain containing proteins that can bind to the phosphor-tyrosine scaffold of BepF. However, we were so far unable to identify a cellular phenotype that is linked to the N-terminal portion of this translocated effector protein. The interference with Rho GTPases to subvert host signaling cascades is a frequent function associated with translocated bacterial effector proteins. Several distinct mechanisms have been reported yet, including bacterial GEF and GAP proteins (SopE, SptP), covalent modification of the target GTPases by AMPylation (VopS, IbpA), glucosylation (TcdA/B) or ADP-rybosylation (C3) as well as the deamidation (CNF1) and partial proteolytic degradation (YopT) of Rho- family G proteins. In this report, we show that BepF can be replaced by constitutive-active CDC42 or Rac1 in the process of invasome formation. The findings that neither constitutive active GTPase was as potent as BepF to contribute to invasome formation and that over-expression of both constitutive active GTPases interfered with BepC/BepF- or *Bhe* wild-type promoted invasome assembly suggests that the tempo-spatial control of Cdc42 and Rac1 activity is important for the establishment of invasome structures. This hypothesis is in accordance with the published data on invasome formation, which showed that the assembly of the massive actin structure is followed by the eventual retraction of the actin arrangement that leads to the release of the bacteria into the host cell. The constitutive activation of Cdc42 and Rac1 that both control processes associated with F-actin filament elongation and cell protrusion formation may be central for the assembly of the invasome structure but rather disadvantageous for the retraction and the disassembly thereof. A BepF-dependent activation of Cdc42 and Rac1 is further indicated the BepF-triggered formation of filopodia- like cell extensions and membrane protrusions on HeLa and NIH 3T3 cells.. Recent work on translocated bacterial WxxxE GEF proteins suggested that the motif itself may have mainly structural roles, in particular by maintaining the conformation of the putative catalytic loop through hydrophobic contacts with surrounding residues. As BID-F1 contains an intact WxxxE motif and its disruption interferes with BID-F1 function, it is tempting to speculate that BepF is a further WxxxE-family bacterial GEF protein. However, sequence alignments of the distinct WxxxE-GEF proteins together with the comparison of available GEF-GTPase co-structures indicate that the WxxxE-GEF proteins share more than only the common WxxxE-motif. They display several key residues that directly contact the GTPase interface and are important for GEF function. In contrast, alignments of BID-F1 and BID-F2 showed that both domains lack all of these described critical residues besides the central WxxxE/WxxxN motif. Thus, BepF is likely to not represent a WxxxE-family GEF protein. However, the detailed mechanism of how BepF may interfere with Cdc42 and Rac1 signaling remains to be investigated. We previously showed that BepC and BepF together mediate invasome formation on various cell types. Further, we showed that this process depends on Cdc42, Rac1 and their subsequent downstream signaling partners. With respect to the results presented on this work, the function of BepF in the process of invasome formation is presumably the activation of Cdc42 and Rac1. BepC consists of a FIC domain and a single C-terminal BID domain. Recently, FIC domains have been demonstrated to reversibly modify Rho GTPases by AMPylation, thereby inhibiting their interaction with downstream partners. Thus, it is tempting to speculate that BepC may negatively regulate Cdc42 or Rac1 by AMPylation, thereby contributing to the proposed dynamic activation/inhibition of Cdc42 and Rac1 during the process of invasome formation. However, further work on BepC and the function of its FIC domain is required to answer that question. In summary, we provide evidence that the *Bartonella* effector protein BepF activates Cdc42 and Rac1 and that this activation functionality is contained in the two BID domains BID-F1 and BID-F2. # Supporting Information We would like to thank C. Mistl for technical assistance and Phillipp Engel for help with. Furthermore we are grateful to M. Düggelin, E. Bieler and M. Dürrenberger from the ZMB for their great SEM service. [^1]: Conceived and designed the experiments: MT PG CD. Performed the experiments: MT PG. Analyzed the data: MT PG. Contributed reagents/materials/analysis tools: MT PG. Wrote the paper: MT CD. [^2]: The authors have declared that no competing interests exist.

# Introduction Viral hemorrhagic fever is a febrile syndrome associated with vascular damage caused by RNA viruses of the families: *Filoviridae* \[*Ebolavirus* and *Marburgvirus*\], *Arenaviridae* \[Lassa fever virus (LASV)\], *Bunyaviridae* \[Rift Valley fever virus (RVFV) and Crimean Congo hemorrhagic fever virus (CCHFV)\], and *Flaviviridae* \[Yellow fever virus (YFV), and dengue virus (DENV)\]. Most are zoonotic, vector-borne, and may cause sporadic, unanticipated, and devastating outbreaks in endemic areas. The Center for Disease Control and Prevention (CDC) has designated *Filoviridae* and *Arenaviridae* as Category A due to their ease of transmission, high mortality, risk to national security, and potential for causing public panic and social disruption. Many agents of VHF are designated as emerging or reemerging pathogens and threaten not only their traditional areas of endemicity in developing countries but new territory in other countries as well. There are few preventative vaccines, and clinical management is largely supportive due to the paucity of effective chemotherapeutic agents. The case fatality ratio for outbreaks of the filoviruses in Africa has ranged from approximately 36–90% and 83–90% for Marburg and Ebola, respectively. Furthermore, the infection of skilled and traditional healers tending the sick complicates care and control measures while abetting nosocomial transmission and the spread of disease. The 2014 West African Ebola outbreak, which is the largest recorded thus far, exemplifies these difficult infection control issues. The filoviruses, CCHFV, and LASV require the highest level of laboratory containment (containment level 4 or biosafety level 4), however, few such facilities are present in resource- poor endemic countries. The worldwide threat of VHF agents to the public health, as well as to veterinary and agricultural communities, is increasingly recognized, as is the possibility of the accidental or malicious release of some of these viruses as agents of bioterror. Human smallpox was eradicated over 30 years ago. Since then, vaccination to poxviruses has largely stopped, leading to a worldwide population of susceptible individuals. Variola virus is legally retained at only two World Health Organization (WHO) Collaborating Center repositories. Reports of covert, undeclared stocks and weaponized virus have fueled fears that variola may be reintroduced as a bioterror agent, an issue of continuing national and international concern. Recognition of index cases with the exotic and geographically restricted VHF viruses, or eradicated smallpox, depends upon the clinical suspicion and diagnostic acumen of first-line physicians. VHF infections due to endemic natural outbreaks, infection in returning travelers, or suspected acts of bioterrorism continue to challenge public health and clinical laboratories. Effective infection control and the implementation of public health containment plans require rapid and effective diagnostic tests.. Rapid detection and specific pathogen identification are key to the control of outbreaks due to VHF and variola viruses. In addition, rapid screening of a large number of samples can be anticipated as was the case in the 2001 anthrax outbreak. There are several molecular detection assays that have been developed for rapid detection of VHF viruses, the majority of which are based on real-time PCR. These assays have improved the sensitivity and the turn-around time for detection of VHF, while significantly reducing the biohazard potential of cultivating these organisms in the laboratory. However, real-time PCR assays often have limited multiplexing capabilities and require several assays to be run for detection of multiple etiologic agents. An additional concern is mutation in these viruses resulting in the alteration of protein sequences and targets used for molecular detection. Sequencing has established that in the recent West African Ebola outbreak, isolates from Sierra Leone had genomes that varied from PCR probes used for four separate assays used for EBOV and pan- filoviral diagnostics. Mutations could potentially impact diagnosis due to mismatch between target and primer/probe, not only in current assays, but also in those that are in the process of being released. Multiplex detection of several viruses using real-time PCR strategy is limited by the choice of fluorescent dyes and their spectral overlap. Palacios et al. have described a multiplex assay for detecting VHF viruses that utilizes PCR primers containing unique mass tags that are then detected by a mass spectrometer. A comprehensive molecular detection panel for high-throughput identification of these viruses, using a single cycling protocol and fluorescent technology, would be a useful addition to current techniques available for molecular identification of these pathogens. PCR/LDR is a versatile technique that has been used in the detection of pathogens in clinical as well as environmental samples. We have previously reported a PCR/LDR universal array- based technique for the simultaneous detection and identification of all four serotypes of DENV and West Nile virus from serum and plasma samples as well as from mosquito pools. Here we describe use of this technology for the multiplex detection and identification of seven VHF viruses (ebolavirus \[Zaire, Sudan and Reston ebolaviruses\], MARV, LASV, CCHFV, RVFV, YFV, DENV) and two orthopoxviruses \[variola virus (VARV) and vaccinia virus (VACV)\] in a single assay. # Methods ## Ethics Statement This study was performed in accordance with a protocol approved by the Institutional Review Board of the Weill Medical College of Cornell University. ## Viral Isolates Vaccine strains of RVFV (MP12) and YFV (17D) were kindly provided as a gift by Dr. Robert Tesh, University of Texas Medical Branch, Galveston, TX. Inactivated viral culture supernatants of ebolavirus \[Zaire virus (EBOV), Sudan virus (SUDV), and Reston virus (RESTV)\] (n = 4), MARV (n = 3), CCHFV (n = 3), RVFV (n = 1) and LASV (n = 1) were obtained from the United States Army Medical Research Institute of Infectious Diseases, Ft. Detrick, Maryland. Genomic DNA from VACV virus (n = 6) was obtained from NIH Biodefense and Emerging Infections Research Resources Repository, NIAID, NIH. The two target amplicons of VARV \[amplicon 1 (RAP94): 421bp; nt: 77,877–78,298 and amplicon 2 (RPO147): 485bp; nt: 82,372–82,856\] representing all VARV sequence variants for these genetic regions were obtained from the CDC Poxvirus and Rabies Branch, Centers for Disease Control and Prevention, Atlanta, GA with the permission of the WHO and in accordance with all applicable regulations. Details of viral cultures and nucleic acid used in the study are provided in Tables and. As previously described, culture supernatants from standard isolates of the four serotypes of DENV were employed. The following viruses or nucleic acids were used as controls: DNA from cowpox and monkeypox viruses obtained from ATCC; St. Louis encephalitis virus (strain MSI-7), Murray Valley Encephalitis virus (strain OR2), Powassan virus (strain M11665), Tick-borne encephalitis virus (strain K23), West Nile virus (NAT-positive plasma samples from blood donors) and Japanese encephalitis virus (strain SA-14-14-2). ## Oligonucleotide Design Virus-specific primers were designed as described previously. Briefly, after alignment of the sequences, areas with relative conservation among different virus strains were chosen for each virus group and, where possible, for several viral groups (*Filoviridae*, *Flaviviridae* and *Poxviridae*) so as to achieve maximum strain coverage with the least number of primers. Primer sets were designed to simultaneously amplify two different regions in each virus or viral group: *NP* and *L* genes of ebolaviruses (EBOV, SUDV and RESTV) and MARV, *S* segment of CCHFV, *M* and *S* segments of RVFV, *L* segment of LASV, *NS5* regions of DEN and YF, and the *RNA pol* (RAP94 and RPO147) genes of VACV and VARV; a total of 57 primers. The amplicons were \~500 bp in length (range 399–685 bp). The primer sequences contained no more than three degenerate positions and had a melting temperature of around 72–75°C. LDR primers were chosen in two to three different conserved regions within each of the two PCR amplicons for the different virus groups. The primers were designed with the intent of achieving the highest possible strain coverage in all the different viruses as well as to differentially identify them individually. A total of 250 LDR primers were designed, with melting temperatures between 75 and 80°C; degenerate bases (no more than three in each primer) were introduced, where required, to account for sequence variations. A complete list of all LDR primers is provided in. The PCR and LDR primers were obtained from Integrated DNA Technology, Coralville, IA. The PCR and LDR primers for each of the target regions for all the virus groups were evaluated in separate assays individually. This was performed to evaluate satisfactory signal detection from each of the regions selected. Primers that failed to produce either PCR amplicon or generated less than two LDR products were replaced, and new primers were designed. The LDR primers were designed such that they were able to differentiate between the three species of ebolavirus tested (EBOV, SUDV and RESTV). The poxvirus primers were designed such that they were able to discriminate between VARV and VACV and would not cross-react or produce false positive signals with other *Orthopoxvirus* species (cowpox and monkeypox viruses). The flavivirus primers were designed to identify and distinguish DENV and YFV without cross-reactivity with a panel of flaviruses. ## Nucleic Acid Preparation and PCR/LDR Assay Nucleic acid extraction and one-step RT-PCR amplification (OneStep RT-PCR kit; Qiagen, Valencia, CA), ligase detection reaction (LDR), and universal array were performed as described previously, with the exception that one-step RT-PCR was performed in a 25 μl final volume using 5 μl of template RNA or DNA. explains the PCR/LDR assay design and detection protocol. To provide a certain degree of redundancy two regions of each virus were amplified, and primers for the subsequent ligation reaction were designed targeting 2 or 3 areas within each PCR amplicon. Consequently, up to six ligation products (only 5 ligation products in the case of VARV and VACV) can be generated for each virus to be detected. However, not all ligation products are required to be present to unambiguously detect and identify each of the viruses. The presence of any 2 ligation products for a given virus, either two ligation products from a single PCR amplicon or at least one ligation product from each of the two amplicons, was considered sufficient for identification. Ligation products bear zip-code sequences and were detected by hybridization to a universal array bearing complementary zip-codes. A signal was considered positive for ligation if the intensity of the corresponding zip-code spot was at least 10-fold higher than the overall average background intensity of the array as determined by the ScanArray Express v 4.0 (Perkin Elmer, MA). No-template controls provided no PCR amplicons and consequently no positive signals for any ligation products. The analyses were repeated in at least 2 different experiments with the exception of *Sudan ebolavirus* for which there is a single experiment. We included the following geographic and genotypic variants: four different strains of three species of ebolavirus (SUDV, RESTV, EBOV); standard strains of DENV serotypes 1–4; VARV (8 genotypically variant isolates of *V*. *major* virus); and CCHFV (3 isolates; Hy-13 from China, and UG3010 and ArD8194 from the Democratic Republic of the Congo and Senegal, respectively). ## Preparation of RNA Standards for Determination of LOD Synthetic RNA fragments of the viruses were prepared for LASV, RVFV, and YFV, MARV and CCHFV. The target regions of LASV, RVFV and YFV viruses to be detected were amplified by RT-PCR, purified using QIAquick PCR purification kit (Qiagen, Inc., Valencia, CA), and cloned into the expression vector pGEM T Easy (Promega, Madison, WI) containing the T7 promoter region. The plasmids were purified, and the presence of complete inserts was confirmed by sequencing the inserts using vector-specific primers. Target regions for MAR and CCHF were synthetically constructed and inserted into the pIDT Blue vector (Integrated DNA Technologies Coralville, IA). The complete inserts of the target regions thus generated were linearized and *in vitro* transcribed using the mMessage *in vitro* transcription kit (Ambion, Austin, TX). Following DNAse treatment, the synthetic RNA was purified using an RNeasy column (Qiagen, Inc., Valencia, CA). The quantity of RNA generated for each transcript was determined using the Ribogreen® RNA Quantitation Kit (Molecular Probes, Eugene, OR) following the manufacturer’s instructions. ## Limit of Detection The limit of detection (LOD) of the different viruses was determined in the following manner. First, 10-fold serial dilutions of the synthetically prepared, previously quantified, RNA transcripts of LASV, RVFV, YFV, MARV and CCHFV (1x10¹⁰ to 1x10⁰ copies/ml) were used in the PCR reaction to determine the limit of detection. Second, the quantified VACV virus DNA stocks obtained from BEI Resources were serially diluted in nuclease-free water and 5 μl of each of the DNA dilutions was used in the PCR reaction to determine limit of detection. Third, serial dilutions (ten-fold) of viral culture stocks were prepared for EBOV (*Zaire ebolavirus* ‘95) from inactivated standard stock cultures. Dilutions were prepared in Dulbecco’s minimum essential medium (DMEM) supplemented with 10% fetal bovine serum (Invitrogen, Carlsbad, California). RNA was extracted from 140 μl of each dilution, and RT-PCR-LDR/universal array was used to determine the analytical sensitivity of the assay. # Results ## Evaluation of the Assay Using Viral Strains The PCR/LDR/universal array was validated on 53 different samples including viral culture filtrates, vaccine strains of viruses, nucleic acid from VACV viruses, and PCR amplicons \<500bp in size of VARV. depicts the ability of the primers to detect and differentiate the viruses from each other. The PCR-LDR assay was able to detect all 32 strains, representing 9 different viruses that were included in the assay. No signals were observed for blank no-template controls and no false positives were seen with 8 other related viruses (data not shown). The assay was designed to provide redundancy and account for sequence drift in the viruses to be examined. Two regions were targeted for PCR amplification and between 2–3 regions within each PCR amplicon were targeted for the ligation reaction such that there were a maximum of 5 or 6 possible signals for each DNA or RNA virus respectively, but only 2 positive signals were required to confirm the identification of any given virus. Using this criterion, the assay detected all 11 viruses/strains with no false identification. Seven of the viruses generated positive signals for all LDR targets (5/5 or 6/6) and three were positive at 5/6 targets. We observed only 3 positive signals in *Sudan ebolavirus* because only one PCR region was successfully amplified. None of these signal drop-outs affected identification of the viral species and were expected given the inherent variability of RNA viruses. Due to restrictions on the possession of VARV and its nucleic acid, DNA extraction and PCR amplification of the RAP94 and RPO147 genes was performed at the CDC. All 5 possible LDR signals from the 8 strains representative of sequence variants in the CDC repository were positive. The experiments were performed in accordance with the regulations and permission of the WHO. All 5 possible signals were detected for the VACV strains tested. No false-positive reactions were observed. The primers for VACV did not produce false signals with VARV DNA and vice versa, nor did either set of primers interact with cowpox virus or monkeypox virus DNA (data not shown). ## Limit of Detection The limits of detection using either synthetically transcribed RNA or dilutions of whole virus are shown in. The LODs for CCHFV and MARV viruses were 190 and 53 RNA copies/ml respectively. The LODs for RVFV and LASV were 7.6 and 100 RNA copies/ml, respectively. The LODs for the whole virus dilutions tested for EBOV and DENV were 1000 FFU/ml and 1PFU/ml respectively. The calculated limits of detection per PCR reaction for EBOV and DENV viruses were \<10 copies/PCR reaction. The LOD for the DNA virus, VACV was found to be 10⁴ genome equivalents/reaction. # Discussion We describe a multiplex detection assay that can simultaneously detect and differentiate 7 VHF viruses, VARV and VACV, in a single assay and is amenable to adaptation to a high-throughput format. The PCR/LDR/universal array was validated on 53 different samples including viral culture filtrates, vaccine strain of viruses, serum specimens from patients with DENV infection, nucleic acid from VACV viruses and PCR amplicons \<500bp in size of VARV. Since the viruses that cause the VHF syndrome include variant viral species with diverse geographical distributions, we attempted to include viruses from several geographic areas and viral species when applicable. The assay has some built-in redundancy, by including 6 possible signals for each virus (only 2 of which are required for identification), to account for the genetic diversity often encountered in RNA viruses. We encountered signal drop-out in some of the viral species but the assay was able to detect all viruses included in the panel. Additionally, the number of signals detected, or the fluorescent intensity of the signals generated, was robust for all samples tested. Early stages of infection for most of these viruses are relatively non-specific. For some viruses, such as RFV and Lassa, the classical hemorrhagic manifestations may be absent, making accurate and timely diagnosis challenging. A multiplexed detection assay such as the PCR/LDR has a potential benefit in being able to screen for several pathogens in an endemic area or in symptomatic returning travelers. In fact, there are several multiplex assays that have been developed in the recent years to address this issue. The analytical sensitivity of our assay was comparable to recently described real-time PCR based assays. The PCR/LDR has the significant advantage of testing for 9 agents simultaneously and is amenable to automation. Additional targets can be added to the repertoire without compromising the performance of the assay. The multiplex capability of PCR-LDR allowed an assay design that can distinguish viral strains that differ in geographic distribution and virulence. There are currently five recognized species of ebolavirus, two of which, *Sudan ebolavirus* and *Zaire* ebolavirus, have been consistently recognized as causing large outbreaks since 1976 with the most recent in 2014. The remaining three species, *Tai Forest ebolavirus* (formerly *Cote d’Ivoire*), *Reston* and *Bundibugyo ebolaviruses* have occurred less frequently. *Tai Forest ebolavirus* has been implicated in a single human infection acquired during the autopsy of a wild chimpanzee. *Reston ebolavirus* causes VHF and death in primates and illness in pigs. Antibody seroconversion in human contacts of infected animals has been documented but without any associated disease. Although *Reston ebolavirus* is thought to be non-pathogenic in humans, the WHO has cautioned that the effect of the infection in the immunosuppressed, pregnant women, and children is unknown. Its detection therefore remains potentially clinically important. In contrast, the relatively new *Bundibugyo ebolavirus* is known to be pathogenic. The outbreak in Uganda (2007–8) resulted in 100 cases with a fatality rate of \~40%. The PCR/LDR assay was successful in the identification of the two major outbreak associated isolates of *Zaire ebolavirus* as well as the isolates of *Sudan* and *Reston ebolaviruses* tested. Signal dropout was noted in *Sudan ebolavirus* for all three LDR products for the nucleoprotein gene. Since 2 out of 6 signals were required for positive identification, this did not affect the sensitivity of the assay. The assay was able to detect MARV from the initial 1967 Ugandan derived German isolate as well as the two Kenyan isolates derived from infections associated with the Kitum Cave in Kenya’s Mount Elgon National Park in 1980 and 1987. Of the CCHFV S segment-defined groups (A, B and C) the assay was able to detect group A viruses from both the African and Asian clades as well as group B virus. Group C was not tested but contains a single virus isolated from a tick in Greece. The assay detects all VARV genetic sequence variants in the CDC repository for the targeted regions. Comparative genomics of 45 geographically diverse VARV isolates obtained over 30 years during the smallpox era indicate low sequence diversity. Hence, we anticipate that our assay would be capable of detecting the virus in the event of a bioterror attack. Since vaccination would be instituted in the event of the intentional release of variola, and vaccine strains can be spread secondarily, the assay is also designed to detect and identify VACV vaccine strains. The current version of the assay does not include specific primers for the detection of zoonotic *Orthopoxviruses*: monkeypox in particular, or the more clinically benign cowpox. It was designed, however, to permit their exclusion. The modular nature of the assay allows for the expansion of the number of organisms and targets that can be identified. Future versions of the assay could be designed to include these viruses, and other hemorrhagic fever viruses such as those of South America, without compromising the sensitivity and specificity of the assay. When our assay was initially developed, the sequence information for *Bundibugyo*, *Tai Forest*, and the 2014 *Zaire ebolavirus* species were unavailable in the public databases, and therefore their detection was not incorporated. Analysis of the *Zaire ebolavirus* implicated in the 2014 West African outbreak indicate that it would be detected and correctly identified by the assay ( and Figs). In addition, analysis of the assay PCR primers for both the NP and L gene amplicons suggest that PCR products would be generated for both *Bundibugyo* and *Tai Forest ebolaviruses*. These analyses however would need to be verified experimentally. Given the modular nature of the assay it would be relatively simple to add additional LDR primers for detection of both *Bundibugyo* and *Tai Forest ebolaviruses* to our existing assay. For example, our prototype PCR/LDR DENV assay was modified to permit the detection of an unusual strain of DENV. The assay described in this study has not yet been used to detect and identify viruses directly from clinical materials nor was it directly compared to other available assays. However, we have used similar PCR/LDR assays to test for DENV in 350 acute phase serum specimens and for WNV in 142 plasma samples from blood donors. The sensitivity and specificity of the WNV assay was 100% and that of the DENV assay was 98.7% and 98.4% respectively. The analytical sensitivity of the assay in its present form (7–200 copies of RNA/ml of sample tested) is equivalent to \<100 viral particles per assay. This is comparable to other techniques used for the detection of RNA viruses. In theory, therefore, it should detect an infection in clinical materials due to viruses in the current assay panel. Nucleic acid extraction for this assay utilized manual extraction methods and thus required 5–6 hours. However, there are several platforms that offer high- throughput automated extraction and are in use in clinical laboratories. The PCR, LDR and hybridization steps have been adapted to automation in a 96-well format on a liquid handling robot in our laboratory. We have also used a 96-well bead-based array platform (BeadXpress, Illumina Inc. San Diego, CA) for downstream detection of ligation products (unpublished data,). The combination of automated extraction and hybridization steps will decrease both hands-on and turn-around time for this assay which would be critical in a high volume situation such as occurred during the US anthrax attacks of 2001. Although the current version of the assay was not designed for point-of-care diagnoses, adaptations currently under investigation are designed to permit this. PCR/LDR has a major advantage over traditional PCR based assays due to the high fidelity of the thermostable ligase enzyme; it is highly specific for a given nucleotide sequence at the site of ligation. The addition of a universal array, spotted with zip-code addresses, has a unique potential to recognize pathogen- specific zip-code complements appended to the LDR primers. This provides an additional technical advantage over existing PCR-based assays, in that it obviates the use of actual genetic sequence on the array for pathogen detection. This is especially useful for multiplexing, as a large number of genomic targets can be detected simultaneously using this array. Such multiplexing capability presents the possibility that an assay panel could be designed to target several pathogens with similar clinical features in a given endemic area, i.e., a “customized” chip. Additionally, the same array can be used for the simultaneous detection of different organisms, as positive hybridization is dependent on the chemistry of the synthesized zip-code oligonucleotides spotted on the array and their complements appended to the primers. In previous studies, we have reported the identification of 20 different bacterial species in blood cultures with a high degree of sensitivity and specificity. The modular nature of this assay allows for the expansion of the number of organisms and targets that can be identified. This is especially important for identification of potentially variable genomes of viral infectious agents that may necessitate additions to the detection primers used. Additions may also be made for the detection of new hemorrhagic fever viruses as they emerge. The assay therefore adds a potential new tool to our armamentarium for the rapid high throughput detection of emerging viral pathogens and potential bioterror agents. # Supporting Information The authors would like to acknowledge; Robert B. Tesh, Center for Biodefense and Emerging Infectious Diseases, University of Texas Medical Branch for providing us with: vaccine strains of Rift Valley fever virus (MP-12) and Yellow fever virus (17D); Inger Damon, Poxvirus and Rabies Branch, Centers for Disease Control and Prevention, Atlanta, GA for valuable discussions; Gene Spier, formerly of Applied Biosystems, for assistance with sequence alignments; and Philip Feinberg for his help in preparing the manuscript. Opinions, interpretations, conclusions, and recommendations are those of the author(s) and are not necessarily endorsed by the U.S. Army. In addition, the content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. [^1]: FB is the holder of multiple patents for methods/primers designs that have been used in the detection and identification of mutations in genetic diseases and cancer as well as infectious agents. A complete list of FB’s patents may be found at: (<http://patft.uspto.gov/netacgi/nph- Parser?Sect1=PTO2&Sect2=HITOFF&p=1&u=%2Fnetahtml%2FPTO%2Fsearch-bool.html&r=0&f= S&l=50&TERM1=Francis&FIELD1=INNM&co1=AND&TERM2=Barany&FIELD2=INNM&d=PTXT>). In the past, FB has received funds from Applied Biosystems (now ThermoFisher) to further develop the aforementioned patents. FB is currently the recipient of a sponsored research grant funded by Roche. The authors would like to confirm that this does not alter our adherence to PLOS ONE policies on sharing data and materials, nor do the current affiliations of MS Rundell or MR Pingle. To the authors' knowledge, neither Roche nor Thermo Fisher has any interest in developing a product from the current study. [^2]: Conceived and designed the experiments: SD MSR AHM MRP KS ARG JP SKS VAO DHL EDS FB LMG. Performed the experiments: SD AHM KS ARG SKS VAO. Analyzed the data: SD MSR AHM MRP KS ARG JP SMK VAO EDS FB LMG. Contributed reagents/materials/analysis tools: MRP ARG JP SMK VAO FB. Wrote the paper: SD MRP DHL EDS LMG. Obtained permits and permissions required for reagents used: SD LMG. [^3]: Current address: Department of Infectious Disease Research, North Shore University Health System, Evanston, Illinois, United States of America. [^4]: Current address: Beckman Coulter, Chaska, Minnesota, United States of America. [^5]: Current address: Center for non-coding RNA in Technology and Health (RTH), University of Copenhagen, Copenhagen, Denmark. [^6]: Current address: Coferon, Inc., Stony Brook, New York, United States of America. [^7]: Current address: Department of Emergency Medicine, Harbor-UCLA Medical Center, Torrance, California, United States of America.

# Introduction Chronic obstructive pulmonary disease (COPD) is a major disease affecting millions of people worldwide. One of the major risk factors for COPD is cigarette smoke (CS), since about 90% of COPD patients are cigarette-smokers **.** The toxic compounds present in CS are partly responsible for the disruption of the alveolar and capillary network in the lung. Specifically, it has been reported that there are greater numbers of apoptotic alveolar epithelial and endothelial cells in lung tissues of COPD patient than in control patients, which in turn can lead to the development of emphysema. Moreover, an increase in cell death mechanisms in the lungs correlates with decreased expression of vascular endothelial growth factor (VEGF), which acts as a mitogenic factor promoting survival and differentiation of endothelial and alveolar epithelial cells. Chronic exposure to CS implies the absorption of a large amount of chemical compounds with important oxidative activity. In addition, inflammatory cells release reactive oxygen/nitrogen species, which also contribute to the sustained oxidative/nitrosative burden in the lungs. In the context of the inflammatory milieu as it occurs in COPD, nitric oxide reacts with reactive nitrogen species resulting in the formation of 3-nitrotyrosine (3NT), which is considered a marker of nitrosative stress and therefore inflammation. Nuclear factor (erythroid-derived 2)-like 2, also known as NFE2L2 or Nrf2, is a transcription factor that regulates the expression of a variety of genes involved in the antioxidant response. Interestingly, Nrf2-deficient mice have early-onset and more extensive CS-induced emphysema compared with wild-type littermates, thus suggesting that Nrf2 protects against the development of emphysema. Currently, treatment of COPD is based on the administration of bronchodilators and corticosteroids, however there are no effective therapies directed towards the regeneration of lost alveoli in emphysema. Cell-cell interactions between alveolar epithelial cells and other cell types are mediated by releasing growth factors including hepatic, keratinocyte, epithelial or vascular endothelial growth factor –. The repair mechanisms to restore the normal airway architecture are inefficient in patients with COPD. Therefore, the use of growth factors as promoters of cell proliferation mechanisms and differentiation have emerged as a promising strategies to stimulate tissue repair under injury conditions. Liver growth factor (LGF) is an albumin-bilirubin complex with mitogenic properties described for the first time in rat liver. The antifibrotic and antioxidant properties of LGF, as well its regenerative effects, have been described in several rodent models such as injured liver, Parkinson's disease, testis degeneration, hypertension and atherosclerosis. In addition, we have previously described the beneficial effects of LGF in a rat model of ClCd₂ induced-lung fibrosis and more recently in AKR/J mice with CS-induced emphysema **.** In this study, our purpose was to deepen the understanding of the regenerative properties of LGF in C57BL/6J mice with CS-induced emphysema and clarify, by the analysis of several markers of lung damage and proliferation, the mechanisms by which LGF promotes the improvement of the emphysematous profile. # Materials and Methods ## Animals C57BL/6J male mice (Charles River Laboratories) 8 weeks old were housed in the Inhalation Core Facility at the IIS-Fundación Jiménez Díaz (n = 40). Protocols were approved by the local Ethical Animal Research Committee at IIS-Fundación Jiménez Díaz. In all cases, the legislation regarding animal treatment, protection and handling was followed (RD 53/2013). ## CS exposure Mice were divided into air-exposed mice and CS-exposed mice. Animals were exposed to a mainstream CS of 4 research non-filtered cigarettes (3R4F, University of Kentucky, Lexington, KY; 11 mg TPM, 9.4 mg tar and 0.73 mg nicotine per cigarette), per day (5 minutes per cigarette with 10 minutes smoke free intervals, 5 days a week) during 6 months. An optimal smoke/air ratio of 1/6 was obtained. Mainstream CS was generated by an exposure system and was drawn into the chambers using a peristaltic pump (KD Scientific, Inc.) reaching concentrations of 200 mg TPM/m³ (Dust Track Model 8520, TSI Inc.). Non-smoking mice were exposed to room air. Body weights were assessed at the beginning of the experiment and at the third, fourth, fifth and sixth months of CS exposure. ## Liver growth factor (LGF) purification and administration LGF was purified from rat serum following the procedure previously reported. Briefly, the purification procedure consisted essentially of three chromatography steps, employing Sephadex G-75, DEAE-cellulose and hydroxyapatite. The absence of other growth factors and/or contaminants in the LGF preparations was also assessed according to standard criteria ,. All LGF preparations showed a single band in sodium dodecyl sulphate polyacrylamide gel electrophoresis (SDS-PAGE. LGF preparations were lyophilized and stored at 4°C. Before use, LGF was dissolved in saline solution (NaCl 0.9%; Braun). The dose of LGF has been optimized in previous pilot experiments. Six months after CS exposure, when emphysema is once established, animals were treated with 4 *i.p*. injections (twice a week for 2 weeks) of 150 µl of a solution containing 1.7 µg LGF per mouse. ## Fluorescence molecular imaging *in vivo* (FMI) Mice were anaesthetized during the imaging acquisition with a mixture of 2% isoflurane, 2 L/min oxygen using an Inhalation Anaesthesia System (Harvard Apparatus). Body hair was removed by shaving and subsequent application of depilatory cream. Each mouse was injected intravenously with 2 nmol of a protease activable fluorescent probe (MMPSense 680; PerkinElmer, Inc.) diluted in 150 µl of phosphate buffer. MMPSense 680 (excitation: 680 nm; emission: 700 nm) emits in the near-infrared when activated by MMP-2, -3, -9 and -13. 24 h after last tobacco exposition and fluorescence probe administration, *in vivo* images were acquired using an IVIS-Lumina Imaging System (Caliper Life Sciences, Inc.) as described previously. ## Lung morphometry Lungs were fixed intratracheally with 4% formalin (Sigma-Aldrich, Co.) at a pressure of 25 cmH₂O overnight. Then, formalin fixed tissues were paraffin embedded and cut into 5-µm-thick sections that were stained with hematoxylin and eosin (H&E) according to standard protocols. Enlargement of alveolar spaces was quantified by measurement of the mean chord length (L_m) and alveolar internal area (AIA) using an image analysis software (LeicaQwin) specifically designed by Leica. After imaging processing, the software displayed the total number of alveolar spaces in the field and the Lm and AIA values for each one. Histological images were selected following random criteria and captured with a videocamera (Leica Microsystems) coupled to an optical microscope (Olympus BX40). Analyses of 18 representative images per mouse were performed in duplicate by two blinded observers. ## Lung function Maximum inspiratory volume (V_max) at a pressure of 30 mbar was the parameter used to evaluate lung function on each mouse. This parameter was registered using a ventilator/respirator device designed especially for small animals. ## Histological Analyses The immunohistochemical staining was performed on paraffin-embedded lung 5-µm- thick sections. Proliferating cells were detected using an anti-PCNA antibody (Sigma-Aldrich). ## Western blot analysis 20 µg of total lung protein extract were denaturalized at 95°C during 5 minutes. Electrophoresis was performed in a 10% sodium dodecyl sulfate–polyacrylamide gel (SDS-PAGE) and transferred to a PVDF membrane. Prior to antibody incubation, membranes were blocked 1 h with 5% non-fat powdered milk in tris-buffered saline (TBS) in order to block non-specific interactions. After incubation at 4°C overnight with VEGFA, Nrf2, tubulin (Novus Biologicals), PCNA (Sigma-Aldrich, Co.) and 3NT (Santa Cruz Biotechnology, Inc.) primary antibodies, membranes were washed three times with TBS-0.5% Tween and incubated at room temperature during 1 h with secondary anti-rabbit antibody conjugated with horseradish peroxidase (Biolegend, Inc.). Finally, bands were visualized using a chemiluminescence detection kit (ECL Plus; GE Healthcare) according to the manufacturer’s instructions and quantified with Quantity-One Software (Bio-rad). Tubulin bands were used to normalize protein loading. ## Enzyme-Linked Immunosorbent Assay (ELISA) Using total protein extracts from lung tissue homogenates, ELISA was performed with a mouse specific VEGF ELISA kit (ref: CSB-E04756m; CUSABIO), as indicated in the manufacturer’s protocol. The same amount of total protein was added on each well. ## Gelatin Zymography Samples were electrophoresed onto a 10% polyacrylamide gel containing 1 mg/mL gelatin as substrate. On each electrophoretic lane, the same amount of total protein supernatants was electrophoresed. The gels were soaked with renaturating buffer (2.5% Triton X-100 in distillated water) at 37°C for 1 h to remove the SDS. After incubating the gels for 24 h at 37°C in the metalloproteinase buffer (50 mmol/L Tris-HCl, pH 7.4, 10 mmol/L CaCl2, 1% Triton-X100, 0.02% NaN3), they were stained for 30 min with 0.4% Coomassie blue to visualize bands of proteolytic activity and rapidly destained with 30% methanol and 10% acetic acid. The relative density of gelatinolytic bands was determined from scanned images of gels using image analysis software (ImageJ). ## RNA Isolation and Real-Time PCR Total RNA was isolated using trizol reagent from frozen lung samples and reverse-transcribed to cDNA. For real-time PCR, we used TaqMan universal master mix and Taqman gene expression assays for the following genes: mmp9 (Mm00442991_m1), mmp2 (Mm00439498_m1), timp1 (Mm00441818_m1) and timp2 (Mm00441825_m1) (Applied Biosystems). 2-ΔΔCT method was applied to get the gene expression data using the Rn18s gene as an internal control to normalize the expression of the target genes mentioned above. ## Statistical analysis Results were expressed as mean ± SEM. Statistical significance was taken as a p-value of less than 0.05 (*P*\<0.05). Mann-Whitney method was performed to test significant differences between experimental groups followed by Monte Carlo's exact methods within each set of comparisons using the Statistical Package for the Social Science software (SPSS, Inc.). # Results ## Body weight gain during CS exposure period Mice were exposed to CS or room air for 6 months. To evaluate the effect of CS inhalation, body weights of CS-exposed mice and air-exposed mice were assessed periodically. We observed that body weight gain in CS-exposed mice (28.9±0.31 gr) was significantly attenuated when compared to air-exposed mice (30.13±0.51 gr) at the third month of the study. Moreover, body weight gain was stabilized in the CS-exposed group until the end of the experiment, whereas in the air- exposed group the mice continued to gain weight. ## Experimental emphysema is reversed morphologically and functionally by LGF L_m, AIA and V_max parameters were used to assess the degree of lung damage and development of emphysema caused by CS inhalation. When comparing representative images of each group, it was evident that the lungs from mice exposed to CS showed alveolar space enlargement. However, the morphology of alveolar spaces in lungs from mice exposed to CS and then treated with LGF was similar to that observed in normal lungs. As measured by mean chord length (L_m) and alveolar internal area (AIA), air-exposed mice presented normal alveolar architecture (L_m = 32.46±0.46 µm; AIA = 794.5±23.39 µm²). In contrast, CS-exposed mice presented enlargement of alveolar spaces (L_m = 41.88±0.64 µm; AIA = 1587.19±59.54 µm²). Lung changes observed in CS-exposed mice were substantially reversed in mice treated with LGF for two weeks after CS exposure (L_m = 32.04±0.35 µm; AIA = 517.18±13.13 µm²). Additionally, analysis of intercepts lengths distribution in each group showed that in normal lungs, the density curve of the raw data had a slimmer and higher peak, whereas in lungs from mice with emphysema-like pathology (CS-exposed mice), the density curve had a wider but lower peak due to the disruption of alveolar septa. The shape of the density curve corresponding to the distribution of the intercept lengths in lungs from mice exposed to CS and then treated with LGF was more similar to that observed in normal lungs, suggesting that LGF was promoting tissue repair. Regarding lung function, CS-exposed mice showed a consequent loss of lung function estimated by a significant increase in V_max (1.39±0.031 ml) when compared to the values in normal lungs (V_max = 1.29±0.058 ml). However, the administration of LGF significantly improved the status of lung function in mice exposed previously to CS (V_max = 1.30±0.044 ml) showing values of V_max similar to those seen in the control group. It is important to note that air-exposed mice treated with LGF presented normal lung architecture compared with the untreated air-exposed group. Thus, we could verify that LGF treatment had no apparent negative consequences in the lungs of healthy animals. ## Lower MMP activity correlates with the reversion of the emphysema MMP activation promotes the destruction of extracellular matrix leading to an enlargement of alveolar spaces. Under normal conditions, tissue inhibitors of metalloproteinases can regulate MMP activity, but this balance can be impaired under tissue damage situations. In our model of experimental emphysema, CS- exposed mice presented a 2-fold increase in MMP activity assessed *in vivo* by FMI. After treatment with LGF, there was a significant decrease in MMP activity in CS-exposed mice when compared to CS-exposed non-treated mice, reaching values similar to those seen in animals exposed to room air. In order to evaluate the activation of MMPs individually, we performed a gelatin zymography assay. The results showed that increased MMP-9 activity was regulated by LGF when it was administered to CS-exposed mice, reaching values similar to those showed in control group, although the variations were not significant. However, LGF administration had no effect on MMP-2 activity. In relation to mRNA expression levels of MMP-9 and MMP-2 and its inhibitors TIMP-1 and TIMP-2, similar results were obtained to those observed by gelatin zymography. ## Tissue levels of VEGF and PCNA In order to assess the effects of LGF administration on cell proliferation, VEGF and proliferating cell nuclear antigen (PCNA) levels were determined in lung tissue homogenates. VEGF levels determined by western blot were diminished ∼30% in CS-exposed mice compared to the control group. But when treated with LGF, the amount of VEGF in lung tissue was similar than that observed in control group. In the same way, VEGF levels determined by ELISA were diminished in CS-exposed mice (2.6±0.8 ng/ml) and restored after treatment with LGF (3.48±0.7 ng/ml). Regarding the amount of PCNA, there was an increase of ∼40% in CS-exposed mice treated with LGF compared to CS-exposed non-treated mice , which can be considered an index of the regeneration wave produced by LGF. Furthermore, a higher number of proliferating cells (PCNA⁺) was determined in CS- exposed and LGF-treated mice (151.3 cells/mm²) when compared to CS- exposed mice without LGF treatment (75.8 cells/mm²). ## LGF ameliorates oxidative stress To study the effects of LGF administration on oxidative stress burden, 3NT levels (marker of oxidative stress) and NF-E2-related factor 2 (Nrf2) levels (marker of antioxidant response activation) were determined in lung tissue homogenates. Long-term exposure to CS induced a significant increase in 3NT (∼20%) compared to the control group. However, when CS-exposed mice were treated with LGF, 3NT expression reached similar levels to those seen in the control group. Moreover, Nrf2 expression did not present variations in CS-exposed mice in relation to control littermates, but increased ∼30% when treated with LGF. # Discussion CS-exposed mice developed lung emphysema as determined by the enlargement of alveolar spaces and loss of lung function, whereas the size of alveolar spaces in LGF-treated mice after CS exposure was similar to that seen in control mice. Moreover, lung changes correlated with an increase in MMP activity and 3NT expression and lower expression of VEGF in CS-exposed mice, while CS-exposed mice treated with LGF showed decreased MMP activity and 3NT expression, and higher levels of VEGF, PCNA and Nrf2. Thus, our study confirms that LGF treatment substantially reverses CS-induced emphysema as recently reported in AKR/J mice with CS-induced emphysema and provides more details about the mechanisms that are regulated by the administration of LGF. In a study performed in C57BL/6J mice, body weight gain was more attenuated in CS-exposed animals compared to the control group. Specifically, we confirmed that the body weight gain in CS-exposed mice was altered from the third week of CS-exposure (more consistently from the third month), whereas mice exposed to room air did not show any change. Similar alterations were also observed in a study performed by our group in AKR/J mice, where mice exposed to CS stopped gaining weight in the third month, and remained stable until the end of the experiment. Over the years, many studies have focused on the use of therapeutic agents to reverse lung tissue disruption. In one study, all-trans-retinoic acid reversed emphysematous lesions in rats instilled with elastase, although not in adult mice. N-acetylcysteine, another agent previously tested by our group, prevented morphometric and ventilation distribution alterations in the lungs of rats exposed to CS. More recently, the use of growth factors has emerged as a new approach for the treatment of COPD. Keratinocyte growth factor ameliorated the enlargement of alveolar spaces in C57BL/6J mice with elastase-induced emphysema. Hepatic growth factor expression also improved the emphysematous changes in rats expressing human hepatic growth factor gene. In our study, we tested the effects of LGF in pre-established CS-induced emphysema in C57BL/6J mice and showed that LGF improved lung changes as demonstrated by a decrease in L_m and AIA compared to CS-exposed and non-treated animals, reaching values similar to those of air-exposed group. Alveolar macrophages activated by the toxic compounds in CS, release MMPs that participate in extracellular matrix degradation and contribute to emphysema. In fact, several studies have shown higher levels of MMPs in patients with emphysema. Consistent with previous studies, here we show that long-term exposure to CS in mice leads to an increase in MMP activity in the lungs. In contrast, LGF treatment after CS exposure promoted a decrease of MMP activity as seen by *in vivo* FMI. The use of imaging techniques to visualize MMP activation in the lungs in a mouse model of CS exposure was previously described by our group. This method allows MMP activity to be assessed in real time, thus serving as a useful tool for evaluating lung damage. According with the in vivo FMI experiments, the results obtained by gelatin zymography revealed that individual MMP-9 and MMP-2 activities, which were slightly increased in CS-exposed mice, were positively regulated by LGF administration. Although differences between groups were not significant, maybe due to the low number of mice per group, a tendency can be observed. Furthermore, the results of mRNA expression determined by real-time PCR correlated with those seen by gelatine zymography suggesting that LGF could regulate the extracellular matrix destruction by promoting the upregulation of MMP inhibitors TIMP-1 and TIMP-2 and the downregulation of MMP-9 but not MMP-2, both involved in the development of emphysema. A study in rats showed that blocking the VEGF receptor increased alveolar enlargement and alveolar septal cell apoptosis. Similarly, lung-targeted ablation of the VEGF gene led to air space enlargement in mice. In this sense, our results showed that the amount of VEGF in lung tissues of mice exposed to CS was reduced, but restored when mice were treated with LGF after long-term CS exposure. Furthermore, it has been demonstrated that LGF is able to stimulate VEGF in rat testis and to promote endothelial cell proliferation in different systems. PCNA was also upregulated in CS-exposed mice treated with LGF, which is considered to be a marker of LGF activity. In fact, several studies that addressed the effects of LGF in other disease models have revealed higher levels of PCNA, thus highlighting the mitogenic properties of LGF. It is important to note that in lungs from mice treated with LGF after CS-exposure, there was a higher number of PCNA-positive cells determined by specific PCNA-staining, which seem to correspond to type II pneumocytes in terms of morphology, size and appearance. Type II pneumocytes are considered the progenitor cells of the alveolar epithelium with capacity to differentiate into type I pneumocytes in response to epithelial injury. In fact, there have been many studies that have highlighted the important role of type II pneumocytes in alveolar epithelium repair. In that sense, the administration of LGF could be promoting lung tissue repair by triggering the proliferation of type II pneumocytes. Oxidative stress triggered by reactive chemicals present in CS may contribute to the abnormalities observed in the lungs of COPD patients such as destruction of the alveolar walls and enlargement of air spaces. In that sense, we observed that in lung tissues of CS-exposed mice, the levels of 3NT were increased when compared to control tissues. Our results are consistent with those described in patients with COPD, where the number of 3NT-positive cells and levels of 3NT increase in COPD airways. Similarly, it has been reported that levels of 3NT are increased in the lungs of mice exposed to CS and negatively correlate with lung function. Thus, we conclude that the decrease of 3NT observed in LGF-treated mice after CS exposure could be indicative of less oxidative burden and reduction of the inflammatory response in the lungs. We also showed that the expression of Nrf2 was not increased in CS-exposed mice. Furthermore, in Nrf2^−/− mice subjected to CS long-term exposure, it has been reported that neutrophil elastase is elevated when compared to wild-type mice, underlying the role of Nrf2 not only in the regulation of oxidant/anti-oxidant balance, but also in the regulation of inflammation and protease/anti-protease balance. Similarly, our results showed that the post-administration of LGF in CS-exposed mice induced an increase in the expression of Nrf2 that correlates with less MMP activity *in vivo* and 3NT in lung tissues. Thus, our results suggest that Nrf2 could be playing an important role against the development of CS-induced emphysema, highlighting the positive impact of LGF in the activation of the anti-oxidant response. In line with this reasoning, LGF is known as a free radical scavenger. Despite this, the mechanism of LGF action is very complex, due to the wide variety of the functions affected by LGF. For instance, LGF is able to stimulate the growth of hepatocytes, endothelial, smooth muscle cells, astrocytes, microglia and stem cells among others. This mitogenic activity of LGF in the liver is mediated by a local, transitory and mild TNF-α stimulation produced by endothelial cells. On the other hand, LGF has also been shown to have antifibrotic activity in the liver and is also a potent free radical scavenger, with both *in vivo* and *in vitro* activity. Lastly, LGF injection promotes overstimulation of a wide number of intermediaries such as sphingosine 1-phosphate, a fundamental compound in cell survival in emphysema. Our results deepen the understanding of the LGF regenerative properties in C57BL/6J mice with CS-induced emphysema and provide evidence, after analysis of lung injury markers and cell proliferation, about the mechanisms by which the LGF promotes improvement of emphysematous profile. Thus, further experiments are necessary to dissect the pathways activated and regulated by the action of LGF and to elucidate which cell types promote tissue repair. Finally, the development of effective therapies to slow COPD progression is critical. Thus, we suggest that LGF treatment may be a promising strategy to reverse the progression of COPD in the future. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: AGM SPR RTE GPB. Performed the experiments: AGM SPR RTE. Analyzed the data: AGM SPR GPB. Contributed reagents/materials/analysis tools: JJDG. Contributed to the writing of the manuscript: AGM SPR GPB NGM.

# Introduction Fragment-based *de novo* protein structure prediction is the current standard for template-free modelling of proteins. This approach, exemplified by ROSETTA, relies on creating a library of fragments extracted from known protein structures. In this context, a structural fragment is a continuous subset of the residues of a protein. Such structural fragments are usually less than 20 residues long. Each fragment in the library represents a specific position in the sequence to be modelled (target). The most common technique used to select library fragments is the sequence similarity between the fragment and the region of the target sequence that the fragment represents. Fragments from the library are pieced together in order to generate complete models of the target structure. There are many fragments in the library for each position in the target (e.g. ROSETTA's fragment libraries contain 200 fragments per position) and many proteins are hundreds of residues long. In order to explore this large combinatorial space, heuristics are needed. Commonly used heuristics rely on statistical and physical potentials to ensure that global structural features of proteins are sustained/respected in the models generated from the combinations of fragments. A major problem for all fragment-based *de novo* approaches occurs when the fragment library for a given target does not contain good fragments for a particular region. In that case, low accuracy models will be generated regardless of the precision of the potentials being used and regardless of the amount of computation time invested in the modelling routine. For that reason, accurate fragment library generation is crucial to the success of template-free modelling. NNMake is ROSETTA’s method for fragment library generation. NNMake extracts fragments from a template database of non-identical high resolution structures (\<2.5 Å). It scores every length nine segment of the target sequence exhaustively against all length nine segments within its template database. NNMake's score is based on sequence similarity, predicted secondary structure and predicted torsion angle. A library for the target sequence consists of the 200 top-scoring fragments per target position. The latest version of NNMake was tested on a set of 62 small globular proteins (all shorter than 150 residues in length). In their test procedure, the authors removed all sequence homologs. In order to assess the quality of generated fragment libraries, ROSETTA was used to generate decoys for each target. The number of decoys generated varied according to target length (ranging from 4,065 to 19,183 decoys). The average coordinate root-mean square deviation (cRMSD) between the native structure of the target and the top 0.1 percentile models were computed (average of the 62 targets cRMSD = 3.75 Å). One of the current limitations of NNMake relates to the use of fragment with a fixed length (nine residues long). This may not be ideal as it has been shown that accurate structures can be built using fragments as short as four residues. It has also been reported that fragments longer than nine residues can generate better results if the modelling routine is adjusted accordingly. Other fragment generation software extracts either longer fragments or fragments with varying lengths. Other fragment library generation software also attempt to increase the amount of local structural information used to generate libraries. For example, FRAGFOLD uses super-secondary structure fragments, which express the relationship between consecutive secondary structure elements. FRazor builds on NNMake and incorporates solvent accessibility and contact capacity into its scoring scheme. The authors of FRazor claim that their scoring scheme improves the precision of NNMake libraries by discarding low quality fragments suggesting that a sequence-based score can benefit from additional structural information. HHFrag selects fragments slightly differently from many other methods: by means of profile hidden markov models (HMM). HHFrag uses the HHpred toolchain to build a profile-HMM of the target sequence and a profile-HMM for each sequence in a pre-defined template database. The template database used by HHFrag is the April, 2010 build of PDBselect25, a subset of 4,824 protein chains with less than 25% identity extracted from the PDB. Sequence and predicted secondary structure information are used in the generation of the profile-HMM. The HMM of the target is divided into a set of overlapping HMM fragments of variable length (6−21 residues). Fragment extraction is performed by HMM-HMM alignment using the HHSearch algorithm. Each of the 6 to 21-long target HMMs is aligned and scored against every HMM profile for the proteins in the template database. All fragments with a probability ≥ 0.2 are accepted. For positions where a minimum of ten fragments are not identified, fragments with lower probabilities are accepted, if possible, until the minimum threshold of ten fragments is fulfilled. HHFrag was tested on a set of 105 proteins. The average length of fragments obtained was 10.3 ± 3.6 residues. In order to assess fragment library quality, two measures were defined: precision and coverage. Precision is defined as the proportion of good fragments (RMSD to native structure \< 1.5Å) in the library. Coverage is the percentage of target residues represented by at least one good fragment in the library. HHFrag reports a higher precision (62 ± 16%) compared to NNMake (38 ± 17%). However, sequence homologs were not excluded from the HHFrag results, which may inflate the method’s precision. HHfrag also reports a coverage of 71 ± 13%, which is far lower than NNMake (\~92%). For some target positions, HHFrag does not output any fragments, which can cause difficulties during the modelling step. A recent fragment library generation programme SAFrag also builds HMM profiles to detect fragment candidates, in an analogous fashion to HHFrag. SAFrag HMM profiles are extrapolated from a 27 state structural alphabet. The extracted fragments vary in length from 6 to 27 amino acids. Fragments are scored based on a profile-profile comparison, using the Jensen Shannon divergence. Two different template databases can be used by SAFrag: pdb25 and pdb50. Pdb25 is the same database used by HHFrag, whereas pdb50 imposes a 50% pairwise sequence identity cutoff. The method was validated on a set of 111 targets. SAFrag reports a higher precision and coverage than HHFrag. However their cutoff for defining a good fragment is less strict (RMSD to native structure \< 2.0Å). They also allowed homologs and the target structure itself to be included in their template database. Further, the method outputs on average less than two fragments per target position, which suggests that homolog structures are dominating the fragment libraries generated by SAFrag. SAFrag is only available as a web-server. As described above, different methods diverge in the number of fragments used per position, in the length of the fragments used, in the selectivity of the template databases from which fragments are extracted, and in the way the extraction is performed. In this work, we investigate these aspects of fragment library generation and how they affect the precision of the library. We also evaluate the impact of including/excluding homologs among the set of known structures that fragments can be extracted to assess its impact on methods such as HHFrag and SAFrag. In a real *de novo* structure prediction method homologs would not be available, so it is important to exclude those fragments during method training and validation. Current methods score all types of fragments using the same methodology regardless of the predominant predicted secondary structure of the fragment. Here we analysed the relationship between the predominant predicted secondary structure of fragments and our ability to accurately predict fragments for that position. We observed that fragments with a predominant predicted secondary structure (e.g. α- helical fragments) can be predicted more accurately than other types of fragments. Based on these analyses, we have implemented Flib, a fragment library generation software that exploits the predominant predicted secondary structure of fragments to increase the precision of generated fragment libraries. We have generated fragment libraries for two validation sets: a set of 41 structurally diverse proteins extracted from the PDB (PDB-representative set) and a set of 275 protein domains that were used in CASP9 and CASP10 (CASP set). Fragment libraries generated by Flib were compared with libraries generated by NNMake and HHFrag and found to obtain the best balance between precision and coverage in both test sets. Finally, we used the Flib libraries for protein structure prediction, using our custom implementation of a fragment-based protein structure prediction algorithm, SAINT2 (based on). We were capable of generating accurate (TM-Score to native structure \> 0.5) predictions for 12 of the 41 proteins in our test set. We compared our modelling results against running SAINT2 with NNMake fragment libraries. NNMake libraries generated accurate models for 8 cases. Of the 13 cases for which accurate models were generated using SAINT2, Flib libraries generated more accurate models in 10 cases. These results indicate that Flib can be used to improve the accuracy of *de novo* protein structure prediction and demonstrate the importance of discriminating between different secondary structure fragments when performing fragment extraction. # Results ## Fragment Library Quality Assessment We assess library quality using two commonly employed metrics: global precision and coverage. Precision is defined as the number of good fragments divided by the total number of fragments in a library (the proportion of good fragments in the libraries). Coverage is defined as the number of residues represented by at least one good fragment divided by the number of residues of the target (the proportion of protein residues represented by a good fragment). Different methods employ different cutoffs to distinguish between good fragment conformations and bad fragment conformations. Instead of selecting a single cutoff, we have varied the good fragment cutoff between 0.1 to 2.0 Å computing the precision and coverage across the range. ## The Rationale Behind Flib Flib extracts fragments from a database of known structures using a target sequence. A framework describing Flib’s pipeline is shown in. ## Template Database Construction The template database, the initial set of structures from which fragments for libraries will be extracted, can be built in many ways. Regardless of how the template database is built, for testing purposes it is important to remove homologs to the target. We use the Flib template database, in which we impose a 90% sequence identity cutoff and a 5.0Å resolution cutoff, as it proved to lead to the most accurate fragment libraries. ## Fragment Extraction The next step in a fragment library generation pipeline is fragment extraction. Most fragment library generation software methods use an exhaustive search approach. We compared the use of an exhaustive approach and a random approach. In our exhaustive approach, every fragment ranging from 6 to 20 residues in the template database is scored against all of the positions of the target. The exhaustive library is composed of the top 1,000 scoring fragments per target position. The score used in Flib is based on a sequence component, a predicted secondary structure component and a predicted torsion angle component (see section for more details). We also compared our exhaustive approach to a random approach, in which 5,000 randomly selected fragments are scored for each target position. Fragments that satisfy a predetermined score cutoff are accepted and fragments that do not are discarded (see for more details). On average, the random library contains 2,000 fragments per position. Surprisingly, we observed that fragment libraries extracted at random present slightly higher precision and coverage than the ones generated exhaustively. This is probably because above a certain score threshold, we observe no correlation between the scores and fragment quality. When comparing our three scores (sequence, secondary structure and torsion angle scores) we observed a higher correlation between predicted torsion angle score and fragment RMSD to native structure compared to the other two. This indicates that the predicted torsion angle score is better suited to rank the fragments in the final ensemble. Within Flib’s pipeline, we combine the exhaustive and random libraries. This combined library has, on average, 3,000 fragments per target position (LIB 3000). We rank all the fragments in this library and output the 20 top-scoring fragments per target position according to the torsion angle score (LIB20). On average, within our test data set, 69% of the fragments in LIB20 are extracted by the random method and 31% by the exhaustive protocol. ## Enrichment Step We observed that the ensemble with the highest scoring fragment per position according to the torsion angle score presented a very high precision, albeit at a loss of coverage. We decided to exploit this by implementing an enrichment step, in which we include fragments from LIB500 (analogous to LIB20, but considering the 500 top-scoring fragments) that present less than 0.5Å RMSD to the highest scoring fragment according to the torsion angle score. On average, 6.5 fragments per position are added to LIB20 during the enrichment step. ## Protein Threading Hits Library (TH Library) The final step in our fragment library generation routine is to add to LIB20 fragments extracted from protein threading hits. Protein threading identifies protein segments that present structural similarity to a given target. As we remove homologs, these protein threading hits are too unreliable to be used as templates for template-based modelling, but may still provide locally similar fragments. We have found that adding such fragments (on average, less than five fragments per position), when they are available, to LIB20 increases the precision of our method. ## Predominant Predicted Secondary Structure Determines Fragment Quality Secondary structures (α-helices and β-strands) have restrictions by definition in the torsion angles of their residues, whilst loop regions are not so constrained and can assume a wider range of conformations. Hence, secondary structure elements have a lower degree of conformational variability. Considering that fragments with a larger number of loop residues will present a higher variability, we hypothesized that they would be harder to predict. In order to test our hypothesis, we have investigated the relationship between the RMSD to the native structure and fragments with different predominant predicted secondary structures. Here a fragment is described as representing a target position N when it represents all the target residues between N and N+L (where L is the length of the fragment). We classify a target position as belonging to one of four distinct classes based on the predominant predicted secondary structure of its residues. The four classes of secondary structure (SS classes) are: *majority α-helical*, *majority β-strand*, *majority loop* and *other* (no predominant predicted secondary structure). We analysed the RMSD to the native structure of fragments extracted at random. The fragments were grouped according to our predicted SS classes. The spread of fragment RMSD for the top 200 scoring fragments is shown for every position in the target 1E6K. THE RMSD spread of the top-200 fragments replicate the results obtained with LIB3000. This figure typifies our general observation that that there is a strong relationship between the RMSD to the native structure and our four SS classes. The RMSDs for *majority α-helical* fragments are significantly lower than the RMSDs for other SS classes. *Majority loop* fragments and *other* fragments show a wider variability and are much harder to predict accurately. The difference between SS classes is important in two ways. Firstly as current methods only offer coverage and precision across all SS classes, very poor *majority loop* and *other* precision may be hidden by high *majority α-helical* precision. Secondly, these results suggest that during fragment library generation, it may improve results if we treat fragments differently according to their predominant predicted secondary structure. For that reason, Flib uses different cutoffs for accepting fragments based on SS class. Less stringent cutoffs for majority loop and other fragments are used as their variability is far higher. We have observed that adopting different cutoffs for each SS class improves the precision of the libraries generated by Flib. The usefulness of the fragments added at the protein threading step also differs between each SS classes. Adding the threading fragments to LIB20 increases the precision for *majority β-strand*, *majority loop* and *other* SS classes, but decreases the precision for *majority α-helical* fragments. Thus, no fragments from threading hits are added to the *majority α-helical* target positions in our final library. ## Cross-comparison between Flib and other software We have compared Flib against NNMake and HHFrag (Figs). A large scale comparison including SAFrag could not be performed because the software is only available as a web-server. In order to perform a rigorous comparison, we have used two distinct validation sets: a set comprised of 275 protein domains from CASP9 and CASP10 (CASP set) and a set of 41 structurally diverse proteins (PDB- representative set). The second set was built to be representative of the PDB, both in terms of protein lengths and distribution of proteins amongst different SCOP classes. In all analyses, fragments extracted from homologs were discarded (the impact of filtering out fragments from homologs is described in the next section). If we compare the overall precision and coverage, Flib presents higher precision compared to NNMake and higher coverage compared to HHFrag. This difference appears to be due to the increase in performance for the SS Classes majority α-helical and majority β-strand. HHFrag coverage is significantly lower than that of the other two methods. At a 1.0Å RMSD cutoff, HHFrag's fragment libraries describe slightly more than half of the target residues correctly (\~55% coverage on PDB-representative set, \~65% coverage on the CASP set). HHFrag failed to produce any fragments for \~13% of the positions. This can become a problem during structure prediction considering that modelling routines generally require at least one fragment representing every target position. When comparing the three programmes, Flib achieves the best balance between coverage and precision. Considering a good fragment cutoff of 1.0Å, HHFrag presents the highest average overall precision, \~43%, compared to Flib, \~35%, and NNMake, \~29.1% (data shown for the PDB-Representative set). But HHFrag's precision is increased due to a reduced number of fragments output per position (see below). HHFrag also boosts its precision by not outputting any fragments for regions that are harder to predict (as stated above, on average, \~13% of the residues are not represented by any fragment in an HHFrag generated library). Not outputting fragments for low confidence regions will improve precision, but will also cause difficulties during protein structure prediction. Data for the CASP set can be found in. On the PDB-representative validation set, Flib outputs, on average, 26 fragments per position, with an average length of \~7.4 residues. HHFrag outputs on average 10 fragments per position, with an average length of 9.1 residues. Generating a smaller number of fragments can improve precision, but can represent a problem during modelling since less conformations will be sampled. NNMake always outputs 200 fragments per position with a constant length of nine residues. The three methods perform well at predicting fragments for α-helical segments, however, at 1.0Å RMSD cutoff, Flib's precision is 74.6%, which is higher than NNMake's 59% and HHFrag's 64.7% (data shown for the PDB-representative set). Flib's precision for β-strand fragments is also higher. At 1.0Å RMSD, Flib presents 41% precision against NNMake's 17.7% and HHFrag's 32.2% (data shown for the PDB-representative set). The precision of Flib for the other two SS classes are comparable to NNMake’s precision (less than 5% difference in precision), but lower than HHFrag's. The coverage of Flib libraries slightly exceeds the coverage of NNMake for all SS classes. Results for the CASP set are included in. To assess the statistical significance of our results, we have compared the distribution of RMSDs to the native structure of all fragments output by Flib and NNMake, for all targets in our PDB-representative validation set. We performed a Kolmogorov-Smirnov test (alternative hypothesis that the cumulative distribution function of the RMSDs of Flib fragments is greater than the cumulative distribution function of the RMSDs of NNMake fragments) and we obtained a p-value of 2.2e^-16. This indicates that Flib generate fragments with statistically significant lower RMSDs compared to NNMake. ## Effect of Homologs on Fragment Library Quality We carried out an analysis to assess the impact of extracting fragments from sequence homologs of the target protein on fragment library quality. It has been shown that when a suitable template (a homolog) can be found for a specific target, template-based modelling is the most accurate way to model the structure of that target. Hence, *de novo* protein structure prediction tends to be used only in cases where no homologs can be found. For that reason, a fragment library that is representative of a real *de novo* protein structure prediction case should not contain fragments extracted from homologs. Nonetheless, not all current methods for fragment library generation exclude such fragments from their outputs or from their published tests \[e.g. 6, 10, 13, 19\]. We have analysed the impact of including/excluding fragments extracted from sequence homologs from fragment libraries generated by two different methods: NNMake and HHFrag. Homologs were extracted from the significant hits output by HHSearch (see section for detail). Results were generated using the PDB- representative set of 41 structurally diverse proteins of lengths varying between 60–500 residues. HHFrag does not provide an option to exclude homologs from their template databases, and fragments resulting from homologs had to be removed from their final outputs in a post-processing step. We have compared the precision and coverage of fragment libraries generated before and after homolog removal. If we consider the cutoff of 1.0Å to define a good fragment, homolog exclusion leads to a loss of precision from \~35% to \~29.1% for NNMake and from \~45% to 43.1% for HHFrag libraries. Homolog exclusion leads to a loss of coverage from \~61.6% to \~54% for HHFrag libraries and from \~96% to 92% for NNMake libraries. Fragments extracted from homologs increase the precision and coverage of fragment libraries. ## Model Generation/Protein Structure Prediction Our results indicate that Flib libraries present higher precision and coverage when compared to NNMake’s. However, in order to determine how well those results translate to protein structure prediction, it is necessary to test the applicability of our libraries within a protein modelling framework. Therefore, we have generated models using Flib libraries first to assess if accurate models could be generated using those libraries and second to compare to the models generated using NNMake’s libraries. We used our custom implementation of the fragment-based *de novo* structure prediction software, SAINT2, to combine the fragments and to sample the conformational space (see for more details). We generated 1,000 decoys for each of the proteins in our PDB-Representative set using SAINT2 and Flib libraries. We compared our results to the results obtained by generating 1,000 decoys with NNMake libraries using SAINT2. The Flib libraries generated accurate models (TM-Score \> 0.5) for 12 of the 41 cases in our test set. The NNMake libraries generated accurate models for 8 of the 41 cases. Of the 13 cases for which accurate models were generated by either method, Flib libraries performed better in 10. Flib failed to generate a correct model in only one case where NNMake libraries produced an accurate result, whereas NNMake libraries failed to generate a correct model in 5 cases where Flib libraries produced good models. # Discussion In this work, we have established that removal of homologs from any fragment library generation pipeline is essential to ensure that the precision and coverage obtained are representative of a realistic *de novo* structure prediction scenario, otherwise overly promising results will be shown. We tested different template databases (subsets of the PDB) in order to understand how database size and selectiveness can affect the quality of generated fragment libraries. Our analysis revealed that larger template databases give marginally better results. This implies that errors introduced by low quality structures are compensated for by the diversity introduced by using more proteins. The correlation between sequence and secondary structure scores and fragment RMSD to the native structure was also investigated. We observed that, once homologs are excluded from template databases, sampling at random from fragments that satisfy a score cutoff produces better results than extracting fragments exhaustively. We opted to employ a combination of both methods (random sampling and exhaustive sampling) in Flib. Exhaustive extraction is useful for finding high scoring fragments that are likely to be good, whereas random methods increase the diversity of the final ensemble. We have observed that ranking fragments according to predicted torsion angles improved results. Previous results suggest that predicted torsion angles perform better than predicted secondary structure in assisting protein structure prediction. Fragments extracted from protein threading hits were also added to our fragment libraries. These fragments improved the accuracy of generated libraries and these fragment ensembles become more consistent because a large number of fragments are extracted from the same template structure. It has been reported that fragment consistency might be more important than target RMSD to generate good models. Fragment consistency is defined as how well a set of fragments representing different target positions can be pieced together. Our analyses have also revealed a strong relationship between library fragment RMSD to the native structure and the predominant predicted secondary structure of the fragments. We have separated fragments into four distinct classes (SS Classes) based on their predominant predicted secondary structure and have shown that more lenient cutoffs lead to higher precision in *majority loop* and *other* fragments, but stricter cutoffs lead to higher precision in the *majority α-helical* and *majority β-strand* fragments. These results also suggest that it is harder to predict good fragments for positions that are represented by *majority loop* or *other* fragments. During model generation, it may be beneficial to concentrate sampling efforts into these harder to predict positions. Flib presents a better balance between coverage and precision when compared against HHFrag and NNMake. Compared to NNMake, Flib can generate fragments with varying lengths. This has been previously shown to improve protein structure prediction. Flib fragments are, on average, 1 residue shorter than NNMake fragments. Considering that RMSD is correlated with fragment length, we investigated whether Flib's higher precision could be explained due to its shorter fragments. We built a new fragment library considering only the first eight residues of each of the fragments output by NNMake. We noticed a slight improvement in the precision of NNMake's libraries, but Flib libraries still presented higher precision and coverage. When compared to HHFrag, Flib presents a higher coverage. Flib also outputs fragments for every target position, which is necessary for structure prediction. We have compared the improvement obtained by using Flib libraries against NNMake libraries in a protein structure prediction framework. Flib libraries generated accurate models in 12 out of the 41 test cases. Further, our libraries outperform NNMake in 10 of the 13 cases where an accurate model was generated. The number of decoys we have generated during our analysis is comparable to the number of decoys that were generated in previous works. However, this number is relatively low and it is hard to assess the statistical significance of our results. For that reason, we compared the RMSD to the native structure of the best fragment for each target position obtained by each of Flib, NNMake and HHFrag. In principle, if the fragment assembly is exhaustive or has reached convergence, it is the best fragment within each window that ultimately determines the outcome of *de novo* structure prediction. Therefore, this comparison describes how well a fragment library can be used to model a target independent of the number of decoys generated. The RMSDs of the best fragments for each target position between Flib and NNMake are comparable and as Flib libraries are nearly 10 times smaller they are better suited for structure prediction. Furthermore, the modelling step in our analysis is computationally intensive. For that reason, we chose to work with a reduced number of targets (41 proteins). We believe that our data set is large enough to assess the impact of using better libraries in a structure prediction context, despite probably not being large enough to be representative of the complete protein fold space. # Materials and Methods ## Training Data Set Our fragment library generation method was trained using a set of 43 structurally diverse proteins extracted from the PDB. A full list of these proteins is given in. These proteins are all single chain, single domain proteins proportionally distributed into the four SCOP protein classes: all alpha, all beta, alpha/beta, and alpha+beta. They are also evenly spread in terms of length, ranging from 50 to 500 residues. Each of the proteins in our dataset belongs to a different Pfam family. Secondary structure for each protein was computed using the software DSSP and predicted secondary structure was computed using PSIPRED. Predicted torsion angles for each protein were computed using the software SPINE-X. ## PDB-Representative Validation Data Set Our fragment library generation method was validated using a set of 41 structurally diverse proteins extracted from the PDB. A full list of these proteins is given in. These proteins are all single chain, single domain proteins proportionally distributed into the four SCOP protein classes: all alpha, all beta, alpha/beta, and alpha+beta. They are also evenly spread in terms of length, ranging from 50 to 500 residues. ## CASP Validation Data Set We have also validated Flib on a set of 275 domains that were used in CASP9 and CASP10. We have used all domains available from both experiments to compose this validation set. ## Homolog Identification Sequence homologue identification was performed using HHSearch. We have used HHSearch with default parameters: database = PDB70_05Jun14, number of iterations = 2, E-value cutoff for inclusion in resulting alignment = 0.001. HHSearch hits with a probability of 99.5% or higher were considered to be homologs. ## Template Databases There are two main criteria used for culling protein structures from the PDB when assembling template databases: pairwise sequence identity and resolution. NNMake accepts what it defines as non-identical sequences (50% identity cutoff) whereas HHFrag imposes a stricter cutoff of 25% pairwise sequence identity. NNmake only uses structure with a resolution better than 2.5Å, whereas HHFrag does not impose any resolution cutoff. We built three protein template databases by culling sequences from the PDB: Database Flib, Database NNMake and Database HHFrag. For Database Flib, we removed any protein that presented a resolution worse than 5Å or that presented more than 90% sequence identity to another protein already in the database. For Database NNMake, we used the same selection criteria defined by NNMake (resolution cutoff of 2.5Å and 50% identity cutoff). Database HHFrag used the same criteria as HHFrag: the April, 2010 build of PDBselect25. These databases were further processed: we precomputed the secondary structure for every entry in each of the template databases using DSSP. We classified each residue in all protein sequences into seven distinct groups based on their backbone torsion angles. These seven groups are based on areas of the Ramachandran Plot as defined by Choi *et al*. These areas define the environments for our environment-specific substitution matrices. Therefore, each entry in a database is represented by three strings: sequence, secondary structure and Ramachandran region identifier. ## Fragment Scores Three main scores were built and tested. All of the scores are defined using a pairwise comparison between fragment and target residues. Ramachandran-specific Sequence Score: The fragment Ramachandran score is defined as the sum of the score of each pair of fragment/target residues. We have defined environment-specific amino-acid substitution matrices to assign scores to a pair of residues. These matrices have been built in a similar fashion to the BLOSUM matrices. They describe the propensity for an amino-acid substitution within a given environment and are extrapolated from amino-acid frequencies encountered in multiple sequence alignments. We defined the environments as the seven Ramachandran plot regions in. This score incorporates additional torsion angle information compared to a standard sequence alignment score (i.e. using the BLOSUM62 scoring matrix). Secondary Structure Score: this score is based on a pairwise comparison between the target fragments’ predicted secondary structures as output by PSIPRED to the database fragments’ known secondary structure as output by DSSP. We used the following scoring scheme: Match = 2, Mismatch = -2. Predicted Torsion Angle Score: the torsion angles (ϕ,ψ) for every database fragment was computed and compared to the predicted torsion angles for the target fragment as output by SPINE-X. We define the predicted torsion angle score as the sum of the absolute differences between predicted and real ϕ angles and between predicted and real ψ angles for each fragment residue. ## Fragment Extraction with Flib Fragments were generated for each of the proteins in our test data set using two extraction methods: random extraction and exhaustive extraction. In all cases, all fragments from homologs to the target were removed. We classified each target position according to its predicted predominant predicted secondary structure into four SS classes: *majority alpha-helix*, *majority beta-strand*, *majority loop* and *other*. For example, if more than half of the residues of a fragment are part of an alpha-helix, then the fragment is classified as a *majority alpha-helix* fragment. If a fragment does not have a predominant SS type, we place it in the *other* category. The random extraction method consisted of scoring 5,000 randomly selected fragments of varying length per target sequence position from the template databases. The length of each fragment was randomized to be between six to 20 residues. Each fragment was scored according to the Ramachandran score and the Secondary Structure score. Every fragment is accepted depending on whether its score satisfies an acceptance cutoff. We have selected the cutoffs that achieve the best precision whilst maximising the coverage. Different cutoffs were determined and used within each fragment SS class. The resulting library presents, on average, 2,000 fragments per target position. In exhaustive extraction all possible fragments from a template database were scored against every position in the target. Analogous to the random extraction, fragments were scored based on the Ramachandran-specific Sequence Score and the predicted secondary structure score. The top 1,000 scoring fragments are selected for each target position as we found that the precision was not increased by the inclusion of more fragments. The top 1,000 fragments per target position obtained by exhaustive extraction are merged with the 2,000 fragments per target position obtained by random extraction. The resulting fragment library presents approximately 3,000 fragments per target position (LIB3000). For each target position, we rank the fragments in LIB3000 according to the predicted torsion angle score. We select the top 20 highest scoring fragments (LIB20) per target position according to the predicted Torsion angle score. We further enrich our final libraries by including any fragment from LIB3000 that presents less than 0.5 Å RMSD to the highest scoring fragment for a given position. In the final step of our routine, we perform protein threading using the target sequence as input to HHSearch. Default parameters for HHSearch were used to perform protein threading. Protein threading hits that originated from homologs, as described earlier, are removed from HHSearch's output. We extract every possible nine-residue fragment from the remaining threading hits (Protein Threading Library). The fragments in the Protein Threading library are ranked according to hit score output by HHSearch. We select a maximum of 20 fragments per target position. Fragments belonging to *majority alpha-helical* positions are removed from the Protein Threading Library in a post-processing step. All fragments in the Protein Threading Library are added to LIB20 to generate the final output of Flib. This final library presents, on average, \~33 fragments per target position. ## Validation Two commonly used metrics to assess fragment library quality are global precision and coverage. Precision is defined as the number of good fragments divided by the total number of fragments in a library (the proportion of good fragments in the libraries). Coverage is defined as the number of residues represented by at least one good fragment divided by the number of residues of the target (the proportion of protein residues represented by a good fragment). The quality of fragments was assessed by superimposing the fragment on to the target's known structure. We have varied the good fragment cutoff between 0.1 to 2.0 Å to compute a curve for precision and coverage. Fragments with an RMSD to the native structure below this varying cutoff are considered to be good fragments. ## HHFrag In order to generate fragment libraries using HHFrag, we have used HHFrag v2.1 with default parameters. ## NNMake We have used NNMake from the Build 3.5 of MiniROSETTA. In order to generate the fragment libraries, default parameters for NNMake were used. ## Model Generation We have generated 1,000 decoys for every protein in our Validation set using two different approaches: Flib’s fragment libraries with SAINT2, NNMake's fragment libraries with SAINT2. ## SAINT2 There is evidence that suggests that co-translational aspects of protein folding could assist protein structure prediction. SAINT2 is a co-translational protein structure prediction software programme. It is a fragment-based approach that relies on sampling the conformational space in a sequential fashion. Unlike other fragment-based approaches, instead of starting with a fully elongated sequence, SAINT2 starts with a short peptide and moves from the heuristic routine are intercalated with an extrusion (a fragment replacement that happens at the end of the nascent chain and that elongates the peptide by one residue). We have incorporated a correlated mutations potential into SAINT2 as it has been reported to improve modelling results. We have used PSICOV to predict protein contacts for each target in our data set. PSICOV generated predictions for 34 of the 43 proteins in our data set (the accuracy of the contact predictions can be found). The predicted contact potential within SAINT2 was based on the contact potential described in. # Supporting Information The authors would like to thank Anthony Bradley for comments on the manuscript. [^1]: One of the authors \[JS\] is currently employed at UCB Pharma. This does not alter the authors' adherence to PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: SHPO JS CMD. Performed the experiments: SHPO. Analyzed the data: SHPO CMD. Wrote the paper: SHPO CMD. Contributed with ideas and with the analyses: JS.

# Introduction Empirical studies evaluating student outcomes across various types of schools can inform decision-making among policy-makers, educators, parents and other education stakeholders. School experiences in primary and secondary schools may be crucial for shaping individuals’ developmental and well-being trajectories in later life, and shaping student well-being is arguably one of the important aims of education. It is, therefore, important to understand students’ long-term achievements and well-being across different aspects of life when comparing various types of primary and secondary schools. Such evidence would further empower decision making among policy-makers, educators, parents and other education stakeholders. While there is considerable variation across individual schools, adolescent schooling can largely be divided into 4 types: public schools, private independent schools, private religious schools and home schooling. According to recent reports, among U.S. adolescents in 2016, approximately 87.0% attended public schools, 8.8% attended private schools and 3.6% were homeschooled. Public schools are mainly funded and regulated by local governments to provide free education to every child. In contrast, private schools primarily depend on private sources of funding (e.g., tuition, donation), are operated by private organizations that are either religiously or non-religiously affiliated, and have relatively high autonomy in decision-making such as student enrollment and curriculum development. Homeschooling involves providing education at home, which is typically led by parents. Homeschooling can follow a predetermined curriculum (i.e., structured homeschooling), or self-directed natural learning without a fixed curriculum (i.e., unstructured homeschooling). These different types of schools often prioritize different educational goals. For example, schools may aim to support students in developing academic knowledge, intrinsic motivation to learn, social skills and networks, civic engagement, a healthy lifestyle, well-being, good character, or a particular religious faith, with different school types emphasizing each of these goals to greater or lesser extents. It is arguably helpful for policy-makers, educators, parents and other education stakeholders to understand associations between school types and student outcomes related to this wide range of educational goals. However, to date, the empirical evaluation of student outcomes across school types has, perhaps understandably, been based primarily on academic achievement. Empirical studies on school types and student outcomes have most often used standardized test scores as the primary outcome for evaluation. The findings from such studies are rather mixed overall, with some studies suggesting that students attending private independent schools, private religious schools and structured homeschooling had modestly higher standardized test scores on some disciplines as compared to their peers at public schools, while other studies did not find such evidence. Beyond academic achievement, studies examining school types and student outcomes related to other educational goals are sparse. There has been some prior research exploring various school types in relation to civic engagement and family formation outcomes, with some research suggesting that attending private independent schools and private religious schools is linked with greater civic engagement and more positive family outcomes than attending public schools, whereas some other research suggested little evidence of such differences. In addition, there has been some prior research on homeschooling versus institutional schooling for a number of student outcomes, with homeschooling associated with greater civic engagement, less alcohol and drug use, better sleep, equal or better mental health and well-being and equal or better social- emotional skills. While these studies have contributed to the literature, several methodological concerns remain. For instance, most of these studies had small samples, limited covariate control, and used cross-sectional data, making it difficult to assess evidence for causal effects. More research is needed to gain a comprehensive understanding of associations between various school types and a diverse array of student outcomes, with longitudinal data and rigorous methodologies. To address these gaps in the literature, we performed an outcome-wide longitudinal analysis to compare adolescents attending various types of schools in the years that followed across a wide range of outcomes in their young adulthood, with extensive control of potential confounders (e.g., family socioeconomic status, family environment). The outcomes include multiple indicators of subsequent psychological well-being, social engagement, character strengths, mental health, health behavior and physical health outcomes. # Methods ## Study population This study used longitudinal data from the Nurses’ Health Study II (NHSII) and the Growing Up Today Study (GUTS). Established in 1989, the NHSII cohort enrolled 116,430 female registered nurses aged 25 to 42 years from across the U.S. In 1996, NHSII participants with children between the age of 9 to 14 years old were invited to have their children participate in another cohort GUTS. Invitation letters and questionnaires were then mailed to the children whose mother provided consent. Of them, 16,882 children returned the completed questionnaires at study baseline, thereby assenting to participate. Since then, NHSII and GUTS participants have been followed up through mail or web-based questionnaires annually or biennially. In this study, school types were assessed in the GUTS 1999 questionnaire wave (N = 12,288, mean age = 14.56 years); thus, this year was considered as the study baseline. Data on outcome variables were taken from the most recent GUTS questionnaire waves, primarily the 2010 questionnaire wave (mean age = 25.10 years); if the outcome was not assessed at the 2010 wave, we used data from the 2013 or 2007 wave; covariates were mostly assessed at or prior to the 1999 wave ( provided the timeline regarding the measurements of all variables). Among participants of the 1999 questionnaire, 1,025 individuals had missing data on school type, another 6,711 participants had missing data on at least one covariate (most covariates had less than 18% of missing data); depending on the outcome, another 681 to 1,510 participants had missing outcome data or were lost to follow-up. A multiple imputation procedure was used handle missing data on all variables. This yielded an analytic sample of 12,288 participants, with 2,432 of them being siblings (some families had multiple children enrolled). This study was approved by the Institutional Review Board at the Brigham and Women’s Hospital. ## Exposure assessment ### School types Participants were asked to report the types of schools that they were attending in response to the question (GUTS 1999): “What type of school do you attend?” The responses were grouped into 4 categories including public schools, private independent schools, private religious schools, and home schooled. Those who reported not in school or attending universities were excluded from all analyses. ## Outcome assessment A wide array of outcomes in young adulthood were assessed (primarily in 2010). Such outcomes included indicators of psychological well-being (i.e., life satisfaction, positive affect, self-esteem, emotional regulation), social engagement (i.e., marital status, community engagement, religious service attendance, educational attainment), character strengths (i.e., volunteering, sense of mission, forgiveness, civic engagement), mental health (i.e., depression, anxiety, post-traumatic stress disorder \[PTSD\]), health behaviors (i.e., current smoking, binge drinking, marijuana or other illicit drug use, prescription drug misuse, number of lifetime sexual partners, early sexual initiation, history of sexually transmitted infections \[STIs\], short sleep duration, preventive healthcare use), and physical health (i.e., overweight/obesity, a number of physical health problems). Details on the measurement of all outcome variables were provided in the. ## Covariate assessment ### Demographic characteristics Demographic covariates included participant age (in years), sex (male, female), race/ethnicity (non-Hispanic white, others), geographic region (West, Midwest, South, Northeastern), and puberty development (assessed with the tanner stage score). Maternal demographic covariates were also considered including mother’s age (in years), race/ethnicity (non-Hispanic white, others), and marital status (married, others). ### Family socioeconomic status (SES) Multiple indicators of family socioeconomic status were adjusted for including maternal subjective SES in the U.S. and in the community (both assessed with validated scales on a 10-point scale), mother’s current employment status (currently employed, unemployed), father’s educational attainment (high school or less, 2-year college, 4-year college, grad school, non-applicable), pretax household income (1: \<\$50,000, 2: \$50,000-\$74,999, 3: \$75,000-\$99,999, 4: ≥\$100,000), census-tract college education rate (used as a continuous variable), and census-tract median income (1: \<\$50,000, 2: \$50,000-\$74,999, 3: \$75,000-\$99,999, 4: ≥\$100,000). ### Family environment factors The following baseline family environment factors were considered including participant family structure (live with both biological parents, live with a stepparent, others), family dinner frequency (never/sometimes, most days, everyday), religious service attendance (never, less than once/week, at least once/week), maternal relationship satisfaction (retrospectively reported by GUTS participants, assessed with a nine-item validated scale measuring parent-child relationship satisfaction), maternal depression (yes, no), and maternal smoking status (never smoker, former smoker, current smoker). ### Prior health status or health behaviors To reduce concerns about reverse causation, the following health characteristics at baseline were adjusted for: depressive symptoms (assessed with the Depression Symptoms Scale of the McKnight Risk Factor Survey), overweight/obesity (yes, no), current cigarette smoking (yes, no), frequent binge drinking (yes, no), marijuana or other illicit drug use (yes, no), prescription drug misuse (yes, no), history of STIs (yes, no), history of early sexual initiation (yes, no), and the number of lifetime sexual partners (a continuous score). ## Statistical analyses All statistical analyses were performed in SAS 9.4 (tests of statistical significance were two-sided). Analysis of variance and Chi-square tests were used to examine baseline participant characteristics across school types. In primary analyses, generalized estimated equation (GEE) models with independent covariance structure were used to regress each outcome on school types separately, adjusting for clustering by sibling status. All continuous outcomes were standardized (mean = 0, standard deviation = 1), so the effect estimates were reported in terms of standard deviations in the outcome variables. To account for multiple testing, Bonferroni correction was performed. All models controlled for sociodemographic characteristics, family environment factors, and health status and health-related behaviors at baseline. Because multiple imputation provides a more flexible approach than many other methods of handling missing data, we performed multiple imputation by chained equations to impute missing data on all variables, with 20 imputed datasets created. As a sensitivity analysis, we also reanalyzed the primary sets of models using complete-case analysis. A number of other sensitivity analyses were performed. First, because public school qualities are often influenced by district- and state-level characteristics, we reanalyzed the primary sets of models 1) stratified by neighborhood SES first, and then 2) restricting to participants from the 10 states with the highest and the 10 states with the lowest public school ranking separately. Second, because some parents might send their children to religious schools for non-religious reasons, we compared students attending religious schools versus public schools, stratified by their frequency of religious service attendance at baseline (considering at least once/week of attendance as a proxy indicator for religiousness). Next, because religious faith is a major reason for homeschooling, we compared the home-schooled with those attending religious schools across the outcomes. Lastly, we evaluated the extent to which the associations between school types and various outcomes were robust to potential unmeasured confounding. For this purpose, we calculated E-values, which represent the minimum strength of association that an unmeasured confounder(s) would need to have with both the exposure and the outcome variables on the risk ratio scale to fully explain away the exposure-outcome associations, above and beyond the measured covariates. # Results ## Participant characteristics At study baseline participant age range was 11–19 years, with a mean age of 14.56 years (SD = 1.62). The participants were higher percentage female, primarily non-Hispanic White, mostly had a high level of family SES, and were generally healthy. The majority reported attending public schools (80.56%), followed by private religious schools (9.67%), private independent schools (8.12%), and homeschooling (1.66%). Compared to those at public schools, participants who attended private independent or religious schools generally had a higher level of family SES. Further, participants at religious schools or in homeschooling were more likely to attend religious services, live with both biological parents, have family dinners frequently, and have lower rates of smoking, binge drinking, drug use, maternal depression or maternal smoking at baseline. Consistent with findings in other samples, homeschoolers in this sample were more common in the South and Midwest, and their mothers were less likely to be currently employed. ## School types and subsequent health and well-being There was little difference in subsequent outcomes between adolescents attending private independent schools versus public schools across various health and well-being outcomes examined, except for some evidence that private school students subsequently reported slightly higher levels of forgiveness (β = 0.08, 95% CI: 0.02, 0.15), though the association did not pass the P\<0.05 threshold after Bonferroni correction for multiple testing. As compared to public schools, there was some evidence that students at religious schools subsequently had a higher likelihood of frequent religious service attendance and becoming registered voters, a lower risk of overweight/obesity and fewer lifetime sexual partners on average (e.g., β_{number of sexual partners} = -0.08, 95% CI: -0.14, -0.02); however, they were more likely to subsequently be frequent binge drinkers (e.g., RR_{binge drinking} = 1.15, 95% CI: 1.04, 1.27), though such associations again did not reach a *p* \<.05 threshold after accounting for multiple testing. Compared to those attending public schools, homeschooled students were subsequently 51% more likely to attend religious services frequently (RR = 1.51, 95% CI: 1.27, 1.80), reported greater frequency of volunteering (β = 0.33, 95% CI: 0.15, 0.52), and had substantially higher levels of forgiveness on average (β = 0.31, 95% CI: 0.16, 0.46), but were 23% less likely to attain a college degree (e.g., RR _{attain a college degree} = 0.77, 95% CI: 0.67, 0.88) in young adulthood; all of these associations also passed the p\<0.05 threshold even after Bonferroni correction for multiple testing. There was also some evidence that homeschooled students subsequently reported a higher level of sense of mission in life, lower risks of marijuana use and fewer lifetime sexual partners, but possibly had a higher risk of PTSD; these latter associations, however, passed conventional, but not Bonferroni-corrected, p-value thresholds. ## Sensitivity analyses for unmeasured confounding E-values were calculated for assessing robustness of the observed associations to potential unmeasured confounding. There was evidence, for example, that the associations of homeschooling with subsequent volunteering, forgiveness, religious service attendance, and educational attainment were at least moderately robust to unmeasured confounding. For instance, to fully explain away the observed association between homeschool and volunteering above and beyond the measured covariates, an unmeasured confounder associated with both homeschooling and greater likelihood of volunteering by 2.04-fold each on the risk ratio scale could suffice, but weaker joint confounder associations could not; and unmeasured confounding risk ratios of 1.54-fold for both volunteering and home-schooling could suffice to shift the confidence interval to include the null value, but weaker joint confounder could not. Similarly strong E-values were observed with homeschooling in relation to lower education attainment, higher forgiveness, and greater religious service attendance. In contrast, for all comparisons of outcomes for public versus private independent schools, and all comparisons of public versus religious schools, the E-values for the confidence interval were at most 1.24, and often considerably smaller, suggesting modest amounts of confounding could suffice to explain away the observed difference. The only moderately robust evidence to potential unmeasured confounding was thus comparing public schools and homeschooling. ## Other sensitivity analyses First, reanalyzing the primary models using complete-case analyses yielded similar results as the primary analyses. Second, the analyses stratified by neighborhood SES also yielded similar results as the primary analyses. Specifically, there was little difference between private independent schools and public schools across outcomes among those residing in areas with either low or high levels of census-tract median income; magnitudes of the effect estimates comparing religious versus public schools across outcomes were also similar to the primary analyses, but the confidence intervals were wider due to the smaller sample size in each stratum. Next, the analyses restricting to participants from states with the lowest and the highest public school rankings again found little difference between private and public schools in those states. Next, the sensitivity analyses stratified by frequency of religious service attendance suggested that the associations of religious schools (versus public schools) with greater likelihood of registered voting status, fewer lifetime sexual partners and lower risk of overweight/obesity, but elevated risks of binge drinking were slightly stronger among those who attended religious services more frequently. Finally, the analyses comparing homeschooling to religious schools provided some suggestive evidence that the homeschooled adolescents may volunteer more frequently and have a lower risk of marijuana use in their young adulthood. # Discussion The present study suggests that for the children of nurses who participated in this study, there was little difference between attending private independent schools versus public schools in subsequent health and well-being outcomes in young adulthood. There was also only modest evidence for differences in subsequent outcomes when comparing private religious schools to public schools. In contrast, there was considerably greater evidence that homeschooling versus public schools was positively associated with several outcomes (e.g., volunteering) but negatively associated with others (e.g., educational attainment). Prior empirical studies comparing student outcomes across various types of schools have primarily used short-term standardized test scores as the outcome for evaluation. This study extends the literature by simultaneously examining multiple long-term health and well-being outcomes using longitudinal data. Below we will comment on relations to prior literature on this topic, but also on the particularities of the sample used in this study. Consistent with some prior studies suggesting little or only modest differences in test scores comparing private and public school students, this study did not find substantial differences in longer-term educational attainment (i.e. college degree). While outcomes beyond academic achievement have been less often investigated, congruent with some of the strongest prior evidence, this study also suggested little difference in social connectedness between private versus public school attendants. Likewise, consistent with some prior evidence, yet contrary to other studies, this study also found little difference in subsequent civic engagement comparing private versus public school students. It is possible that private and public schools may differ in outcomes that were not examined in this study, such as students’ subjective schooling experiences, opportunities for parental involvement and parental satisfaction. It is also possible that there may be greater variations within, rather than between, these types of schools. For instance, some important factors that contribute to school performance such as teacher quality, teacher experience, and the availability of after-school programs may vary considerably across individual schools. This study found only relatively modest health and well-being associations comparing attending religious schools versus public schools concerning overweight/obese and lifetime sexual partners. Attending religious schools was associated with a slightly higher risk of frequent binge drinking in young adulthood in this sample. This was surprising as prior research has suggested that religious service attendance during childhood and adolescence is associated with subsequently healthier behaviors in general. However, it may be religious service attendance (rather than religious schooling) that is the primary driver of the overall associations with religious upbringing. Our analyses adjusted for, and stratified by service attendance, while this has not often been accounted for in prior studies of religious schooling. It is, therefore, possible that the associations between religious schooling and health in some prior studies may in fact reflect confounding by religious service attendance, which again evidence suggests is related to subsequent health and well-being. However, if service attendance is itself a part of religious schooling (and possibly the only source of service attendance for some students) then it is also possible that control for service attendance is over-adjustment and may in fact be an integral part of the effects of religious schooling. In any case, the present analysis suggests that it may be religious service attendance, however it is experienced, rather than other aspects of religious schooling that have the more substantial associations with outcomes later in life, at least for the outcomes examined here. Religious knowledge and literacy, which may be the primary motivation for religious schooling for some parents, was not assessed in this study. The largest differences in our study in subsequent outcomes were between homeschooling and public schools. Congruent with prior studies, homeschoolers in this sample (versus those at public schools) were more likely to report subsequently greater character strengths and fewer risky health behaviors. However, homeschooled students were less likely to attain a college degree. While educational attainment may differ between structured and unstructured homeschooling, this study did not have data on such subtypes and found that, averaging across these subtypes, and overall homeschoolers had a lower likelihood of attaining a college degree in young adulthood. This may in part reflect lower attainment in learning or less interest in attending college, but it may also reflect the status quo that some U.S. universities have restricted admission policies for the homeschooled. Contrary to prior evidence that homeschoolers (versus public school attenders) typically have equal or greater psychosocial and emotional well-being, this study suggested that homeschoolers may have a higher risk of probable PTSD in young adulthood. These contrasting results might in part be attributed to the longitudinal design and the covariate control strategies in this study as compared to prior studies; we were examining outcomes in young adulthood, rather than while the children were still at school and associations could potentially differ for outcomes assessed in the short-run versus the long-run. There have been controversies over regulations concerning homeschooling and also over whether and what types of public-school services should be made accessible to the homeschooled, with many of the discussions centered around academic resources and extracurricular activities. With the growth in internet use, homeschooling has becoming increasingly easier and more popular in the United States. The Covid-19 pandemic has also forced some parents into home-schooling and this may itself alter long-term practices. Although the associations in our study warrant further investigation in future studies, the results here provide some suggestive evidence that support for the psychological well-being of homeschoolers may be worthwhile. This study is subject to certain limitations. First, the participants were mostly non-Hispanic White and were all children of nurses. Findings of this study may not be generalizable to other populations. Specifically, because all of the students were children of relatively well-educated mothers, this group may have been more able than most to ensure high quality schooling for their children regardless of school type and also more likely to change school type if the particular public or private or religious schools in their area were deemed to be inadequate. The comparisons in this paper pertain to the schools attended by students in this particular sample; they are not comparisons across all U.S. schools. The findings may therefore be most relevant for families who are facing decisions and school dynamics relatively similar to this sample, rather than representative of the general U.S. population. Second, while there may be substantial variation within types of schools, we were unable to account for characteristics of individual schools due to the lack of data. However, the homogeneous feature of this sample (all participants were the children of nurses) and the sensitivity analyses stratified by multiple sociodemographic characteristics helped reduce such concerns. Third, the various school types can be further divided into subtypes that may be associated with different outcomes in certain cases, we could not explore such subtypes here due to a lack of data. For example, we could not examine charter schools separately, which are publicly funded schools with relatively high levels of autonomy in curriculum design, budgets and personnel hiring, though these are more common now than when school type in this study was assessed. Likewise, we could not examine the subcategories of structured and unstructured homeschooling independently. Further, the sample size of homeschoolers was relatively small (n = 187) in this study, which may have limited our statistical power. However, we nevertheless found associations between homeschooling and several outcomes, even with this more limited statistical power; moreover, we found few differences among any of the other school types, even though the sample sizes were larger. Despite these limitations, this study provides important evidence concerning associations between school types and a wide range of long-term outcomes. To our knowledge this is the first study that has prospectively examined a wide range of long-term health and well-being outcomes across multiple types of adolescent schooling. Further, this study rigorously accounted for a wide array of covariates that helps reduce concerns about potential confounding, selection bias and reverse causation, which are major methodological concerns in prior studies. School choice is certainly shaped by a variety of factors, such as beliefs, values, and logistical considerations, in addition to a desire for academic learning and educational achievement. A broad range of outcomes, considering numerous aspects of a child’s long-term well-being, is therefore arguably relevant for decision-making. The results of this study might thus help inform policy-makers, educators, parents and other education stakeholders in their decisions by consideration of the evidence on this broader range of educational goals and outcomes. # Supporting information We thank the Channing Division of Network Medicine, Department of Medicine, Brigham and Women’s Hospital, and Harvard Medical School for their support in conducting this study. [^1]: The authors have declared that no competing interests exist.

# Introduction Perennial crops promise sustainable production and increased environmental benefits relative to annual cropping systems. For example, perennial species allocate more resources to belowground productivity than annuals, which may lead to increases in soil carbon, nutrient retention, and hydraulic conductivity. These benefits result from a simultaneous reduction in soil tillage and by shifting the succession of agricultural systems to establish perennial crops that interact with their soil ecosystems for several years or longer. With increased crop longevity and lack of crop rotation, perennials have prolonged interactions with their soil microbiome. For this reason, it is important to understand how perennial crops respond to the biological context of agriculture, including soil pathogens, mutualists, and plant cropping diversity. Many perennial plants, including perennial crops, are strongly responsive to mutualistic relationships with arbuscular mycorrhizal (AM) fungi and are more sensitive to AM fungal identity than annuals. Therefore, productivity of perennial crops is likely to be influenced by the composition of AM fungi present in soils. As in annual systems, new perennial plantings typically occur in recently disturbed soils, where land manipulation such as tilling, crop monocultures, and the use of soluble fertilizers and biocides can lead to degraded AM fungal diversity, composition, and abundance. Past work has shown that new perennial crops benefit from being planted with native AM fungal amendments isolated from undisturbed soils. Productivity of perennial crops may also depend upon interactions with pathogens. While perennials can be better defended against pathogens than annuals, the longer duration of their plantings makes them more likely to accumulate host-specific pathogens than annual plantings. These host-specific pathogens may cause greater declines in productivity over time when compared to annual systems, where rotations of different crops may reduce dominance of crop- specific pathogens. Identifying the relative importance of AM fungal and pathogen components in perennial cropping systems is critical to help leverage plant-microbe interactions for sustained production in perennial agriculture. Crop diversification, by planting multiple species simultaneously (intercropping), can be an important component for sustainable agriculture that can help mitigate some of the pathogen accumulation as well as abiotic changes predicted to occur with perennial crops. Intercropping can increase agricultural productivity, as diversified mixed species plantings commonly have greater yield than monocultures. This phenomenon, known as overyielding, is predicted to occur when different crops are able to use different resource pools in space or time. Different crops can have disparate growth patterns above and below ground to optimize resource capture (i.e. light, nutrients, and water), which can reduce competition and increase net resource utilization relative to monoculture plantings. For example, the different seasons of activity in wheat and maize intercrops can result in overyielding. This type of overyielding based on reduced resource competition explicitly depends on the availability of specific resources, but the direction of resource effects on overyielding is not always consistent. For example, increasing resource availability can increase or decrease overyielding. In addition, crop identity, functional group, and phylogenetic distance may all be important factors to minimize resource competition, enable facilitation, and create compatible crop mixtures. Thus, understanding species-specific crop companion interactions is essential to predict outcomes and sustainability of crop diversification for landscape scale plantings. Overyielding via crop diversification can also be mediated by interactions with the soil community. Microbial mediation of resource partitioning underlies the classic expectation of overyielding between cereals with high demand for nitrogen and legumes, whose symbiosis with rhizobia allows them to access atmospheric nitrogen. In addition, symbioses with AM fungi can ameliorate resource deficiencies for hosts (nutrients and water) and can alter the strength of interactions between species. For example, AM fungi mediation of resource partitioning is supported by increased complementarity observed between maize and faba bean when in association with AM fungi. Soil pathogens could also mediate overyielding, as accumulation of species-specific pathogens may limit yield in monoculture plantings. Substitutive planting with another crop lowers host density in mixture, resulting in decreased pathogen abundance and a reduction in this deleterious effect. While both pathogen accumulation and microbially-mediated resource partitioning have been observed to generate overyielding in perennial and annual systems, these overyielding mechanisms themselves may be context dependent. The benefits of intercropping legumes may be reduced in soils with high nitrogen availability, and the positive effect of AM fungi on overyielding may be decreased when phosphorus is abundant. While AM fungi mediated impacts on overyielding under different levels of water availability are less known, pathogen impacts on hosts do vary with water availability. This could cause the magnitudes of overyielding to vary between wet and dry conditions. Understanding the biotic and abiotic contexts and mechanisms driving overyielding can help predict compatible perennial crop pairs, and potentially illuminate ways to increase sustainability of perennial crop plantings. The objective of this experiment was to determine the compatibility and potential overyielding in mixtures of three perennial crop species under different abiotic (water availability) and biotic (changes in soil biota) conditions. We chose perennial crop candidates that have cereal, oilseed, and forage production potential and also represent three distinct functional groups (cool-season grass, forb, and legume). The cool-season grass *Thinopyrum intermedium* produces the novel perennial grain Kernza^® and it has been selected for many desired agronomic traits at The Land Institute in Salina, KS. Throughout the rest of the manuscript, Kernza will be used to describe the entire crop plant not just the grain. The forb *Silphium integrifolium* (Rosinweed) is also being studied at The Land Institute. It is a warm -season forb native to the tallgrass prairie and has potential as a perennial oil seed crop. The commonly farmed alfalfa (*Medicago sativa*) was used as the perennial legume. Mixtures of these species have the potential to increase a number of ecosystem services, yet more research is needed to understand the interactions of these crops under different abiotic and biotic contexts. In this greenhouse study, we ask these questions to better understand the interactions of these crops: 1. How do the planted crop community, soil community, and water availability influence the performance of the perennial crop species? 2. Do mixtures of these perennial crops overyield relative to their component monocultures? 3. Is any overyielding mediated by the soil community, water availability, or their interaction? # Materials and methods ## Experiment location The pot experiment was conducted in the west campus greenhouse at the University of Kansas in Lawrence, Kansas U.S.A. Greenhouse temperature controls were set to allow a temperature range of 65 to 85°F and no supplemental lighting was used. ## Soil inoculum Pots (7 L) were partially filled with a steam sterilized (twice at 174° F) 50:50 sand:soil mixture. The nutrient content of the sterilized soil was 15.8 ppm phosphorus via Melich extraction and 26.55 ppm nitrate (NO3-N) and 5.8 ppm ammonium (NH4-N) via KCl extractions. One of four soil inoculum was added (280 cm³ total, 4% by volume), and then the pots were filled the rest of the way with the sterile sand:soil mixture. Each inoculum consisted of two components (140 cm³ each): live whole soil and live prairie AM fungi (LWLF), live whole soil and sterilized prairie AM fungi (LWSF), sterile whole soil and live prairie AM fungi (SWLF), or sterile whole soil and sterile prairie AM fungi (SWSF). The small volume of inoculum was used to minimize potential differences in abiotic properties among the inoculum, which may be due to soil conditioning effects or nutrient release after sterilization. The whole soil (LW) was collected from long-term (established in 2002) monoculture plots of intermediate wheatgrass (*T*. *intermedium*) at The Land Institute in Salina, KS as part of the Agroecology Research Trials (38.767690°, -97.572539°). We chose to use a soil community with a history of long-term soil conditioning by *T*. *intermedium*, without disturbance (no tillage), to test *T*. *intermedium-*specific pathogens and mutualists (i.e. AM fungi), which have been shown to be important in mediating overyielding in perennial systems. Whole soil was collected from the top 10 cm, sieved (1 cm), and stored at 4°C for less than one week prior to inoculating the experiment. The prairie AM fungi inoculum was isolated and cultured from a native Kansas remnant prairie (39.044991°, -95.191569°) with Oska silty clay loam and Pawnee clay loam soil types. Undisturbed remnant prairies contain unique AM fungi communities not found in highly disturbed agricultural systems, and studies have shown differential responses of plant species to fungi isolated from remnant prairies relative to disturbed fungi. In a previous experiment, alfalfa and *Silphium* were shown to be highly responsive to AM fungi. This inoculum was used to test the differential responsiveness of the crop communities to the whole soil and prairie AM fungi inoculum. The prairie AM fungi inoculum consisted of seven AM fungi species with high spore abundance at the time of sampling: *Scutellospora dipurpurescens*, *Gigaspora gigantea*, *Funneliformis mosseae*, *Funneliformis geosporum*, *Glomus mortonii*, *Rhizophagus diaphanous*, and *Claroideoglomus claroideum*. Each species of AM fungi was cultured independently on native prairie plants for one growing season in a sterilized 50:50 sand:soil mixture (10.15 ppm P via Melich extraction, 7.375 ppm NO3-N, 22.2 ppm NH4-N via KCl extractions) under greenhouse conditions (see for a detailed description of isolation and culturing). A community mixture of these cultures was homogenized and used as our native AM fungi treatment ("LF" for living cultures). All biota from the live whole soil and live fungi were sterilized via autoclaving (2 X 60 minutes at 121°C) to create the sterile whole soil (SW) and sterile AM fungi (SF) treatments so that each pot had similar additions of whole soil and cultured fungal inoculum, whether living or dead. The sterilized SWSF inoculum was used to test the responsiveness of the crop communities in the absence of soil biota. ## Crop community Six crop communities were designed to test overyielding that included all possible combinations of monoculture and biculture plantings for the three perennial crop species, *Silphium integrifolium* (henceforth referred to as *Silphium* or S), *Medicago sativa* (henceforth referred to as alfalfa or A) and *Thinopyrum intermedium* (henceforth referred to as Kernza or K). Any combination of two letters represents a biculture (i.e. KA represents a Kernza/alfalfa biculture plant community). Kernza and *Silphium* seeds were obtained from The Land Institute’s breeding program, and The Land Institute granted permission for seed use. Alfalfa (Kansas Common variety) seeds were purchased from a commercial supplier. *Silphium* seeds were cold moist stratified two months prior to germination. Alfalfa was inoculated with commercially produced rhizobia (Exceed Superior Legume Inoculant for alfalfa/true clover, Visjon Biologics, Wichita Falls, TX, USA). Seeds of all crop species were germinated and grown for one week at the end of March in 2018 on a sterilized (2 X autoclaved as above) sand:soil mixture. We planted four conspecific seedlings (one week old) into each pot for monocultures, and two conspecific plants were planted diagonally from each other in each biculture. ## Water availability Pots were randomized via split block where half the block was well-watered, and the other half was given a drought treatment. All plants were well-watered for 18 days before drought treatments were applied by watering twice daily for two minutes (266.7 ml/day) via a drip irrigation emitter to prevent splashing of soil microbes. On day 19, drought pots were watered twice per day every other day for 1 minute (133.3 ml/day), while well-watered pots received no change in water volume for the duration of the experiment. The full experiment design included 7 replicates of each crop community, water regime, and inoculation combination (2 levels of water treatment x 6 levels of crop community x 4 levels of inoculum x 7 replicates = 336 pots). ## Data collection Crops were grown for 8 weeks, and then aboveground biomass was collected by cutting at 4 cm above the soil surface line, separated to species, dried at 60° C, and weighed. Crops were allowed to regrow an additional 5 weeks and a second harvest was performed. A second harvest was conducted to assess the context dependency of biotic and abiotic effects on crop regrowth, as aboveground biomass of perennial systems may be cut multiple times in one growing season. Ten plants out of 1344 (0.7%) died before the second harvest. These plants were recorded as 0.0 g at harvest 2. After the second harvest, root tissues were collected from a subset (4 blocks) of the monoculture pots to confirm AM fungal colonization. Root subsamples from each pot were cleared and stained with Trypan Blue. Hyphae and arbuscules were counted using the magnified intersections method. The results from the root analysis showed that the presence of AM fungal hyphae and arbuscules was greater in monoculture pots with live soil inoculum (LWLF, LWSF, and SWLF) than in pots inoculated with sterile whole soil and sterile prairie AM fungi (SWSF) (See supporting information for detailed results; ; – Figs). The mean hyphal and arbuscule presence for the sterile inoculum (SWSF) was close to zero. ## Statistical analyses ### Crop-specific responses To gain insight on the crop-specific responses and uncouple a three way interaction between the water, inoculum, and crop species nested within crop community treatments, a separate mixed model for each crop was analyzed with yield per individual as the response and block, water, crop, and inoculum set as fixed factors in SAS (proc mixed, SAS v9.4, SAS Institute, Cary, NC, USA). To account for the spatial separation of the watering treatment within each block, a block x water interaction and its’ higher order interactions were included as random effects. Yield per individual was natural log transformed to meet statistical assumptions. Tukey’s HSD multiple comparisons test was used to determine differences among groups within a significant treatment effect. Because results were similar for each harvest response, only total harvest responses are presented. ### The effects of water, crop community, and inoculum on overyielding For analysis of overyielding, we calculated the average individual yield of each crop species in each pot. We used this as a response in a mixed model (proc mixed) in SAS (version v.9.4, SAS institute, Cary, NC, USA) with pot designated as the subject. Average individual yield was natural log transformed to meet statistical assumptions. Block, water, crop community, crop species nested within crop community, and inoculum treatments were designated as fixed factors. The block x water x crop x inoculum x pot interaction was included as a random effect to account for multiple samples taken from the same pot (in mixtures), and the block x water interaction and its’ higher order interactions were included as random terms to account for the spatial separation of the watering treatment within each block. To test for overyielding and its abiotic and biotic context dependency, we designed four set of contrasts to compare monoculture vs mixture performance within each crop community, inoculum, and water combination. There were four sets of contrasts: 1) all possible combinations of mixtures versus component monocultures overall and for all three possible combinations of crop community designs (KA, KS, AS), 2) all possible combinations of the overall and crop species specific interaction of mixtures versus monocultures when grown among living soil (LWLF, LWSF, SWLF vs. SWSF), AM fungi (LWLF, SWLF vs LWSF, SWSF), or whole soil (LWLF, LWSF vs. SWLF, SWSF), 3) all possible combinations of the overall and crop species specific interaction of mixtures versus monocultures by water treatment, and 4) all possible combinations of the overall and crop species specific interaction of mixtures versus monocultures when grown among living soil, AM fungi, or whole soil by water treatment. We analyzed crop performance with data from the first and second harvest as well as the combined harvests for a total harvest. Results were similar for the first, second, and total harvest, so here we present total harvest results only. # Results ## Crop-specific responses to crop community, inoculum, and water Kernza growth was 25% better in mixture than monoculture (crop treatment main effect;), and the growth of Kernza was inhibited 26 – 30% by the presence of live whole soil and live prairie AM fungi relative to sterile soil (inoculum treatment main effects;). Increasing water availability increased Kernza growth by 50% (water treatment main effect;). Kernza performed the best when planted in mixture with sterile soil (SWSF) inoculum (crop x inoculum interaction; ;), and removal of prairie AM fungi (LWSF) only significantly decreased growth relative to removal of whole soil (SWLF) when Kernza was planted with alfalfa in wet pots (water x inoculum x crop interaction;) *Silphium* growth was reduced 11% in mixture with Kernza relative to being planted in monoculture (crop treatment main effect;), and alfalfa growth was reduced by 34% in mixture with Kernza in dry pots (water x crop interaction;). Alfalfa and *Silphium* had greater growth (increases ranging from 600 to 1000%) in the presence of whole soil (LWSF) and live prairie AM fungi (SWLF) or their combination (LWLF) relative to non-inoculated (SWSF) (inoculum main effects; ; ). Increasing water availability increased alfalfa growth by 163% and *Silphium* growth by 95%. Water availability also moderated inoculum effects on alfalfa and *Silphium* growth (water x inoculum interaction;). Inoculation with native prairie AM fungi (LWLF and SWLF) increased alfalfa growth most in wet pots ( vs ), and increased *Silphium* growth most in dry pots ( vs). ## Overyielding depends on crop pairs Crop mixtures overyielded relative to monocultures (P = 0.0004; contrast set 1; Overall;), but the level of overyielding depended significantly on the crop pairing (P \< 0.0001; contrast set 1; Mix vs mono x crop community;). Comparing each crop community (KA, KS, AS) individually to their respective monoculture components, we found significant crop community overyielding for KA (10.5% overyielding, P \<0.0001; contrast set 1; KA) and KS (8.9% overyielding, P = 0.0006; contrast set 1; KS). Crop communities of AS did not overyield (P = 0.9829; contrast set 1; AS). The presence of living inoculum (LWLF, LWSF, SWLF) vs sterile inoculum (SWSF) had significant effects on overyielding across crop communities (contrast set 2; Mix vs mono x live vs sterile (P = 0.0114)) and in crop specific mixtures (contrast set 2; Mix vs mono x live vs sterile KA (P = 0.0398), KS (P = 0.0004), and AS (P = 0.7513)). Living inoculum substantially reduced overyielding in mixtures of KA (61% reduction) and KS (86% reduction), but not in mixtures of AS. When looking at the effects of specific inoculum, AM fungi (LWLF, SWLF vs LWSF, SWSF) significantly lowered overyielding in mixtures of KS (87% reduction; P = 0.0067; contrast set 2; Mix vs mono KS x AM fungi), while no significant effects were found for whole soil inoculated pots (LWLF, LWSF vs SWLF, SWSF; contrast set 2; Mix vs mono x whole soil Overall, KA, KS, AS). We found no significant contrasts for any combinations of AS by inoculation treatment, but in general, A and S monocultures performed extremely poorly without living biota (SWSF;). Crop community and inoculum effects on overyielding were consistent across watering treatments (; contrast set 3 and 4), except for a marginal 11.6% increase in KA overyielding in wet pots relative to dry pots (P = 0.0583; contrast set 4; KA). # Discussion We found very strong effects of inoculation, watering regime, and plant diversity on crop productivity, and that these effects varied markedly across the crop species. Notably, we found that two crops, *Silphium* and alfalfa, were very responsive to the presence of arbuscular mycorrhizal (AM) fungi. Each of these species benefited more from native AM fungi, but this benefit depended on water availability. While native AM fungi was particularly beneficial to *Silphium* in drought conditions, native AM fungi benefited alfalfa most under well-watered conditions. In contrast, Kernza did not benefit from AM fungi and grew best in sterile soil. Crop mixtures that included Kernza overyielded, but this overyielding depended on the presence of soil biota (live vs sterile). Unlike previous work in annual and perennial systems, overyielding in our system was greatest in sterile soil, where Kernza compensated for the poor growth of *Silphium* or alfalfa. While water availability had large effects on total productivity, it had only marginal effects on overyielding. These results reinforce the importance of soil biota, especially AM fungi, in crop-specific performance and overyielding. ## Crop-specific responses to water and inoculum Both biotic and abiotic context strongly affected the growth of alfalfa and *Silphium*, especially the presence of AM fungi. While the whole soil inoculum increased growth, both species tended to perform better in the presence of native prairie AM fungi. Other studies have shown that AM fungi dependent plant species perform better with inoculations of AM fungi from locally adapted undisturbed systems. This is because AM fungal communities in agricultural systems tend to differ in composition and be less beneficial following anthropogenic manipulations such as crop tillage and chemical application. It should be noted that our whole soil inoculum contained all components of the soil community, including AM fungi, bacteria, nematodes, pathogens, etc. Thus, the reduced benefit found for alfalfa and *Silphium* for whole soil versus prairie AM fungi could be attributed to less beneficial AM fungi and/or the presence of these other soil biota inhibiting crop productivity. Regardless, this work suggests that native, locally adapted mycorrhizal amendments may boost the growth of mycorrhizal-dependent plant species in perennial agricultural plantings. Future work should isolate the effects AM fungal composition and the broader microbiome in promoting perennial cropping systems. Not only was the growth of our mycorrhizally sensitive crop species dependent on soil inoculum composition, but the response of alfalfa and *Silphium* to the presence of whole soil biota varied with water availability. We expected the presence of AM fungi to boost plant resistance to drought conditions. *Silphium* supported this pattern as it performed better with native AM fungi in water limited conditions. In contrast, native AM fungi increased alfalfa performance in wet conditions. This could be linked to native AM fungi enhancing facilitation of phosphorus uptake, and phosphorus being more limiting in well- watered conditions. However, the contrasting effects of water on AM fungal inoculation effects further highlights the importance of abiotic and biotic context dependency in polyculture systems. Apart from AM fungi effects depending on water, effects of pathogens present in whole soils may also vary with water since pathogens often proliferate under well-watered conditions. Based on this we expected susceptible plants to perform more poorly with whole soil biota under well-watered conditions. However, we did not observe growth inhibition due to pathogen accumulation under well-watered conditions in this study. Despite our whole soil inoculum being sourced from a long-term field trial of the Kernza progenitor, the lack of responsiveness to soil biota—positive or negative—may be attributed to Kernza being a mid-successional introduced cool- season grass. Mycorrhizal responsiveness tends to increase with plant successional stage, is stronger for native than non-native plant species, and C₃ grasses (cool-season) are less responsive than other plant functional groups. Thus, we anticipated that Kernza would not demonstrate strong mycorrhizal responses. Past work has shown a lack of or reduction in mycorrhizal responsiveness for introduced plant species, and this difference may grant novel crops an edge as they are introduced into new agricultural environments. Novel environments may also give introduced crop species an edge because they may also be less susceptible to species-specific pathogens because the pathogens may not have been co-introduced with the host. ## Crop-specific responses to abiotic and biotic conditions resulting in overyielding While previous studies have found evidence of pathogen-mediated overyielding in annual and perennial systems, we did not find support for this mechanism in this system. In retrospect, this might not be surprising, as our soil collection targeted potential pathogens of Kernza by using inocula from a mature Kernza field, but Kernza is a relatively newly introduced species in Kansas and newly introduced plant species often do not suffer negative effects of host-specific pathogens. Moreover, the soil from the Kernza field may not have abundant host- specific pathogens of *Silphium* or alfalfa. Both *Silphium* and alfalfa do suffer heavy losses from host-specific pests in the mid-western US, and it is possible that pathogen mediated overyielding could have been observed with a different initial soil inoculum. Moreover, overtime non-native plant species accumulate pathogens and therefore, as Kernza is planted more widely, intercropping may become important to managing pathogen accumulations and sustaining Kernza yield in the future. While individual studies have found evidence for mycorrhizally mediated overyielding, several studies have found less overyielding with AM fungi alone compared to whole soil. In our case, we did not find overyielding with AM fungi, but did find overyielding in sterile soil in mixtures that include Kernza, as Kernza compensated for the very poor growth of the AM-dependent *Silphium* and alfalfa. This compensation was largely independent of water treatment. This context dependence is not consistent with prior expectations of AM mediation of overyielding. Moreover, we did not see evidence of symbiotic N-fixation mediating overyielding in mixtures that include legumes. This is surprising given that it is a commonly invoked mechanism of microbially-mediated resource partitioning and facilitation. Longer experiments including those in the field may have generated greater N-limitation and more context for symbiotic N facilitation, as enhanced benefits of polycultures of Kernza and alfalfa may take as long as four years to develop. This study reinforces that soil biotic effects on perennial polycultures are context dependent and gives insight into interactions among specific perennial crops. Longer term field studies and studies that include potential host- specific beneficial and pathogenic microbes of all crops would enhance our understanding of overyielding in perennial systems, particularly because the relative importance of biotic and abiotic factors may change over time. Given the life cycle of perennial crops and our ultimate goals for sustainable production, long term monitoring is even more essential than in annual systems. ## Perennial polycultures as a model for future cropping systems Our study found consistent yield across the 6 different crop communities treatments, whether crops were grown in mixture or monoculture. While overyielding was only found in sterile soil conditions that are absent in the field, our work suggests that bi-culture plantings can result in similar field production yields as monocultures, while providing other beneficial ecosystem services. For instance, incorporating a companion crop such as alfalfa or *Silphium* can improve pollinator abundance, increase forage and habitat quality, or create a new revenue stream. Moreover, these consistent yields across planting were also present at each level of water availability. So although well-watered plants grew better than drought plants, we found that bi- cultures persisted and provided as much mass as monocultures when water was limited. The findings of our study should also be considered in new plantings when agricultural landscapes are converted from annual systems to perennial systems. Although we did not find strong evidence of overyielding due to biotic conditions, polycultures with the highest per capita yields tended to be inoculated with whole soil and native mycorrhizal amendments. These data highlight that choosing or manipulating the biotic conditions to meet the needs of plant species grown together can help achieve the greatest yields when planting of future polyculture perennial crops. # Supporting information We would like to thank Emily Cady, Maci Harford, and Laura Kemp for assistance with experiment setup, monitoring, and data collection. We also thank Katie Nus and KU greenhouse staff for help in maintaining the experiment. This project was funded by the Perennial Agricultural Project sponsored by the Malone Family Foundation Land Preservation (<http://themalonefamilyfoundation.org/index.html>) and the National Science Foundation (DEB‐1556664, DEB-1738041, OIA 1656006). [^1]: Authors have no conflict of interest

# Introduction Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has high human-to- human transmission capacity and is classified within the *Coronaviridae* family, specifically the *Betacoronavirus* genus. It includes severe acute respiratory syndrome coronavirus (SARS-CoV, identified in 2002) and Middle East respiratory syndrome coronavirus (MERS-CoV, identified in 2012). Comparative analysis has shown that the SARS-CoV-2 spike protein has more than 10-fold higher binding activity compared to SARS-CoV. These viruses infect bats and are transmitted to humans via zoonotic transmission. The first cases of coronavirus disease 2019 (COVID-19) were identified in Wuhan, Hubei Province, China, in December 2019, after which the infectious disease has become distributed globally. Currently, variants of SARS-CoV-2, the causative agent of COVID-19, continue to proliferate rapidly. Although China attempted to reduce the distribution of SARS-CoV-2 within the country, the infection quickly was spread in parts of Hubei Province neighboring Wuhan. Later, SARS-CoV-2 has become distributed worldwide, representing a serious global issue; in particular, Europe and Latin America have suffered more than other countries in the beginning of coronavirus pandemics. The World Health Organization (WHO) officially declared a pandemic on March 11, 2020. Overall, 161,513,458 confirmed cases of COVID-19 with 3,209,109 deaths, including 92,320 confirmed cases and 655 deaths in Uzbekistan, have been reported to the WHO as of May 4, 2021. Microscopic vision of SARS-CoV-2 is a spherical, enveloped particle, positive- sense, single stranded RNA. Its genome contains 29,903 nucleotides. The SARS- CoV-2 genome contains the open reading frame (ORF) proteins, spike (S), envelope (E), membrane (M), and nucleocapsid (N) genes in a 5’-3’ orientation. The replication ORF1ab gene of SARS-CoV-2 is the longest among other genes, 21,291 nt in length, and contains 16 predicted non-structural proteins, followed by (at least) 13 downstream ORFs. Moreover, it shares a highly conserved domain (amino acids 122–130) in nsp1 with SARS-CoV-2. The other genes, such as the S, ORF3a, E, M, and N genes of SARS-CoV-2, are 3,822, 828, 228, 669, and 1,260 nt in length, respectively. The SARS-CoV-2 S protein binds to the host receptor, enabling entry into the cell. The gene encoding this protein is an important part of the viral genome, along with high sequence variability as a key point for new mutations. Full and partial genome sequences of SARS-CoV-2 obtained from different countries and laboratories are now available at the Global Initiative on Sharing All Influenza Data (GISAID), the National Center for Biotechnology Information (NCBI) and the Virus Pathogen Database and Analysis Resource (ViPR). Notably, SARS-CoV-2 variants have been studied by analyzing 48,635 complete SARS-CoV-2 genomes obtained from GISAID. Additionally, SARS-CoV-2 mutations were annotated by comparison with the sequence of the Wuhan-Hu-1 SARS-CoV-2 isolate, a reference genome NC_045512.2, and an average of seven mutations per sample was detected. The analysis has shown that major mutational types worldwide are caused by single-nucleotide transitions. To date, at least three clades have been characterized based on geographic and genomic specificity of SARS-CoV-2 variants. The first is the V clade, defined in the GISAID EpiCoV portal, with limited variability. This clade members were isolated in Asia, and their genomes are similar to the reference coronavirus genome NC_045512.2. The second is the G clade (branching into its two offspring subclades, GH and GR), whose members are scattered across Europe and on the East Coast of the USA. This clade members carry a D614G amino acid mutation in the S protein and four characteristic point mutations: C241T, C3037T, C14408T, and A23403G. The third is the S clade, whose members have mostly been isolated in the USA, especially on the West Coast, and is characterized by C3037T and T28144C mutations. The global concern related to novel virus variants is based on their rapid mutation rates, helping viral evasion of immunity formed by approved vaccines or previous variant(s) and cause reinfection. Among all mutations in the SARS-CoV-2 genome, S gene mutations are important because the spike protein defines the viral host range and is often the target of neutralizing antibodies. For example, the D614G substitution in the S protein occurred due to the A23403G missense mutation resulting in an amino acid change from aspartate to a glycine residue at position 614 (D614G). This mutation has emerged as a predominant G clade in Europe (954 of 1,449 (66%) sequences). Since the beginning of the pandemics, the coronavirus variant carrying the G614 mutation has been rapidly distributed across the world and has become the dominant global strain. Even in local epidemics, the shift has occurred by converting the D614 to G614 variant. These changes were highly statistically significant, and the G614 variant may have a fitness and selective advantage. On the other hand, if the mutation occurred early in the growing population, a genetic drift could also help to increase the frequency of a specific variant without selective advantage. Moreover, other mutations (N439K in RBD) were observed in the S protein. The major frequent mutations, such as D614G, N439K, and S477N, in the S protein are causing the increased viral transmissibility. A variant containing D614G-associated mutations in the RBD became more infective and increased resistance to some neutralizing antibodies, with obvious implications for the recovery of COVID-19 patients. SARS-CoV-2 cases were first identified on March 15, 2020, in passengers returning from Europe to the capital city Tashkent, Republic of Uzbekistan. The government of Uzbekistan has begun a massive testing of people who had symptoms of COVID-19 infectious disease. This event was the beginning of the epidemic in Uzbekistan that required effective diagnostics and monitoring of rapidly spreading coronavirus genotypes among population. Therefore, with the main goal of identifying virus genotypes distributed in our territory as well as studying genomic diversity, types of mutations and possible emergence of new variants of SARS-CoV-2, we initiated the whole genome sequencing of COVID-19 samples. Here, we present the first whole-genome sequence data from COVID-19-infected patients in the Republic of Uzbekistan. We successfully assembled the 18 high-quality sample genome sequences for coronavirus genotypes and profiled 128 mutations with nonsynonymous and synonymous types. Comparative analysis using globally known genomes of SARS- CoV-2 grouped Uzbekistan sample genomes into two major clades of S and GR in the global phylogenetic tree. The sequenced genomic data of coronavirus genotypes, described herein, should be useful in fighting against coronavirus threats in Uzbekistan. # Materials and methods ## Sample collection In the middle of October and beginning of December of 2020, samples were collected several times from one hundred symptomatic patients with high temperature and occasional cough using nasopharyngeal and oropharyngeal swabs sticks (Huachenyang Technology, Shenzhen, China) and immediately placed in viral transport medium. Patients with possible COIVID-19 infection were from the diagnostics laboratories of the Tashkent Region Epidemiological Centre and the private BiogenMed COVID-19 testing center, Tashkent, Republic of Uzbekistan. Biological samples were collected randomly from PCR-positive patients after laboratory testing for SARS-CoV-2. The research study has been approved by the Ethics Committee under the Ministry of Health of the Republic of Uzbekistan (#6/20-1582). All the experiments were carried out in accordance with the relevant guidelines and regulations. Samples were renumbered and de-identified so no one, even researchers could know the identity of the patients. For the reporting purpose only anonymous data including age and biological sex were kept. We received a verbal consent for a voluntary participation from all patients involved for sample collection. Fall participants we have explained the use of collected samples for a sequencing experiment only without disclosing their personal identity or disturbing them in future. Verbal consent was preferred to written consent in this study because patients were not in a mood of signing any written document because of worriedness about COVID-19 infection at that time and had a hesitation because of not full-understanding of genome sequencing experiment. No minors were involved for the sample collection. Because of non- invasiveness of sequencing experiment, non-involvement of participants to any further downstream clinical procedures as well as anonymity of the personal identity of samples in this study, there was no requirement to obtain a consent document approval by the Ethics Committee. Total RNA was extracted from the collected samples using a MagMax Viral/Pathogen kit (Life technologies Corporation, Austin, Texas, USA). Each sample was tested by PCR systems (threshold– 0,050) to the presence of SARS-CoV-2 using a CFX Connect Real-Time PCR System (BioRad, Hercules, California, USA). RT-PCR were implemented in 40 μl of final volumes. PCR amplification was carried out by the following steps: for cDNA synthesis at 35°C for 20 min, a first denaturation at 95°C for 5 min followed by 50 cycles of 94°C for 15 sec and 64°C for 20 sec (set fluorescence measurement for Fam, Hex, Rox, and Cy5 channels at 64°С) and a SARS-CoV-2/SARS CoV multiplex real-time PCR assay targeting the nucleocapsid (N), envelope (E) and region of SARS-CoV-like viruses (DNA Technology, Moscow, Russia). Among all tested patients, 32 PCR-positive samples (18 females and 14 males) were selected for further studies, which were randomly selected. ## SARS-CoV-2 sequencing Complementary DNA (cDNA) was synthesized from 5 μl of RNA sample using a SuperScript VILO with DNAse cDNA Synthesis Kit (Life Technologies, Carlsbad, California, USA) and a ProFlex^TM Base (Life Technologies holding Pte Ltd, Mapletree, Singapore). Libraries were constructed manually using the Ion AmpliSeq SARS-CoV-2 Research Panels, Ion Xpress Barcodes, and an Ion AmpliSeq Library Kit Plus Life Technologies Corporation, Frederick, Maryland, USA) following the manufacturer’s recommendations; the process included using amplification cycles from 12- to 24,000-plex in a single well based on viral load. Template amplification and enrichment as part of the manual workflow for the Ion S5 systems were performed with the Ion OneTouch 2 system using an Ion 540 Kit (Life Technologies, Carlsbad, California, USA). The thirty-two samples were multiplexed on an Ion 540 chip and sequenced using an Ion GeneStudio S5 Semiconductor Sequencer (Life Technologies Holdings Pte Ltd., Singapore). ## Sequenced data analysis Sequenced reads were aligned with the Wuhan-Hu-1 Reference Genome (NC_045512.2) on the Torrent Suite v. 5.12.2 (Life Technologies, Carlsbad, California, USA). Plugins were used as follows by order coverage analysis (v5.12.0.0) and Variant Caller (v.5.12.0.4), both with default parameters and COVID19AnnotateSnpEff (v1.3.0.2; Life Technologies, Carlsbad, California, USA). To predict the effect of a base substitution, a plugin specifically developed for SARS-CoV-2. VCF files generated by Torrent Suite (Life Technologies, Carlsbad, California, USA). Variant caller was filtered to remove variants with read depths less than 1000 and ion torrent quality scores less than 400 to keep reliable variants only. The filtered variants were used for sample clustering with Maximum Likelihood Tree in Molecular Evolutionary Genetics Analysis (MEGA, <https://www.megasoftware.net>) software. The consensus for each SARS-CoV-2 genome sequence were then submitted to the GISAID under the accession numbers of EPI_ISL_1402423 to EPI_ISL_1477049 (available for registered users) and NCBI under the accession numbers GI:2021275696 to GI:2021275839 (or MW853559.1 to MW853569.1) databases. # Results ## Sample selection for sequencing One hundred symptomatic patients with high temperature and occasional cough were collected for this study at the Tashkent Region Epidemiological Centre and a private COVID-19 testing center. Among them, 32 PCR-positive samples were selected for sequencing. There were seven men and eleven women with an average age of 47. Out of 32 SARS-CoV-2 samples sequenced, 14 samples were excluded from further analysis due to the low quality of sequencing coverage, which resulted in several gaps in the consensus sequence. The remaining 18 samples (seven men and eleven women) were selected with an average of mapped reads per sample of 3.9 million, with an average mean read depth of 23,203. However, the average uniformity of coverage in selected samples was low (75.31%;). ## Analyzing the most reliable mutations among all sequenced samples The analyzed mutations in these selected samples were generated by Variant Caller. According to the results, a number of mutations (Figs and) varied from five (samples 1,2 and 5) to 20 (samples 25 and 37). Most viral genomes contained between 8 and 15 mutations when compared with the NC_045512.2 reference genome, with mutations in sample 11 (15 mutations), sample 13 (14 mutations) and sample 10 and 15 (11 mutations), sample 8 (8 mutations) and sample 14 (9 mutations). The most common nucleotide substitution detected was from cytosine to thymine (52/128 mutations), followed by guanine to thymine (36/128 mutations) and thymine to cytosine (10/128 mutations). All the shared mutations were homozygous. We observed one unique frameshift mutation (21574; c.13delC), seven shared mutations (G21850T, G22335T, A23403G, G23438T, G23593T, A24078G and G24410C), five unique (T22020C, T22478C, G22484T, C23634T and G24872T) missense mutations, two shared (G21724A and T25219G) and five unique (A23503T, C23758T, C24023T, G24199T and C24442T) synonymous mutations and five unique (A29676G, G29692T, C29708T and C29733T) with large deletions (`4373delTCACCGAGGCCACGCGGAGTACGATCGAG`) and one shared (G29742A) downstream region mutations in the gene encoding the S protein. One missense mutation was identified in the E (envelope) region; two synonymous mutations were found in M (matrix), and one synonymous and eleven missense mutations were found in N (nucleocapsid). Furthermore, 29 missense mutations, 36 synonymous mutations and three upstream gene mutations were found in the ORF1ab region. One synonymous mutation and seven missense mutations were detected in the ORF3a region. The ORF6 region showed one upstream mutation while the ORF7a region exhibited two missense and two synonymous mutations. Finally, five missense mutations and one synonymous mutation were found in the ORF8a region. Overall, we identified a total of 128 mutations, consisting of 45 shared and 83 unique mutations representing one unique frameshift mutation, four upstream region mutation, six downstream region mutation, 50 synonymous mutations, and 67 missense mutations. ## The phylogenetic tree was drawn based on viral sequences We aimed to analyze the major mutations in all sequences to determine the difference between our cases and those worldwide. For this, a phylogenetic tree was generated in MEGA X based on the 18 viral sequences using the maximum likelihood method. One hundred twenty-eight mutations were obtained from eighteen SARS-CoV-2 viral genome sequences in samples from COVID-19 patients by Variant caller (v.5.12.0.4). The sequences have clustered into two large clades, including four sequences with five shared mutations that corresponded to the S clade (C8782T, G11230T, T28144C, G28167A and G28878A) and 14 samples with four mutations grouped in the GR subclade (C241T, C3037T, C14408T and A23403G). Two serial CA-TC substitutions (C28253T and A28254C) in some of the S clade viruses were identified (samples 25 and 27), and three serial GGG-AAC substitutions (G28881A, G28882A and G28883C) were identified in GR subclade viruses. In the GR subclade, two out-grouped samples (samples 2 and 5) and smaller clusters that corresponded to cluster 2 (samples 1 and 7) and cluster 3 (samples 4, 8, 10, 11, 12, 13, 14, 15, 17, and 32) were also found. In the S clade viruses, we observed one small cluster that corresponded to cluster 1 (samples 3, 6, 25 and 27). Global phylogenetic tree was produced using local COVID-19 patient-derived sequences and the reference genome NC_045512.2, using [nextstrain.org](nextstrain.org) website. According to this analysis, the viruses in samples 3, 6, 25 and 27 belong to the S clade of SARS-CoV-2 and have sequences very similar to that of the reference genome, even though they harbor more substitutions in each gene, particularly in the S region. These variants also have grouped with African and Near East variants in the global phylogenetic tree provided by [nextstrain.org](nextstrain.org). The remaining 14 sequences belong to the G clade, which possibly originated in Europe and North America. # Discussion Our results revealed that whole-genome sequences of isolates obtained from the 18 symptomatic COVID-19 patients represent important nucleotide diversity (Tables and –). These 18 sequenced samples have grouped in two major clades of SARS-CoV-2 on the public database of the GISAID named clade G (or GR subclade) and clade S according to the similarity of mutation signatures. The G clade, whose members are scattered across Europe and on the East Coast of the USA and carry a D614G amino acid mutation in the S protein and four characteristic point mutations namely C241T, C3037T, C14408T, and A23403G. Among the 18 sequenced samples, fourteen samples have grouped into subclade GR. This branch contained an important deleterious trinucleotide mutation GGG28881AAC in the N gene of the coronavirus, inducing an ArgGly203LysArg change. This mutation later was found in many subsequent studies, with a frequency in deceased (39.45%) and recovered patients (31.38%), making them globally dominant in the coronavirus genomes. These 14 samples also contained well defined C28311T mutation of proline to leucine amino acid change in the N gene, playing a key role in the formation of replication–transcription complexes In addition, twenty shared mutations in these fourteen samples (G61T, C920T, C1878T, C2536T, G3340T, C3373A, C5654T, C6883T, C9118T, T10150C, C17135T, C19220T, G21724A, G23438T, G23593T, T25219G, G25912T, G27703A, G28237A, G28300T) were new and have not been described in any other studies before. The remaining four samples out of 18 sequenced have grouped into the S clade and contained S clade- specific mutations such as C8782T and T28144C, including Leucine (L) to Serine (S) amino acid change (L84S). These mutations were found in many studies. In addition, G11230T, G28167A, and G28878A mutations found in all four samples of our S clade were similar to Lui et al., and were found in United Emirate Arabic. The G29742A mutation, shared by three samples out of four S clade samples (except Sample 3) was also found in United Emirate Arabic. Other shared mutations found in four samples (C5812T, C8950T, C9443T, G15543T, G18624T, C19017T, C19602T, G20580T, G21850T, G22335T, A24078G, G24410C, 28253\*, C28311T) were new and have not mentioned by any other studies. Moreover, we found a number of unique mutations in each sample that were specific for our SARS-CoV-2 genome sequence data. Sequencing of the virus genome is fundamentally important for identifying SARS- CoV-2 strains and investigating local and global spread. In addition, the full- genome sequence of any virus that causes the infection can be helpful for investigating outbreak dynamics, such as changes in the size of the epidemic over time as well as spatiotemporal spread and transmission routes (WHO COVID report 2021). This is the first step attempt to sequence the full genome of SARS-CoV-2 from COVID-19-positive patients of the Republic of Uzbekistan. The first identified COVID-19 cases in Tashkent are believed to be of foreign origin due to international travel. Indeed, the results of this study showed that many of the infections originated from European (GR subclade) and Near East (S clade) countries. The cause of spreads was the result of international travelling. The mutations between samples showed how virus structure changed itself over time when conditions became different. The genomic sequence data generated from these 18 samples, submitted to global databases such as NCBI and GISAID, should be helpful for public health and research organizations to observe the dynamics of disease spread in the region. Moreover, results will assist in the design of diagnostic assays, medications and vaccines. In this context, several institutions, including our centre, have already started working on a national vaccine based on these SARS-CoV-2 genome sequences found in symptomatic Uzbek patients. Furthermore, the sequence data presented herein should add new sequence and mutational profile data from our region to a COVID sequence database (GISAID), useful for future molecular epidemiology and evolutionary phylogenetic studies by health and research organizations. In particular, our whole-genome sequence data, reported herein, should be helpful for tracking the origin and source of the currently spreading SARS-CoV-2 variants and for identifying and comparing the emerging new variants in Uzbekistan and beyond. Here, we only provided the first effort of SARS-CoV-2 genome sequencing obtained from infected Uzbek patients using samples collected at the end of 2020. They were the early phase samples of the coronavirus disease pandemic that were spread to the whole country by that time. To date, the mutations multiplied by human-to-human transmission, and new strains have been scattered again in the country through international travels. The new mutations and variants require further sequencing efforts and analysis, which is in progress. # Supporting information We thank the Sanitary-Epidemiological and Public Health Department of Tashkent Region, Ministry of Health of Uzbekistan, and the private clinic of BiogenMed, Tashkent, Uzbekistan, including but not limited to Mr. Abdukhakim M. Sotvoldiev, Mrs. Ra’no M. Abidova, Mrs. Larisa E. Alieva, Mr. Botir B. Sattorov and Ms. Saule A. Karimova, for their help in collection of the samples from symptomatic patients. [^1]: The authors have declared that no competing interests exist.

# Introduction People with end-stage kidney disease (ESKD) on chronic dialysis experience high rates of hospitalisation, including the need for surgery, and have been reported to have longer hospital stays and higher in-hospital mortality compared to patients with normal kidney function. It is unclear whether these heightened risks of postoperative complications are attributable to ESKD per se or the fact this group tend to be advanced in age and have a greater comorbid burden. In addition, published reports of postoperative outcomes in dialysis patients are limited by small studies, inconsistent adjustment for confounding factors, and lack of accounting for type and urgency of surgery. As a result, chronic dialysis patients are frequently deemed to be ‘high risk surgical candidates’; potentially affecting their access to surgical care. Given that the number of people worldwide with ESKD requiring chronic dialysis is projected to double to 5.4 million by 2030, an increasing number of older, multimorbid chronic dialysis patients will contemplate elective surgery in the foreseeable future. Therefore, ascertaining the excess postoperative risk for chronic dialysis patients would assist in planning elective surgery, mitigating risk and inform shared decision making between patients and clinicians in relation to potential benefits and harms of surgery. The purposes of this study were to estimate the excess odds of fatal postoperative outcomes in dialysis-dependent patients compared to patients with normal kidney function, and to examine the influence of comorbidities on mortality. # Methods This systematic review adhered to the recommendations of the Meta-analysis Of Observational Studies in Epidemiology (MOOSE) and the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) checklists, with a protocol registered in PROSPERO (CRD42017076565). ## Study eligibility criteria All cohort studies that measured and reported postoperative mortality in adult (aged 18 years or older) chronic dialysis patients (haemodialysis and peritoneal dialysis) and in a control group of patients who had normal kidney function, as defined by serum creatinine less than 110μmol/l or absence of International Classification of Disease Coding (ICD) coding of chronic kidney disease and ESKD, were considered for inclusion. Non-dialysis dependant chronic kidney disease patients were excluded. All types of surgery requiring a general anaesthetic were considered, including general, orthopaedic, cardiac, vascular and urology/gynaecological surgery. Kidney transplant surgery and haemodialysis \[HD\]/peritoneal dialysis \[PD\] access surgery were excluded, as the former is performed in highly selected patients following rigorous cardiovascular evaluation, and the latter is considered minimally invasive surgery. More so, patients with normal kidney function would be ineligible for these surgeries. Since many studies reported data over a time period of several years, the data were assigned as close as possible to the median year in which the patients were recruited. Studies in which more than 20% of the procedures were emergent (defined as an acute illness leading to an emergency presentation or an unplanned admission requiring a surgical procedure) were excluded because emergent procedures have an inherently higher risk of perioperative complications. Studies reporting outcomes involving patient receiving continuous renal replacement therapy (CRRT) for acute kidney injury in the perioperative period were also excluded. ## Data sources and searches MEDLINE, Embase and the Cochrane Controlled Register of trials (CENTRAL) were searched from inception to January 10th 2020, without language restriction using a combination of relevant keywords including *surgery*, *dialysis*, *postoperative*, *perioperative mortality* and their variants (and Figs). Exploded MeSH terms for perioperative medicine and chronic dialysis patients were also used. Search terms were modified to correspond to the tree structure and descriptors of the two databases. Full-text articles obtained were hand searched for further references. Tangential electronic exploration using links to related texts was also performed for additional materials. Case-control studies, animal studies, opinion papers, case reports and editorials were excluded. Database of Abstracts of Reviews of Effects (DARE), the Cochrane Database of Systematic Reviews (CDSR), National Institute for Health and Clinical Excellence (NICE) and the NIHR Health Technology Assessment (NIHR HTA) programme websites were all searched for existing reviews. ## Data extraction and quality assessment Two authors (D.P. and A.N.) independently reviewed all titles and abstracts identified in the initial search to assess study eligibility, and any disagreements were resolved by a third reviewer (M.F.). Type of surgery, numbers of patient on chronic dialysis and patients with normal kidney function, summary statistics for baseline characteristics (including cardiovascular disease, peripheral vascular disease, diabetes mellitus, hypertension and smoking status), and frequency of postoperative outcomes in each group were extracted from full-text manuscripts of eligible studies using an electronic data extraction form. Additional data from corresponding authors was requested when required. The primary outcome was all-cause mortality, defined as either 30-day mortality or death within the same hospitalisation as the index surgery. The methodological quality of each study was assessed in duplicate by 2 investigators (DP and AN) using the Newcastle-Ottawa Scale (NOS), which employs a star system to evaluate the selection of the study groups (0–4 stars), comparability of the groups (0–2 stars), and ascertainment of the outcome of interest (0–3 stars). Six reviewers (D.P., E.P, J.C., D.J., C.M., and M.F.) discussed overall quality of evidence and graded it on the basis outlined by Grading of Recommendations Assessment, Development and Evaluation (GRADE) working group. ## Data synthesis and analysis Comparative effect sizes were expressed as either relative risks, hazard ratios or odds ratios \[OR\] in the original articles. Unadjusted mortality ORs and 95% CIs were calculated using the number of events in each group. If no events occurred in one of the groups, then 0.5 was added to all cells in the 2 x 2 table to avoid computational difficulties with both the inverse variance and Mantel-Haenszel methods. In addition, if studies performed a multivariable analysis adjusted for age as a minimum, then the reported adjusted OR and 95% CI were recorded. Random effects meta-analysis was used to estimate the mean weighted effect within the different surgery types. Heterogeneity was assessed using both *I*^*2* to measure the percentage of variation in odds ratio estimate due to unobserved differences across studies. *I*^*2* tends to 100% as the number of patients included in a meta-analysis increases, and hence Tau² is also provided to reflect absolute variation in odds estimates. Unadjusted and adjusted ORs for the primary outcome in dialysis versus patients with normal kidney function were pooled by surgical discipline. A pooled odds estimate across surgical disciplines was not calculated as postoperative risk is inherently related to the type of surgical procedure such that pooling across disciplines was not clinically relevant. The influence of comorbidities on the odds ratio estimate of postoperative mortality was assessed by a series of weighted univariable and multivariable random-effects meta- regression analyses. These analyses included two categories of predictor variables: study characteristics, including overall study quality (as per NOS), single versus multicentre cohorts, study continent (United States vs all other countries), median year of recruitment, surgery duration, single procedure studies versus composite procedures, and patient characteristics, including age and prevalence of diabetes mellitus or ischaemic heart disease among dialysis patients compared to patients with normal kidney function. The R² index was used to quantify the proportion of variance in odds ratios explained by the covariates. L’Abbé plots were generated to identify studies with divergent results as well as the study groups that were responsible for such differences. Sensitivity analysis using meta-influence analysis was also performed to evaluate the influence of each study on the overall meta-analysis summary estimate and identify outlier studies that may have affected the validity of the conclusions. Inter-rater reliability of study selection was assessed using Cohen’s kappa. A funnel plot and Egger’s test for funnel plot asymmetry were used to assess publication bias. Statistical analysis was performed with Stata 14.0 for Windows. Statistical significance was defined as a two-sided p-value \<0.05. # Results ## Description of included studies In total, 5135 abstracts were reviewed, from which 115 full-text articles were retrieved and evaluated (See). 49 studies, involving 10,476,321 patients with normal kidney function and 41,822 chronic dialysis patients, satisfied the inclusion criteria. lists the characteristics and designs of the 49 studies. The definition of chronic dialysis varied across studies, with 22 studies using registry-based definitions, seven using ICD coding and the remaining studies confirming chronic dialysis status by medical chart reviews. Twenty-one studies defined normal kidney function by serum creatinine. Non-emergent cardiac surgery was the most commonly reported type of surgery (15 studies, 31%), followed by general surgery (12 studies, 25%), vascular surgery (9 studies, 18%), orthopaedic surgery (9 studies,18%) and urologic/gynaecologic surgery (4 studies, 8%)(62–65). Twenty-two of the 49 studies assessed a single surgical procedure, while the remaining 27 studies examined a combination of discipline- specific surgical interventions. Thirty six studies did not report dialysis modality, eleven studies specifically examined haemodialysis patients only, and two studies reported on outcomes for patients on peritoneal dialysis separately. Only four studies recorded the cause of death. Of the 49 studies, 19 reported findings from a single centre and only three collected data prospectively. Twenty-seven studies extracted information from existing data registries while the remaining extracted information from re- examined health records. Thirty-four studies were reported from North America, \[,,, –,,,, –\] 11 from Asia and 4 from Europe. Thirty four studies were published after 2010. \[, ,,,,, –\] All 49 studies reported age and gender, but comorbidities were less consistently described, with 37 (76%) studies reporting the prevalence of diabetes mellitus, 31 (63%) reporting ischaemic heart disease(IHD), 21 (43%) reporting smoking status, and 15 studies (31%) reporting all three comorbidities. ## Meta-analyses: All-cause mortality The incidence of in-hospital and/or 30-day mortality ranged from 0 to 19.3% in dialysis patients and from 0 to 7.7% in patients with normal kidney function, depending on surgical discipline and study. All studies consistently demonstrated an increased odds ratio of post-operative mortality in dialysis patients compared to patients with normal kidney function regardless of surgical discipline with the OR ranging from 3.96 (95%CI 3.23–4.87) to 10.76 (95%CI 7.30–15.86). The highest reported absolute median mortality rate was for chronic dialysis patients undergoing cardiac surgery (8.7%), followed by vascular surgery (7.8%). The largest odds ratio for postoperative mortality was observed for orthopaedic surgery (9 studies, 8014 dialysis patients, OR 10.76, 95% CI 7.30–15.86, I² 77.1%, p for heterogeneity \<0.001, Tau² = 0.04, low certainty evidence), followed by general surgery (12 studies, 13 798 dialysis patients, OR 6.67, 95% CI 4.11–10.83, I² 90.3%, p for heterogeneity \<0.001, Tau² = 0.56, low certainty evidence) and cardiac surgery (15 studies, 11 557 dialysis patients, OR 4.23, 95% CI 3.21–5.56, I² 81.9, p for heterogeneity \<0.001, Tau² = 0.14, low certainty evidence). The lower bound of the 95%CI was greater than 1.0 in 73% of included studies. Vascular surgery carried the lowest odds of postoperative mortality for chronic dialysis patients (9 studies, 7010 dialysis patients, OR 3.96, 95% CI 3.23–4.87 I² = 68.1%, p for heterogeneity 0.003, Tau² = 0.05, low certainty evidence). Twenty-three studies examined risk of postoperative mortality after adjusting for covariates, with age as a minimum. Subgroup meta-analysis from the 23 studies showed that dialysis patients remained at higher risk of postoperative death compared to patients with normal kidney function (OR ranged from 2.48 to 4.88), but lower than the relative risk observed in studies reporting unadjusted analyses. Except for a single study by Barbas et al, which observed no significant differences in postoperative mortality odds between dialysis patients and patients with normal kidney function following pancreatic procedures (OR 0.97, 95% CI 0.26–3.60, p = 0.99, low certainty evidence), the ORs for death in all other individual studies showed a significantly greater mortality odds for dialysis patients compared to those with normal kidney function. Pooled adjusted ORs by surgical discipline showed that orthopaedic surgery (4 studies, 5485 patients, OR 4.88, 95% CI 3.35–6.40, I² 53.5%, p for heterogeneity = 0.091, Tau² = 0.09, low certainty evidence) retained the highest mortality odds ratio for dialysis patients, followed by vascular surgery (4 studies, 4692 patients, OR 3.74, 95% CI 3.21–4.28, I² 53.0%, p for heterogeneity = 0.094, Tau² = 0.02, low certainty evidence) and cardiac surgery (7 studies, 9877 patients, OR 3.54, 95% CI 3.23–3.85, I² 62.4%, p for heterogeneity = 0.014, Tau² = 0.10, low certainty evidence). Cause of death was poorly reported, thus formal analyses to evaluate cause- specific mortality were not performed. ## Explaining variation in relative risk estimates A series of weighted univariable and multivariable random-effects meta- regression analyses were performed using study characteristics, including overall study quality (as per NOS), single versus multicentre cohorts, study continent, median year of recruitment, single procedure studies versus composite procedures, and cardiac versus non-cardiac surgery, none of which explained the variation in observed effect estimates. However, meta-regression with patient characteristics as predictor variables, including weighted mean age, prevalence of diabetes mellitus and ischaemic heart disease, did explain some of the variation in the observed mortality odds ratios. Univariable meta-regression of postoperative mortality odds ratio with weighted mean age of each study as a predictor variable showed that the logarithmic odds of postoperative mortality had an inverse linear relationship with mean age (slope -0.04 95%CI -0.06 – -0.01; p = 0.018;). A similar inverse relationship was found for diabetes (slope -0.02 95% CI -0.03–0.01; p = 0.022) but not ischaemic heart disease (slope -0.01 95% CI -0.01–0.00; p = 0.156, see, respectively). However, in multivariable meta-regression with all 3 factors, only diabetes mellitus was significant (slope -0.02 95% CI -0.03–0.00; p = 0.045). ## Risk of bias and certainty of evidence As per the Newcastle-Ottawa Scale, reporting of outcomes was of good quality, but comparability of patient groups on the basis of analysis was poor in 26 (53%) studies due to the absence of multivariable adjustment for patient demographics and co-morbidities. \[–, ,, \] L’Abbe plots and meta-influence analysis did not identify studies with outlying results or patient groups that may have been responsible for differences in estimates of relative mortality risk ( and Figs). There was no significant evidence of publication bias, as determined by funnel plot and Egger’s test (p = 0.328). Inter-rater variability between the two independent reviewers was strong (κ = 0.81). The certainty in the quality of evidence used to estimate postoperative mortality odds was deemed to be low. The quality of evidence was downgraded due to serious concern with risk of bias and inconsistency in the magnitude of odds ratio estimates. The large and consistent direction of risk estimates improved the strength of evidence. # Discussion This systematic review and meta-analysis demonstrated an increase in postoperative mortality odds in patients with ESKD requiring dialysis compared to patients with normal kidney function following all types of elective surgery (OR range 4.0 (95%CI 3.2–4.9)– 10.8 (95%CI 7.3–15.9)). Sensitivity analyses including studies only reporting multivariable adjusted risk estimates reduced the odds (OR range 2.5 (95%CI 2.1–2.8)– 4.9 (95%CI 3.4–6.4)), highlighting the importance of judicious interpretation of odds ratios where comparator patients are not matched for important demographic and comorbid characteristics. Furthermore, our meta-regression analyses demonstrated that the excess odds for postoperative mortality attributable to receiving chronic dialysis was attenuated by increasing patient age and prevalence of diabetes, which are themselves established independent risk factors for post-operative mortality. Having said that, a number of non-traditional risk factors associated with end- stage kidney disease including accelerated vascular calcification, mineral bone disease, anaemia, increased oxidative stress and impaired immunity, are all likely contributing factors. Another potential reason for the observed elevated mortality odds, is the definition of dialysis dependent ESKD patients used in the studies. Major databases from which a number of studies were undertaken, including the American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP), Vascular Quality Improvement Program and The Society of Thoracic Surgeons Database (STS), do not record the presence of ESKD per se but rather the requirement of dialysis at the time of surgery to indicate the presence of ESKD without distinguishing between an acute kidney injury requiring dialysis and a patient with ESKD. Patients with severe acute kidney injury requiring dialysis tend to be sicker and have a substantially increased mortality odds at baseline, and therefore the inclusion of these patients may have potentially exaggerate the findings. Nevertheless, patients considering kidney transplant surgery undergo rigorous cardiovascular assessment to identify occult coronary artery disease. However, no such recommendations exist to guide clinicians when contemplating elective surgery. In addition, studies comparing patient outcomes of surgical treatment to those of continued medical management are lacking. In a prior meta-analysis of 31 cohort studies involving 125 930 patients with normal kidney function and 27 955 with non-dialysis-requiring chronic kidney disease, kidney dysfunction was identified as an important risk factor for post- operative death (OR 2.8; 95% CI 2.1–3.7), similar to that observed for diabetes, stroke and coronary artery disease. Subsequent meta-regression demonstrated a graded relationship between declining glomerular filtration rate and post- operative mortality. The results of our meta-analysis with meta-regression support these findings and extend them by demonstrating that the odds of post- operative mortality remain elevated when patients transition to dialysis. Studies did not report cause of death thereby precluding further exploration of mechanisms of heightened risk. Indeed, adverse effects associated with ESKD and dialysis treatment itself on the cardiovascular and immune system may be important factors in the causal pathway. ## Strengths and weaknesses A comprehensive search strategy was used to identify published studies, ultimately pooling a large number of dialysis patients across all elective surgical types, allowing for greater generalisability to clinical care. Rigorous assessment of methodologic quality using a validated tool, robust ascertainment of patient-level outcomes, and use of meta-regression to explore heterogeneity and interactions are key strengths of this review. Nine studies did not provide statements on the adequacy of follow-up and eight did not explicitly state at the start of the study that the outcome of interest was not present. Meta-regression identified that age and presence of diabetes attentuated the association between dialysis status and postoperative mortality, rather than specific attributes of the studies themselves. Furthermore, not all studies reporting adjusted results adjusted for the same covariates, so that adjustment by itself varied. Despite adjustment for potentital confounding variables, such as age, indication bias with residual confounding could not be excluded. Many of the studies did not report potentially important confounding variables, such as primary kidney disease, dialysis modality, vintage, and dialysis access type, residual kidney function or use of immunosuppression. The majority of studies occurred in North America thereby potentially limiting the generalisability of the review's findings. The definition of dialysis dependency varied across the studies, such that the possibility of misclassification of acute kidney injury requiring dialysis could not be excluded. This review highlights the urgent need for future prospective studies to be more comprehensive in reporting patient baseline dialysis treatment characteristics (e.g. aetiology of kidney disease, dialysis modality, time on dialysis, access type, etc), procedural information and cause of death to allow for more informative analyses with adjustment for confounding, to help direct clinician’s perioperative risk assessment and efforts to minimise risk. In conclusion, patients on chronic dialysis have a two- to fivefold increased odds of postoperative mortality following elective surgery. The magnitude of the excess odds attributable to dialysis dependent chronic kidney disease may be lower among older patients with diabetes. # Supporting information [^1]: The authors have read the journal’s policy and have the following potential competing interests: DP has received speaking honoraria from the Australian Medical Forum. DJ is a current recipient of an Australian National Health and Medical Research Council Practitioner Fellowship. DJ has previously received consultancy fees, research grants, speaker’s honoraria and travel sponsorships from Baxter Healthcare and Fresenius Medical Care. CH has received funding from Janssen and GlaxoSmithKline to her institution for trial steering committee roles and research grant support to her institution from Shire, Baxter, Fresenius, and Otsuka and travel sponsorship from Otsuka. This does not alter our adherence to PLOS ONE policies on sharing data and materials. There are no patents, products in development or marketed products associated with this research to declare

# 1. Introduction The shade plant, *Epimedium pseudowushanense* B.L.Guo belong to the genus *Epimedium* (Chinese name, Yin Yang Huo) from the Berberidaceae family. This genus contains 58 species. Among them, *Epimedium brevicornum* Maxim, *Epimedium sagittatum* (Sieb. et Zucc.) Maxim, *Epimedium pubescens* Maxim, *Epimedium wushanense* T.S.Ying, and *Epimedium koreanum* Nabai were considered authentic sources of pharmacological products (2015 Chinese pharmacopeia). Materials from *Epimedium* plants have been used to invigorate sexuality and to strengthen muscles and bones. They are of significant economic importance as the annual sale value of medicinal products containing active components of *Epimedium* is estimated to exceed 1.1 billion Chinese Yuan in China (personal communications). *E*. *pseudowushanense* B.L.Guo is one the species most similar to *E*. *wushanense* in terms of morphology and chemical components. Due to its many favorable agricultural properties, *E*. *pseudowushanense* has been cultivated widely and used extensively as a substitute of *E*. *wushanense*. Improvement of its cultivation efficiency remains an active area of research. Active components of *Epimedium* plants largely consist of flavonoids, particularly prenylated flavonol glycosides. Well-known compounds include epimedin A, epimedin B, epimedin C, and icariin. Previous studies have revealed significant therapeutic effects of these compounds on breast cancer, liver cancer, and leukemia. With the increased demand of active components from *Epimedium* and the low recycling rate of these plants, increasing the production of the active compounds through valid commercial culture and metabolic engineering has become an active area of research. Based on the previous study we found that light could influence the content of *Epimedium pseudowushanense* B.L.Guo. So we should research the molecular mechanisms underlying the effects of light intensity on flavonoid production of it. This study could help us to know why the flavonoid content changed under different light conditions. Flavonoids are a remarkably large group of plant secondary metabolites that are derived from phenylalanine. The flavonoid biosynthetic pathway is one of the best most studied pathways of plant secondary metabolites. Many structural gene encoding enzymes involved in this pathway have been isolated and well characterized from several model species such as *Arabidopsis*, maize, and grape. Our study intends to investigate the effects of one of the most important environmental factor, light, on the production of its active components, flavonoids in *E*. *pseudowushanense*. Furthermore, we would like to identify the optimal light intensity for maximal flavonoid accumulation. Last, we exploited RNA-seq technology to understand the underlying molecular mechanisms. The success of this study would not only determine the optimal conditions for cultivation and flavonoid production, but also identify the genes responsible for flavonoid biosynthesis and regulation. RNA sequencing (RNA-seq) technology uses next-generation sequencing (NGS) to reveal the presence and quantity of [RNA](https://en.wikipedia.org/wiki/RNA)s in a biological sample under a particular condition. Given its high-throughput capability, RNA-seq can detect low-abundance genes with sufficient sensitivity. RNA-seq has been widely used for gene discovery, differential gene expression analysis, single nucleotide polymorphism discovery, and SSR discovery. NGS technology has been applied to identify genes in *Epimedium* species in recent years. For example, analysis of the leaf transcriptome of *E*. *sagittatum* through 454 GS-FLX pyrosequencing led to the discovery of many genes involved in flavonoid biosynthesis. Light is an important environmental factor that can induce plant growth, development and the biosynthesis of secondary metabolites and stimulate the accumulation of these compounds in plants. Changes in light intensity may influence flavonoid content because the flavonoid hydroxyl groups on the A and B rings vary in number and position. Several studies have shown that high light irradiance promotes the biosynthesis of flavonoids, such as dihydroxy B-ring- substituted flavonoids (luteolin 7-O- and quercetin 3-O-glycosides) but does not influence the biosynthesis of monohydroxy B-ring-substituted flavonoids (pigenin 7-O- and kaempferol 3-O-glycosides). Pacheco reported that *Piper aduncum* grown under 50% natural light irradiance had higher total flavonoid concentration than those grown under 100% natural irradiance. Deng and others found that *Cyclocarya paliurus* under 100% natural light had higher kaempferol, quercetin and isoquercitrin than 50% and 15% natural light. The effects of light are likely to be mediated through the upregulation of the expression of genes involved in the secondary metabolite biosynthesis. For example, light can promote the upregulation of genes involved in the biosynthesis and accumulation of flavonoids in *Catharanthus roseus* and *Ligustrum vulgare*. In the study of Azumaet, light treatment led to induced higher expression levels of CHS, CHI, F3H, flavonoid 3’,5’-hydroxylase (F3’5’H), DFR, O-methyltransferase (OMT) as well as UFGT compared to dark grown berries. Pacheco reported that *Piper aduncum* grown under 50% natural light irradiance had higher PAL expression than others. Leyva also found that the regulation of CHS was up with the increased light intensity in Arabidopsis thaliana. Based on the information described above, we hypothesize that (1) the accumulation of flavonoid is induced by light in an intensity dependent manner; (2) the induction is mediated by the differential expression of genes involved in the biosynthesis of the active components, flavonoids. To test this hypothesis, we first treated the plants with different light intensity levels. Second, we determined the abundance of the flavonoid contents with HPLC. Third, we compared the flavonoid abundance against the light intensity to identify the optimal levels. Forth, we selected plant materials treated at three levels with lowest, middle and highest levels of flavonoids for RNA-seq analysis. Fifth, analysis of the RNA-seq results identified genes involved in flavonoid biosynthesis and differential expressed genes (DEGs) between different light treatment groups. Last, models were proposed to explain the light-induced flavonoid accumulation. # 2. Materials and methods ## 2.1 Plant materials and growth conditions Ninety 2-year-old healthy *E*. *pseudowushanense* plants were collected from Lei Shan County (16° N, 108° E) in Guizhou Province. The plants were transferred to plastic pots (10 cm × 10 cm for inner diameter and height, 1 plant per pot) filled with a substrate mixture of 75% peat and 25% vermiculite, and then placed in the greenhouse of the Institute of Medicinal Plant Development on March 1, 2015. The plants were randomly subjected to radiation with five level I1 (5.5 ± 2.5 μmol· m⁻²·s⁻¹), I2 (14.5 ± 2.5μmol· m⁻²·s⁻¹), I3 (18.2 ± 2.5 μmol· m⁻²·s⁻¹), I4 (54.6 ± 2.5 μmol· m⁻²·s⁻¹), and I5 (90.9 ± 2.5μmol· m⁻²·s⁻¹) light intensities for 16 h per day (T5-fluorescent lamps were used as the light resource, and there were 30 pots per level). A 20–21°C temperature range was set for entire cultivation, and humidity was maintained at 60%. Except for the light intensity, the other culture conditions are same at each pot. To control the light intensity is the same for all plants in each light treat level, the thin paper were used which eliminated the effect of light from outside. The light conditions were confirmed by Li-6400 external quantum sensor (LI-COR, Lincoln, NE, USA) system. After treatment for 30 days, the plants in each group were further divided into three subgroups with 10 plants each. Fresh leaves from plants belonging to the same subgroups were randomly collected, pooled, and then stored in liquid nitrogen until use. ## 2.2 Profiling of chemical compositions using HPLC *E*. *pseudowushanense* leaf powder (200 mg) was passed through a No. 3 pharmacopoeia sieve (Each treatment group had 30 plants, they were divided into three sub groups, with 10 plants. The sub group leaves were mixed and each treatment group had 3 biological replications) and then extracted with 50 mL of 70% EtOH by ultrasonication at room temperature for 30 min. The solution was passed through a 0.45 μm microfiltration membrane, and a 20 μL aliquot of the filtrate was injected into HPLC for analysis. HPLC separation was performed on a Zorbax SB-C18 column (Agilent Technologies, Palo Alto, CA, USA) (5 μm, 250 mm × 4.6 mm). Eluents A and B were water and acetonitrile, respectively. The gradient elution program was as follows: 0–17 min (25%–26% B) and 17–26 min (26%–100% B). The column was washed with 100% eluent B for 15 min between every two testing samples and then re-equilibrated with 25% eluent B for 10 min. The elution was performed under the following conditions: flow rate, 1.0 mL/min; column temperature, 25°C; and detection wavelength, 270 nm. Data processing was performed using PerkinElmer ChemStation software (version 6.3.1). ## 2.3 RNA isolation and quantification For RNA-seq experiments, plant samples from two subgroups of each treatment group were subjected to total RNA extraction using the RNAprep Pure Plant Kit (Polysaccharides and Polyphenolics-rich) (Cat No. DP441, TianGene, China). RNA degradation and contamination were monitored using GeneGreen-stained 1% agarose gels, and RNA purity was determined using a NanoPhotometer^® spectrophotometer (IMPLEN, Westlake Village, CA). RNA concentration was measured using Qubit^® RNA Assay Kit in Qubit 2.0 Fluorometer (Life Technologies, Foster City, CA), and RNA integrity was assessed using the RNA Nano 6000 Assay Kit of a Bioanalyzer 2100 system (Agilent Technologies, Santa Clara, CA). ## 2.4 RNA-seq library construction and sequencing The sequencing libraries were constructed using the NEBNext^® Ultra^™ RNA Library Prep Kit for Illumina (NEB, USA) in accordance with the manufacturer’s protocol. In brief, mRNA was purified from total RNA using poly-T oligo-attached magnetic beads. Fragmentation was carried out using divalent captions under elevated temperature in the NEBNext First-Strand Synthesis Reaction Buffer (5×). First-strand cDNA was synthesized using a random hexamer primer and M-MuLV Reverse Transcriptase (RNase H). Subsequently, second- strand cDNA was synthesized using DNA Polymerase I and RNase H. The remaining overhangs were converted into blunt ends via exonuclease/polymerase activities, and the enzymes were removed. After adenylation of 3′ends of DNA fragments, NEBNext Adaptor with a hairpin loop structure was ligated to the cDNA fragments, which were then purified, end-repaired, A-tailed, and then ligated to index adapters (NEB). The templates were amplified by PCR and then sequenced on an Illumina Hiseq^™ 2500 platform, which led to the generation of 125 bp paired-end reads. Data analysis and base calling were performed using Illumina instrument software. DNA sequencing was performed at Beijing Ori-Gene Science and Technology Co., Ltd. Raw data had been deposited in the Short Read Archive of GenBank with the accession numbers: xxx (to be provided). ## 2.5 De novo assembly and function annotation Raw sequencing reads were processed with SolexaQA (<http://solexaqa.sourceforge.net/>) to filter out low-quality reads with default parameters and short reads with length ≤ 60 bp. The resulting high- quality RNA-seq data from the libraries were assembled using the computer program Trinity. In case several transcripts were identified for the same gene, the longest transcript was selected as the representative sequence of the gene and will be called unigene sequence in the following text. For functional annotation, all unigene sequences were searched against several databases, including the NCBI non-redundant protein sequences (Nr, <ftp://ftp.ncbi.nlm.nih.gov/blast/db/FASTA/nr.gz>), Gene Ontology (GO <http://www.geneontology.org/>), Swiss-Prot/Trembl (<http://www.uniprot.org/>), Pfam (<http://pfam.xfam.org/>), and Kyoto Encyclopedia of Genes and Genomes (KEGG; <http://www.genome.jp/kegg/>), by using the program BLASTX with E value ≤ 1e⁻⁵ and percentage of similarity ≥ 30%. ## 2.6 Gene expression quantification and differential gene expression analysis To estimate the abundance of the transcripts, all transcripts assembled by Trinity were treated as the reference sequences. The clean reads were then mapped to the reference sequences using TopHat (version 2.0.10, <http://tophat.cbcb.umd.edu/>) with default parameters. The program Cuffdiff (version 2.2.1,(<http://cuffdiff.cbcb.umd.edu/>) was used to calculate the expression levels of genes and transcripts in terms of reads per kilobases per million reads (RPKM) and the p-value for differentially expressed genes (DEGs) based on two-tailed unpaired Student’s t-test. Genes with the number of mapped reads ≥ 10, fold change ≥ 2, and uncorrected p ≤ 0.05 were deemed significant DEGs. ## 2.7 Enrichment analysis GO enrichment analysis was conducted using GOseq. We identified the significantly enriched GO term of DEGs with corrected p ≤ 0.05. For KEGG analysis, we used the KEGG pathway as a unit and applied the hyper geometric test to find significantly enriched pathways. We identified the significantly enriched KEGG pathway of DEGs with corrected p ≤ 0.05. ## 2.8 Identification of transcription factors in *E*. *pseudowushanense* Gene-encoding transcription factors were identified by comparing all unigene sequences against the plant transcription factor database (PlnTFDB; <http://plntfdb.bio.uni-potsdam.de/v3.0/downloads.php>) using BLASTX with a cutoff E value of 1e^-5. ## 2.9 Validation of RNA-seq experiments The RNA samples used for RNA-seq analyses were subjected to reverse transcription quantitative real-time PCR (RT-qPCR) analysis. Each experiment was conducted with three technical replicates. For each sample, reverse transcription was performed on 1 μg total RNA by TransScript One-Step gDNA Removal and cDNA Synthesis SuperMix (TransGen) in a 20 μl volume with anchored oligo(dT)18 primer. The reaction was carried out at 42°C for 15 min and 80°C for 5 s using an ABI 7500 Fast instrument (Applied Biosystems). Gene-specific primers were designed using PrimerQuest (<http://www.idtdna.com/Primerquest/Home/Index>). The primers used in this study are listed in. The actin gene was chosen as the endogenous control. Each qPCR reaction contained 10 μL of 2× TransStart^® Top Green qPCR SuperMix (TransGen), 25 ng of cDNA sample, and 200 nM gene-specific primers in a final volume of 20 μL. The cycling conditions were 94°C for 30 s, followed by 40 cycles of 94°C for 5 s and then 60°C for 34 s. Melting curve analyses were performed to verify the specificity by ABI 7500 Fast instrument. The relative expression levels were calculated using the 2^–ΔΔCt method. ## 2.10 Sequence analysis For selected proteins, homologous sequences were retrieved from Genbank with an E value cutoff of 1e⁻⁵. The sequences were then aligned with ClustalW software. Phylogenetic trees were constructed using the neighbor-joining algorithm with MEGA 7.0. The bootstrap score was calculated based on 1000 replications. ## 2.11 Statistical analysis Correlation coefficients among flavonoid contents, gene expression levels of related enzymes, and transcription factors were calculated using Excel. All values are presented as the mean standard error of the mean. Statistical significance of differences was evaluated using Student’s t-test or ANOVA in SPSS10 software. The significance of pearson correlation was calculated as described by VassarStats (<http://www.vassarstats.net/>). # 3. Results ## 3.1. Effects of light intensities on flavonoid content The methodology validated in our previous study was applied to analyze the flavonoid content by HPLC at five light levels. shows the changes in the contents of four different flavonoid glycosides in *E*. *pseudowushanense* under different light intensities. Interestingly, epimedin A showed different changes from epimedin B, epimedin C and icariin at I4 and I5 treatments. Epimedin A content increased as light intensity increased from I1 to I5. Thus, I5 increased epimedin A by 360.6% (p\<0.05) compared with I1. Furthermore, epimedin B, epimedin C and icariin amounts showed similar changes. Epimedin B, epimedin C and icariin increased when light intensity increased from I1 to I4, whereas all decreased under I5. The highest epimedin B, epimedin C and icariin contents were observed under I4. Epimedin B, epimedin C and icariin contents were 421.9% (p\<0.05), 624.0% (p\<0.05) and 659.9% higher, respectively, than under I1. ## 3.2. RNA-seq analysis of E. pseudowushanense treated with different light intensities In order to explore the molecular mechanism of light-induced flavonoid synthesis and accumulation in *E*. *pseudowushanense*, six cDNA libraries constituting two biological repeats were constructed from three treatment groups which the flavonoid contents were found most significantly different (i.e., the low I1, the middle I3 and the high I4) and sequenced using Illumina high-throughput sequencing platform. Six samples were named L1, L2 (I1); M1, M2 (I3); H1 and H2 (I4). The RNA-seq results were summarized in. For the six samples, the total number of raw reads ranged from 45 to 55 million. After removing the adapters, low-quality sequences, and reads shorter than 35 bp, the numbers of clean reads were 40.3, 27.3, 27.0, 28.9, 32.9, and 32.9 million for the six samples, respectively. All the clean reads were combined and then assembled into 57,962 contigs by using Trinity. For each unigene, the longest transcript (in case of multiple transcripts) was selected as the representative and was called “unigene sequence.” A total of 43,657 unigene sequences with lengths ranging from 224 bp to 17,683 bp, with an average length of 837 bp and an N50 of 1383 bp, were obtained. To assess the quality of our assembly, the clean reads were mapped to unigenes. The ratios of all mapped reads ranged from 80.27% to 90.38%, whereas the ratios of uniquely mapped reads ranged from 71.67% to 83.44%. We then examined the length distribution of these unigene sequences. In contrast to 12,127 unigene sequences that were longer than 1000 bp, 18,591 unigene sequences had lengths between 200 and 400 bp. We then compared our assembly with those for *E*. *sagittatum* based on information provided in the manuscript. Our assembled transcript was, on average, 1.23 times longer than those of *E*. *sagittatum*. Moreover, the number of genes was 15.3% greater than that of the *E*. *sagittatum* dataset. To determine the potential functions of these unigene sequences, they were searched against the databases Nr, Nt, Trembl, Swiss-Prot, and Pfam by using BLAST with an E value cutoff of 1e⁻⁵. The ratios of annotated unigene sequences ranged from 35% to 61%. Among the 43,657 unigenes, 25,989 (59.5%) and 15,441 (35.4%) had at least one significant match with an E value ≤ 1e⁻⁵ against the Nr and Nt databases. The mapping rates of unigene sequences to the Swiss-prot, Trembl, and Pfam protein databases were 45.2%, 60.1%, and 47.8%, respectively. In terms of the species source of top hit sequences, sequences from *Nelumbo nucifera* represented 36.5% of the top hits of our unigene sequences, followed by *Vitis vinifera* (10.7%), *Ricinus communis* (3.4%), *Theobroma cacao* (2.9%), *Jatropha curcas* (1.9%), and *Populus trichocarpa* (1.6%). This distribution suggests that *N*. *nucifera* is the closest species that has a large number of sequences in the Nr database. ## 3.3. Functional classification of unigenes We mapped the transcripts to GO terms and KEGG pathways; 23,553 unigene sequences were assigned GO terms. These terms belong to 57 functional groups, which were distributed under three main categories: molecular function, biological process, and cellular component. In the molecular function category, “binding,” “catalytic,” and “transporter” were the most mapped terms. In the biological process category, “biological regulation,” “cellular process,” “metabolic process,” “response to stimulus,” and “single-organism process” were the most mapped terms. In the cellular component category, “cell,” “cell part,” “organelle,” “membrane,” and “organelle part” were the mainly mapped terms. Furthermore, a few unigenes were mapped to terms “cell killing,” “extracellular matrix component,” “other organism,” “other organism part,” “nutrient reservoir activity,” “protein tag,” “translation regulator,” and “metallochaperone activity.” For the KEGG pathways, 6085 unique sequences were assigned to the pathways. The top 10 most mapped pathways were “Ribosome” (322 sequences), “Carbon metabolism” (213 sequences), “Biosynthesis of amino acids” (196 sequences), “Purine metabolism” (157 sequences), “Spliceosome” (146 sequences), “Protein processing in endoplasmic reticulum” (146 sequences), “Oxidative phosphorylation” (135 sequences), “RNA transport” (130 sequences), “Huntington’s disease” (123 sequences), and “Pyrimidine metabolism” (121 sequences). In particular, KEGG analysis showed that 21 unigene sequences were involved in flavonoid biosynthesis. Compared with those described for the *E*. *sagittatum* dataset, 54% of unigene sequences in our dataset were mapped to GO terms, whereas only 29.2% transcripts were mapped to GO terms for the *E*. *sagittatum* dataset. ## 3.4. Determination of gene abundance and identification of differentially expressed genes The abundance of unigene sequences was quantified using the program cuffdiff and represented by FPKM. The pearson correlation coefficients of gene expression levels between biological replicates are 0.76, 0.79 and 0.81 for the three treatment groups respectively. A total of 39,380, 38,103, 38,696 expressed genes were identified in groups L, M, and H, respectively. As shown in, a total of 34731 genes were expressed in all three treatments. Among them, there were 1140, 1022 and 870 genes expressed only in the L, M and H treatment, respectively. Boxplots showing the abundance distribution are shown in and. It appears that the overall deistrubtion of gene expression levels are similar for the three treatment groups. Similarly, the differentially expressed genes (DEGs) were also identified using cuffdiff. The volcano plots showing the distribution of fold changes and p values are shown in. A total of 4008 DEGs were identified between groups L and M, of which 1928 were upregulated and 2080 were downregulated. By contrast, 5260 DEGs were found between groups M and H, of which 2468 were upregulated and 2792 were downregulated. Lastly, 3591 DEGs were detected between groups L and H, of which 1289 were upregulated and 2302 were downregulated. Details for these DEGs can be found in and. These DEGs are potentially involved in the light-induced accumulation of flavonoids. ## 3.5. Functional enrichment analysis of DEGs To further narrow down the genes that are involved in light-induced flavonoid biosynthesis, the DEGs were first mapped to GO terms. The distribution of mapped GO classifications is shown in. The details for the mapping can be found in. The most mapped terms of DEGs for the categories of biological process and cellular component were “defense response” and “integral component of membrane.” In the category of molecular function, the most mapped term for DEGs between groups L and M was “Metal ion binding.” By contrast, the most mapped term for DEGs between groups M and H was “ATP binding.” Furthermore, the most mapped term between groups L and H was “protein serine/threonine kinase activity.” In parallel, the DEGs were mapped to KEGG pathways. The most enriched pathways between groups L and M included a two-component system in environmental information processing and signal transduction (22 DEGs), phenylpropanoid biosynthesis (14 DEGs), and glyoxylate and dicarboxylate metabolism in carbohydrate metabolism (14 DEGs). By contrast, the most enriched pathways between groups M and H included starch and sucrose metabolism (39 DEGs), amino sugar and nucleotide sugar metabolism (27 DEGs), and phenylpropanoid biosynthesis (20 DEGs). The results confirmed that light-induced flavonoid accumulation is mediated through the increased expression levels of genes involved in the biosynthesis of phenolic acids and flavonoids. Furthermore, a dose-response relationship exists between light intensity and gene expression levels. ## 3.6. Enzyme genes involved in the biosynthesis of active compounds in *E*. *pseudowushanense* The flavonoid pathway can be divided into three pathways leading to the production of anthocyantin, proanthocyanin, and flavonol, respectively. Basing on the structural characteristics of the compounds, we proposed a pathway for the biosynthesis of flavonoids in *E*. *pseudowushanense*. In this proposed pathway, L-phenylalanine is first converted to trans-cinnamic acid by phenylalanine ammonia-lyase (PAL, EC: 4.3.1.24) and subsequently to p-coumaric acid by trans-cinnamate 4-hydroxylase (C4H, EC: 1.14.13.11). p-Coumaric acid can be converted into p-coumaroyl-CoA by 4-coumarate-CoA ligase (4CL, EC: 6.2.1.12) and then catalyzed by chalcone synthase (CHS, EC: 2.3.1.74), chalcone isomerase (CHI, EC: 5.5.1.6), and flavanone 3-hydroxylase (EC: 1.14.11.9). As the product of these steps, dihydrokaempferol can be further converted into kaempferol by flavonol synthase (FLS, EC: 1.14.11.23), which is then converted to prenyl- flavonoids such as icariin by UGT, OMT, and some unknown methoxy transferase and isopentenyl transferase. Alternatively, kaempferol can be either hydroxylated by flavonoid 3′ hydroxylase (EC: 1.14.13.21) to produce dihydroquercetin. Furthermore, kaempferol can be converted successively by dihydroflavonol 4-reductase (DFR, EC: 1.1.1.219) and leucoanthocyanidin dioxygenase (EC: 1.14.11.19) to generate anthocyanin. In our study, 21 unique sequences that encode 14 enzyme families involved in the flavonoid biosynthetic pathways were identified. The short and full gene names were shown. A prefix “epps” was added to the short gene name to indicated that it is derived from *E*. *pseudowushanense*. Multiple sequence alignments of the identified proteins and their homologous sequences were conducted to determine if the full-length sequences have been obtained. Furthermore, phylogenetic trees were constructed to examine the relationship of the following proteins: PAL, 4CL, caffeoyl-CoA O-methyltransferase, CHS, CHI, leucoanthocyanidin dioxygenase, FLS, flavonoid 3′-monooxygenase, anthocyanidin reductase, naringenin 3-dioxygenase, bifunctional dihydroflavonol 4-reductase/flavanone 4-reductase, trans-cinnamate 4-monooxygenase, shikimate O-hydroxycinnamoyltransferase, and coumaroylquinate(coumaroylshikimate) 3'-monooxygenase. As shown in the Figures, all *E*. *pseudowushanense* genes are highly similar to their homolgous sequences. The phylogenetic relationship among these genes are consisitent with the those of the species. Based on the length of their homologs, it is likely that all these sequences containing the full-length codoing sequences. ## 3.7. Correlation between the expression profiles of biosynthetic genes and flavonoid contents To determine which flavonoid biosyntheis genes were most strongly induced by light, we examined the differential expression of these genes as well as the correlation of the expression profiles of these genes and those of the flavonoid contents across the three treatment conditions. The expression level of FLS was upregulated between groups L and M. The expression levels of CHS and FLS were upregulated between groups L and H. Lastly, the expression levels of C4H, CHI, and FLS were upregulated, whereas that of caffeoyl-CoA O-methyltransferase (COMTEC: 2.1.1.104) was downregulated between groups M and H. The expression profiles of four of the twenty-one unigenes were found to be highly correatled with those of the flavonid contents, with pearson correlation coefficents ≥ 0.9. In summar, the genes FLS, CHS, C4H and CHI seemed to be most strongly associated with the light-induced flavonoid accumulation. To study the co-expression patterns of these genes, we performed hierarchical clustering of the expression profiles of these 21 flavonoid biosynthesis genes. Three main clusters were readily discernable, the first cluster contained 15 genes that showed the highest expression levels in the “H” group. The second cluster contain 3 genes that were expressed at the highest levels in the “M” group. The remaining three genes belonged to the third cluster and had the highest expression levels in the “L” group. Genes in the cluster I have expression profiles that were better corrleated with the flavonoid contents comparing to those of the cluster II and III. It should be pointed out that the four genes FLS, CHS, C4H and CHI all belonged to cluster I. ## 3.8. RT-qPCR validation To validate the RNA-seq data, 15 genes were selected and subjected to RT-qPCR analysis. These genes include the 2 genes that are highly correlated and differentially expressed, two genes (TR2108\|c0_g1, TR575\|c1_g1) belonged to the PAL family, one gene (TR6321\|c0_g1) belonged to the FLS family, three genes (TR1945\|c0_g1, TR10614\|c0_g1, TR9038\|c0_g1) belong to the 4CL family, three genes (TR11916\|c0_g1, TR9672\|c0_g1, TR18393\|c0_g1) belonged to the CHS family, one gene (TR19880\|c0_g1) belonged to the DFR family and five genes (TR11207\|c0_g7, TR11560\|c0_g1, TR1989\|c0_g1, TR19575\|c0_g1, TR21768\|c4_g3) belongd to the UGT family. Except for TR575\|c1_g1, TR9672\|c0_g1, TR19880\|c0_g1 and TR9038\|c0_g1, the expression profiles determined by RNA-Seq experiments correlated well with those obtained from RT-qPCR experiments for 11 out of 15 (73.3%) genes with pearson correlation coefficients (r) \> 0.9. And 15 of 16 pairs of expression profiles were found to be significantly correlated (p \< 0.05). This finding suggests that the results of our RNA-seq experiments are reliable. To see if any correlations existed between flavonoid content and expression patterns of the flavonoids biosynthesis genes, we analyzed transcript abundance of four related genes (*C4H*, *CHS1*, *FLS*, and *CHI*) by RT-qPCR from 5 different light intensity described in. The relative expression level of the four genes showed similar changes from I1 to I5 light intensity. The expression of four related genes under I1 and I2 are lower than I3 to I5. Interestingly the changes of *CHS1* and *FLS* showed similarly changes to epimedin B, epimedin C and icariin. The I4 treatment showed highest expression level at *C4H*, *CHS1* and *FLS* while the I5 treatment showed highest expression level at *CHI* gene (**)**. ## 3.9. Transcriptional factors involved in the light-induced flavonoid accumulation in *E*. *pseudowushanense* To understand how the expression of genes involved in flavonoid biosynthesis was regulated in response to light, we first identified all unique sequences encoding the transcription factors in our RNA-seq dataset by comparing to sequences in the plant transcription factor database using BLAST with an E value cutoff of 1e⁻⁵ ( and Tables). We identified 4621 unigene sequences that likely encode transcription factors. The lengths of unigene sequences representing these transcription factors varied from 224 to 13,144 bp, with an average length of 1241.5 bp and an N50 value of 1863 bp. The length distribution of these putative transcriptional factor genes is shown in. In terms of types, the identified transcription factors were distributed in 59 families, such as C3H, bHLH, FAR1, WRKY, NAC, MYB-related, and so on. The differentially expressed transcription factors after light treatment mainly belong to the families FAR1, WRKY, bHLH, and MYB-related families. To select further the transcription factors that are involved in the light- induced flavonoid accumulation, we first collected the sequences of all transcription factors from *Arabidopsis thaliana*, *Oryza sativa*, *V*. *vinifera*, and *E*. *sagittatum*, based on (1) similarity to known transcription factors involved in flavonoid biosynthesis; (2) p value for differential gene expression in any contrast group; and (3) correlation between gene expression profiles and flavonoid contents. Transcription factors including 31 FAR1, 17 MYB, 12 bHLH, and 5 WRKY are likely involved in light-induced flavonoid accumulation. ## 3.10. Light signalingl factors involved in the light-induced flavonoid accumulation in *E*. *pseudowushanense* To select the light signaling factors that are most likely involved in the light-induced flavonoid accumulation, we collected them, based on (1) similarity to known light signal factors involved in flavonoid biosynthesis; (2) p value for differential gene expression in any contrast group; and (3) correlation between gene expression profiles and flavonoid contents(\>\|0.9\|). Light signal factors including 3 COP1, 1 pif, 1 HY5, 1 SPA, 1 DET, 3 phy and 3 cry are likely involved in light-induced flavonoid accumulation. # 4. Discussion ## 4.1 Enzymatic genes involved in flavonoid biosynthesis Previous studies demonstrated that light treatment of grape and kale could influence gene expression, leading to the accumulation of specific flavonol glycosides. Further studies in grape berries reported that flavonol levels are sensitive to changes in light conditions; flavonols accumulate with increased expression of FLS. These studies suggest that the expression levels of genes involved in flavonoid biosynthesis are regulated by light. In the present study, we found that C4H, CHS, CHI, and FLS were all upregulated under the different light treatments, partially explaining the light-induced flavonol accumulation in *E*. *pseudowushanense*. ## 4.2 Transcription factors involved in light-induced flavonoid biosynthesis Transcription factors regulate the secondary metabolite biosynthesis and accumulation of flavonoids. Several families of transcription factors play roles in the production of flavonol compounds. Qiu et al. identified a WRKY protein (OsWRKY13) as a transcriptional regulator of flavonoid biosynthesis in *O*. *sativa*, which could induce the expression of CHS. WRKY transcription factors are defined by the presence of the DNA-binding domain WRKY. The identified WRKY genes are significant regulators involved in plant developmental processes and responses to biotic and abiotic signals. The inducible expression patterns of WRKY genes suggest that they are involved in the regulation of plant secondary metabolis. As for flavonol biosynthesis, several specific regulators belonging to the MYB transcriptional factor family have been identified in model species. MYB proteins are characterized by the presence of one or many MYB repeat (R) DNA- binding domains. In *A*. *thaliana*, AtMYB12 activates the expression of AtFLS and AtCHS. In grape, VvMYBF1, orthologous to AtMYB12, markedly upregulated the expression levels of VvFLS and VvCHI. In *E*. *sagittatum*, some MYB members have been isolated and characterized, among which EsMYBF is homologous to AtMYB12 that is related to flavonol synthesis. In grape, light induces the expression of an array of MYB transcription factors, such as VvMYBF1 and VvMYB12, which are positive regulators of the general flavonoid biosynthesis pathway as well as those specifically responsible for flavonol biosynthesis. MYB transcription factors can directly and specifically interact with MYB recognition element (MRE). MRE is part of the light regulatory unit, which also contains bZIP recognition element (ACE). MREs can be found in the promoter regions of light-induced structural flavonoid genes, such as CHS and FLS in Arabidopsis and grapevine. The expression levels of these MYB are also regulated by other transcription factors, such as Elongated Hypocotyl 5 (HY5). HY5 is a bZIP transcription factor that can promote photo-morphogenesis by recognizing ACE. In particular, HY5 has been linked to the activation of MYB and key structural genes (CHS and FLS) of the flavonoid pathway as well as the accumulation of flavonoids in response to light in Arabidopsis. Located further upstream of the regulatory pathway, HY5 is a direct target of RING-finger-type ubiquitin E3 ligase Constitutive Photo-morphogenic 1 (COP1). COP1 acts as a negative regulator of light signaling directly downstream of the photoreceptors and controls different light-regulated plant development processes by adjusting its subcellular localization. In the presence of light, the interaction of the COP1/Suppressor of PhyA (SPA) complex with activated photoreceptors inhibits COP1/SPA function through the dissociation of COP1 from the complex and exportation from the nucleus. The downregulation of COP1 in the nucleus allows nuclear-localized transcription factors, such as HY5, to accumulate and induce the expression of genes responsible for flavonoid biosynthesis. Aside from the transcription factors described above, other important classes of transcriptional factors that might be involved in flavonoid biosynthesis include the Far-red impaired Response 1 (FAR1) and Far-Red Elongated Hypocotyl 3 (FHY3) families. FAR1 and FHY3 participate in diverse developmental and physiological processes and are essential for PhyA signaling in A. thaliana. HY5 physically interacts with FHY3/FAR1 through their respective DNA binding domains in *A*. *thaliana*. ## 4.3 Other pathways related to light-induced flavonoid accumulation Enrichment analysis showed that DEGs are significantly enriched for those involved in the two-component regulatory system, suggesting that this pathway might be involved in light-induced flavonoid accumulation. A two-component regulatory system is a basic stimulus-response coupling mechanism to allow organisms to sense and respond to changes in different environmental conditions. Two-component systems typically consist of a membrane-bound histidine kinase that senses a specific environmental stimulus and a corresponding response regulator that mediates the cellular response, mostly through the differential expression of target genes. Two-component regulatory systems are also commonly found in plants. How this system is involved in light-induced flavonoid accumulation in *E*. *pseudowushanense* represents an interesting research question in the future. ## 4.4 Model proposed To date, the mechanism by which light induces the biosynthesis of specific flavonoids in *Epimedium* is unknown. However, analysis of our transcriptome data implies that the mechanism of flavonoid accumulation in *E*. *pseudowushanense* is rather complex. Basing on previous studies, we proposed a model explaining light-induced flavonoid accumulation. In this model, light signals are received either by photoreceptors such as phytochrome or the two- component regulatory system through downstream signaling pathways, leading to the upregulation of genes involved in flavonoid biosynthesis and ultimately resulting in the accumulation of these compounds. This model will serve as a central hypothesis for the light-induced flavonoid biosynthesis that will be tested in the future. # 5. Conclusions This study represents the first comprehensive investigation of the genetic makeup responsible for the flavonol biosynthesis in *E*. *pseudowushanense*. Firstly, we find I4 light intensity is optimal for flavonoid ingredient accumulation. Then, we identified 43,657 unigene sequences in *E*. *pseudowushanense* from samples treated with light at three intensity levels by using RNA-seq technology. We determined the full-length sequences of 21 enzymatic genes involved in the flavonol biosynthesis. Among them, the FLS, CHS1 genes were strongly associated with light-induced flavonoid accumulation. We also found 65 transcription factors, including 31 FAR1, 17 MYB-related, 12 bHLH, and 5 WRKY, which might participate in light-induced flavonoid accumulation. A model was proposed to explain the underlying molecular mechanism. This work provides valuable resources for further studies on flavonoid production in *Epimedium*. These information can help us to know why the flavonoid content changed under different light conditions. Besides in vitor experiments could be conduct to examine the fouction of FLS and CHS1 under diferent light intensities. # Supporting information We are very grateful for the experiment support from Xiangbo Yang and Li Li of TongJiTang (GuiZhou) Pharmaceutical Co. LTD., a subsidiary of SinoPharm Groups. [^1]: Pharmaceutical Co. LTD provided material support in the form of plant materials for this study. There are no patents, products in development or marketed products to declare. This does not alter our adherence to PLOS ONE policies on sharing data and materials.

# Introduction Matrix-assisted laser desorption / ionisation time-of-flight mass spectrometry (MALDI-TOF MS) is used in clinical microbiology for the identification of bacterial or fungal strains isolated from positive blood cultures. The mass spectra produced from isolates are reproducible, quickly generated, and libraries of spectra have been created from the analysis of known strains allowing the rapid identification of unknown isolates. Sample preparation can simply involve the transfer of isolated culture (e.g. with a toothpick) to the target plate and overlaying with HCCA matrix (α-Cyano-4-hydroxycinnamic acid)—a direct sample application. Chemical extraction methods can also be performed and have been shown to increase the number of successful identifications, especially for Gram-positive bacteria. On a target plate 1 μL volumes of chemical extract are dried, overlain with 1 μL matrix, and tested. Routine use of MALDI-TOF systems in hospital laboratories typically involves the use of a direct method to apply culture to target plates on the laboratory bench. MALDI-TOF systems are generally not operated within primary biological safety containment (i.e. a microbiological safety cabinet). The vast majority of organisms isolated from hospital blood cultures are classed in the UK as Advisory Committee on Dangerous Pathogens (ACDP) Hazard Group (HG) 2 or lower, and therefore pose a low risk to the operative. However, there is a danger that an operative may inadvertently apply an ACDP HG3 bacterial agent directly to a target plate, especially in a geographic region where ACDP HG3 organisms are endemic. This may incur an unacceptable risk of exposure to the operative. MALDI-TOF chemical extraction methods have been reported to inactivate ACDP HG3 bacteria, though the addition of filtration sterilisation steps has been reported to be required to ensure all viable bacteria are removed from the extract, especially samples which may contain bacterial spores. A recent inadvertent release of *B*. *anthracis* in a US laboratory was attributed to the production of an unfiltered MALDI-TOF chemical extract. In this paper we report on the inactivation efficacy of a MALDI-TOF chemical extraction method (using ethanol, formic acid, acetonitrile, and filtration), on the viability of high concentrations of *Bacillus anthracis* vegetative cells and spores. This method was designed to be able to be used in a clinical laboratory, within a small microbiological safety cabinet or isolator, with the MALDI-TOF MS system being situated on the bench, outside of primary biocontainment. The method was required to allow generation of mass spectra which could be compared with databases supplied with the Bruker Biotyper MALDI- TOF MS system, and therefore allow routine identification of common clinical strains. # Materials and Methods ## Chemical extraction method for MALDI-TOF MS analysis The Formic Acid Extraction Method, as found in the Biotyper manufacturers (Bruker) user manual, was adapted to include a filtration step. In brief: 10 μL loops of culture were added to 1 mL of 70% ethanol. The suspension was mixed thoroughly, centrifuged (13 000 rpm; 2 min), and the supernatant discarded. The pellet was then allowed to air dry. The pellet was then re-suspended in 50 μL of 70% formic acid and mixed thoroughly. Fifty μL of 100% acetonitrile was then added, the suspension mixed thoroughly, and centrifuged (13 000 rpm; 2 min). The supernatant was then passed (7 000 rpm; 10 secs) through a 0.2 μm Micro-Spin, Regenerated Cellulose membrane, spin column (Chrom Tech Inc, MN). The resulting filtrate was then passed (7 000 rpm; 10 secs) through a fresh 0.2 μm Micro-Spin spin column and the filtrate retained for MALDI-TOF analysis. Regenerated cellulose filter membrane spin columns were chosen for their low protein binding characteristics and reported resistance to solvents, plus their ability to allow filtration of low sample volumes (\< 100 μL). All reagents (Sigma-Aldrich UK) were reagent grade or HPLC grade. The method was tested using overnight cultures of the avirulent (pXO1^-; pXO2^-) *B*. *anthracis* UM23CL2 strain. With input amounts in the range of 10⁷ to 10⁸colony forming units (cfu) resulting protein extracts were shown to be able to be identified as *B*. *anthracis* when resulting mass spectra were compared against spectra in the Bruker Security Relevant database (data not shown). *E*. *coli* MRE 162 and *S*. *aureus* ATCC 6538 strains, species commonly isolated from blood culture, were also successfully identified from extracts produced using the method (when spectra were compared with those in the Bruker Taxonomy database). ## Initial evaluation of inactivation efficacy of method (with and without filtration) against gram-negative and gram-positive bacteria Under ACDP containment level (CL) 2 (BSL-2 in other countries) conditions *Escherichia coli* MRE 162, *Klebsiella pneumoniae* ATCC 35657, and *Bacillus anthracis* UM23CL2 were incubated (37°C) overnight on Luria (L) agar plates. In each species specific experiment 10 μL microbiological loops of culture were extracted using the above method (3 reps with filtration / 3 reps without filtration). Entire protein extracts (typically comprising final volumes of 80–90 μL) were pipetted into the wells of a sterile 6-well tissue culture plate (Corning Costar), and spread around the base of each well with a sterile cell scraper. Extracts were then allowed to dry (typically for 20–30 mins) in order to ensure the formic acid and acetonitrile evaporated and was not carried forward into culture. Initial work had shown that even low concentrations (0.05% each) of these chemicals could interfere with L-broth bacterial culture (data not shown). Each experiment (i.e. each species) also included a positive and negative control. Positive controls comprised cell suspensions prepared following the above method—until the filtration steps–but with substitution of all reagents with 1 × Dulbeccos Phosphate Buffered Saline (1 x DPBS). The final bacterial pellet (in 100 μL of 1 x DPBS) was re-suspended and added to a well, as before, and allowed to dry. A sterile swab was applied to the dried protein extract / positive control visible in the bottom of each well in a repetitive crossing pattern and then streaked onto L-agar plates. Negative controls comprised a swab applied to a naïve plate well. Plates were incubated (48 hours; 37°C). Resulting colonies from the *B*. *anthracis* extracts were re-isolated onto fresh L-agar plates and incubated (24hr; 37°C). One μL of resulting culture was added to 1 mL sterile distilled water and heated (99°C; 15 mins). One μL aliquots of the resulting lysed cell suspensions were tested by PCR on the chromosomal target Ba chr-MGB -. ## High stringency evaluation of the inactivation efficacy of the method on *Bacillus anthracis* Vollum vegetative cells and spores Under ACDP CL3 conditions the virulent (pXO1⁺; pXO2⁺) *B*. *anthracis* Vollum strain was cultivated overnight (37°C) in either L-broth or on L-agar plates. None of the cultures used were more than 24 hours old to help prevent widespread sporulation on agar plates. Cell pellets or plate cultures were re-suspended into 1 mL of sterile 1 × DPBS, and 100 μL aliquots of the resulting suspension were added to 900 μL of 70% Ethanol. MALDI extracts were then prepared using formic acid and acetonitrile, and double filtration, as described above. Suspensions were enumerated (to determine the number of bacterial cells in 100 μL aliquots), by the production of a 10-fold dilution series and plating of appropriate dilutions (100 μL aliquots; 3 reps per dilution) onto L-agar plates. These plates were incubated (min. 48 hours) and colonies counted. Extracts were also prepared from a pre-existing, enumerated (1.2 × 10⁹ cfu·mL^-1), suspension of *Bacillus anthracis* Vollum spores. In total eighteen MALDI-TOF extract replicates were produced from each cell type. Each experiment comprised 3 MALDI-TOF protein extracts (from an overnight plate culture or the pre-existing spore suspension), one positive control and one negative control. Different experiments were initiated on different days. Extracts / positive controls were applied to individual wells of sterile tissue culture plates, allowed to dry and then re-suspended by adding 100 μL aliquots of sterile 1 × DPBST (DPBS with 0.01% Tween) to each well. The aliquot was applied across the well by repeated pipetting (30 secs) to ensure coverage of the entire well base and re-suspension of as much protein extract / positive control as possible. All 1 × DPBST was then removed from each well and put into culture (see below). A negative control (addition of 100 μL of sterile 1 × DPBST to a naïve well) was also prepared and cultured in each experiment. To provide a qualitative indication (QUAL experiments) to the inactivation efficacy of the method nine reps of each cell type protein extract (re-suspended into 1 × DPBST) were added to 900 μL of Luria broth and incubated (7 days; 37°C). Entire broths were then plated onto L-agar (250 μL × 4 plates) and incubated (a further 7 days; 37°C). To provide a quantitative indication (QUANT experiments) on the inactivation efficacy of the method, nine reps of each cell type protein extract (re- suspended into 1 × DPBST) were directly plated onto L-agar plates and incubated (4–7 days; 37°C). Resulting colonies from vegetative cell extracts were re-isolated onto fresh L-agar plates and incubated (24hr; 37°C). Where there were too many cells to count (TMTC) a 1 μL loop was streaked across the centre of the plate and used to inoculate a fresh L-agar plate. A 1 μL loop of each resulting culture was added to 1 mL sterile distilled water and heated (99°C; 15 mins). One μL aliquots of the resulting lysed cell suspensions were tested by PCR on the pXO1 target pXO1-MGB. Resulting colonies on L-agar plates from spore extracts were either tested by the same method or assumed to be *B*. *anthracis* on the basis of typical colony morphology. # Results ## Initial inactivation experiments (with and without filtration) Results from the initial inactivation experiments are summarised in. No input strain was recovered from filtered or non-filtered extracts generated from *E*. *coli* MRE162 or *K*. *pneumoniae* ATCC 3657 strains. *B*. *anthracis* UM23CL2 was recovered from 2/3 non-filtered extracts, as confirmed by Ba chr-MGB PCR, but not from any filtered extracts. ## *B*. *anthracis* Vollum inactivation experiments Results from vegetative cell experiments are summarised in. *B*. *anthracis* was recovered from 3/18 replicates and only from QUAL experiments where replicates had undergone a L-broth recovery incubation (7 days), and a further L-agar plate incubation (a further 7 days). In one replicate (QUAL MALDI 1: Extract 3), seven individual *B*. *anthracis* colonies were recovered from one plate replicate, rather than the multiple *B*. *anthracis* colonies recovered from all four plates in the other two positive replicates (QUAL MALDI 3: Extracts 2 and 3). Individual colonies and plate cultures from these experiments were confirmed as *B*. *anthracis* by pXO1 MGB-PCR. Results from spore experiments are summarised in. *B*. *anthracis* was recovered from 10/18 replicates, including quantification (QUANT) experiments where extracts were not incubated in L-broth. Plate cultures from QUAL MALDI S2 and QUAL MALDI S3 experiments were confirmed as *B*. *anthracis* by pXO1-MGB PCR. In QUANT experiments the total number of colonies from 4/9 extracts comprised 5, 3, 5, and 107 colonies respectively. These colonies were either confirmed to be *B*. *anthracis* by pXO1-MGB PCR (QUANT S1 experiment) or assumed to be *B*. *anthracis* due to typical colony morphology (QUANT S2 experiment). # Discussion *Bacillus anthracis* Vollum was chosen for the high stringency inactivation experiments based on the results from the initial inactivation experiments. These indicated that viable *B*. *anthracis* was harder to inactivate / remove in extracts than two gram-negative bacterial species (*E*. *coli* and *K*. *pneumoniae*). In addition it was possible to test *B*. *anthracis* Vollum in its spore form, considered to be one of hardest organism types to inactivate, which allowed as much information on the inactivation efficacy of the method to be obtained as was possible. It is unlikely that neat preparations of *B*. *anthracis* spores would be tested in a clinical laboratory as fresh plate cultures (from positive blood cultures) are typically tested. These are not cultivated on defined sporulation (nutrient deficient) media. Previous research has indicated the requirement for filtration of MALDI-TOF extracts to ensure inactivation or removal of *Bacillus anthracis*, especially extracts that could contain spores—a finding confirmed in a recent Centers for Disease Control and Prevention report. The stringent *B*. *anthracis* Vollum inactivation experiments conducted in our study differed in methodology from previous studies in that the entire protein extract was tested for the presence of viable *B*. *anthracis* and the formic acid and acetonitrile matrix was evaporated off to prevent interference with culture techniques carried out in viability testing. During method development we determined that even 0.05% concentrations of acetonitrile and formic acid in L-broth could interfere with the growth of *B*. *anthracis* (data not shown). In addition, in half of the experiments, a broth culture recovery incubation was applied in order to facilitate the best chance of the resuscitation of cells stressed or damaged during the inactivation method, but which retain viability. In total, from all vegetative cell experiments, viable *B*. *anthracis* were recovered from 3/18 experimental replicates (QUAL and QUANT experiments). Viable *B*. *anthracis* cells were only recovered from replicates which underwent L-broth recovery incubations (QUAL experiments). It is hypothesised that the recovery of *B*. *anthracis* cells in these experiments represent the survival of individual cells that were sub-lethally injured or induced into a viable but not-culturable state (VBNC), by the inactivation method but which were able to be resuscitated during two weeks of culture. This hypothesis is supported by the isolation of 7 individual *B*. *anthracis* cells from one L-agar plate in the QUAL MALDI 1 (Rep 3) experiment. This suggests that these cells were sub- lethally injured to the point where they could not replicate during 7 days of broth culture but which were subsequently able to grow on L-agar plates. In quantitative (QUANT) vegetative cell experiments, where no broth culture incubation was employed, no *B*. *anthracis* colonies were recovered. A previous study has shown that disinfectant treated *Bacillus atropheaus* spores can be more readily identified during broth culture incubation than direct plate culture. Our study indicates that this may also be true of vegetative *B*. *anthracis* cells. If this is true, and the recovery of *B*. *anthracis* in the QUAL MALDI 1 (input 10⁷ cfu) and QUAL MALDI 3 (input 10⁸ cfu) experiments does represent the survival of sub-lethally injured or VBNC individual cells, then a six to seven log reduction in viable cell counts was observed. It should be noted that in 10 experimental replicates with an input of 10⁸ cfu (vegetative cells), no viable *B*. *anthracis* were recovered. It is possible that recovery of *B*. *anthracis* from vegetative cell MALDI extracts might be a function of sporulation on L-agar plates prior to extraction, even with fresh cultures being processed. As might be expected, *B*. *anthracis* spores were shown to be more difficult to inactivate by the chemical extraction method with viable *B*. *anthracis* being recovered from 10/18 spore extract replicates. In quantification (QUANT) spore experiments, where inactivation was not demonstrated, the highest colony count from an extract was 107, suggesting that in this instance a six log reduction in viable spores was achieved. However, if individual spores are able to survive an extraction process, whatever the input amount, this may account for the recovery of *B*. *anthracis* extracts in the vegetative cell experiments. It is striking that from extracts generated from both cell types a range of positive and negative *B*. *anthracis* recovery responses were generated. The multiple replicates tested in our study help elucidate this phenomena, as it has in other pathogen inactivation studies, but we have not determined what survival mechanism(s) are being employed by *B*. *anthracis*. We use terms such as ‘sub- lethally injured’ and ‘VBNC’ only as labels to hypothesise how *B*. *anthracis* might be surviving the extraction process without determining which, if either, of these labels might apply. Various stress responses are known to be employed by bacteria entering a VBNC state, including modifications to cell walls and membranes. Significantly, bacterial cells are also known to shrink during a stress response. Although filtration was indicated to be an important component of *B*. *anthracis* inactivation in this study (as it has in others) it appears that spores and vegetative cells were able to pass through two 0.2 μM filters. A gram-positive rod, *B*. *anthracis* vegetative cells range in size between 1–1.2 μM in diameter and 3–5 μM in length. *B*. *anthracis* spores have been measured with mean diameters 0.81–0.86 μM and mean lengths of 1.26–1.67 μM. It is possible that the action of ethanol, formic acid, and acetonitrile could be shrinking the bacterial cell / spore allowing passage through filters, though other factors such as chemical action on the filter membrane or inconsistencies in the pore size could also play a part, as in fact could a combination of all these factors. In our study the choice of a 0.2 μM filter size was mandated by an aspiration to use spin columns (for ease of use and the ability to filter low volumes), and the membrane material was chosen for its low retention of protein and for its reported resistance to solvents. We could not source a 0.1 μM spin column filter and had we been able to do so the action of the solvent / acid test matrix on regenerated cellulose membranes is still unclear. With filtration identified as an important part of the sterilisation of MALDI extracts a greater understanding of these factors (i.e. bacterial shrinkage; chemical effect on filter membranes; even operational factors associated with spin column architecture) could aid in the development of improved and robust inactivation protocols. As a pointer for future research we did not test filtrates from *B*. *anthracis* suspensions in water. This could help determine if there is a chemical action on filter membranes. A further factor in the inactivation efficacy of the method is the volume of formic acid / acetonitrile required to produce a protein extract that is able to allow confident agent identification by MALDI-TOF MS. The manufacturer provides guidelines that different amounts of these reagents can be used for different amounts of bacterial culture (i.e. 20–40 μL each for 1 μL culture loops; 40–80 μL each for 10 μL culture loops). In our study 50 μL volumes of formic acid were used as we had determined that amounts of *B*. *anthracis* of 10⁷ to 10⁸ cfu were required for a reliable identification by MALDI-TOF MS, and that these amounts were collected by these loops (data not shown). In terms of inactivation studies carried out under ACDP CL3 conditions it is expensive to test multiple different volumes against multiple different agent amounts. This is a potential difficulty in developing and validating a one-size-fits-all extraction method, allowing identification of common clinical strains and ACDP HG3 pathogens, but still inactivating all bacteria species and culture amounts required to be tested. In this study 6 to 8 log reductions in viable bacterial counts were demonstrated when using the devised chemical extraction method. However, the highest quantified number of cells recovered from a 80–90 μL protein extract suspenion was 107 cells (QUANT MALDI S2: Extract 2), and this was from a neat spore suspension which is unlikely to be tested in a clinical laboratory. We have hypothesised that the recovery of *B*. *anthracis* from vegetative cell extracts represent the survival of sub-lethally injured or VBNC individual cells. With the requirement to only test a 1 μL volume of an extract (which is dried onto target plates and overlain with a further 1 μL HCCA matrix aliquot–with the HCCA being reconstituted in an acetonitrile / trifluoroacetic acid (TFA) solution), it is possible that methods using the findings of this project could be developed in clinical laboratories to provide only a very low residual risk to operatives. The risk is further lowered by the fact that the vast majority of clinical blood cultures do not contain ACDP HG3 pathogens. With MALDI-TOF being postulated as a driver for total laboratory automation then an automated chemical extraction method could also further reduce the risk to the operative. If so then some of the findings in this study should be considered. In summary, viability tests analysing the entire extract, with multiple replicates and extended broth and plate culture steps, has provided valuable information on the inactivation efficacy of the approach. In a separate study similar principles (i.e. testing of the entire extract; extended culture steps) were used to develop inactivation methods for Ebola Virus. A recent report on the inadvertent shipment of *B*. *anthracis* spores by the US Department of Defense has identified differences in viability testing methodology between institutes; e.g. some testing 10% of an extract, some omitting a broth culture stage. In previous reports on the bacterial inactivation efficiency of MALDI-TOF chemical extraction methods similar differences in methodology are also apparent, or have not been clearly defined. The findings of our study could therefore help inform the development of common Standard Operating Procedures for viability testing from many inactivated sample types–a recommendation in the US Department of Defense report—in addition to improving bacterial inactivation within MALDI-TOF protocols. The authors thank Andy Maule (Dstl) for supply of *B*. *anthracis* Vollum spores and advice during experimental work. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: SAW. Performed the experiments: SAW MGMS. Analyzed the data: SAW RAL. Contributed reagents/materials/analysis tools: RAL. Wrote the paper: SAW.

# Introduction Considered one of the main causes of global environmental change, land use change (LUC) is generally defined as human modifications of land and the ways it is used (e.g. clearing forests for agriculture, infrastructure development, timber harvesting, among others). Changes in land use have multiple drivers, ranging from population growth, migration, and changes in governmental policies, to cultural changes influencing attitudes towards land, dietary transitions and incentives for forests conservation. Usually, the goal of LUC is to obtain natural resources to fulfill human needs, which can result in negative impacts on the environment and human health. There is evidence that LUC has affected global water and carbon cycles, and global climate. Worldwide, agricultural expansion, a component of LUC, remains the most substantial driver of deforestation rate: about 40% of deforestation in the tropics and subtropics is for large-scale commercial agriculture. From 2010 to 2015, loss of forest area occurred mainly in the tropics, reaching 1770 million ha in 2015. During the same time period, Brazil ranged first in net tropical forest loss, accounting for 984,000 ha of net forest loss per year. Deforestation in the Amazon, while not as prevalent as it used to be, increased by 29% between 2015 and 2016. 7,989 square kilometers of Brazil’s jungle were lost from August 2015-July 2016, and between 2001 and 2015, 1,809,553 ha of forest have been lost in Peru. Changes in ecosystems can have harmful consequences to human health. It is estimated that almost one quarter of the global burden of disease can be attributed to the environmental changes (including LUC) that contribute to air, water, and soil pollution. Forest fragmentation and degradation due to LUC is threatening the Amazon Basin, one of the most biodiverse regions on earth and provider of key environmental services that contribute to global and local well- being. These impacts in turn affect human health through different pathways, for instance by changing vector population dynamics and pathogen transmission dynamics, potentially contributing to increased disease burden relating to vector borne diseases. A variety of studies conducted in the Amazon region have described the health effects of LUC, both directly (e.g. measuring incidence of specific diseases on human populations) and indirectly (e.g. measuring changes in vector population). However, there is no comprehensive review of the current state of evidence for the whole region. Scoping reviews aim to broadly examine the existing literature on a topic in order to review concepts, report on the type of studies available regardless of their methodological quality, and identify gaps in knowledge to further inform research practice. As populations continue to move into previously uninhabited areas, and as more research continues to reveal potential health effects of human activities, having comprehensive reviews on LUC and health topics will become vital to planners, policymakers, and those who work in the health industry. The aim of this scoping review is to identify articles that focus on LUC practices and their potential health outcomes within the Amazon region and identify gaps in knowledge to further inform research practice and policy makers. # Methods The steps included for our review followed the framework outlined by Arksey and O’Malley, taking into account recommendations made by Levac: (1) identification of the relevant research question, (2) identification of relevant articles, (3) article selection, (4) charting of the data, and (5) collecting, summarizing and reporting of the results. The optional sixth step proposed in the framework, conducting a consultation exercise, was not carried out. ## Identifying the research questions This review answered the following question: What is the current state of evidence on the link between LUC and human health in the Amazon region? ## Identification of relevant articles We conducted a comprehensive scientific literature search in the electronic databases PubMed (biomedical sciences) and Web of Science (multidisciplinary). We developed and mapped key search terms with online databases prior to the article search. The research query included terms related to the Amazonian region, different aspects and activities related to LUC, and various types of diseases. We carried out the search on September 6, 2016, where all articles were uploaded to a Zotero database. Our analysis for inclusion was a multi-step process. Our initial search yielded 780 papers, and after evaluating for the search criteria’s (LUC link, health link, etc.) we analyzed 14 articles. On March 16, 2017, from articles in our Zotero database, we created an initial list that met eligibility criteria by evaluating their titles and abstracts. At this step, we recognized that there was already a wealth of information on malaria and mercury (64 papers, including other reviews); therefore, we opted to exclude them from our paper. Selected articles were then read in full and evaluated for inclusion. Two reviewers independently conducted all stages of the scoping review, from relevance screening to data extraction. The two selectors individually selected the papers for each phase of elimination. Before moving on to the next stage, the selectors met and discussed each paper they chose to eliminate or keep until selectors reached agreement. Any discrepancy was discussed in a meeting with all co- authors to debate whether the article met our selection criteria or not. ## Article selection We defined relevant publications as: any peer-reviewed article (except for reviews) published between January 1 2000 and August 31 2016; in English or Spanish; focused on or presenting data from the Amazon region (Brazil, Peru, Colombia, Venezuela, Ecuador, Bolivia, Guyana, Suriname, and French Guiana); that refer to a specific health issue; and include the topic of LUC. Our inclusion criteria for the topic of LUC was defined as either papers that refer to the existence of one or more aspects of LUC, or that partly or fully fulfilled the concept of land use change (LUC) as defined by the Intergovernmental Panel on Climate Change (IPCC): “…human activities which: (a) Change the way land is used (e.g., clearing of forests for agricultural use, including open burning of cleared biomass), or (b) Affect the amount of biomass in existing biomass stocks (e.g., forests, village trees, woody savannas, etc.)". We used papers that studied causative agents for health problems (e.g. vectors, arsenic exposure, air and water quality, etc.) for analysis, even if no health measurements were taken on humans. While it is not always true that the presence/absence of these factors will result in negative health outcomes, we felt this was a safe assumption based on existing literature. Finally, we excluded articles on LUC that focused on topics that may be distally related to health outcomes, such as immigration (distally related to LUC, however after settlement). ## Data management and characterization/charting Authors created an Excel spreadsheet in which data extracted from the selected articles, including authors, year of publication, title, research objectives, location, study design, type of LUC and measures, and health outcome were recorded. In an additional column, authors recorded the summary of the findings from the articles. ## Analyzing, summarizing, and reporting the results The analysis and synthesis of literature included quantitative analysis (e.g. descriptive statistics) and qualitative analysis (i.e. content analysis). A narrative approach allowed reviewers to extract common themes that emerged from the findings. # Results ## Literature profile Of the 26 articles reviewed in full, we included 14 in the analysis after exclusion criteria were applied. These articles pertain only to two Amazonian countries, Brazil and Peru, with the former providing the most articles (11). Health data referred to in these articles came from direct and indirect measurements of a disease in humans, vectors, or animals. Only one paper addressed disease as a risk through a measurement of air pollutant exposure. Three articles were based on epidemiological data (e.g. incidence, prevalence, mortality), 5 of them on data collected from vectors (e.g. relative abundance and richness of species), and one addressed local people’s perceptions of their wellbeing. Most papers did not describe specific methods for measuring changes in land use or forest cover. Five papers used mapping and satellite images to quantify forest cover; one used data on fire outbreaks related to deforestation, and one was based on governmental data on the average amount of square kilometers deforested per year. Over three quarters of reviewed articles were published from 2010 and on, reaching a peak with four publications in 2016. All were written in English. The data collection table we used is shown in, and includes our results. Study characteristics are summarized in. depicts a conceptualization we developed to portray how we defined what we considered to be direct and indirect LUC and health topics; it describes how various specific LUC activities, human activities, and health effects are related. outlines the methods used in each of the papers selected for review. ## The link between land use change and health Eight out of 14 papers described negative relationships between LUC and health, one depicted positive, and five demonstrated neutral/mix correlations. The most common health issue was lung health, followed by mosquito-borne illnesses and rabies. ### Road building Road building is considered a particularly damaging form of LUC in tropical forests and was the most commonly addressed in the analyzed articles. Andrade et al. (2016) found that the presence of highways was related to a higher risk of contracting bovine and human rabies in the state of Pará, Brazil. The presence of the Brazilian BR-163 highway (Cuiabá-Santarém Highway) was associated with an increased prevalence of specific antibodies against hantaviruses, increased incidence of hantavirus pulmonary syndrome, and emergence of new hantavirus lineages. Authors ascribed this to an increased probability of contact between humans and rodents (vectors of hantaviruses) due to activities related to the highway pavement and construction itself, and deforestation fostered by economic activities (e.g. agriculture, farming, wood exploitation) developed along the influence area of the highway. Similarly, Salmon-Mulanovich et al., (2016) reported a perceived increase in rodent population after the paving of the Interoceanic Highway (IOH) along eight communities in the state of Madre de Dios, Peru, where participants mentioned seeing rodents in their *chacras* (small agricultural plot of land), communities, and houses. By easing access to healthcare services, roads were associated with a decrease of diarrhea and acute respiratory infections (ARI) in Brazil. The same paper also found an increase in malaria cases due to roads, likely due to ecosystem disturbances that led to more human-mosquito interactions. There have been mixed positive and negative results when examining highway construction and child/maternal health. Despite the construction of the Pacific Highway in Assis, Brazil, women still had low numbers of prenatal appointments. While the women in the community had access to more appointments than the rest of the state, the access was lower overall compared to the rest of Brazil. The addition of a road allowed for easier access to cesarean procedures, but there was no increase in them, suggesting that they were only used when medically necessary. There also continued to be a low number of institutionalized births (compared to the national average), suggesting the continued use of in-house childbirth.. ### Biomass burning Multiple papers looked at lung diseases. Brazil has implemented burning and deforestation restrictions in many areas in order to minimize poor air quality impacts and to help control escaped fires. While these laws can result in improved air quality and reduction of premature adult deaths, other studies have found that people are still potentially exposed to significant pollution in affected areas. In Theobroma, Rondônia, measurements taken during burning season found particulate matter (PM₁₀ and PM_2.5) levels above Brazilian and/or U.S National Ambient Air Quality standards, CO levels comparable to urban areas (despite sampling taking place in a rural village), and elevated benzene and HCHO levels when compared to other rural areas around the world. Another paper reported hospitalizations for children for asthma in the Brazilian Amazon increased in the arc of deforestation, though this could be explained by the rainy season and the resulting increase in fungi and mites. In contrast, when assessing the number of admissions due to pneumonia and its relationship with fires in the Mato Grosso, one study found that hospital admissions were randomly distributed through the municipalities, showing no apparent correlation with biomass burning. ### Dam power plants Only one paper discussed LUC related to dam power plants—these were associated with abundance of potential vector species of snail. These snails had the potential to spread *Centerocestus formasus*, schistosomiasis, and the parasite responsible for cercarial dermatitis. ### Livestock farming When assessing papers with LUC due to livestock farming and health outcomes, one paper reported that bat bites on humans were more common where there was no livestock, suggesting that livestock could have a protective effect on bat bites on humans. Bats fed on both humans and wildlife in areas of the Amazon affected by deforestation, which shows potential dietary switching. The results from these studies give rise to concerns of increased bat-to-wildlife transmission of rabies in the amazon, which may potentially create more intermediate hosts or reservoirs for the disease and maintenance of wild rabies in the region. Similarly, the Marajó region, an area with a low and widely dispersed population, had a large number of human rabies cases in 2004 that coincided with a reduction in cattle herds, most likely causing an increase on human feeding by bats as they sought new food sources. There is more bat-borne rabies in humans in the Brazilian Amazon than there is in the rest of Latin America, which could be explained by the dispersed populations and proximity to bat food sources. The overall high-risk areas for both human and bovine rabies are deforested areas, areas with livestock, and highways. ### Agriculture LUC related to agriculture was associated with multiple diseases. A study of different Brazilian localities found that the prevalence of *Trypanosoma cruzi* infection was higher in the Genipaúba locality (characterized by low human occupation and less environmental degradation than Ajuaí), and the Ajuaí locality (characterized by low human occupation, and some degradation through fruit harvest and plantations), as compared to other districts that have high human occupation, nearby secondary vegetation, and areas of only farm and pasture. There was also higher prevalence of detected antibodies among mammals and domestic mammals in Ajuaí compared to other study sites. *Phlebotomine* sandflies, potential carriers for leishmaniasis, were found in equal abundance across areas with different degrees of human occupations and land degradation. All sand fly species in this study were found in similar numbers through the land use categories suggesting high adaptability of sandflies, and the high adaptability could increase human-vector interactions, leading to more disease. Deforestation, and the resulting land use changes, were found to increase the presence of potential mosquito habitats, either through more fallen plant parts or through plant axils that could serve as refuge. ### Forest conservation The changing of land from unprotected forest to protected areas was associated with a decrease in malaria, diarrhea, and ARI. There are many possible explanations for this: the halt of deforestation decreasing interactions between people and forested areas that contain or can spread the diseases; a decrease in smoke emission from fires for deforestation; air filtration by trees; more clean water for hygiene; and filtration and water purification through forested areas by natural processes. # Discussion The scoping review utilized a systematic approach to explore associations—both positive and negative—between health problems and LUC in the Amazon forest of Latin America. The most striking observations were the lack of clear definitions for LUC, the lack of qualitative articles, a lack of studies exploring the potential positive health effects of LUC, and the predominance of studies coming from the Brazilian Amazon. Slightly more than half of the articles examined in this study revealed that LUC can lead to negative health consequences (beyond those excluded from this analysis that are associated to malaria and mercury), but other articles in our review also focused on the fact that there are often positive consequences associated with LUC. Gaps in knowledge include lack of agreed definition for LUC, a lack of representation for many different health outcomes, and the need for more research beyond the Brazilian Amazon. ## The concept of land use change LUC is becoming a growing interdisciplinary field of study, receiving increasing attention from scholars within both environmental and social sciences. Academic and non-academic organizations are designating LUC as a core concern, thus potentially fostering research to better understand its causes and results. There are multitudes of ways to examine LUC, from physical observations of cover change to studies on the ways humans interact with their ever-changing world. The present review focused on LUC within the tropical rainforest, and indeed there is a great deal of research on these topics in places such as Southeast Asia and Central America. However, many other types and categories of LUC exist, and having clear definitions of land use and land use cover classes is important in order to collect data more efficiently and to use and build modeling tools. With such a breadth of focus in this field, the use of “land use change” as a keyword, and defining it, would make seeking information within these varied focuses easier. It could aid in helping investigators to explain complex, interrelated factors that may affect their research. Another important task for the research community would be to work on the development of a framework—or various frameworks—that portray the complex linkages associated with LUC—whether some are causal relationships or co- emerging issues. For example, one of our difficulties was determining what articles to include in this review, since the linkages between LUC and health are sometimes direct, and at other times more indirect. Having a framework that is used by the various researchers working on this topic would allow us to contribute to building the knowledge about these linkages, modifying the frameworks based on findings if appropriate, and exploring complex topics regarding attribution and causality. Another reason to have a clear framework is that it could aid in helping direct us on issues to examine—information that would serve vulnerable populations who often have the most at stake in LUC activities. Rural populations, especially indigenous ones, are at a higher risk for exploitation by the government and corporations when it comes to LUC activities, through both the co-opting of land and through the unsafe and unhealthy conditions that can often result from LUC. These underserved populations have the most to gain from LUC and health research. Underserved and rural populations can also be major drivers of land use change themselves as they expand into new territories for property, food, and other resources. It is vital that stakeholders and organizations work with the local communities to also help them understand the cost-benefits of LUC so they can make informed decisions about their land. Though most of reviewed papers address changes in land use by reference to deforestation, none of them defined, and therefore measured, LUC explicitly. Streiker & Jacob (2016) and Bauch et al. (2015) present the concept in their titles and/or abstracts, and approach LUC beyond deforestation, providing multiple examples and/or references to both causes and consequences of LUC (such as urbanization, agricultural intensification, climate change and ecosystem change). Other articles, while listing LUC as keyword, do not provide an explicit definition. Thus, we found most of the papers (10) via the specific LUC topic keywords, instead of by the term “land use change” in and of itself. ## Health outcomes Health effects related to smoke exposure and air quality were the most frequently addressed topic in the reviewed articles, especially in Brazil. This could possibly be explained by the multitude of laws regarding biomass burning that are now in effect in Brazil, and the desire to see their effectiveness. In contrast, there were no studies addressing food security and nutrition, mental health, and accidents or injuries, demonstrating a lack of representation of these issues among studies that link LUC to wellbeing and health. Many of the papers we examined studied human-vector interactions. This is unsurprising, as vectors are a major cause for disease and can be relatively easily measured. There are also many examples of diseases transmitting as a result of increased contact between human population and wildlife, such as Ebola and yellow fever. Only one qualitative research satisfied our inclusion criteria. This should not be interpreted as an absence of qualitative research—although very few studies using qualitative methods came across in the initial search—but rather that they are not well represented when researching the link between LUC and health. The drivers of LUC are human-based, and usually revolve around people seeking better opportunities through livelihoods. Some studies show that while people living in newly settled areas know that land degradation is bad, they believe that the economic opportunities outweigh the negatives. Knowing this can help organizations shape their outreach programs and policy recommendations to ensure that people feel economically stable while still securing a healthy environment. ## Importance of more geographical variety in studies The Amazon Basin encompasses 8 different countries (Brazil, Bolivia, Peru, Ecuador, Colombia, Venezuela, Guyana, Suriname and French Guiana); however, nearly all the papers we reviewed were about studies in Brazil. The majority of the Amazon rainforest is in Brazil, so its importance in management cannot be ignored. However, the differences between countries and areas of the Amazon in land use, human demographics, economic use, environment, and regulations call for more geographical diversity in Amazonian studies. The differences in ecosystems across the Amazon basin make it difficult to apply one study to the entire region. While it is tempting to characterize the Amazon as one giant rainforest, it is made up of multiple ecosystem types, including mangroves, moist forest, mountain forest, and swamps. This diversity in ecosystem leads to dramatic biodiversity, which can lead to different diseases spread to humans via the vast differences in plant and animal species. When working with any health topic, the diversity of people is important to consider. The indigenous people of the Amazon are made up of about 350 ethnic groups, some of which are still isolated. Amazonian tribes that are geographically close can sometimes have very different diets, as the foods harvested and consumed are a means of cultural expression. This can lead to not only nutritional differences, but differences in disease exposure. ## Drivers of land use change The drivers of LUC, and subsequent relationships with the government, also can vary. The illegal logging and mining industries of the Amazon are a looming presence. For example, in Colombia, almost all forest clearing is illegal. However, Colombia has made strides in its protection of indigenous lands, giving ownership of the majority of the remaining Amazon to its indigenous peoples. Brazil, having the higher GDP and more of the Amazon, has a higher potential to impact the rainforest than other states. Part of this is apparent in the construction of the trans-oceanic highway, which had significant foreign influence from important economic partners. In Suriname, indigenous communities are working to create protected areas that make up almost half of Suriname, and contains much of the regions forest. ## Limitations One of our barriers was language. There were 24 papers that were written in Portuguese. Because we had no Portuguese-speakers on our team, we were unable to include these papers, limiting our results. Spanish publications were underrepresented. The addition of grey literature would have enriched our findings; however, due to limited human resources and time constraints, we were unable to search these. As described previously, another problem the researchers encountered when conducting this search was drawing specific conclusions whether an article covered a specific LUC issue, and to what extent the LUC occurred, making drawing conclusions difficult. # Conclusion LUC will continue to be an important research topic, both globally and locally, as human expansion continues and shifts. These changes have the potential to do great harm to humans, but with continued research and advocacy, those harms could be mitigated. Our research found that most papers lack a clear definition of LUC, demonstrate mostly negative impacts on human health, and few papers study qualitative aspects of LUC. Moving forward, the authors recommend offering clear definitions of LUC and the way it is measured, exploring the social dimensions of LUC, and performing more qualitative studies to assess these better. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Neuropeptide Y (NPY) is one of the most common peptides in the brain and is an abundant neurotransmitter in the peripheral sympathetic nervous system (SNS). NPY has been shown to play a role in energy metabolism, appetite regulation, cardiac rhythm, blood pressure, smooth muscle contraction and relaxation. Cumulative evidence also suggests that NPY acts as a metabolic signal that may contribute to obesity, hyperinsulinemia, and hyperglycemia. Intracerebroventricular injection of NPY and overexpression of NPY in the paraventricular nucleus (PVN) of the hypothalamus, for example, have been shown to temporarily increase food intake and promote insulin release. The activity of NPY in cellular metabolism appears to be mediated through its ability to bind the transmembrane domain G protein coupled receptors NPY Y1–Y5. These receptors are found in a broad array of tissues including those involved in metabolism, like adipose tissue and liver. There is evidence suggesting that NPY influences metabolic function in peripheral tissue mostly via Y1 and Y5 receptor signaling. However, in a stress-induced obesity model, NPY induction in fat correlated with insulin resistance that could be attenuated by blockage of the Y2 receptor. The mechanism by which NPY contributes to insulin resistance in adipose tissue, therefore, is not well understood. In this study we investigated the molecular mechanisms of NPY that contribute to peripheral insulin resistance. Our hypothesis was that NPY contributes to peripheral insulin resistance in adipose tissue via the Y5 receptor. To test this hypothesis, we created an insulin resistant *in vivo* model by injecting NPY in the PVN of the hypothalamus in rats. Eight weeks after injection, euglycemic-hyperinsulinemic clamp and intravenous glucose tolerance tests confirmed that the rats had developed insulin resistance in adipose tissue. As the main glucose metabolism pathway, PI3K-AKT signaling changes in adipose tissue were assessed in our study. Glycogen synthase kinase 3 (GSK3) α and β were also detected, since there is evidence suggesting that GSK3 contributes to the induction by insulin resistance independently of insulin receptor signaling or PI3K-AKT activity. NPY Y1 and Y5 receptor antagonists were used to unravel the mechanism disorders induced by NPY in adipocytes. # Methods and Procedures ## Reagents and antibodies Dexamethasone, 3-isobutyl-1-methylxanthine (IBMX), bovine insulin, human NPY, and 2-deoxy-D-glucose were purchased from Sigma-Aldrich (St Louis, MO, USA). The Y5 receptor antagonist L-152,804 was purchased from Tocris Bioscience (Bristol, UK). The Y1 receptor antagonist BIBP-3226 (Diphenylacetyl-D- Arg-4-hydroxybenzylamide) was purchased from Bachem (San Carlos, CA, USA). 2-deoxy-D-\[³H\] glucose (2-\[³H\] DG) was obtained from Amersham Life Sciences (Buckinghamshire, UK). Antibodies for immunoblot and immunofluorescence assays included: anti-NPY, anti-GSK3α, anti-GSK3β, anti-pGSK3α^Ser21, anti-pGSK3β^Ser9 (Santa Cruz Biotechnology, CA, USA); anti-PI3K, anti-PI3K p85, AKT, anti- pAKT^Ser473 (Cell Signaling Technology, MA, USA). Secondary antibodies conjugated to HRP and Alexa Fluor dyes for immunoblotting and immunofluorescence were purchased from Life Technologies (Grand Island, NY, USA). Antibodies were used according to manufacturer's instructions. ## Animals 6-week old male Sprague-Dawley rats (240–260 g) were purchased from the Research Institute of Surgery Experimental Animal Center of the Third Military University (Chongqing, China) and housed individually. Initially, all rats were maintained under a controlled environment (temperature 20 ± 3°C, humidity 60 ± 5%, 12 h dark-light cycle) with regular chow consisting of 5% fat, 55% carbohydrate, 23% protein, 7% ash and 10% fiber with a total caloric value of 3.2 kcal per gram. Food and water were available *ad libitum*. Rats were used in the experiment at 9 weeks of age (290–310 g). At this time, rats were fed either the regular chow (low-fat diet, LFD) or a high-fat diet (HFD), consisting of 50% fat, 17% carbohydrate, 25% protein, 3% ash and 5% fiber with a total caloric value of 4.7 kcal per gram. Rats were handled and cared for according to the Guide for the Care and Use of Laboratory Animals, and all procedures were approved by the Ethics Scientific Committee of the Third Military Medical University. ## Lentivirus NPY production and hypothalamic PVN injection The chemically synthesized rat NPY gene (RefSeq ID: NM_012614) was cloned into the lentiviral vector pUbi-IRES-Cherry with Agel and Nhel restriction enzymes cutting sites, the final commercial recombinant lentiviral particles containing NPY (LV-NPY-Cherry) were purchased from Shanghai GeneChem Co., Ltd. China. The empty lentiviral expression vector pUbi-Cherry (LV-Cherry) was used as a negative control. All lentivirus batches used for experiments had comparable titers ranging from 2×10⁸ to 3×10⁹ transducing units (TU)/mL. Viral suspensions were stored at -80°C until use and were briefly centrifuged and kept on ice immediately before they were injected into the PVN of rats. At the age of 9 weeks, rats were anaesthetized with an intraperitoneal injection of sodium pentobarbital (Sigma, 36 mg/kg) and placed on a stereotaxic frame (David Kopf Instruments, Tujunga, CA) connected to a nanoliter syringe pump. After exposure of the skull surface, a burr hole was made in the skull, and the nuclear injection of LV-NPY-Cherry or LV-Cherry was carried out using a 10 μl syringe (Hamilton, Switzerland). The rats were allocated randomly to each of the indicated treatment groups and received LFD + LV-Cherry injection (LFD), HFD + LV-Cherry injection (HFD), LFD + LV-NPY Cherry injection (LFD+NPY), or HFD + LV- NPY Cherry injection (HFD+NPY). Each rat (n = 8 per group) was subjected to two injection sites, chosen according to the Rat Brain Atlas as follows: point 1, 1.8 mm posterior to the bregma, 0.5 mm left lateral, 8 mm deep; point 2, 1.8 mm posterior to the bregma, 0.5 mm right lateral, 8 mm deep. 2 μL of lentiviral suspension containing 2 × 10⁸ TU/mL was injected in each point at a rate of 0.2 μL/min. After completion of the injection, the needle was left in place for 5 min before withdrawal to ensure lentiviral diffusion into the tissue. Rats received an injection of 0.05 mg/kg buprenorphine (Schering-Plough, Maarssen, Netherlands) subcutaneously for analgesia before the surgery, and then every 8 h post-surgery for the next two days. Furthermore, rats were monitored every 12 h and allowed to recover for 7 days after surgery. 3 rats died during or after the surgery because of serious bleeding and respiratory depression, and other 3 rats received lentiviral injection to ensure the maintenance of 8 rats in every group. ## Evaluation of insulin resistance Intravenous glucose tolerance test (IVGTT): 8 weeks after LV injection, IVGTT was conducted on unanesthetized animals. Animals were fasted overnight and a 27-gauge butterfly catheter was placed in the saphenous vein for infusion of a bolus dose of 500 mg/kg body weight of a 50% dextrose solution. Blood samples were collected from the tail at 0, 5, 10, 20, 30, 45, 60, 90, and 120 min after glucose administration. The homeostasis model assessment (HOMA-IR) also was used to detect insulin resistance as described previously. HOMA-IR (mM × μU/mL) = fasting glucose (mM) × fasting insulin (μU/mL) / 22.5. Euglycemic-hyperinsulinemic clamps: Briefly, 8 weeks after LV injection, 3 rats per group were anesthetized with sodium pentobarbital for jugular and artery catheterization as previously described. Under aseptic conditions, the right jugular vein and the left carotid artery were cannulated with PE-50 cannulas attached to a 1 mL syringe containing 5 U heparinized saline. 50 U heparinized saline were used to flush the cannulas every 24 h to avoid clotting during the recovery period. For analgesia, buprenorphine was used as described above and rats were allowed to recover for 48 h. After that, rats were fasted 8 h and the euglycemic-hyperinsulinemic clamps were performed while they were awake and unrestrained. Continuous infusion of 5 mU/kg min insulin (HumulinR insulin, Eli Lilly, Indianapolis, IN) was performed to acutely increase and maintain high levels of insulin for 2 h. The arterial blood glucose concentration was measured every 10 min using a Freestyle Lite glucose meter (Abbott Diabetes Care, Alameda, CA USA) and clamped at 5–6 mM by using a variable rate of 20% glucose infusion delivered via the jugular cannula. The glucose infusion rate during the second hour of clamp (GIR_60-120) was used as the response parameter of potency of whole body insulin action. Glucose metabolic rate in individual tissues was estimated using a technique which has been described elsewhere. 2-\[³H\] DG (50 μCi) was administered as a bolus 45 min before the end of the clamp. Blood samples (100–500 μL) were harvested for plasma glucose, insulin and 2-\[³H\] DG estimations during the clamp experiment. At the end of clamps, rats were anesthetized with sodium pentobarbital (100 mg/kg), and tissues were rapidly removed and frozen in liquid nitrogen for subsequent analysis. Plasma 2-\[³H\] DG concentration and the tissue accumulation of phosphorylated 2-\[³H\] DG were estimated as described previously. The glucose utilization index (Rg’), an estimate of tissue glucose uptake, was calculated as described by James et al. ## Basal metabolism indexes Rats were fasted for 12 h then individually housed in metabolic cages, where food and water were available *ad libitum* for exactly 24 h. After this, rats were monitored every day for eight weeks after LV injection. Food intake, body weight, and rectal temperature were recorded once a week. At the end of experiment, the body weight and naso-anal length (cm) of the animals was measured, and the Lee index was used to assess obesity by calculating the ratio between the cube root of the body weight (g) and the naso-anal length (cm) of the animals multiplied by 10. ## Tissue and sample preparation Rats were sacrificed 8 weeks after LV injection by administering sodium pentobarbital, were intracardially perfused with 4% paraformaldehyde (PFA), and the brains were removed. Blood, skeletal muscle and white adipose tissues (WAT), including subcutaneous, epididymal and retroperitoneal adipose tissue were collected prior to perfusion. Fasting venous blood was collected and serum was separated by centrifugation (2000 *x g*, 15 min), adipose tissue was isolated, weighed, and immediately frozen in liquid nitrogen. Perfused rats’ brains were dehydrated gradually with cane sugar, embedded in optimal cutting temperature (OCT) compound (Sakura Finetek, Tokyo, Japan), and stored at -80°C until they were cut into sections. ## Measurements of blood indexes Blood glucose levels were measured using a Freestyle Lite glucose meter. Serum glycated hemoglobin A1c (HbA1c) concentrations were determined using a commercially available ELISA kit (R & D Systems, Minneapolis, MN). Insulin, triglycerides and total cholesterol concentrations were measured by radioimmunoassay and enzymatic assays, respectively. ## Cell culture The mouse fibroblast cell line 3T3-L1 (Cell Bank of the Chinese Academy of Sciences, Shanghai, China) was maintained and differentiated as previously described. Briefly, differentiation was induced with high glucose Dulbecco’s modified Eagle’s medium containing (DMEM) 10% fetal bovine serum (FBS), 10 μg/mL insulin, 1 μM dexamethasone, and 0.5 mM IBMX. After 2 days, insulin, dexamethasone and IBMX were removed, cells were maintained in 10% FBS media with 10 μg/mL insulin for another 2 days, then the insulin was removed, and cells were cultured in high glucose DMEM with 10% FBS until the day of experiments (days 7–8). NPY, insulin, Y5 receptor antagonist L-152,804 and Y1 receptor antagonist BIBP-3226 were dissolved in DMSO, and then diluted in the appropriate test solvent. ## Glucose consumption and glucose uptake Glucose consumption was conducted as described previously. The differentiated 3T3-L1 adipocytes were plated into 96-well plates, incubated with DMEM containing 0.2% bovine serum albumin (BSA) for 12 h and treated with insulin for 2 h or various concentrations of NPY for 12 h for detecting basal glucose consumption\[–\]. To test the effect of receptor antagonists on insulin- stimulated glucose consumption intervened with NPY, the adipocytes were pre- treated with L-152,804 or BIBP3226 8 h, then subsequently treated with NPY for 12 h with insulin for 2 h. The glucose concentration in the culture medium was determined using the glucose oxidase method. The amount of glucose consumption was calculated by subtracting the glucose from the control well. 2-\[³H\] DG uptake was measured according to a method described previously. In brief, insulin-simulated glucose uptake was studied in differentiated monolayers. The cells were cultured in 12-well plates and starved 12 h before treatment. Then cells were pre-treated with L-152,804 or BIBP-3226 8 h and subsequently treated with NPY for 12 h and insulin for 20 min in KRH buffer (20 nM HEPES, 136 mM NaCl, 4.7 mM KCL, 1.25 mM MgCl₂, 1.25 mM CaCl₂, pH 7.4) at 37°C for 30 min. The cells were subsequently incubated with 2-deoxy-D-glucose (0.1 mM) and 2-\[³H\] DG (0.5 μCi/mL) for 10 min, reaction was stopped quickly with cold PBS, and cells were solubilized in 0.4 mL of 0.1 M sodium hydroxide. Radioactivity of 2-\[³H\] DG was determined in the whole cell lysates using a Beckman LS6500 scintillation counter. ## Immunofluorescence OCT-embedded brain tissues were cut into 4 μm serial sections. Sections were incubated with 0.3% Triton X-100 at 37°C for 20 min, 5% normal goat serum at RT for 10 min, then incubated with anti-NPY primary antibody (1:100 dilution) at 4°C overnight; as a negative control, sections were incubated without primary antibody. After they were washed three times, sections were incubated with the appropriate secondary antibody directly conjugated with FITC for 1 h, stained with DAPI for 5 min at RT, then washed extensively, rinsed in ddH₂O and mounted using gelvatol. Confocal laser imaging was performed using a Leica TCS-SP5 confocal scanning laser microscope (Leica Microsystems, Germany), and the level of NPY overexpression in PVN was quantified by ImageJ software. ## Western blotting 3T3-L1 adipocytes were plated into 6-well plates, incubated with DMEM containing 0.2% BSA for 12 h and pre-treated with 100 μM L-152,804 or 100 μM BIBP-3226 8 h, then subsequently treated with 100 nM NPY for 12 h with 100 nM insulin for 2 h. Whole cell lysates were obtained using a commercial cell protein extraction buffer (Thermo Scientific Pierce, USA) with 0.5 mM phenymethanesulfonyl fluoride (PMSF) and 10 mM Ser/Thr phosphatase inhibitor. Retroperitoneal adipose tissue was also homogenized in tissue protein extraction buffer with PMSF and phosphatase inhibitor. Samples were centrifuged at 10,000 *x g* at 4°C for 15 min, which resulted in the separation of a fat layer. The supernatants were removed making sure there was no any residual fat and centrifuged at 10,000 *x g* at 4°C for 15 min again. The resulting supernatant solutions were used for Western blot analysis. Protein concentrations were measured using a NanoDrop2000 (Thermo Scientific Pierce, USA). Equal amounts of protein (50 μg) were subjected to SDS-PAGE and transferred onto polyvinylidene fluoride (Millipore, Bedford MA) membranes. Membranes were blocked with 3% BSA/PBS-Tween20 0.1% at RT for 1 h and incubated with primary antibodies (1:500–1:1000) at 4°C overnight. The membranes were then incubated with the appropriate HRP-linked secondary antibodies at RT for 1 h. Blots were developed using an enhanced chemiluminescence method (Thermo Scientific Pierce, USA). ## Statistical analyses For analysis of body weight, daily food intake and body temperature of rats, two-way ANOVA with repeated measures was applied with factors of group and time. Tukey's post hoc test was used when ANOVA indicated significance. One-way ANOVA was applied to analyze other indexes in *in vivo* and *in vitro* experiments. Unpaired Student’s t tests were used to compare the means of two groups in both *in vivo* and *in vitro* experiments. All results were analyzed using the GraphPad Prism software (version 6.0; Graphpad, San Diego, CA; [www.graphpad.com](http://www.graphpad.com)). Statistical significance was defined as *P*\<0.05. # Results ## Constitutive overexpression of NPY in the PVN of rats The LV-Cherry (vehicle control) or LV-NPY-Cherry was injected into the PVN of rats. Eight weeks after injection, the expression of the vectors was measured using immunofluorescence. Rats were only included in the study when the reporter protein (Cherry) expression was located exactly in the PVN. As shown in, both vectors were still expressed in the PVN 8 weeks after injection. Additionally, the expression of NPY and Cherry co-localized, as shown in the merged image. Comparing with LFD group, LV-NPY-Cherry injection in PVN induced a 3.89-fold increase of NPY protein expression in LFD+NPY group (*P*\<0.05); and a 3.09-fold increase of NPY protein expression in HFD+NPY group (*P*\<0.05) compared to HFD group. These findings suggest that we were successfully able to overexpress NPY in the PVN of rats, and overexpression of NPY can be sustained for more than 8 weeks in the hypothalamic PVN of rats by injection of recombinant lentivirus- mediated NPY particles *in situ*. ## Overexpression of NPY contributes to insulin resistance in rats The euglycemic-hyperinsulinemic clamp assay showed that HFD and NPY overexpressing rats had lower GIR_60-120 than LFD rats, which indicated that peripheral insulin resistance was induced in rats fed with HFD or rats overexpressing NPY for 8 weeks. Otherwise, even though the plasma glucose levels did not increase, the insulin levels measured during a 120 min intravenous glucose tolerance test (IVGTT) increased in the three groups, which is consistent with the result of euglycemic-hyperinsulinemic clamp assay. Moreover, HOMA-IR test confirmed previous results, demonstrating that rats overexpressing NPY with or without a HFD develop insulin resistance. Interestingly, although compared with the LFD group, the other three groups had higher fasting insulin levels, a HFD did increase fasting glucose levels in both the control and NPY-overexpressing rats, but constitutive overexpression of NPY in the PVN alone did not induce obvious higher fasting glucose levels. ## Overexpression of NPY induces obesity and insulin resistance in adipose tissue The body weight of HFD, LFD+NPY and HFD+NPY rats were significantly higher than those of LFD rats, especially during the last period of the experiment. Noticeably, however, the food intake per day of LFD+NPY and HFD+NPY groups was greater than the LFD and HFD groups at the beginning of the experiment but leveled out by the week 4. A diet high in fat or overexpression of NPY did not alter the rats’ body temperature. The serum HbA1c, triglyceride, and cholesterol concentrations were not different in the control versus NPY-overexpressing rats; however, they were significantly elevated in rats fed a HFD, and in rats overexpressing NPY who were fed a HFD. This finding correlates with the results presented in, showing that a HFD also increases fasting glucose levels. A HFD and chronic overexpression of NPY in the hypothalamic PVN of rats increased the ratio of total adipose tissue weight to body weight, which is consistent with the Lee index, the indicator of obesity. These results indicated that both HFD and NPY overexpression induced obesity, but the latter didn’t increase glucose, HbA1c, or blood lipid levels. On the other hand, peripheral insulin resistance had been confirmed by the euglycemic-hyperinsulinemic clamp test and IVGTT. To further reveal the contribution of individual tissues to peripheral insulin resistance, the glucose utilization index (Rg’), an estimate of tissue glucose uptake, of adipose tissue and muscle was examined. The results indicated that the 2-\[³H\] DG uptake of adipose tissue decreased in HFD, LFD+NPY and HFD+NPY groups, while there were not obvious differences in 2-\[³H\] DG uptake in muscle among all groups. ## NPY modulates the PI3K-AKT and GSK signaling pathways Compared with rats fed a LFD, the PI3K protein levels in the adipose tissue of rats overexpressing NPY fed a LFD or HFD decreased slightly, although the AKT, GSK3β, and GSK3α protein levels did not change. Specifically, constitutive overexpression of NPY in PVN of rats dramatically decreased the phosphorylation of GSK3β, PI3K and AKT independently of the diet they were fed. However, the phosphorylation of GSK3α significantly decreased in the HFD, LFD+NPY and HFD+NPY groups. ## NPY inhibits glucose consumption and 2-\[³H\] DG uptake in 3T3-L1 adipocytes via the NPY Y5 receptor In *in vitro* experiments we found that NPY can directly decrease the glucose uptake in 3T3-L1 adipocytes. High dose NPY (100 nM) reduced basal glucose consumption, whereas lower dose NPY (1 and 10 nM) failed to affect the basal glucose consumption in 3T3-L1 adipocytes. Although NPY can also reduce the insulin-simulated glucose consumption, BIBP-3226, a NPY Y1 receptor antagonist, did not reverse the effect of NPY treatment in the adipocytes. NPY Y5 receptor antagonist L-152,804 (1–100 μM) obviously reversed the restrain of NPY on insulin-simulated glucose consumption. Furthermore, 2-\[³H\] DG uptake experiments demonstrated that the NPY Y1 receptor antagonist did not contribute to the insulin-simulated 2-\[³H\] DG uptake of 3T3-L1 adipocytes incubated with 100 nM NPY, whereas the NPY Y5 receptor antagonist does, suggesting that the NPY Y5 receptor is responsible for the observed NPY effects. ## NPY changes the PI3K-AKT and GSK signaling pathways in 3T3-L1 adipocytes via the NPY Y5 receptor NPY inhibited the phosphorylation of GSK3α, GSK3β, PI3K and AKT significantly, although NPY did not change the AKT, PI3K, GSK3α and GSK3β total protein levels in 3T3-L1 adipocytes. Treatment with the NPY Y5 receptor antagonist reversed the suppression, whereas the NPY Y1 receptor antagonist did not, corroborating the previous findings that the NPY Y5 receptor is responsible for the effects of NPY in adipocytes. # Discussion In this study, we found that overexpression of NPY can be sustained for more than 8 weeks in the hypothalamic PVN of rats by injection of recombinant lentivirus-mediated NPY particles *in situ* with a \>3-fold protein induction compared with vector injection. NPY overexpression temporarily increased daily food intake after injection and then leveled out by the week 4. A similar phenomenon was found in other comparable studies. Tiesjema et al. have shown that NPY overexpression in PVN significantly induced daily food intake during the 6 weeks after injection of recombinant adeno-associated viral particles containing NPY cDNA in rats when allowed to eat *ad libitum* until “a certain body weight” not specified by the authors was achieved. Noticeably, at the beginning of our experiments the rats’ body weights (290–310 g) were significantly higher than those of their experiment (220–250 g). This could partly explain the fact that we found an increase in food intake only up to week 3 because our heavier rats could be closer to that “certain body weight”. At the end of the experiment, the body weights of the NPY groups increased significantly compared with the LFD group, which was consistent with other experiments. Moreover, the relative weight of adipose tissue and Lee index increased in HFD and NPY overexpression groups, which suggested that either HFD or long-term NPY expression in PVN result in obesity. However, we failed to find any obvious differences in body weight, adipose tissue weight, and Lee index among rats in the HFD group and those in the NPY-overexpressing groups, which might suggest that the effect of NPY on adipose tissue does not synergize with the effect caused by feeding a HFD. We further found both NPY overexpression in PVN and HFD can decrease the average GIR_60-120 of euglycemic-hyperinsulinemic clamps, increase total plasma insulin AUC 0 to 120min of IVGTT, HOMA-IR and fasting insulin levels, which indicated that both NPY overexpression and HFD induce peripheral insulin resistance. To reveal the contribution of individual tissues to the peripheral insulin resistance, the glucose utilization index (Rg’) of adipose and muscle was detected. The obvious decreased in glucose utilization inadipose tissue was demonstrated in insulin resistance groups, and even there was a trend of decreased in glucose utilization in muscle. These results suggested that adipose tissue might have a higher contribution to the peripheral insulin resistance seen in HFD and NPY overexpressing rats. In this study, insulin resistance was established in NPY overexpressing rats independently of the diet, and there was no difference in body and adipose tissue weights between HFD and LFD+NPY groups. HFD induced higher fasting glucose levels, serum HbA1c concentration and dyslipidemia while NPY overexpression did not induce any effects on blood glucose or lipids. However, it could be difficult to conclude from the present data whether insulin resistance in the NPY overexpression group was a result of direct or indirect effects through obesity. Even in a pair-feeding study, restriction of food intake also could not eliminate the increase in adiposity induced by NPY. Interestingly, a short-term NPY overexpression study suggested that NPY influences leptin and insulin secretion independently from food intake and obesity. Additionally, activation of the parasympathetic nervous system by NPY and the abundance of autonomic nerves in adipose tissue may partly explain how NPY directly contributes to the establishment of insulin resistance in adipose tissue in rats. However, *in in vivo* experiments we also found that NPY can directly decrease the glucose uptake in adipocytes. Our results indicate that the IR in adipose tissue might be driven by central NPY action, but it is hard to explain whether the development of IR is a direct consequence of the central or the peripheral Y5 signaling or both. More *in vivo* experiments need to be performed to explore the definite ways by whichthe central NPY affects the adipose tissue, such as throughthe sympathetic nerves, the adrenal medulla, platelets and various cell types within WAT. However previous studies suggest a plausible mechanism that explains how central NPY exerts a function on WAT. It is well known that the hypothalamus is a major source of forebrain input into the SNS. Immunostaining after pseudorabies virus (a transneuronal tract tracer) injection into fat revealed that PVN was one of the sites that modulated SNS outflow to the WAT. Recently, the definition of hypothalamus-SNS-adipose tissue was used to clarify the feasibility of crosstalk between hypothalamus and adipose tissue. Moreover, there is data that suggests that the pituitary-adrenal ensemble and other endocrine cues may be engaged by prolonged central administration of NPY. It is verified that increased hypothalamic NPY inhibits sympathetic nerve system outflow and suppresses catecholamine release, mainly norepinephrine (NE). Furthermore, NPY colocalized with NE in peripheral SNS as an adrenergic cotransmitter which exerts pleiotropic activities either synergistic or antagonistic with NE. Recent data also demonstrated that the release of NPY as a sympathetic neurotransmitter directly into the WAT leads to abdominal obesity with the depletion of NE in the adipose tissue.Moreover, the increased Y5 receptor expression in adipose tissue of obese animals and differentiated adipocytes\[–\] proved valuable cue to explain why NPY overexpression in PVN induced obesity and insulin resistance partly via Y5 receptor in this experiment. Furthermore, we also observed that the phosphorylation of PI3K and AKT was inhibited by NPY overexpression and HFD, and NPY aggravated the down regulation on PI3K-AKT pathway by HFD in adipose tissue. Recently, more evidence suggests that GSK3 contributes to the development of peripheral insulin resistance and type 2 diabetes. Moreover, high glucose induced insulin resistance by decreasing the phosphorylation of GSK3β, triggering the ubiquitination and degradation of insulin receptor substrate (IRS1), which did not require insulin receptor signaling or PI3K/AKT activity. In the current study, NPY overexpression dramatically decreased the phosphorylation of GSK3β. However, HFD failed to induce any obvious change in the phosphorylation of GSK3β in adipose tissue, which implies that the differential phosphorylation of GSK3β may be one important factor to explain a different mechanism of insulin resistance in adipose tissue induced by NPY overexpression and HFD. NPY-induced insulin resistance of adipose tissue was consistent with dexamethasone-induced insulin resistance in muscle, which also decreased insulin-stimulated AKTand GSK3 phosphorylation, and completely blocked the ability of insulin to dephosphorylate and activate glycogen synthase without reducing expression of these proteins. This study also demonstrated that high dose (100 nM) NPY can inhibit basal, insulin-simulated glucose consumption, and insulin-stimulated 2-\[3H\] DG uptake in 3T3-L1 adipocytes, suggesting that NPY can directly affect glucose uptake and consumptions in adipocytes. This result is consistent with other reports that found that NPY 100 nM inhibited fat metabolism in adipocytes. Y1 and Y5 receptors are most likely the receptors that contribute to the metabolic effects of NPY on adipocytes. In this study, the NPY Y5 receptor antagonist L-152,804 specifically reversed the inhibition of NPY on glucose consumption and 2-\[³H\] DG glucose uptake in 3T3-L1 adipocytes. Moreover, previous studies report that L-152,804 did not show significant affinity for Y1, Y2, and Y4 receptors even at a dose of 10 mM. Therefore we conclude that the effect of NPY on glucose consumption depends mostly on the Y5 receptor. In addition, our study further proved that NPY not only inhibits phosphorylation of PI3K and AKT proteins, but that it also decreases the phosphorylation of GSK3α and GSK3β protein significantly in 3T3-L1 adipocytes. Once again, the NPY Y5 receptor antagonist L-152,804 relieved the suppression, but the Y1 receptor antagonist BIBP-3226 did not contribute to this relief. However, NPY did not change the AKT, PI3K, GSK3α and GSK3β total protein levels in 3T3-L1 adipocytes. In summary, this study has demonstrated the establishment of insulin resistance in rats’ adipose tissue, after long-term constitutive overexpression of NPY in the PVN of hypothalamus. NPY may affect glucose metabolism by modulating the phosphorylation of PI3K, AKT and GSK3 in adipose tissue and adipocytes, and it is possible that the NPY Y5 receptor contributes to glucose metabolism disorder induced by NPY. This study might provide more direct pharmacological evidence that using NPY receptors antagonists could prove beneficial in the treatment of diseases induced by NPY disorders. We thank Jaramillo MC Ph.D and Rojo de la Vega M Ph.D (University of Arizona, USA) for assistance on writing and grammar modification. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: ML JZ ZX SZ. Performed the experiments: ML JZ DL LZ. Analyzed the data: ML LZ. Contributed reagents/materials/analysis tools: ML JZ DL ZX SZ. Wrote the paper: ML JZ DL ZL ZX SZ.

# Introduction Anemia, a condition characterized by insufficient hemogolobin (Hb) concentration to meet the oxygen demand of the tissue, affects nearly one-quarter of the world’s population. Distinct Hb cut-offs to define anemia varies by age, sex, altitude, smoking, and pregnancy status and are available in guidelines put forward by the World Health Organization (WHO). For non-pregnant women anemia is defined as Hb\<12 g/dL. Whereas, during pregnancy the cut-offs varies by gestation ageas a result of increased blood volume and plasma expansion, and is \<11g/dL during the first trimester, \<10.5 g/dL during the second trimester and \<11 g/dL during the third trimester. In sub-Saharan Africa (SSA) anemia is prevalent among women of reproductive age and young children. In year 2011, it was estimated that 496 million (29%) non-pregnant and 32 million (38%) pregnant women had anemia globally, and the World Health Organization (WHO) set a target to reduce by 50% anemia among women of reproductive age by year 2025. The etiology of anemia is multifactorial with the prevalence and causes varying considerably in different areas of the world by population group, region, residence (urban or rural), socioeconomic status (SES), and general environmental factors. Iron is a key component of Hb, and its deficiency is estimated to be responsible for 50% of all anemia cases in SSA. Iron demands are high during pregnancy and increased parity and gravidity and short interpregnancy interval can therefore substantially reduce maternal iron reserves. Additional risk factors include other micronutrient deficiencies (folate, vitamins B₁₂ and A), malaria, soil transmitted helminths (STH), chronic infections like HIV, genetic disorders including sickle cell anemia, glucose-6-phosphate dehydrogenase (G6PD) deficiency and, α-thalassemias as well as other inflammatory conditions. The consequences of anemia and ID vary depending on the severity, the population groups and living conditions but includes reduced work capacity in adults and neurocognitive impairment in young children. Furthermore, it is well established that anemia during pregnancy is associated with intrauterine growth retardation, preterm delivery, low birthweight, and perinatal mortality.To address this, pregnant women in Tanzania are prescribed iron and folate supplements but compliance is poor due to side effects. Vitamin B₁₂ deficiency causes megaloblastic anemia and infant vitamin B₁₂ deficiency, but it is not included in the micronutrients supplementation program because of insufficient data supporting its implementation. Although anemia during pregnancy has been associated with adverse pregnancy outcomes, the proportion of women who are already anemic or had depleted iron stores just before conception has not been well investigated in SSA. Therefore, the aim of this study was to determine the prevalence, types and risk factors associated with preconception anemia among women of reproductive age in a rural setting of northeastern Tanzania. # Materials and methods ## Ethical statement The study received ethical approval from the Medical Research Coordinating Committee of the National Institute for Medical Research (reference number NIMR/HQ/R.8a/Vol. IX/1717).Written informed consent or thumbprint (for illiterate women) was obtained prior to enrolment. All study procedures were performed according to good clinical and laboratory practices and the Declaration of Helsinki. ## Study design and setting This cross sectional study was conducted as part of a community-based epidemiological study entitled “Foetal exposure and epidemiological transition: the role of anemia in early life for non-communicable diseases in later life” (FOETALforNCD) fromJuly 2014 to December 2016 in Korogwe and Handeni districts, Tanga region, Tanzania. The aim of theFOETALforNCD project was toevaluate fetal growth alterations, placental development, and newborn susceptibility to non- communicable diseases in later life, following exposure to maternal anemia before and during pregnancy. The study population composed of women of the reproductive age. The analyses presented here utilized baseline data from women enrolled before they became pregnant. Inclusion into this study was based on their likelihood to conceive during the study period. To be included, women had to be aged 18–40 years, not be using modern contraceptive methods (except condom), or not be sub-fertile (defined as failure to conceive for two or more consecutive years for women who were trying to become pregnant), or not be pregnant at the time of enrolment (negative urine pregnancy test, HCG Vista Care Company, Shandong China), or not have a baby less than nine months old and live in an accessible area, and be willing to receive antenatal care and deliver at Korogwe District Hospital. ## Participant identification and recruitment Different stakeholders including village leaders, health care providers, opinion makers as well as community members were sensitized about the study goals and aims through village and health facility meetings prior to the implementation of the FOETALforNCD study. The primary means of identifying and recruiting eligible women was through contact at the household level within each village. Trained field workers made door-to-door visits to each household to explain the study, enumerate all women of reproductive age, and issue invitation cards for them to visit the nearby health facility for screening and enrolment. Other awareness and recruitment strategies included regular home visits by trained field workers (to identify new women moving into established households) and screening women as they sought other health care services. Eligible women were informed that after conception, the intention was to follow them throughout pregnancy until delivery. Upon conception transabdominal ultrasound (5–2 MHz abdominal probe, Sonosite TITAN and Sonosite Turbo, US High resolution, Sonosite, Bothell, WA, USA) was used to estimate gestational age (GA). Gestational age estimation was based on measurement of crown rump length in the first trimester and head circumference in the second trimester. From July 2014 to December 2015, 2629 women were screened for eligibility for inclusion into the FOETALforNCD study and 1415 were enrolled. Of the 1214 exclusions, 313 (25.8%) were not eligible by age, 322 (26.5%) were still using modern family planning methods, 116 (9.6%) were sub-fertile, 208 (16.6%) were already pregnant, 34 (2.8%) refused, 51(4.2%) migrated out of study area and 93 (7.7%) had a child \<9 months old, while 77 (6.3%) were excluded due to other reasons. Of the 1415women included, 72 were later excluded because venous blood was not collected, and 11 did not fulfil the inclusion criteria. Furthermore, 84 women who conceived during the follow up were excluded because; 34 were already pregnant at enrolment based on the ultrasound estimated GA and 50 had a miscarriage before the GA could be ascertained, leaving 1248 women for the present analysis. ## Data collection and tools Socio-demographic data including age, educational level, marital status, and economic (household size, house ownership and type of roofing materials, main source of drinking water and its ownership (private or public),type of toilet facility) and lifestyle factors (smoking, alcohol and tea consumption)were collected using a structured questionnaire. Previous medical histories, including gynecological and obstetric details, were documented. In order to define SES, a principal component analysis was applied and the variables which showed relevant contribution (greater than 10%) to the combined SES score were regarded as the ones which sufficiently described the SES of a woman. Variables included in the final principal component analysis were educational level, occupation, type of house ownership, roofing materials, source of domestic water and its ownership as well as the type of toilet facility. The respective SES scores were categorized in tertiles as low, medium and high. Weight (in kilograms) was measured while on barefoot and wearing light clothes (precision 0.1kg, digital weighing scales, SecaGmbh& Co. Kg, Hamburg, Germany). Height in centimeters (cm) was measured with a stadiometer (precision 1 cm). Mid-upper arm circumference (MUAC) was measured on the upper right arm at the midpoint of the acromion process and the tip of the olecranon (precision 1mm). For measurement of skinfold thickness, trained staff pinched the skin above triceps muscle group to raise a double layer of skin and the underlying adipose tissue without the muscle. The HARPENDEN skinfold caliper (BATY International, England) was then applied 1 cm above and at right angle to the pinch, and a reading in millimeters (mm) taken after a few second. Waist circumference was measured just above the iliac crest in the horizontal plane, and hip circumference was measured at the point yielding the maximum circumference over the buttocks, all using a standard measuring tape to the nearest 1mm. At enrolment, 15ml of venous blood was collected in ethylenediamine tetra acetic acid coated and plain serum tubes, transported at 2° to 8°C to the NIMR Korogwe Research Laboratory and processed within two hours of collection. To avoid photo degradation during transportation, all plain tubes were wrapped in aluminium foil. Separated serum samples were stored at -80°C and later shipped in dry ice to University Hospital Sealand, Denmark for micronutrients analysis. Hemoglobin level was measured by using Sysmex KX-21N hematological analyzer (Sysmex Corporation Kobe, Japan).According to WHO’s definition, anemia was defined as Hb\<12.0 g/dL, and further categorized as mild (10.1–11.9 g/dL), moderate (8.0–10.0 g/dL) and severe (\<8.0 g/dL). Microcytosis was defined as mean corpuscular volume (MCV) value \<80 fL and hypochromic as mean cell hemoglobin concentration (MCHC) value \<32 g/dL. Anemia was further classified as normocytic-normochromic (Hb\<12 g/dL, MCV 80–100 fL and MCHC\>32 g/dL), microcytic hypochromic (Hb\<12 g/dL, MCV\< 80 fL and MCHC\<32 g/dL), megaloblastic (Hb\<12g/dL, MCV≥100) or as mixed types (normocytic-hypochromic, microcytic-normochromic macrocytic-normochromic and macrocytic-hypochromic) anemia. For clinical care, Hb levels were measured using HemoCue 301 Hb analyzer (HemoCue AB, Angelholm, Sweden). Anemic women received treatments as follows: mild anemic (Hb10.1–11.9 g/dL) women with no symptoms received dietary counseling whereas women with symptoms were offered one combination tablet of 200 mg ferrous sulfate (\~ 43 mg elemental iron) and 400μg folate per day (Ferrolic–LF, Laboratory and Allied LTD, Mombasa, Kenya). Moderately anemic patients with Hb 9.1–10.0g/dL received 2–3 combination tablets of iron and folic acid (Ferrolic–LFLaboratory and Allied LTD, Mombasa, Kenya) per day and monitored at each scheduled visit. Those with Hb 8.0–9.0 g/dL received a daily dose of 20 mL Hemovit multivitamin syrup (200 mg Ferrous sulfate, 0.5mg B₆, 50 μg B₁₂, 1500 μg Folic acid and 2.33mg Zinc per 5mL, Shelys Pharmaceuticals, Dar es Salaam, Tanzania) and monitored at each scheduled visit. Severely anemic (Hb\<8g/dL) women were referred to Korogwe District Hospital for further evaluation. Serum ferritin, vitamin B_12, folate, alanine aminotransferase (ALT) and bilirubin levels were measured using Dimension Vista 1500 biochemical analyzer (Siemens Healthcare Diagnostics, Inc, New York, USA). C-reactive protein (CRP) was measured by using Afinion AS 100 analyzer (Axis Shield PoC, AS, Oslo Norway) and inflammation defined as serum CRP \>5 mg/L and/or ALT \>45 U/L. To account for elevated serum ferritin due to sub-clinical infection and other inflammatory conditions, three approaches were applied to define ID and results compared. The first approach utilized arithmetic correction factor (CF)as proposed by Thurnham*et al*. to adjust for the increased serum ferritin levels due to inflammation. In this approach CF of 0.67 was applied only for samples that had evidence of inflammation (CRP\>5 mg/L), and a cut-off of \<15 μg/L was then applied to the adjusted ferritin levels to define ID. If CRP was not available, serum ferritin was coded as missing. In the second approach, a higher ferritin-cutoff (\<30μg/L) was applied to the subset of individuals with inflammation (CRP \>5 mg/L) to define ID, as proposed by the WHO. In this approach ID was defined as serum ferritin \<15μg/L (no inflammation) or 30μg/L (inflammation present). The third approach utilized the higher serum ferritin cutoff (\<30μg/L) if CRP\>5mg/L and/or ALT\>45 U/L which is considered as a sign of liver disease. In the core analyses presented here, ID anemia was defined as Hb\<12g/dL in the presence of ID based on Thurnham approach. Vitamin B₁₂ deficiency was defined as serum cobalamin \<150 pmol/L, and folate deficiency as serum folate\<10 nmol/L without adjusting for inflammation. Malaria was diagnosed using malaria rapid diagnostic test (mRDT) kit, ParaHIT (span diagnostics, Gujarat, India) or CareStart Malaria Pf (HPR2), ACCESS BIO, New Jersey, USA) according to manufacturer instructions. In addition, thick and thin blood films were prepared for the detection and quantification of parasitemia. Malaria patients received oral artemether-lumefantrine, (Lumartem 20mg/120mg (Cipla Ltd, Patalganga, India), quinine or artesunate injections according to Tanzanian standard treatment guideline. Human immunodeficiency virus infection was tested by using DetermineHIV-1/2 test kit (Alere ltd, Stockport, UK) and seropositive cases were confirmed using Unigold test kit (Trinity Biotech Plc, Wicklow, Ireland) according to the manufacturers’ instructions. Newly diagnosed HIV patients were referred to the nearby care and treatment clinics for long-term care. Considering low prevalence of STH infestations in north eastern Tanzania, stool samples were collected from a subgroup of 434 women at the time of enrolment and preserved in 10% neutral buffered formalin solution. Formol-ether concentration technique was used to detect presence of STH infestations. All confirmed (on stool samples)or clinical suspected cases of STH infestation received a single dose of albendazole (400mg) or mebendazole (500mg) tablets according to the existing Tanzanian standard treatment guideline. ### Statistical analysis Microsoft Access software 2007 (Microsoft corporation, Redmond’s, USA) was used for data entry and validation. Stata version 13 (StataCorp, Lake Way drive, College station, USA) software was used for statistical analyses. Continuous variables were visually inspected for normality using histograms and described using mean and standard deviation if normally distributed or median (interquartile range—IQR) for skewed data. Univariate analysis was done using Student’s t-test or Mann-Whitney test for continuous parametric and non- parametric variables, and Chi-square (χ²) or Fisher's exact test for categorical variables. Factors associated with preconception anemia were determined using logistic regression analysis and expressed as unadjusted odds ratio (OR) and adjusted odds ratios (AOR). All variables with *P*-value \<0.20 in the univariate analysis were entered into the multivariate models. Using a stepwise backward elimination approach final models including variables with a *P*-value \<0.10 were obtained. A *P*-value of *\<*0.05 was considered statistically significant. Due to missing data on HIV infections in 350 (28.8%) women and considering HIV infection being an important confounding factor, two different models with and without adjusting for HIV infection were generated and compared. Finally, in order to illustrate the association between a risk factor and anemia, trendline figures were generated for each continuous risk factor found to be statistically significant or borderline significant in the multivariate model. # Results The baseline characteristics of the women enrolled in the study are shown in below. The median (IQR) age of the women was 28.2 (IQR = 22–34) years. Few women were underweight (8.0%), while 60.5%, 20.2% and 11.2%, were normal weight, overweight and obese, respectively. Majority (71.5%) had previously used at least one type of hormonal contraceptive. Malaria infection was detected in 8.1% (95% CI 6.7–9.8) of the women while 5.7% (95% CI 4.4–7.3), 3.0% (95% CI 1.7–5.1) and 15.2% (95% CI 13.3–17.3) had STH infestations, HIV and inflammation, respectively. The mean±seHb level was 12.2±0.04 g/dL and 36.7% (95% CI 34.1–39) women had anemia. Severe, moderate, and mild anemia were found in 1.8%, 95% CI 1.8–2.7), 16.1% (95% CI 14.2–18.3 and 18.8% (95% CI 16.8–21.1) of women, respectively (data not shown). Comparing anemic and non-anemic women, there was a statistically significant difference in age, ethnicity, MUAC, waist and hip circumference, BMI, and self-reported or malaria infection at the time of enrolment. Overweight (BMI of 25-\<30kg/m²) and obesity (BMI≥30 kg/m²) were observed in 20.2% (95% CI 18.1–22.6%) and 11.2% (95% CI 9.6–13.1) of the women, respectively. The coexistence of overweight and obesity with anemia was found in 81/449 (18.0%) and 39/449(8.6%) of women, respectively while79/451 (17.5%) and 42/451 (10.0%) of iron deficient women had concurrent overweight and obesity, respectively (data not shown). When using the internal CF approach by Thurnham*et al*., 37.6% (95% CI 34.9–40.4) women had ID. Using the WHO definition (serum ferritin \<15 μg/L or \<30 μg/L in the presence of inflammation), 39.2% (95% CI 36.5–41.9) were classified as having ID. After further adjustment to account for liver disease (serum ferritin \<15μg/L or \<30μg/L in the presence of inflammation (CRP \>5 mg/Land ALT \>45 U/L), 40.2% (95% CI 37.5–43.0) women were classified as having ID. Of the 458 women with anemia 58.8% (95% CI 54.2–63.3), 3.1% (95% CI 1.8–5.1) and 0.4% (95% CI 0.1–1.7) had concurrent iron, folate and vitamin B₁₂ deficiencies, respectively. Based on Thurnham approach, among 461 women with ID, only 2.2% (95% CI 1.2–4.1) had combined iron and folate deficiency and 0.4% (95% CI 0.1–1.7) had combined iron and vitamin B₁₂ deficiencies, while none had combined folate and Vitamin B₁₂ deficiency (data not shown).Finally, concurrent ID with malaria, HIV and inflammation were found in 2.8% (95% CI 1.7–4.8), 4.7% (95% CI 3.0–7.5), and 11.7% (95% CI 9.1–15.1), respectively (data not shown). The majority of the women (56.3%, 95% CI 53.6–59.3) had normocytic-normochromic RBC morphology, while 17.5% (95% CI 15.5–19.8), 19.9% (95% CI 17.7–22.2), 5.9% (95% CI 4.7–7.4) and 0.2% (95% CI 0.1–0.8) had microcytic-hypochromic, microcytic-normochromic, normocytic-hypochromic and megaloblastic RBC morphology, respectively. However, for the anemic women, only 32.8% (95% CI 28.6–37.2) had normocytic-normochromic while 38.2% (95% CI 33.9–42.8), 19.0% (95% CI 15.6–22.9), 9.8% (95% CI 7.4–12.9) and 0.2% (95% CI 0.03–1.5) had microcytic-hypochromic, macrocytic normochromic, normocytic-hypochromic and megaloblastic RBC morphology, respectively. Interestingly, among the ID anemic women (Hb\<12g/dL and serum ferritin \<15μg/L), only 48.7% (95% CI 42.6–54.7) had microcytic-hypochromic RBC morphological pattern, whereas 25.5% (95% CI 20.5–31.1), 20.6% (95% CI 17.1–24.6) and 6.5% (95% CI 4.5–9.2) had normocytic- normochromic, microcytic-normochromic and normocytic-hypochromic RBC morphological types, respectively (data not shown). The results of the univariate and multivariate logistic regression analysis with all variables entered into the model except HIV infection are shown in below. In the univariate analysis, women from the minority tribes, increased MUAC, waist and hip circumferences, BMI as well as previous use of hormonal contraceptives, were statistically significantly associated with reduced risk of anemia. Increasing age, length of menstrual periods, malaria infection before or at the time of enrolment, inflammation and ID were significantly associated with increased risk of anemia. Skinfold thickness showed similar associations with risk of anemia as MUAC but data was missing on 25% of the women and were therefore not included in further analyses (data not shown). In the multivariate logistic regression analysis, women from the minority tribes (adjusted OR (AOR) 0.63, 95% CI 0.45–0.87) and increased MUAC (AOR 0.90, 95% CI 0.84–0.96) remained statistically significantly associated with reduced odds of having anemia while previous use of hormonal contraceptives was marginally protective against anemia (AOR 0.74, 95% CI 0.54–1.00). Increased age (AOR 1.05, 95% CI 1.03–1.07), malaria infection at the time of enrolment (AOR) 2.21, 95% CI 1.37–3.58), inflammation (AOR 1.77, 95%, CI 1.21–2.60), and ID (AOR 4.68, 95% CI 3.55–6.17 were all statistically significantly associated with increased risk of anemia. A model including HIV was also generated. The AORs for anemia were similar for increased age, ethnicity, and malaria infection at the time of enrolment, inflammation and ID as in the model without HIV. Human immunodeficiency virus infection led to a three-fold increased risk of anemia (AOR2.57, 95%CI 1.37–4.81). Increased MUAC was no longer significantly associated with increased risk of anemia after adjusting for HIV infection. The risk of anemia gradually and in linear manner decreased with increasing MUAC and hip circumference but it increased with increasing age and CRP level. However, with a MUAC\<28cm, hip circumference\<95 cm, age \>28yearsand CRP \>10mg/L the prevalence of anemia was higher than the average prevalence of 36.7%. Finally, the risk of anemia steadily decreased with increasing ferritin levels up to 100μg/L and then flattened out, but the prevelence of anemia was above the average only at ferritin concentrations below 20μg/L. # Discussion The present study assessed the prevalenceand risk factors of preconceptionanemia in a cohort of rural Tanzanian women who were likely to conceive. In this study, a significant proportion of women had anemia and ID before conception. The prevalence of anemia in this study was similar to other community based studies among non pregnant women of reproductive age in Tanzania, Ethiopia and Sierra Leone, which reported prevalence of anemia of 39.5%,30.4% and 44.8%, respectively,but with a wide variation in the prevalence of ID. In this study all women who were currently using modern contraceptives were excluded but the observed prevalence of anemia and ID were comparable to what Haile *et al*. reported among non-pregnant Tanzanian women of reproductive age including hormonal contraceptive users. Since all women in this study were not using modern contraceptive methods and hence were likely to conceive, over one third are at risk of having anemia and/or suboptimal iron levels at the time of conception without interventions. The ID and anemia may further deteriorate during pregnancy in response to the increased physiologic demands of gestation, thereby increasing the risk of adverse pregnancy outcomes. The risk of anemia decreased with increasing MUAC. Previous studies have also showed a negative correlation between anemia and MUAC. In low resource settings, MUAC has been shown to correctly identify undernourished or anemic women, but there is no consensus on appropriate anthropometric cut-off points for risk prediction.Despite increasing adiposity being protective against anemia, pregnant women who are overweight or obese may face increased risk of numerous other complications, including gestational diabetes, pre-eclampsia, pregnancy- induced hypertension, stillbirths and preterm births. Moreover, clustering of obesity and ID or anemia within the same individual poses a great challenge for public health interventions. In the current study, we found that nearly one third of overweight/obese women had concurrent anemia or ID. Furthermore, not until a MUAC \>28cm was the prevalence of anemia lower than the average of 36.7% indicating that some women are still anemic despite having a normal weight or only slight overweight. Therefore, specific interventions aimed at preventing anemia before conception should target all women of reproductive age, irrespective of their nutrition status. Iron deficiency was common among both non-anemic as well as anemic women, but ID was associated with an almost five-fold increased risk of anemia. This finding is comparable to the recent study by Haile *et al*. and underscores that adequate iron stores among women of the reproductive age is important and might be ensured by routine provision of iron and folic acid supplementation to all menstruating women, as recommended by the WHO. Iron deficiency was characterized based on both RBC morphology (MCV and MCHC) and serum ferritin and compared, but a large proportion of the women with true ID would have been misclassified as not having ID based on RBC morphological patterns alone. This is not surprising as up to 40% of pure ID anemia may present with normocytic-normochromic pattern. Moreover, the presence of microcytosis does not necessarily imply ID and can be present due to other causes such as thalassemias, or anemia of chronic disease. Likewise, dimorphic anemia can occur in partially treated ID. Serum ferritin is an acute phase reactant and can be increased if inflammation is present. We applied the CF to get a more precise estimate of ID in all women who had inflammation. Interestingly, the prevalence of ID using the CF did not differ significantly as compared to the two other used approaches (higher ferritin cut-off and or/ adjusting for high ALT). This was in contrast to a recent study by Namaste *et al*. who reported a substantial underestimation of the burden of ID when only a higher ferritin cut-off approach was used as compared to adjusted approaches. One possible explanation for this difference could be the smaller sample size in our study compared to Namaste *et al*. study. Haile *et al*. excluded all malaria patients and used serum transferrin receptor (sTfR) as a marker of ID. However, sTfR can also be affected by acute inflammation, and recently adjustment of sTfR using internal CF has been proposed. Likewise, exclusion of all women with malaria would have introduced a selection bias. A much higher prevalence of ID was observed as compared to Wirth *et al*. who reported a prevalence of ID of 8.3% among rural non-pregnant women of reproductive age in Sierra Leone despite a high prevalence of anemia (44.3%). The difference could be explained by a higher prevalence of malaria in Sierra Leone (23%) as compared to 8.1% in our study. Likewise, a recent meta-analysis has shown that the contribution of ID to the overall burden of anemia in women of reproductive age is much lower in regions with high burden of malaria and inflammation. Although both malaria and HIV prevalence was below 10%, both were statistically significant risk factors for anemia and public health strategies should target these conditions. This should include screening and treatment to reduce the consequences for those already infected as well as preventive measures such as bed nets and condom usage in order to reduce the risk of infection. We also observe a trend towards a reduced risk of anemia for women who had previously used hormonal contraceptives. Similar, to our study, Haile *et al*., found that hormonal contraceptive use was significantly associated with a reduced risk of anemia and ID among non-pregnant women in Tanzania. This might partially be due to reduced menstrual bleeding, as well as indicative of a higher socio-economic status and hence better health status among hormonal contraceptives users. For example, a recent report has shown that that 35% of women in the wealthiest households used modern contraceptive methods as compared to 20% among women from low income households. The prevalence of STH infestations was very low and among the anemic women, none had infestations. This finding is similar to a previous study in this region which showed very low prevalence of STH (0.3%) among children in an urban settings. The area participates in annual mass drug administration for control of lymphatic filariasis using ivermectin and albendazole as well as school deworming program. These programs may have shown additional impact on reduction of STH infestations. Dietary iron absorption from the human gut is dependent on physiological requirements, but may be restricted by the quantity or bioavailability. Whether IDin this population was primarily the result of dietary insufficiency or problems with absorption was beyond the scope of this paper. Understanding these patterns could provide important information about the relationship between food consumption and the risk of ID and anemia before conception in this setting. However, previous studies from a neighboring district of Muheza has shown that 73% of total iron intake came from staple food mainly cereals, fats and nuts and green leafy vegetables.In addition to low bioavailability, cereal staples contain iron absorption inhibitors, particularly tannins and phytates. Therefore, high prevalence of ID without concurrent folate deficiency in our study could signify poor bioavailability rather than dietary insufficiency. The Tanzanian government has strengthened different interventions to reduce the burden of anemia during pregnancy, including iron and folic acid supplementation, deworming, intermittent preventive treatment for malaria (IPTp) with sulfadoxine pyrimethamine as well as scaling up the coverage and usage of insecticide treated bed nets. Furthermore, the Tanzanian National Multisectoral Nutrition Action Plan (2016–2021) aims to reduce the proportion of women of reproductive age with anemia from 45% to 33% through increased uptake of iron and folate supplements during pregnancy, increased acess to micronutrient powders from 10%-35% and increased proportion of iron fortified flour produced in Tanzania from 36%-50%. However, despite anticipated benefits, the initiative has been compromised by high prevalence of pre-existing anemia and ID and short period of supplementation.Daily supplementation with 30–60 mg of elemental iron taken for three consecutive months in a year, is recommended by the WHO as a public health intervention to prevent anemia and ID for menstruatng women and adolescent girls living in settings where the prevalence of anemia 40% or higheror intermittent iron and folic acid supplementation in settings where prevelence of anemia is 20–40%. However, despite its proven benefit, and recommendation by the WHO,preconception iron and folic acid supplementation has not been included in the current Tanzanian National Multisectora Nutrition Action Plan. It is therefore worrisome that many anemic women willreceive these supplements only after conception and after having booked for antenatal care. The strengths of the study include a large number of socio-demographic, economic, and clinical variables collected and stringent data monitoring system to ensure the quality of collected data. Furthermore, we could be rather certain that the cohort was truly preconception. For women who conceived during follow up, 85% had GA estimation in the first trimester. Pregnant women were excluded from analysis if the ultrasound examination demonstrated that they were already pregnant before conception or the GA was uncertain. On the other hand, this study had several limitations. Firstly, the design of the main study was to identify women who would potentially conceive during the study period, especially women who were not using modern contraceptives. Women who were still using traditional contraceptive methods were included and their behavior and SES could be different from women who were actively trying to become pregnant. Secondly, serum ferritin level, which is also an acute phase protein, was used to detect ID. Serum ferritin levels were adjusted to account for the effect of inflammation by using internal CF and exclusion of all women without CRP measurement but this might not have corrected the entire effect. Further adjustment by using alfa1 acid glycoprotein in addition to CRP as marker of inflammation as well as using internal regression approach have also been proposed. An alternative was to use sTfR, which is less affected by inflammation, but it was not available in this study. Thirdly, our study utilized serum sample to measure folate and B₁₂ levels. The sensitivity of serum folate and vitamin B₁₂ measurements for the diagnosis of clinically significant deficiency is uncertain because of the lack of a gold standard diagnostic method. However, this may have had little effect on our estimate of the proportion of women with folate or vitamin B₁₂ since previous studies have shown these deficiencies are responsible for a small proportion of anemia among women and children in SSA. Finally, the present study did not determine the burden of vitamin A deficiency is also a risk factor of anemia. # Conclusion The prevalence of preconception anemia and ID in this study population was high, while vitamin B₁₂ and folate deficiencies were rare. Iron deficiency, malaria and HIV infections as well as inflammation were significantly associated with preconception anemia. This signifies that many women will enter pregnancy with depleted iron stores and/or anemia putting themselves and their growing fetus at risk. Therefore, concerted efforts, including intermittent iron and folic acid supplementation and nutrition counseling to improve iron status before conception for optimal fetal growth and newborn short and long-term health should also be considered in this population. # Supporting information We thank the women who participated in the FOETALforNCD study, field workers, study nurses, clinicians, laboratory technicians, drivers and data entry clerks for their hard work. Others, who provided assistance during data collection and quality control, include Lise Egholt Søgaard, Cecilie Bøge Paulsen, Hannah Kousholt, Anna Mathilde Yde, and Josephine Roth Eckmann. Special thanks to head of NIMR Korogwe Research Laboratory and Korogwe District health system officials as well as local community leaders for their valuable assistance, which made this work possible. [^1]: The authors have declared that no competing interest exist.

# Introduction Oncolytic viruses are live replicating viruses that selectively infect cancer cells and kill them. Healthy cells are largely spared. The idea is to inoculate the virus into a cancer patient, and let the virus spread throughout the tumor, thereby driving it into remission. Selectivity for cancer cells occurs because cancer cells tend to lack important genes that normally shut down the replication cycle of the virus. For example, the adenovirus ONYX-015 has been engineered such that it only replicates in p53−/− cells, a characteristic of many cancers. Certain animal viruses by chance have the ability to replicate in human cancer cells, while healthy human cells are not permissive. An example is Newcastle disease virus, which can replicate in tumor cells that lack interferons. In general, a wide array of viruses is being explored as potential oncolytic viruses. Oncolytic viruses have shown promising results in clinical trials. Cancers have been found to respond to treatment, leading to tumor remission in some cases. Consistent and sustained eradication or control of cancers has, however, been very difficult to achieve. This is caused in part by our lack of understanding regarding the dynamics that underlie the spread of oncolytic viruses through tumors. Without such a rigorous understanding, much of the work is based on trial and error. In such scenarios, mathematical models can be very useful to complement empirical work. Mathematical analysis allows us to see the whole spectrum of possible outcomes, and provides a means to logically suggest ways to optimize treatment. Limited mathematical analysis of oncolytic virus therapy has been performed in the past. This work is largely qualitative in nature, examining how variation in viral and host parameters influences the outcome of treatment. For example, it has been suggested that maximizing the virus-induced rate of tumor cell killing is not going to lead to the best treatment outcomes. Instead, an intermediate and optimal rate of virus-induced cell death optimizes treatment success. This work was based on the analysis of the equilibrium properties of the model. That is, the lower the total number of cancer cells that remain as the dynamics converge to steady state, the better the predicted outcome of therapy. While such steady state analysis can provide some valuable qualitative insights, it has limitations. The main problem is that in such infection dynamics models, the population of cells and viruses can show extensive oscillations before converging to a steady state. During these oscillations, the populations of cells and viruses can potentially go extinct, and the system might never reach equilibrium. Therefore, it is important to understand these oscillatory dynamics, and how they relate to the chances that the cancer cell population is driven extinct. This paper aims to analyze these dynamics in detail in an attempt to provide a more realistic description of oncolytic virus dynamics. This is a difficult task because these infection dynamics, and in particular the occurrence of population oscillations, can be dependent on particular details of the models that are of a biologically uncertain nature. To address this issue, we avoid concentrating on a particular model, but take a more general approach. Through specific restrictions about biological assumptions, we analyze a class of mathematical models that aim to describe viral spread through a tumor in different settings. We seek to determine conditions under which the virus is successful at eliminating the tumor, and the conditions when virus therapy fails. In order to underline the insights that we gain from this general framework, we also consider specific models that are examples of the general framework. This modeling framework provides the basis for experimental validation and testing procedures, which will allow us to accurately predict the time course of cells and viruses at least in relatively simple scenarios, such as in vitro experiments or simple in vivo scenarios. In this context, we fit the models to previously published experimental data and discuss implications for model testing, model selection, and experimental work. A predictive model of a complex in vivo situation (e.g. including immune responses) will obviously be more difficult to attain, but can arise from a thorough understanding of the simpler in vitro scenario that we examine here. # Results ## The modeling framework We will model the dynamics of oncolytic virus replication by ordinary differential equations that describe the development of the average population sizes of cells and viruses over time. This approach is based on very well established mathematical models that describe the general dynamics of virus spread both in vivo and on an epidemiological level,. Instead of considering a specific model, however, we will take a generalized approach and consider a class of models. The general modeling framework used in our study is as follows. We take into account two populations: uninfected tumor cells, *x*; and infected tumor cells, *y*. The population of free viruses is not modeled explicitly. Because the turnover of free viruses is much faster than that of infected cells, we simply assume that the free virus population is in a quasi-steady state and proportional to the number of infected cells. The basic model is given as follows: The function *F* describes the growth properties of the uninfected tumor cells, *x*, and the function *G* describes the rate at which tumor cells become infected by the virus. These functions are unknown and can potentially take a variety of forms, which will be discussed below. The coefficient *β* in front of the infection term represents the infectivity of the virus. Finally, virus- infected cells die with a rate *ay*. We will not include immune responses in our considerations. While immune responses will certainly be an important factor for oncolytic virus dynamics in vivo, our goal is to first understand those dynamics in a simpler setting without the presence of immune responses. These models would be suitable to describe the growth of oncolytic viruses in relatively simple in vitro or in vivo settings. Once an understanding of such simple systems has been achieved, additional biological complexities (such as the presence of immune responses) can be added to the model. This class of models is characterized by the existence of equilibria, the number and nature of which depends on the tumor growth term *F* and the infection term *G*. In the most general sense, the equilibria of the system are defined by the following two equations:We will explore the equilibria and their properties depending on the tumor growth term, *F*, and the infection term, *G*. The term *F* reflects the growth properties of an uninfected tumor. It comprises both division and death rates. The simplest assumption that can be made about the term *F* is that growth is exponential (or, more precisely, the division and death happen according to an exponential law, and the division rate is higher than the death rate). While this can be true during early stages of tumor growth, tumor growth certainly deviates from an exponential pattern at larger sizes for a variety of reasons, for example space or nutrient limitations. Therefore, more complicated tumor growth terms involving some form of saturation must be considered. In this respect, we can distinguish between two basic scenarios: First, while the rate of tumor growth saturates and slows down at higher tumor sizes, the tumor has the potential to keep growing towards infinity. Growth would stop once the tumor has reached a lethal size. Second, it can be assumed that growth not only slows down, but comes to a halt as the tumor size reaches a critical level, which can be called the carrying capacity of the tumor. This could happen when the division rate equals the death rate of the cells. Regarding the infection term, the assumption used most often in mathematical models is that it is directly proportional to the number of infected and uninfected cells. This, however, assumes mass action or perfect mixing of populations, which is unrealistic, especially in the context of tumors. Instead, virus spread is likely to be slower, limited by spatial constraints. Since the virus released from one infected cell cannot reach all susceptible tumor cells in the population, the infection rate must be a saturating function of the number of susceptible tumor cells. Similarly, not all infected cells present in the population will be able to contribute to the generation of newly infected cells, for example if they are spatially separated from susceptible cells. In the following section, we will define different classes of infection terms that have biologically reasonable characteristics, and investigate how they influence the properties of the model. These are based in part on mathematical work done in the context of infectious disease epidemiology. Subsequently, we will examine how changing the tumor growth term influences the model predictions. ## Different classes of infection terms and their properties Let us consider two different classes of viral growth, see. Tumor-virus systems belonging to class I are characterized by the following property: if the number of uninfected tumor cells is high relative to the number of infected cells, virus growth does not slow down as the number of infected cells rises. Virus growth is exponential. Biologically, this can be interpreted as virus replication in a non-solid tumor where cells mix relatively freely. In other words, infected cells are not clustered together in a mass but are interspersed among uninfected cells. This is shown schematically at the top of (the white circles represent uninfected cells, and the black circles - infected cells). In this case, if the number of uninfected cells is relatively large, then every infected cell is likely to be surrounded by uninfected cells to which the virus can be passed on. Alternatively, a similar picture can be achieved by a very high motility of the virus. In either case, all infected cells contribute to viral spread and growth is exponential. We call this “fast virus spread”. On the other hand, with tumor-virus systems that belong to class II, the virus growth rate decreases as the number of infected cells rises, even if the number of uninfected cells is very large. The biological interpretation is that infected tumor cells are clustered together. This can occur in solid tumors, which typically show a high degree of spatial arrangement. In this case, as the number of infected tumor cells increases, most infected cells will be surrounded by other infected cells and not by uninfected cells. Hence, they cannot pass on the virus and cannot contribute to virus spread. Only cells at the periphery of the infected cell mass have uninfected cells in the neighborhood and can contribute to new infection events. We refer to this model of infection as “slow virus spread”. Next let us connect this classification with the mathematical model, and in particular, with the infection term, *βyG(x,y)*. The function *G(x,y)* is related to the proportion of the total population of the infected cells which participates in the infection process. It is plotted in as a function of the number of tumor cells, *x*, and we examine the shape of these plots. Let us take a closer look at the schematic at the top of. Because of the geometrical arrangement of the cells in this case, only the infected cells on the surface of the black core will be able to infect other cells (it is 6 out of the 7 cells in the smaller colony presented). Now, let us increase the system size, such that the number of infected and uninfected cells grows in the same proportion. Again, only the infected cells close to the surface of the infected core will participate in the infection process. However, now the proportion of the surface cells is much smaller (11 out of 20 cells). As the size of the system increases, the proportion of such “active” cells (that is, cells capable of infecting other cells) decreases. This is what is depicted in the graph in, where the function *G(x,y)* declines following the peak. (For very small system sizes, the proportion of cells participating in infection is formally zero because of the lack of uninfected cells, therefore the graph of the function *G* starts at zero, reaches a peak, and then declines for high values of *x*). Next, we take a look at the cell arrangement at the top of. Here, the populations are well- mixed, and as the system grows, a constant fraction of infected cells will be able to infect new cells. This is reflected in the corresponding graph of *G(x,x/a)*, which reaches an asymptote and does not decline. In we list several examples of fast and slow growth laws. In general, we can prove that the two scenarios above are the only possible outcomes, given the biological requirements imposed on the function *G*. As *x* increases, this function increases, and can either approach zero or a nonzero level. If it approaches a non-zero level, this does not necessarily need to occur via a monotonic approach to the asymptote. It is possible that the function *G* first increases, peaks, and then converges to a non-zero asymptote. For intermediate values of *x* the function *G* may have a more complicated structure than that shown in, but in the absence of any biological evidence of that it is a safe bet to assume the simplest shape with a minimal number of local extrema. How does the shape of *G* help us draw meaningful conclusions about the behavior of the biological system? It turns out that the function *G* is essential in determining the number and the stability properties of the equilibria of the system, and thus it will help us reason about long-term predictions on the treatment outcome. Equations (3–4) can be combined in a single equation,where the function *y(x)* is a relationship between the number of infected and uninfected cells at equilibrium as the total system size grows; it is obtained from equation (3) and depends on the exact rate of cancer growth, *F*. If the cancer growth is exponential (*F = 1*), we have *y(x) = x/a*, that is, at equilibrium, the infected cells comprise a fixed fraction of uninfected cells. Thus the function *G(x,x/a)* depicted in is just the left hand side of the equation for the equilibria, equation (5). The right hand side is represented by horizontal dashed lines, whose level decreases with the viral replication rate *β*. The number of intersections corresponds to the number of equilibria in the system. We can see that the two graphs in exhibit different numbers of equilibria. First we consider, fast virus spread. In this case, the model always contains a parameter region in which exactly one equilibrium exists. If the viral replication rate, *β*, lies below a threshold (*β*\<*β_c*) then no equilibrium exists. If the viral replication rate lies above that threshold, the following is observed. As shown in exactly one equilibrium is found. In other cases, it is possible that there are two or more equilibria for intermediate viral replication rates. (For example, if the function *G(x,x/a)* first rises and achieves a maximum before descending to its horizontal asymptote, or if it goes through a number of local extrema before approaching a horizontal asymptote.) The most important universal feature in all fast growth scenarios is that for sufficiently high values of *β*, there is exactly one equilibrium. Next, consider, slow virus spread. Again, for an equilibrium to exist, the viral replication rate needs to lie above the threshold *β*\>*β_c*. If this is the case, the system is always characterized by the presence of not one, but two equilibria. Again, in some cases, it is possible that the intermediate values of *β* correspond to more than two equilibria. The biological interpretation of this analysis is as follows. We saw that for both modes of infection, if the values of the viral replication rate *β* are small, no equilibria exist. This translates into an uncontrolled cancer growth. This is an intuitive result: for low viral replication rates, treatment is impossible. A less intuitive result is connected with the number of equilibria once *β* is above its threshold value. The cancer-virus system displays a fundamentally different behavior depending on whether it is characterized by one or two equilibria. If there is only one equilibrium, then the dynamics will be governed by the properties of this equilibrium only. Because the number of tumor cells is relatively low at this equilibrium, this outcome corresponds to containment of the tumor by the virus. For convenience, we call this internal equilibrium *E_I*. On the other hand, the situation is more complicated if the system is characterized by two equilibria. The first equilibrium, at which the number of tumor cells is lower, is again the internal equilibrium, *E_I*, and can be interpreted as containment of the tumor by the virus. The second equilibrium can be shown to be an unstable saddle node equilibrium, call it *E_S*. The presence of the saddle equilibrium means that the dynamics are qualitatively different depending on the initial conditions. If the initial number of tumor cells is relatively low and close to the internal equilibrium, then the dynamics are governed by this internal equilibrium, *E_I*, leading to a degree of tumor control. If the initial number of tumor cells is higher and around or above the saddle node equilibrium *E_S*, then the number of tumor cells increases in an uncontrolled fashion. Hence, in this regime, uncontrolled cancer growth is always a possible outcome. We conclude that our biologically defined modes of virus spread correspond to very different mathematical properties. Models of class I (fast virus spread) contain a parameter region (of high enough *β*) in which only a single equilibrium is observed. In this case, the model contains a parameter region in which uncontrolled cancer growth is impossible. Models of class II (slow spread) never have only one equilibrium and the saddle node equilibrium *E_S* is present whenever the internal equilibrium *E_I* exists. In this class of models, no matter how high *β* is, uncontrolled cancer growth is always a possibility. ## Effect of the tumor growth term For the purposes of classification of the virus spread terms, we looked at the changes in *G* as the number of infected and uninfected cells grew in the same proportion. This led to a direct evaluation of the number of equilibria for exponential cancer growth (*F = 1*). While mathematically the simplest scenario, exponential growth is an unrealistic assumption, because the growth of cells is bound to saturate as the tumor grows. Our methods allow to study any reasonable cancer growth law in a very natural way. Let us model a slow-down of the tumor growth rate as the number of tumor cells increases. This can be done in two different ways. On the one hand, we can assume that while tumor growth slows down, it never stops, such that the tumor can grow towards infinity over time. That is, there is no upper limit to the number of tumor cells; in practical terms growth will stop when the organism dies. An example is what we call “surface growth”, where only the cells around the surface of the tumor can give rise to viable daughter cells and can contribute to tumor spread. This can apply to solid tumors that have a high degree of spatial structure. Surface growth in 2D and 3D are listed in. The parameter *η* determines the tumor size at which saturation comes into play. Another possibility that falls into this category is that the rate of tumor growth becomes linear as the number of tumor cells increases. In this case, tumor growth is even slower; we refer to it as “linear growth”. On the other hand, it is possible that there is a natural limit or carrying capacity, *W*, that limits tumor growth. Thus, we will assume that growth slows down and eventually stops as the number of tumor cells increases. This can occur in a variety of ways. Tumor growth can be exponential until the number of cells approaches carrying capacity and the rate of tumor cell growth becomes zero. For example, this can be described by the logistic growth term. Alternatively, we can assume that tumor growth first saturates according to the surface growth or linear growth patterns described above, and only reaches the carrying capacity once the tumor has grown to a significantly larger number of cells. Another example of a growth with a carrying capacity is a Gompertzian type growth. As mentioned before, the term *F* reflects implicitly both division and death properties of uninfected tumor cells. For example, an exponential growth is characterized by a net expansion rate resulting from exponential division and death processes. The logistic growth is a consequence of saturation of the division rate while the (exponential) death rate remains constant. In fact, any process with a sub-exponential division rate and an exponential death will be characterized by a finite carrying capacity. On the other hand, an unlimited (but saturated) growth (such as surface growth) implicitly includes death which happens slower than exponentially. If we were to add an exponential death term to a surface growth, it would lead to a limited growth with a carrying capacity. Our framework includes all these and any other reasonable functional forms of cellular growth. In the following, we will examine the effect of different types of tumor growth terms on the properties of the model. We will do this first in the context of the faster virus infection terms that belong to class I, and then in the context of the slower infection terms that belong to class II. Note that our analysis is quite general and the particular growth laws listed in are merely an illustration; the results are not restricted to these particular growth laws. ### Effect on fast virus growth With this class of virus infection term, we found that in the context of exponential tumor growth, *G(x,y(x))* with *y(x) = x/a* approaches a nonzero asymptote for large values of *x* (note that it can either rise monotonically to the asymptote, or first go through one or more local maxima before declining towards the asymptote). In either case, for any equilibrium to exist, the viral replication rate needs to lie above a threshold *β*\>*β*_c, and there exists a parameter region (characterized by values of *β* greater than a threshold) in which only the internal equilibrium *E_I* is present. In this parameter region, tumor control is the only outcome. Introducing saturated tumor growth (or changing the function *F* in any way) will lead to a different functional form of *y(x)* in equation (5). A universal feature is that any tumor growth slower than straight exponential growth will lead to smaller values of *y(x)* and thus to higher values of *G*. Therefore, as a result of tumor growth saturation, the asymptote becomes higher for slower tumor growth terms. This means that only the internal equilibrium *E_I* can exist, as with exponential growth. The only difference lies in the viral replication rate threshold beyond which this equilibrium can exist and beyond which tumor control is possible. The slower the tumor growth, the lower the viral replication rate threshold required for virus-mediated control. If we assume saturated but limited tumor growth (i.e. growth stops at carrying capacity *W*), then the picture is similar for the most part, with one difference. After the term *G(x,y(x))* has approached the asymptote, the curve *G* takes an upward turn in the vicinity of *x = W*, i.e. when the number of cells approaches carrying capacity. This means that the model acquires an additional equilibrium, which corresponds to the cancer growing to its carrying capacity *W*. In the systems with unrestrictive growth, this was equivalent to unlimited growth of the cell population to infinitely large sizes. This is illustrated with the dotted line in. We can see that for *x*≪*W*, the curves for limited and unlimited growth laws look identical, and near the carrying capacity *W* they deviate. So far, we have concentrated on the case where *G(x,y(x))* increases monotonically towards an asymptote. Alternatively, the term *G(x,y(x))* can rise to a peak and then decline toward a non-zero asymptote. In this case, including saturation into the tumor growth term *F(x,y)* leads to similar consequences. However, the hump in the function can disappear, eliminating any parameter region in which both equilibria can exist. In other words, with slower tumor growth, there is no parameter region anymore in which the tumor can escape the effect of the virus and grow out of control. Whether this occurs or not depends on the relative size of the two spatial scales involved. The first scale is defined by the tumor size at which the virus infection function *G* saturates and peaks in the context of exponential growth; this is entirely dependent on the properties of the viral growth term. Let us call this scale *s_v*, where the subscript refers to “viral”. The second scale is given by the colony size at which the tumor growth law starts to deviate from exponential; we will call this scale *s_t* (where the subscript refers to “tumor”). When *x_t*≤*s_v*, the asymptotic value of *G* becomes sufficiently large such that the hump disappears. The disappearance of the hump makes treatment easier, and this occurs if tumor growth slows down before virus growth does. ### Effect on slow virus growth Here, we assume slower virus growth terms that belong to class II. In the context of exponential growth, the function *G(x,y(x))* first increases, and then declines towards zero. This means that if equilibria exist, both the internal equilibrium *E_I* and the saddle node equilibrium *E_S* are aways present. Consequently, the possibility always exists that the tumor can out-run the virus infection and grow uncontrolled. Taking into account saturated tumor growth has the following effect. *(i)* The function *G* can remain qualitatively the same; that is, it rises to a peak and then declines towards zero. *(ii)* Alternatively, the picture can change such that it does not decline towards zero, but towards a non-zero asymptote, while remaining a one-humped function. *(iii)* Finally, the picture can change further such that the function *G* increases monotonically towards an asymptote. Which outcome is observed depends on the exact nature of the functions *F* and *G* and also the relative size of the two spatial scales involved: the tumor size at which the virus infection term *G* saturates and peaks (*s_v*), and the size *s_t* at which the pattern of tumor growth starts to deviate from exponential. Lowering the value of *s_t* relative to *s_v* shifts the outcome from scenario *(i)* to *(iii)*. As the value of *s_t* becomes similar to the value of *s_v*, the model contains parameter regions in which only the internal equilibrium *E_I* exists and in which uncontrolled tumor growth is impossible. If s_t≪s_v, then the hump in the function *G* disappears, and the saddle node equilibrium *E_S* is never present. In this case, virus-induced tumor control is the only outcome, and uncontrolled tumor growth cannot be observed. In biological terms, saturation of tumor growth at lower sizes promotes successful virus therapy. These arguments apply to all saturated tumor growth scenarios. With saturated but unlimited tumor growth, the function *G* approaches an asymptote for large tumor sizes *x*. For tumor growth that is limited by a carrying capacity *W*, the function *G* eventually deviates from the asymptote and rises again, indicating the presence of an equilibrium that describes tumor growth towards carrying capacity rather than towards infinity. Lowering the carrying capacity *W* has the same effect as lowering the parameter *s_t* that determines the tumor size at which growth starts to saturate: it shifts the outcome from scenario *(i)* to *(iii)*. ## Summary of model properties In summary this analysis has provided the following insights. We examined two types of infection terms and found that they strongly influence the dynamics of oncolytic virus spread. In the first class of models, virus spread was fast because infected cells are mixed among uninfected cells. In this case, tumor control is always observed if the viral replication rate lies above a threshold. In these parameter regions, loss of tumor control is not observed. In the second class of models, virus spread was assumed to be slow, because infected cells are clustered together in space. In this situation, the model can be characterized by bistability. If the initial number of tumor cells lies below a threshold, tumor control is observed. If the initial number of tumor cells lies above this threshold, uncontrolled tumor growth is observed. If tumor growth only saturates at high numbers of tumor cells or not at all, then uncontrolled tumor growth is always possible in parameter regions in which tumor control is possible. If tumor growth saturates at lower levels, there are parameter regions in which only the tumor control outcome is observed and in which uncontrolled tumor growth is not possible. If tumor growth saturates at even lower levels, then the bistability and the dependence on initial conditions vanishes completely. ## Properties of the internal equilibrium The above analysis concentrated on the equilibria. By examining which equilibria exist under different conditions, we can obtain information about the ability of the virus to control the cancer, and about the possibility that the cancer grows despite the presence of the virus. If the dynamics are governed by the internal equilibrium *E_I*, then the virus keeps the tumor cell population at relatively low levels and prevents uncontrolled tumor growth. We have discussed the conditions under which this can be achieved and interpreted these conditions from a biological angle. If the virus does control the tumor, however, additional questions arise. The virus can either control a persisting tumor at low levels, or the virus can drive the tumor cell population extinct. Because we are considering ordinary differential equations that describe the average behavior of the cell and virus populations, true extinction cannot occur in this model. The number of cells can, however, drop to very low levels. If the average number of cells is below one, we can assume that tumor extinction is a likely event. Therefore, if the number of tumor cells at equilibrium lies below one, we can say that the virus is likely to drive the tumor extinct. However, even if the equilibrium number of cells lies above one, the tumor cell population can still go extinct during oscillatory dynamics that can occur before the dynamics reach equilibrium. Therefore, we need to understand the properties of the internal equilibrium in more detail. We will examine this in the context of both fast and slow virus growth. We will only assume saturated tumor growth and not consider straight exponential tumor growth. ### Fast virus growth One of the most important parameters that influence the properties of the internal equilibrium is the replication rate of the virus *β*. In general, the faster the replication rate of the virus, the lower is the equilibrium number of tumor cells. Further, it can be shown that if the viral replication rate *β* crosses a threshold, the behavior near the equilibrium becomes oscillatory. Both promote the eradication of the cancer. In general, the internal equilibrium can either be stable or unstable, depending on the particular model under consideration as well as parameter values. Let us first consider the case where the equilibrium is stable. Then we can distinguish between two parameter regions. Denote the size at which tumor growth slows down and deviates from exponential by *s_t*. In the first parameter region, the value of *s_t* is large compared to a value related to the virus scale, *s_v* (for the exact definition see the). In this parameter region, we observe a viral replication rate threshold, at which the equilibrium number of tumor cells drops sharply from relatively high values to values of the order *1*. This replication rate threshold can be defined for individual models that belong to this class and defines the condition for cancer eradication. If the tumor size at equilibrium drops to small values (of the order of *1* cell), stochastic effects are very likely to lead to extinction. This is further supported by changes in the oscillatory approach to the equilibrium, which we have investigated in the context of individual models. At this viral replication rate threshold, the amplitude of the initial oscillations can increase sharply, as can the time it takes for the dynamics to approach the stable equilibrium (the real part of the eigenvalues of the Jacobian matrix rapidly approaches zero). Since pronounced oscillations reduce the number of tumor cells well below one, tumor eradication is the likely outcome. Note, however, that this drastic change in the oscillatory pattern is not observed in all models that belong to this class. The sharp drop in the equilibrium value is, however, a universal feature of models that belong to this class. Now assume the other parameter region in which the scale *s_t* is small. In this case, no such viral replication rate threshold exists. Instead, the equilibrium number of tumor cells declines proportional to the viral replication rate *β*. Numerical simulation of individual models, however, indicates that the minimum number of tumor cells can decline exponentially with an increase in the viral replication rate, although this could not be proved in general. Taken together, these findings indicate that in the parameter regions where virus replication is fast enough such that there is an oscillatory approach to the equilibrium, tumor eradication is the likely outcome. As mentioned above, it is also possible that the internal equilibrium *E_I* is unstable. In this case, we observe oscillations that diverge away from the equilibrium if the viral replication rate *β* is sufficiently fast. That is the amplitude of the oscillations increases over time. This is likely to correlate with extinction of the tumor, especially if the number of tumor cells at equilibrium is relatively low. This is because the oscillations will reduce the number of tumor cells well below the equilibrium value over time. Thus we conclude that for sufficiently large values of *β*, the cancer will be driven extinct by the virus through (convergent or divergent) oscillations. ### Slow virus growth In this case, the tumor size at the internal equilibrium is again negatively correlated with the viral replication rate *β*. Similarly to fast virus growth, the internal equilibrium can be stable or unstable depending on the individual model and on the parameter values. The dynamics will be discussed for both stable and unstable equilibria *E_I*. If the equilibrium is stable, the approach is again oscillatory if the viral replication rate is sufficiently large. Numerical simulations of individual models indicate that the minimum tumor size during these oscillations can decline exponentially with the viral replication rate *β*, although again this could not be proved in a general setting. These results indicate, however, that if oscillations are observed it is likely that the cancer is eradicated by the virus. Note that this assumes that the initial number of tumor cells is sufficiently small such that the population is in the region of attraction of the internal equilibrium. If this is not the case, the virus fails and unlimited virus growth occurs because the long-term outcome depends on the initial conditions as discussed above. In addition to these dynamics, the following can occur. Assume that the tumor cell population is reduced to low levels during the initial oscillations, but not to extinction. As the tumor cell population rises again, it can actually cross over to values larger than the saddle equilibrium *E_S*. Consequently, the cancer will grow uncontrolled and virus therapy will fail. On the other hand, if the internal equilibrium is unstable, then the following is observed. If the viral replication rate is fast enough, the populations show diverging oscillations away from the equilibrium, i.e. the amplitude of the oscillations increases over time. During these diverging oscillations, the minimum number of tumor cells declines over time. Hence, the tumor is likely to hit extinction. Again, there is the possibility that during the oscillations the tumor cell population crosses over to values larger than the saddle equilibrium *E_S*. In this case, the tumor cell population would grow uncontrolled to ever increasing levels. To summarize, for slow virus growth oscillations around the internal equilibrium have the potential to drive the tumor cell population extinct. However, the bi- stability of this system causes problems since there is always the possibility that the populations can escape to large numbers, leading to uncontrolled tumor growth. ## Application of models to experimental data Here we fit our models to previously published experimental data and discuss implications for model validation, model selection, and further experimental work. We examined data published by. This study considered A549 human lung cancer nude mouse xenografts, and infected them with the wild-type adenovirus Ad309 and a mutant virus Ad337 (characterized by a deletion in the E1b-19kD gene). The resulting dynamics were investigated under two conditions. (i) Under the first condition, the cancer cells were used to establish subcutaneous tumors in the mice. When the tumors reached a certain size, the virus was injected into the tumor. (ii) In a second scenario, infected cells were first mixed with uninfected cells, and the mixture was injected into the mice. The first scenario corresponds to spatially more restricted virus growth, while in the second scenario there is a higher degree of mixing between infected and uninfected tumor cells due to the experimental protocol. For both scenarios, we fitted models that differ in the infection term *G* and the tumor growth term *F*. We performed non-linear least squares regression, using standard software. The exact models that were used are provided in and. The parameter estimates obtained for all fits are tabulated in the. We first fitted the control tumor growth in the absence of the virus. Both exponential growth and saturated growth models (logistic, gompertzian, surface and linear, see) were applied. The saturated growth models fit the data better than exponential growth. All saturated growth models fit the data well, the logistic growth yielding the lowest (by a small margin) root mean square (RMS) error. For convenience, we chose logistic tumor growth as the basis for analyzing the effect of virus infection. First, consider experimental condition (i), where the tumor was allowed to grow in the mice before the virus was inoculated. Only the wild-type virus Ad309 is considered. In this experiment, tumor growth was significantly reduced by the virus. However, tumor size reached a plateau by day 50, despite the persistence of the virus, leading to the conclusion that the virus failed to eradicate the tumor cell population. For fitting purposes we considered one fast and one slow virus spread term, the first and the third in. shows that both a fast and a slow virus growth model can fit the data. However, extrapolating beyond the experimental time frame, very different long term outcomes are observed. The fast model predicts that the tumor remains at relatively low levels, controlled by the persisting virus infection. With the slow model, we show two parameter combinations which both fit the data well, but which are characterized by different long term outcomes. For one parameter combination (slow 1), damped oscillations are observed that lead to persistence of both the tumor and the virus at relatively low levels. For the second parameter combination (slow 2), the tumor cell population escapes control and grows to high levels. The virus population persists at low and ineffective levels (not shown). Therefore, not only do different models predict different long term dynamics; within one model, different parameter combinations that describe the data equally well can give rise to different predictions regarding the long-term dynamics. Our discussion of this and other results is postponed until the end of this section. Next consider experimental condition (ii), in which infected cells were mixed with uninfected cells at a ratio of 1∶1000 before the tumor was injected into the mice. In this case, the viruses were generally more effective. Tumor growth was prevented, and the number of tumor cells declined to low levels. fits a slow and a fast model to data that document infection with the wild type virus Ad309 and the mutant virus Ad337. Consider the wild type As309 virus first. All models fit the data well. Again, the predictions about the long-term dynamics vary, not only between models, but between different parameter combinations of the same model. Two qualitatively different outcomes are depicted in. On the one hand, the cancer can grow out of control following the initial reduction in the number of cancer cells. On the other hand, the virus maintains control of the cancer, which persists but is suppressed to relatively low levels. Thus, the encouraging but limited trend shown by the data cannot be used to conclude efficient virus- mediated tumor control. Longer experimental studies are needed in order gain insights into the eventual outcome of treatment, and to differentiate between the various model predictions. shows the same analysis for the mutant virus Ad337. As before, the slow and fast model can both fit the data well, and within one model, different parameter combinations are possible. The long-term dynamics show different outcomes, depending on the model and the parameter combinations. They include long term virus-mediated cancer control, as well as uncontrolled cancer growth. In the context of the experimental data, however, these long-term dynamics will not be observed, as the cancers regressed completely in the experiments. During the initial decline of the cancer cell population in the model, the number of cells drops to such low levels that extinction is actually the likely outcome in practical terms. However, what this tells us is that if by chance the cancer cell population does not hit extinction in the experiments, it is entirely possible that the cancer cell population rebounds and grows to high levels, depending on the model and its parameters. As mentioned above, the experiments include both a highly spatial setting where the virus was inoculated into an already established tumor, and a mixed setting where infected and uninfected cells were mixed before the tumor cells were placed into the mouse. Therefore, it can be tempting to examine whether the relative goodness of fit for the fast and slow models is different in these two situations. As explained, however, each model can fit the data with several alternative parameter combinations. There are many more solutions to the least squares regression than shown here. Therefore, it does not make sense to compare the goodness of fit for slow and fast models. For instance if the fits obtained for the slow model are slightly better than those obtained for the fast model, it is quite possible that there exists another parameter combination in the fast model that is better yet, and that has not been encountered so far. This brings us to the fundamental problem of nonlinear data fitting and model validation, which is an interesting issue in itself and will be discussed briefly here. As with many (and perhaps most) other nonlinear models, the parameter space where the minimization of the RMS error is performed, is multidimensional and is characterized by many shallow local minima. Most standard fitting routines get “stuck” at local minima, and even more sophisticated algorithms aimed at finding the global minimum are not very useful, because the difference between the global minimum and many runner-ups is usually insignificant and cannot serve as an indicator of the “right” or “best” fit. Therefore, to attack the fitting problem, one is required to repeat the minimization procedure multiple times, either by performing an exhaustive span of the space of the initial guesses, or by implementing a Monte-Carlo method. The statistics of the outcomes are then analyzed in the hope to find clusters of good fits, which are then assumed to be indicative of the solution of interest. The unfortunate part is that most of the times, these sophisticated statistical techniques are not very useful because the data sets that the fitting is applied to are simply too sparse, and they probably do not contain enough information to distinguish between models. Some of the deficiencies of experimental data sets are (i) an insufficient number of time-points, (ii) a very large experimental error at each time point, due to the experimental difficulties as well as a small sample size, and (iii) the long-term dynamics is often not captured due to the time-constraints of the experiment. In other words, no sophisticated statistical data manipulation can help distinguish between models if the data set is too sparse, short and contains large scatter. What can we conclude from these considerations and our own attempts to validate the models based on published experimental data? The good news is that at least some of the models contain parameter combinations which describe the existing data reasonably well. The bad news is that model validation/rejection was not possible in the particular system that we used. If data were collected over longer periods of time, and with a larger sample size, then the number of parameter combinations that can fit the data would be significantly reduced, and allow for more meaningful model comparison. # Discussion In this paper we presented the first modeling approach that tries to analyze the dynamics of oncolytic viruses in a general setting, going beyond particular models in which results can easily depend on mathematical terms chosen. Previous approaches to modeling oncolytic virus dynamics, and virus dynamics in general, have been based on particular models that include uncertain and unrealistic assumptions. The most striking is the assumption about the infection term, which usually assumes perfect mixing of populations, and which is certainly violated in any biologically realistic setting. Our method can be considered a hybrid between such space-free, mass-action approaches, and much more complex methods involving spatial network ideas, e.g.. The former approach fails to capture spatial and geometric constraints which play an important role in infection spread. The latter approach is only analytically tractable to a certain degree; also, it usually relies on a particular, given, set of rules that govern the infection spread. Our investigation aims to capture general trends that arise from different assumptions on the infection mechanism. It combines the analytical tractability of simple dynamical systems with a more realistic modeling of infection spread. We found that based on the infection term, we can divide models into two categories with fundamentally different behavior. In one group, virus growth is relatively fast because the infected cells are dispersed among the uninfected cells rather than being clustered together. In this case most infected cells contribute to virus spread. In these models, there is a clear viral replication rate threshold beyond which the number of cancer cells drops to levels of the order of one or less, corresponding to extinction in practical terms. Under this parameter region, this is the only outcome in this class of model. In the other category, infected cells are assumed to be clustered together to some degree in a mass, which might be realistic for solid tumors. In this case, only the infected cells located at the surface of the cluster contribute to virus spread because they are in the vicinity of uninfected cells. The infected cells located in the center of the cluster are surrounded only by other infected cells and therefore do not contribute to virus replication. The larger the number of infected cells, the smaller the proportion of cells that can pass on the virus. In this scenario, virus therapy is more difficult. If tumor growth saturates only at relatively large sizes or does not saturate, then even in the parameter regions where the dynamics can converge to tumor control or eradication, there can be the possibility that the cancer can outrun the virus if the number of cancer cells lies above a threshold at the start of virus therapy. This is because of the existence of the saddle node equilibrium which ensures dependence of the outcome on initial conditions. This might be problematic in clinical settings, because there is only a relatively small window between the size at which the tumor becomes detectable (about *10¹⁰* cells) and the size at which it can induce mortality (around *10¹³* cells).Tumor growth saturation at lower levels introduces a parameter region in which only the tumor control outcome is possible. A further reduction in the number of tumor cells at which growth saturation occurs can abolish the existence of the saddle node equilibrium altogether. In this case, the only outcome is tumor control. This result makes intuitive sense: earlier saturation of tumor growth slows down the cancer, and makes it easier for the virus to gain the upper hand. It also means that if the tumor is found early, it might be possible to slow down tumor growth by means of more conventional drug therapy, enabling the virus to control the cancer and to prevent runaway growth. There is indication in clinical data that a combination of chemotherapy and oncolytic virus therapy leads to better results than either approach alone. Another important finding of our study is that the basic results regarding the outcome of oncolytic virus therapy do not depend on the particular tumor growth terms used in the model. The exact kinetics of tumor growth are still poorly understood and a source of uncertainty. We examined straight exponential growth, as well as a number of more realistic options, including saturated but continued growth at high numbers of cancer cells, as well as cessation of growth as the number of tumor cells approaches an upper limit. While there are minor differences (such as the existence of a stable equilibrium at large tumor sizes vs continues slow growth), the properties of the tumor control equilibrium are largely independent from the exact way in which tumor growth is modeled. Throughout this paper we discussed the ability of the virus to eradicate the tumor in the context of our mathematical model that aims to describe oncolytic virus growth in relatively simple settings. It is important to point out that even simple scenarios could be characterized by complicating conditions which are not captured in the model and which make actual tumor extinction difficult to achieve. For example, tumor cells might become resistant to the virus by for example down-regulating the receptor required for viral entry. Related to this, cells could temporarily become resistant to virus-induced effects depending on the stage of the cell cycle. Such effects can be easily incorporated into our framework, if data suggest that they play a role in determining the dynamics of oncolytic virus growth. The framework presented here aims to bring us closer towards predictive computational models of oncolytic virus replication in vitro. Both additional computational and experimental work will be necessary to advance this framework. On the theoretical side, it will be important to also explore spatially explicit and stochastic models. The ordinary differential equations are desirable because they can be applied to experimental data in a relatively straightforward way. At the same time, however, spatial aspects of population growth can only be captured in a phenomenological way. Hence, it will be important to consider a spatially explicit model and to compare its properties to the results obtained here. On the empirical side, it will be important to run experiments that document the growth of specific oncolytic viruses e.g. in a culture of specific tumor cells. These data can be fitted to the various models explored here to determine which model describes the data best and which models can be rejected. This can be done in a variety of setting: a culture where cells and viruses can mix well; a 2D tissue culture which imposes a degree of spatial constraints; and a 3D tissue culture which can impose further spatial constraints. Different models will apply to these different scenarios. This will allow us to test the theoretical notions presented here, and to obtain a set of models that are predictive for the relevant scenarios. Of course, for clinical relevance, oncolytic virus replication needs to be considered in the context of more complex settings. Most importantly, the virus is immunogenic, and immune responses can inhibit the spread of the virus and can even drive it extinct. Such components will have to be incorporated into a mathematical model that describes the replication of an oncolytic virus in vivo. However, before we have obtained a solid understanding of the principles that govern the dynamics of oncolytic viruses in simpler settings, it is unlikely that modeling can contribute much to understanding the more complicated in vivo scenario. The modeling framework discussed here provides a basis to incorporate increasing amounts of biological complexity in the future, and thus to gradually improve our understanding of the key factors that determine the outcome of oncolytic virus therapy. # Materials and Methods The results described in this paper are based on the analysis of ordinary differential equations. Extensive mathematical details are provided in the. # Supporting Information [^1]: Conceived and designed the experiments: DW NK. Performed the experiments: DW NK. Analyzed the data: DW NK. Wrote the paper: DW NK. [^2]: The authors have declared that no competing interests exist.

# Introduction The microphthalmia-associated transcription factor (*MITF*) was discovered as the gene mutated in mice carrying the coat color mutation microphthalmia (*Mitf*). *Mitf* mutant mice lack melanocytes, resulting in pigmentation defects and deafness, and they have small eyes and some alleles show osteopetrosis (reviewed in). In humans, *MITF* mutations have been linked to the rare dominant pigmentation disorders Waardenburg Syndrome type 2A (WS2A) and Tietz Syndrome (TS) as well as the more serious COMMAD syndrome in compound heterozygotes. In addition, the *MITF* germline mutation E318K, has been linked to melanoma. MITF is regarded as the master regulator of the melanocyte lineage as it regulates the expression of various genes required for melanocyte development, proliferation and survival (reviewed in). The *MITF* gene is expressed in multiple isoforms that differ in their first exon and promoter usage. In most isoforms, the variable first exon is spliced to exon 1B1b which is then spliced to exon 2 and the following common exons which encode for the functionally important motifs necessary for DNA-binding, protein dimerization and transactivation ability. Among the exon 1B1b-containing isoforms, MITF-D is expressed in the human retinal pigment epithelium (RPE), whereas isoforms MITF-A, MITF-B, MITF-E and MITF-H are more ubiquitous. The shortest isoform, termed MITF-M, is predominant in melanocytes and contains a short exon 1M directly spliced to exon 2. Together with transcription factor EB (TFEB), TFE3 and TFEC, MITF forms a subfamily of related bHLHZip proteins, sometimes termed the MiT-TFE family. TFEB and TFE3 participate in the biogenesis of lysosomes and autophagosomes and the clearance of cellular debris upon starvation or lysosomal stress, through the activation of the CLEAR (Coordinated Lysosomal Expression and Regulation) network of target genes. More recently, a role has been described for MITF in regulating the starvation-induced autophagy response. The bHLHZip transcription factors, including the MiT-TFE subfamily, form homo- and/or heterodimers that bind to DNA and activate target genes. *In vitro* translated MITF has been shown to be able to form stable DNA-binding heterodimers with TFEB, TFE3, and TFEC. These proteins have nearly identical basic regions and very similar HLH and Zip domains. However, they fail to dimerize with other related bHLHZip factors such as c-Myc, MAX or USF due the presence of a 3 amino acid sequence in the MiT-TFE proteins, which limits dimerization within that family. Interestingly, the MITF-M isoform, which is predominantly expressed in melanocytes, is constitutively nuclear. This is in stark contrast to TFEB, TFE3 and other MITF isoforms, which have been shown to be located in the cytoplasm under normal conditions. The predominant nuclear localization of MITF-M has been explained by the absence of an N-terminal domain important for cytoplasmic retention, encoded by exon 1B1b. This 30 amino acid cytoplasmic retention domain is present in TFEB, TFE3 and all isoforms of MITF except for MITF-M. This domain allows for interactions with active Rag-GTPase heterodimers, which are required for the localization of these factors to the lysosome. In the case of TFEB, phosphorylation of TFEB by kinases such as mTORC1 and ERK2 promotes interaction with 14-3-3 proteins leading to subsequent cytoplasmic retention of TFEB. In this study, we show that the MITF, TFEB and TFE3 transcription factors are all produced in melanoma cells, albeit at different levels. MITF and TFEB can regulate each other’s expression, and the mTOR signaling pathway further regulates this cross-regulatory relationship by modulating their subcellular localization and transcriptional activity. Thus, pharmacological interventions that modulate the expression or activity of these factors may affect the balance in their expression and the cellular processes that they regulate. # Results ## MITF and TFEB modulate each other’s expression Previous analysis of RNA sequencing data from 368 metastatic melanoma tumors from The Cancer Genome Atlas (TCGA) showed that the mRNA expression of *TFE3*, *TFEB*, and *TFEC* was 4-, 14-, and 40-fold lower than that of *MITF*, respectively. Furthermore, analysis of gene expression in 23 human melanoma cell lines as well as in normal human epidermal melanocytes (NHEM) using a microarray platform revealed that the expression of *TFEB* and *TFE3* was roughly 50-fold lower than that of *MITF*, whereas expression of *TFEC* mRNA was about 850-fold lower than that of *MITF*. Particularly, in the 501Mel and Skmel28 human melanoma cell lines used in this study, *MITF* is highly expressed whereas *TFEB* and *TFE3* are expressed at considerably lower levels. *TFEC* is virtually undetectable in both cell lines. Although the expression level of transcription factors may not be directly related to their importance, we decided to focus on MITF, TFEB and TFE3 in the remaining analysis. The finding that the MiT-TFE factors are co-expressed in melanoma cells and tumors, together with recent evidence pointing to a role of MITF in the regulation of lysosomal and autophagy-related genes, similar to TFEB and TFE3, may suggest an overlap in function or cooperation between these factors. This led us to investigate whether MITF and TFEB could transcriptionally regulate each other’s expression. Publicly available ChIP-seq data for MITF in 501Mel and Colo829 human melanoma cells were analyzed in order to determine if MITF might be involved in directly regulating its own expression or that of TFEB and TFE3. A number of ChIP-seq peaks containing CACGTG or CATGTG elements were observed for MITF in the *MITF* gene, in both 501Mel and Colo829 cells. In addition, both ChIP-seq datasets showed that MITF binds to a region within intron 1 of *TFEB* containing an E-box CAGCTG sequence. In contrast, no ChIP-seq peaks were found within or near the *TFE3* gene. The peak closest to the *TFE3* gene is an E-box element located approximately 35 kb upstream of the transcription start site (TSS), between the neighboring genes *WDR45* and *PRAF2*, and was not considered further (10.6084/m9.figshare.12568646). These data suggest that MITF can regulate the expression of both *MITF* and *TFEB* through binding to putative DNA regulatory elements, whereas MITF is less likely to directly regulate the expression of *TFE3 in vitro*. In order to determine whether MITF and TFEB are able to influence each other’s expression, we investigated the effects of transiently overexpressing the two individual factors on their expression in 501Mel and Skmel28 human melanoma cell lines. We first separately overexpressed MITF or TFEB in 501Mel cells containing the doxycycline (DOX)-inducible piggybac (pBac) vectors pBac-MITF or pBac-TFEB prior to evaluating the mRNA levels of each factor by qRT-PCR. The overexpressed MITF protein was the isoform lacking exon 6a that encodes six amino acids located immediately upstream of the basic domain. We used primers specific for the MITF(+) isoform to determine the expression of endogenous *MITF* mRNA and universal MITF(+/-) primers to detect expression of total *MITF* mRNA; *TFEB* and *TFE3* were assayed using gene-specific primers. Overexpressing MITF significantly increased *TFEB* mRNA expression whereas *TFE3* levels remained unchanged. In addition, MITF overexpression significantly reduced the expression of endogenous *MITF* as detected using primers specific to the *MITF(+)* isoform. Overexpression of TFEB in the 501Mel cell line resulted in reduced *MITF* expression whereas *TFE3* mRNA levels were unaffected. These results were confirmed at the protein level by Western blot analysis—overexpression of TFEB resulted in reduced MITF protein expression. Ectopic expression of MITF led to decreased levels of endogenous MITF and increased TFEB protein levels. Quantification of the changes in protein expression were performed by normalizing each protein’s band intensity to the expression of Actin and are presented as a fold-change relative to the samples overexpressing an empty vector. To further examine the effect of MITF on TFEB expression, we treated 501Mel cells overexpressing FLAG-tagged MITF in a DOX-inducible pBac vector with increasing doxycycline concentrations and assayed for TFEB expression using Western blot analysis. As expected, we observed a dose-dependent increase in MITF-FLAG protein expression after 24 hours of doxycycline treatment. Importantly, increasing expression of the MITF protein led to a significant increase in TFEB protein expression, thus supporting the positive effects of MITF on TFEB. Since the MITF-FLAG pBac system overexpressed the MITF(+) isoform, it is unlikely that there are differences between the MITF (+) and isoforms with respect to effects on TFEB expression. In order to investigate this further, we transiently overexpressed the MITF(+) isoform in the 501Mel cell line and analyzed mRNA expression of *TFEB* and endogenous *MITF* using 3’UTR-specific primers. The effects were comparable to those observed after overexpression of the isoform, further validating that these six amino acids do not have additional effects on the expression of MITF or TFEB. We also performed overexpression of MITF(+) and TFEB in Skmel28 cells and observed similar effects as observed in the 501Mel cells. In order to validate our overexpression experiments, we used two sets of pooled siRNAs to separately knock down MITF or TFEB in 501Mel cells followed by RT-qPCR and Western blot analysis. The smart pool of siRNAs utilized were effective at targeting each individual factor as confirmed at both mRNA and protein levels. We observed that MITF knockdown dramatically reduced expression of both *TFEB* mRNA and protein to a degree comparable to that observed upon TFEB knockdown. On the other hand, TFEB knockdown increased expression of the MITF mRNA and protein. Quantification of the changes in protein expression were performed as the mean of three independent experiments by normalizing each protein’s band intensity to the expression of Actin and are presented as a fold-change relative to the samples treated with control siRNA. Of note, TFEB protein could not be detected with our TFEB-specific antibody in Skmel28 cells. However, in Skmel28 cells, siRNA-mediated silencing of MITF reduced the mRNA expression of *TFEB*. To further characterize the effects of siRNA-mediated silencing of MITF on *TFEB* expression, we employed an Skmel28 cell line expressing a doxycycline- inducible miRNA targeting MITF. The miR-MITF-mediated knockdown of MITF correlated with a significant reduction in *TFEB* mRNA levels 24 hours after addition of doxycycline into the cell culture medium. Importantly, after 24 hours of MITF knockdown, we washed off doxycycline treatment, and observed that *TFEB* mRNA was gradually restored at 72 and 96 hour doxycycline-free time points along with *MITF* mRNA. Collectively, these data indicate that MITF and TFEB are able to regulate each other’s mRNA and protein expression in human 501Mel and Skmel28 melanoma cells. ## MITF directly regulates *TFEB* expression We then asked whether the effects of MITF on the expression of *TFEB* are direct transcriptional effects involving DNA-binding by MITF. To address this, we cloned the DNA sequences in *TFEB* shown to be bound by MITF according to ChIP- seq analysis, into a pGL3-Promoter luciferase reporter plasmid. More specifically, we chose a fragment encompassing an 853-basepair (bp) sequence of intron 1 of *TFEB* containing the region between bases -29,373 and -28,521 located upstream of exon 2 of *TFEB*. This sequence contains a CAGCTG sequence, a potential MITF binding site. A mutated version of this construct was generated where the potential binding element was mutated to CCCTTT. The resulting reporter constructs were transfected into HEK293T cells, which express low levels of the MiT/TFE transcription factors endogenously, together with a construct expressing a FLAG-tagged wild type MITF-M protein. We used the *Tyrosinase* (*TYR*) promoter as a positive control, a classical MITF target. The results revealed transactivation of the *TYR* promoter by MITF, showing that our assay worked as intended. Expression from the wild type *TFEB* intron 1 element was significantly increased upon expression of FLAG-tagged MITF-M. However, this enhanced transactivation was abrogated when the 6-bp sequence was mutated, suggesting that MITF binds to this sequence and activates expression of *TFEB*. We further tested whether the effects on *TFEB* expression upon ectopic expression of MITF involved direct transcriptional regulation by using a mutant MITF protein, which lacks four arginines in the DNA binding domain that are essential for both DNA binding and nuclear localization of MITF. R214-217A mutant MITF exhibits constitutive cytoplasmic localization and presumably is transcriptionally inactive. Overexpression of this R214-217A MITF in HEK293T cells showed no reporter transactivation from the *TYR* promoter or the wild type *TFEB* intron 1 element, suggesting that the reporter transactivation previously observed requires the presence of transcriptionally active MITF protein. Interestingly, the MITF binding site in intron 1 of *TFEB* is a CAGCTG element and not a canonical CACGTG E-box or a CATGTG M-box, which have previously been described as target sequences for MITF transcriptional regulation. In order to validate that MITF can efficiently bind the CAGCTG sequence, we used electrophoretic mobility shift assay (EMSA) to determine whether an *in vitro* translated MITF-M protein can effectively bind a radiolabeled probe containing the above sequence. In this assay, we included a radiolabeled E-box (CACGTG) as a positive control for MITF-DNA binding. As expected, the MITF protein can shift the canonical E-box. It can also shift the CAGCTG probe with similar efficiency. The DNA-protein complexes were both further shifted with the anti-MITF C5 antibody, indicating that the complexes observed specifically bind MITF. The CAGCTG 6-mer is not a canonical E-box or M-box, both of which have been previously described as MITF binding sites. We thus analyzed the MITF ChIP-seq dataset (p\< 0.05) in order to find whether there are more occurrences of the CAGCTG regulatory element among MITF target genes. In the ChIP-seq dataset, 11,173 sequences under the peaks were found to contain putative MITF binding sites. The CACGTG motif was present at least once in 9,092 of the 11,173 sequences. The CATGTG element, core to the M-box regulatory element associated with several melanocyte-specific genes, was present in 8,328 sequences. The CAGCTG motif (the motif bound by MITF in the TFEB promoter) was found in 8,153 sequences. In contrast, the CCCTTT motif used as a scramble control was present in only 4,375 peaks. These results indicate that MITF directly regulates its target genes through direct binding to CANNTG motifs, including the transcription of *TFEB* through a CAGCTG motif located in *TFEB* intron 1. ## mTOR signaling affects the subcellular localization of MITF and TFEB in melanoma cells The subcellular localization of TFEB and TFE3 as well as of the MITF-A isoform has been shown to be regulated by the mTOR pathway. Phosphorylation of TFEB at Ser211 by mTORC1 promotes its cytoplasmic localization. In contrast, MITF-M, the main isoform of MITF in melanocytes and melanoma cells has been shown to be primarily nuclear. We analyzed the endogenous expression of MITF and TFEB in the 501Mel and Skmel28 cell lines by immunostaining and confocal microscopy. In addition, we used the mTOR pan-inhibitor Torin-1 to determine whether their subcellular localization responds to inhibition of this signaling pathway. Both endogenous MITF and TFEB were detected in 501Mel cells using specific antibodies. While TFEB was located in the cytoplasm and nucleus of 501Mel cells, MITF showed a major nuclear presence although a fraction of the protein was present in the cytoplasm. Treating 501Mel cells with the mTOR pan-inhibitor Torin-1 (1 μM, 3 hours) resulted in increased nuclear localization of the endogenous TFEB protein, suggesting that mTOR activity contributes to the cytoplasmic retention of TFEB in melanoma cells. Interestingly, the mTOR inhibitor also increased the fraction of MITF located in the nucleus. The same results were observed when constructs containing GFP-fusions of the MITF-M and TFEB proteins were overexpressed in 501Mel cells using a doxycycline-inducible pBac system, and subsequent treatment with Torin-1. The GFP-fusion of TFEB was mostly cytoplasmic whereas that of MITF was mostly nuclear, suggesting that overexpression does not affect the nucleocytoplasmic distribution of these factors. Torin-1 treatment led to nuclear localization of TFEB and a reduction in the cytoplasmic presence of MITF. This shows that the cytoplasmic portion of the MITF-M isoform is affected by mTOR signaling and that the cytoplasmic signal detected with an antibody against endogenous MITF is neither an artifact nor the result of cross-reactivity and is not due to the presence of alternative isoforms of MITF in these cells. Immunostaining of Skmel28 cells showed that MITF is expressed in these cells whereas TFEB protein was not detectable. Similar to 501Mel cells, MITF was mostly nuclear in this cell line and its cytoplasmic presence was reduced after treatment with the mTOR inhibitor. Next we hypothesized that the mTOR-induced changes in the subcellular localization of MITF and TFEB may affect their respective transcriptional regulation. We have shown that both MITF and TFEB negatively affect the expression of endogenous MITF in both melanoma cell lines used in this study (Figs,) and would therefore expect their increased nuclear presence upon Torin-1 treatment to further repress MITF expression. 501Mel cells containing the doxycycline-inducible pBac-MITF or pBac-TFEB vectors were treated with Torin-1 or a vehicle control (DMSO) prior to evaluating the mRNA levels of each factor by qRT-PCR. Overexpression of either MITF or TFEB inhibited the expression of endogenous *MITF* mRNA in 501Mel cells, as previously shown. The expression of endogenous *MITF* in pBac-MITF overexpressing cells was not further reduced upon Torin-1 treatment, however it was further decreased in Torin-1 treated cells containing pBac-TFEB as compared to vehicle-treated cells. These results fit the fact that MITF is consistently nuclear whereas TFEB is both cytoplasmic and nuclear but becomes predominantly nuclear upon treatment with Torin-1 and therefore can further reduce *MITF* gene expression. As expected, the pBac-MITF cells showed increased *TFEB* mRNA expression; this induction was not further enhanced by Torin-1 treatment. Taken together, these data indicate that the mTOR signaling pathway is active in melanoma cells and can affect import of TFEB into the nucleus. Increased nuclear localization of TFEB may subsequently enhance its transcriptional repression of *MITF*. In contrast, the melanocyte and melanoma cell-specific MITF-M isoform is mostly nuclear under basal conditions and highly expressed in several melanoma cell lines and tumors. Therefore, the subtle mTOR- dependent effects on its subcellular localization may not translate into significant modulation of its transcriptional activity. # Discussion MITF, TFEB, TFE3 and TFEC constitute the MiT-TFE subfamily of transcription factors featuring high structural homology. Their basic domains are identical and, unlike other members of the bHLHZip family, they share a three amino acid sequence in the HLHZip domain that enables them to restrict the formation of DNA-binding heterodimers to the MiT-TFE subfamily. MITF, TFEB and TFE3 are expressed to some extent across melanoma tumors and cell lines, whereas TFEC is not. Similarly, TFEB and TFE3 mRNA expression can be detected in normal melanocytes, albeit at lower levels than MITF. Although the expression of TFEB is low at the mRNA level, there may be sufficient protein in the cells to have major effects, especially taking into consideration the biological role of TFEB and recent findings regarding low affinity vs high affinity binding sites across the genome for a given transcription factor. Evidence for a degree of interplay between these factors comes from previous studies that have highlighted a role of the MiT-TFE factors in renal cell carcinoma. While the four MiT-TFE subfamily members are expressed at comparable levels in healthy human kidney, in subsets of renal cell carcinoma tumors the expression of these factors is significantly imbalanced. A fusion of *TFEB* located on chromosome 6 with the *Alpha* gene on chromosome 11 resulting in an *AlphaTFEB* fusion gene, links *TFEB* with the regulatory regions upstream of the *Alpha* gene, leading to promoter substitution and a 60-fold increase in expression. Additionally, translocations of the Xp11.2 region involving *TFE3* have been reported in up to 30–50% of pediatric papillary renal cell carcinoma cases. The most commonly occurring translocations of this locus fuse the *TFE3* gene with the *ASPL* or *PRCC* genes, resulting in chimeric proteins with aberrant function. Interestingly, TFE3 depletion inhibited proliferation of a renal carcinoma cell line, whereas ectopic overexpression of MITF in the TFE3-depleted cells rescued proliferation. Likewise, the proliferative defects induced by MITF knockout in an MITF-driven renal clear cell carcinoma model was restored by transfecting TFE3, indicating that MITF and TFE3 may have partially redundant roles in regulating proliferation and survival. Redundancy between MITF and TFE3 was also suggested by the observation that loss of function mutation in either gene leads to normal bone development whereas simultaneous knockout of both factors resulted in the development of severe osteopetrosis in mice. We show that overexpression of TFEB or MITF itself reduced the expression of endogenous MITF at both the mRNA and protein levels in two different melanoma cell lines. This indicates that MITF represses its own expression. This is consistent with ChIP-seq data showing that MITF binds to sequences within the *MITF* gene. MITF expression is regulated by multiple signaling mechanisms and transcription factors, including potential self-regulation. The peptide hormone α-MSH activates the MC1R receptor at the melanocyte membrane triggering cAMP signaling, which regulates the MITF-M promoter due to the cooperation of CREB with SOX10, a transcription factor that is expressed in several neural crest- derived cell lineages. Various other transcription factors, such as beta- catenin, LEF1 and PAX3 have been reported to regulate MITF transcriptionally. Previous studies using Northern blot analysis showed that α-MSH treatment resulted in increased *MITF* mRNA expression, peaking at two hours and entering a declining phase beyond this time point, indicating a homeostatic regulatory mechanism that was coupled with a decrease in protein expression beginning at 4 hours. Of note, MITF has been shown to induce the expression of both HIF1-alpha and miR-148a, which in turn inhibit the expression of MITF, resulting in oscillatory levels of MITF expression that are required for an adequate physiological response to UVB exposure. Whether MITF alone is sufficient for this negative feedback mechanism or if HIF1-alpha, miR-148a or other factors are also involved remains to be determined. Our results show that MITF positively regulates the expression of TFEB. This is consistent with ChIP-seq data indicating that MITF binds intron 1 of *TFEB*. Using reporter gene assays, we demonstrated that the increase in TFEB expression mediated by MITF is through direct binding to the CAGCTG motif located under the MITF ChIP-seq peak in intron 1 of *TFEB*. Our analysis of the MITF ChIP-seq dataset revealed that the CAGCTG 6-mer is as likely to be found under the peaks containing putative MITF binding sites as the canonical E-box and M-box, but more likely to be found than the CCCTTT motif used as a scramble control. These data as a whole suggest that the non-canonical E-box CAGCTG is a DNA regulatory element that is bound by MITF and mediates the regulation of TFEB expression by MITF. Previous analysis of the MITF structure bound to the CACGTG and CATGTG sequences showed that Arg217 of MITF forms specific bonds with the two central bases of the E-box (CACGTG) motif, whereas it does not form base-specific bonds with the two central bases of the CATGTG M-box motif. Instead, MITF forms specific bonds with the two bases flanking the 6-bp motif, -4 and +4, and with -3, -2 and +3, counting from the center of the motif, in the CATGTG M-box motif. This suggests that the two central bases within the subset of 6-bp E-box motifs are not always required for MITF binding, allowing a certain degree of flexibility for MITF’s function as a transcription factor. Recent studies have shown that the acetylation status of MITF impacts genomic occupancy as a means to modulate its transcriptional activity. Non-acetylated high DNA-binding-affinity MITF is able to bind a large pool of DNA loci including non-canonical degenerate motifs. In contrast, K243-acetylated MITF or the acetyl-mimetic K243Q mutant has low DNA-binding-affinity, yet robustly activates expression of melanocyte and melanoma target genes. It is possible that acetylation of MITF affects binding to the non-canonical CAGCTG motifs found in TFEB. Furthermore, mTORC1 has been shown to phosphorylate and positively regulate the p300 acetyltransferase, which in turn acetylates MITF, suggesting that the mTOR pathway might be capable of modulating not only the subcellular localization of the MiT-TFE factors, but also shift their genomic occupancy towards high-affinity sites. By immunostaining of cell preparations and confocal imaging we showed that in 501Mel cells TFEB is present in both the cytoplasm and the nucleus, whereas MITF-M is mostly nuclear, consistent with previous studies showing predominantly nuclear location of MITF-M. Inhibition of mTOR promoted nuclear shuttling of TFEB. The effects of blocking mTOR activity on the subcellular location of TFEB have the potential to modulate the autophagy response, which has been linked to increased vesicle trafficking and chemoresistance in melanoma. Consistent with this, in pancreatic ductal adenocarcinoma (PDA), but not in non-transformed human pancreatic ductal epithelial cells, the MiT-TFE factors have been shown to escape cytoplasmic retention mediated by mTOR regulation under fully fed conditions. This may enable PDA cells to constitutively induce autophagy activity, thus maintaining a high supply of amino acids. Moreover, a subset of genes involved in endolysosomal trafficking and autophagy have been found to be overexpressed in melanoma, suggesting that in some tumor types and under certain conditions, high levels of autophagy activity can be beneficial for tumor survival and/or progression. Altogether, our data is descriptive of a transcriptional cross-regulatory mechanism between MITF and TFEB that adds another layer of regulatory interactions beyond their ability to heterodimerize. In addition, changes in the subcellular localization of TFEB, such as those mediated by mTOR, affect its transcriptional activity and its regulation of *MITF* expression. The physiological relevance of their cross-regulation may be largely dependent on the relative abundance of each factor in a given tissue. Considering the prominent role of the MiT-TFE transcription factors in regulating various basic process as well as their roles in cancer, characterizing their expression patterns and how they are being modulated is instrumental for improving our understanding of their role in healthy tissue and in pathogenesis, and how best to identify targetable vulnerabilities in their action. # Materials and methods ## Cell culture Two human melanoma cell lines were used in this study, 501Mel cells (generously donated by Ruth Halaban) and SkMel28 cells (#HTB-72, ATCC). HEK293T human embryonic kidney cells (#CRL-3216, ATCC) were used for transactivation assays. All cells were grown in DMEM medium (#10569–010, GIBCO) supplemented with 10% fetal bovine serum (FBS \#10270–106, GIBCO). Cells were grown at 37°C and 5% CO₂ and medium was changed two to three times per week. ## Plasmid constructs and cloning In order to induce expression of the different transcription factors at will, we used an inducible piggyback (pBac) system. We generated inducible 501Mel cells by transfecting the cells with three pBac vectors, one containing GFP-tagged human MITF, TFEB, TFE3 or GFP alone, one containing a reverse-tetracyclin transcription activator, and one containing transposase. The pBac vectors were a gift from Dr. Kazuhiro Murakami (Hokkaido University, Japan). The GFP-tagged MITF and TFEB cDNAs were amplified from plasmids pEGFP-N1-MITF-M (Addgene plasmid \# 38131) and pEGFP-N1-TFEB (Addgene plasmid \# 38119), using the primers listed in (pBac-EGFP), and then introduced into the pBac vector by restriction digestion using *Mlu* I and *Not* I sites and ligation at a 3:1 insert to backbone ratio using Instant Sticky-end Ligase Master Mix (M0370S, NEB). For the generation of the inducible 501Mel cells carrying the MITF-M-FLAG- HA pBac plasmid, a FLAG-tagged MITF-M cDNA was amplified from the p3XFLAG- CMV^TM-14 plasmid expressing mouse Mitf-M, kindly provided by Colin Goding (Ludwig Institute, Oxford University, UK), using the primers listed in (pBac-MITF-M-FLAG-HA), and then introduced into the pBac vector by restriction digestion with *EcoR* I and *Spe* I. Cells transfected with the pBac plasmids were cultured in DMEM supplemented with 10% FBS and kept under G418 selection (#10131–035, GIBCO) for 8 days to obtain stable cell lines that are inducible by adding 0.2 μg/mL doxycycline to the culture medium. The p3XFLAG- CMV^TM-14 construct expressing mouse Mitf-M (MITF-M-FLAG) was the one used for MITF-M overexpression in the transactivation assays. The R214-217A mutation was introduced into MITF using the Q5 Site-directed Mutagenesis Kit (New England Biolabs, Ipswich, MA) according to the manufacturer’s instructions. The pGL3-Basic vector (#E1751, Promega) was used as a control for *Tyrosinase* (pTYR) transactivation, whereas pGL3-Promoter vector (#E1761, Promega) was used as a control for the *TFEB* intron 1 enhancer (TFEB-int1) and mutated *TFEB* intron 1 (TFEB-int1-mut) promoter constructs. The *TYR* promoter in a luciferase reporter plasmid was constructed using a 380 bp region at the promoter of the human *TYR* gene (bases -382 and -3 upstream of the TSS at location +1). This sequence was then cloned into a pGL3-Basic vector upstream of the luciferase reporter gene and verified by Sanger sequencing. The fragment containing TFEB- int1 was amplified from human genomic DNA using the primers listed in. The 853 bp fragment was blunt-ligated into Nhe1-site of the pGL3-Promoter vector. TFEB- int1-mut was generated from the TFEB-int1 luciferase reporter plasmid by mutating CAGCTGA to CCCTTTA using *in vitro* mutagenesis. All the constructs were confirmed by sequencing (Genewiz, Essex, UK). All the mutagenesis primers are listed in. ## MITF knockdown cell lines In order to be able to induce knockdowns of MITF at will, we generated two piggybac (pBac) constructs containing miRNAs under the regulation of an inducible promoter. Skmel28 doxycycline-inducible MITF knockdown cell lines were generated using a piggybac transposable vector pPBhCMV_1-miR(BsgI)-pA-3 obtained from Dr. Kazuhiro Murakami (Hokkaido University). MicroRNAs target sequences were selected using the BLOCK-iT RNAi Designer, specifically targeting both MITF-M exon 2 (miR(MITF-X2) and 8 (miR(MITF-X8). The non-targeting control miR- NTC was used as a negative control. The BLOCK-iT RNAi Designer was also used to design primers for inserting the pre-miRNA (including a mature miRNAi sequence, terminal loop and incomplete sense targeting sequence required for the formation of the stem-loop structure) RNAi into the murine miR-155 cassette in the pBac vector pPBhCMV_1-miR(BsgI)-pA-3 containing the reverse tetracycline transcription activator. Sequences of the mature miRNAs and the primers used for the generation of the pre-miRNAs are listed in. Primers were annealed by initial denaturation at 95°C followed by slow cooling in a water bath forming a short double stranded DNA with overhangs matching a BsgI overhang. The backbone vector was digested with BsgI (#R05559S, NEB) and the vector DNA purified after running the DNA on an agarose gel. The backbone vector and the annealed primers were ligated at 15:1 insert to backbone molar ratio using Instant Sticky-end Ligase Master Mix (M0370S, NEB). The ligation products were transformed into high- competent cells and the plasmid DNA isolated from the individual clones were screened as described above in Mutagenesis and cloning. For generation of miR- MITF cell lines, Skmel28 cells were transfected with the transposase-containing plasmid pA-CAG-pBase, the plasmids pPBhCMV_1-miR(MITF_X2)-pA and pPBhCMV_1-miR(MITF_X8)-pA encoding miRNA targeting two different exons of MITF and the plasmid pPB-CAG-rtTA-IRES-Neo (10:5:5:1) that confers resistance to neomycin. For miR-NTC cell lines, Skmel28 cells were transfected with pA-CAG- pBase, pPBhCMV_1-miR(NTC)-pA encoding a non-targeting miRNA and pPB-CAG-rtTA- IRES-Neo (10:10:1). After 48 hours of transfection, miR-MITF, miR-NTC vector and non-transfected cells were selected with 0.5mg/ml G418 (#10131–035, GIBCO) for 2 weeks and 1μg/ml of doxycycline was used for induction. ## Transfection of plasmids Cells were cultured in 6 or 12-wells plates one day before transfecting them with 2 μg of plasmid DNA and 6 μL transfection reagent FuGENE HD (#E2311, Promega) in 100 μL of serum-free culture medium per mL of cell culture medium. The medium with the transfection complexes was removed after 24 hours and replaced with fresh culture medium. 48 hours after transfection, cells were harvested for RNA or protein extraction. ## RNAi treatment Cells were cultured in 6 or 12-wells plates one day prior to transfection with the appropriate siRNA pools. Cells were transfected with 10 nM siRNA pools and 1 μL Lipofectamine RNAiMAX (#13778075, Invitrogen) transfection reagent in 100 μL of Optimem Pro (#31985–062, GIBCO) transfection medium per mL of cell culture medium. Cells were harvested for RNA or protein extraction 2 days after transfection. The siRNAs used for the procedure were the following: siRNA pool for human MITF (#4390824, ID S8792, Ambion), siRNA pool for human TFEB (#M-009798-02, Dharmacon) and a control siRNA (#4390843, Ambion). ## Protein extraction and immunoblotting For total protein extraction, cells were cultured in 6 or 12-wells plates and lysed in Laemmli sample buffer and boiled at 95°C for 5 minutes. The samples were then run on 8% or 10% gels and blotted onto a 0.2 μm PVDF membrane (#88520, Thermo Scientific). The membranes were blocked with 3% BSA in TBS-T (0.1% Tween 20 in TBS) for 1 hour at room temperature, and stained overnight at 4°C with 3% BSA in TBS-T and one of the following primary antibodies: MITF (MS771-PABX, Thermo Scientific), TFEB (#4240, CST) (#2775, CST), GFP (#ab290, Abcam), FLAG (#F3165, Sigma Aldrich) and Actin (MAB1501, Millipore). Membranes were washed with TBS-T and stained for 1 hour at room temperature with fluorescent secondary antibodies: anti-mouse IgG(H+L) DyLight 800 conjugate (#5257, CST) and anti- rabbit IgG(H+L) DyLight 680 conjugate (#5366, CST). The images were captured using Odyssey CLx Imager (LI-COR Biosciences). ## Real-time quantitative PCR for gene expression analysis TRIzol reagent (#15596–026, Ambion) followed by isopropanol precipitation was used for total RNA extraction. The cDNA was generated according to the manufacturer’s instructions with High-Capacity cDNA Reverse Transcription Kit (#4368814, Applied Biosystems). Primers for RT-qPCR were designed using NCBI Primer BLAST, and RT-qPCR performed with SensiFAST SYBR Lo-ROX Kit (#BIO-94020, Bioline) using a CFX384 Touch Real-Time PCR Detection System (Bio-Rad). The RT- qPCR reactions were performed using 1 ng/μL cDNA per reaction in technical triplicates and the fold change in gene expression was calculated with the 2(-Delta Delta C(T)) method, normalized to the geometrical mean of Actin and human ribosomal protein lateral stalk subunit P0 (RPLP0) expression. Standard curves for primer efficiency were calculated using the formula E = 10^(-1/slope) for each primer pair. ## Immunostaining and confocal imaging 501Mel and Skmel28 cells (3x10⁴ per well) were cultured for 48 hours in 8-well chamber slides (#354108 from Falcon). For the inducible 501Mel cell lines, 0.2 μg/mL of doxycycline were added to the cell culture medium. At day 2, cells were fixed for 2 min with 2% paraformaldehyde (PFA) in cell culture medium and then for 15 min with 4% PFA in PBS. For imaging of the overexpressed GFP- tagged factors, cells were washed 3 times with PBS, followed by DAPI staining (#D-1306, Life Technologies) and two additional washes with PBS. The chambers were then removed and the samples allowed to dry prior to mounting in Fluoromount-G™ (#00-4958-02, Invitrogen). For immunostaining of the endogenous MiT/TFE factors, following the treatment with the Torin-1 (#4247, Tocris) inhibitor, cells were washed once with PBS after fixation, then permeabilized for 8 min with 0.1% Triton X-100 in PBS, followed by three washes with PBS. They were then blocked with blocking buffer (5% normal goat serum, 0.05% Triton X-100 and 0.25% BSA in PBS) for 1 hour at room temperature, and stained overnight at 4°C with 0.25% BSA in PBS antibody buffer containing the primary antibodies: MITF (MS771-PABX, Thermo Scientific) and TFEB (#4240, CST) (#2775, CST). The cells were then washed three times with PBT (0.1% Tween-20 in PBS) and stained for 1 hour at room temperature with the Alexa Fluor 546 goat anti-mouse IgG(H+L) (#A11003, Invitrogen) or the Alexa Fluor 488 goat anti-rabbit IgG(H+L) (#A11070, Invitrogen) fluorescent secondary antibodies diluted in PBT. Subsequently, cells were washed twice with PBT and once with PBS and finally stained with DAPI and prepared for imaging as previously described. Imaging was performed using a FluoView FV1200 laser scanning confocal microscope (Olympus) equipped with a PlanApo N 60X/1.40 ∞/0.17 Oil Objective. ## Transcription activation assays HEK293T cells (1.5x10⁴ per well) were seeded in white 96-well plates (#781965, BRAND) and cultured for 24h prior to transfection (FuGENE, Promega) with 33 ng of a construct carrying the relevant regulatory region (Tyr, TFEB- int1 or TFEB-Int1-mut) coupled to the luciferase reporter, 33 ng of an MITF-M construct and 33 ng of a pRL Renilla control reporter vector. Cells were assayed 24 hours after transfection using the Luciferase DualGlo kit (E2940, Promega) as described by the manufacturer. The luminescence activity was measured in a multimode microplate reader (GloMax, Promega) with a 300-millisecond reading per well. The luciferase signal of each sample was normalized to the renilla signal for transfection efficiency and cell viability. The pGL3-Basic (#E1751, Promega) or pGL3-Promoter (#E1761, Promega) vectors were used in order to calculate the fold induction of each respective regulatory element activity. Three technical replicates per sample were included and the assay was performed in at least three biological replicates. Error bars indicate SEM and statistical significance was assessed with student’s t-tests. ## Electrophoretic Mobility Shift Assay (EMSA) Two DNA fragments were generated for the EMSA studies by synthesizing the oligos E-box-Fw (containing the sequence `5’-AAA GTC AGT CAC GTG CTT TTC AGA-3’`) and E-box-Rv (`5’-GTC TGA AAA GCA CGT GAC TGA CTT T-3’`) for the canonical E-box (containing the CACGTG motif). CAGCTG-box-Fw (containing the sequence `5’-AAA GTC AGT CAG CTG ATT TTC AGA-3’`) and CAGCTG-box-Rv (`5’-GTC TGA AAA TCA GCT GAC TGA CTT T-3’`) were synthesized for the CAGCTG motif-containing probe. Subsequently, the oligos were allowed to anneal and labeled with α-³²P-dCTP, (#BLU013H100UC, PerkinElmer). The labeled probes were purified on Sephadex G-25 Quick Spin columns (#11273922001, Roche). The EMSA was performed according to Pogenberg et al.. Briefly, the MITF protein was expressed from a plasmid containing wild type MITF-M under a T7 promoter, using the TNT-T7 Quick Coupled Transcription/Translation System (#L1170, Promega). 2 ul of TNT-translated MITF was preincubated in a buffer containing 20 ng of cold probe poly(dI–dC), 10% fetal calf serum, 2 mM MgCl2, and 2 mM spermidine for 15 min on ice. For supershift assays, 0.5 mg of mouse monoclonal antiMITF (C5) antibody (#ab12039, Abcam) were added and incubated on ice for 30 min. Then, 50,000 counts per minute (cpm) of each of the two ³²P-labeled probes in a binding buffer containing 10 mM Tris (pH 7.5), 100 mM NaCl, 2 mM dithiothreitol, 1 mM EDTA, 4% glycerol, and 80 ng/mL salmon sperm DNA were added to the preincubated MITF protein solution in a total reaction volume of 20 ul and incubated for 10 min at room temperature. The resulting DNA–protein complexes were resolved on 4.2% non-denaturing polyacrylamide gels, placed on a storage phosphor screen, and then scanned using a Typhoon PhosphorImager 8610 (Molecular Dynamics). ## ChIP-seq data analysis and motif scan Raw FASTQ files for MITF ChIP-seq were retrieved from GEO archive under the accession number GSE50681 and GSE61965 and subsequently mapped to the human reference genome hg19 using bowtie 2. Then the aligned reads from bowtie2 were used to call peaks with MACS. Following, wig and bed files were generated using the p-value\<0.05. For visualization of the peaks, wig files were uploaded on the IGV genome browser. For scanning the motifs, DNA sequences corresponding to MITF peaks were extracted from the UCSC genome browser. The R package Biostrings was used to scan each motif and the resulting number of motifs for each peak sequence were plotted in R. ## Statistical analysis Results are represented as the mean from three or more independent experiments with standard error of the mean (SEM). Graphpad Prism 7 was used for all the statistical analyses. Analyses of RT-qPCR and Western blot upon overexpression of the pBac factors were performed using multiple t-tests comparing each cell mean to the control cell mean and multiple comparison correction by Holm-Sidak method with statistical significance set as \*P\<0.05. Analysis of the transactivation assays was performed using one-way analysis of variance (ANOVA). All the remaining statistical analyses were performed using two-way ANOVA. ANOVA analyses were performed comparing each cell mean to the control cell mean and multiple comparison correction by Dunnett method for one-way ANOVA and correction by Sidak method for two-way ANOVA with statistical significance set as \*P\<0.05. # Supporting information We thank Colin Goding for critical comments on the manuscript. 10.1371/journal.pone.0238546.r001 Decision Letter 0 Roemer Klaus Academic Editor 2020 Klaus Roemer This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. # Transfer Alert This paper was transferred from another journal. As a result, its full editorial history (including decision letters, peer reviews and author responses) may not be present. 19 Jun 2020 PONE-D-20-15262 MITF and TFEB cross-regulation in melanoma cells PLOS ONE Dear Dr. Steingrimsson, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please respond to all critique, point-by-point. In particular: \- Consult databases that provide insight into the expression of the MiT-TFE gene family. \- Discuss pool of low-affinity bdg. sites of MITF and the possible influence of acetylation. Please submit your revised manuscript by Aug 03 2020 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols> We look forward to receiving your revised manuscript. Kind regards, Klaus Roemer Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. PLOS ONE now requires that authors provide the original uncropped and unadjusted images underlying all blot or gel results reported in a submission’s figures or Supporting Information files. This policy and the journal’s other requirements for blot/gel reporting and figure preparation are described in detail at <https://journals.plos.org/plosone/s/figures#loc-blot-and-gel- reporting-requirements> and <https://journals.plos.org/plosone/s/figures#loc- preparing-figures-from-image-files>. When you submit your revised manuscript, please ensure that your figures adhere fully to these guidelines and provide the original underlying images for all blot or gel data reported in your submission. See the following link for instructions on providing the original image data: <https://journals.plos.org/plosone/s/figures#loc-original-images-for-blots-and- gels> In your cover letter, please note whether your blot/gel image data are in Supporting Information or posted at a public data repository, provide the repository URL if relevant, and provide specific details as to which raw blot/gel images, if any, are not available. Email us at <[email protected]> if you have any questions. 3\. We note that you have included the phrase “data not shown” in your manuscript. Unfortunately, this does not meet our data sharing requirements. PLOS does not permit references to inaccessible data. We require that authors provide all relevant data within the paper, Supporting Information files, or in an acceptable, public repository. Please add a citation to support this phrase or upload the data that corresponds with these findings to a stable repository (such as Figshare or Dryad) and provide and URLs, DOIs, or accession numbers that may be used to access these data. Or, if the data are not a core part of the research being presented in your study, we ask that you remove the phrase that refers to these data. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The manuscript is interesting and is well structured in general. The Authors conducted a number of assays using different techniques to reveal MITF and TFEB cross-regulation in melanoma cells. Moreover, in my opinion it could be interesting for a reasonable number of scientists since the studied cross-regulatory interactions need to be taken into account when considering pathways regulated by these proteins - e.g. as a potential mechanism underlying antimelnoma effect of the new treatment options. I recommend publication of this study in its present form. Reviewer \#2: This manuscript sets out to investigate the potential for cross regulation of the MiT-TFE family of genes, specifically how the melanocyte master regulator MITF may be influenced by the other family members. Previous studies have shown that the MITF mRNA is expressed at much higher levels than the TFE3, TFEB and TFEC genes in human melanoma cell lines and tumors. The authors have chosen to look at two melanoma cell lines 501Mel and Skmel28 to conduct a series of experiments, and first exclude the potential of TFEC for regulation as its expression is undetectable. Through the examination of ChIP- seq data it was found that MITF can bind the MITF and TFEB loci regulatory regions but less likely interact with TFE3. From these initial experiments the focus was then on the cross regulation of MITF and TFEB. The interaction of these genes at both the mRNA and protein level was studied by transient overexpression of each protein in these melanoma lines. This found that the MITF protein induced TFEB but reduced the endogenous MITF transcript, with TFEB reducing MITF. The effect of TFEB on endogenous TFEB is not addressed? These results were validated using an siRNA KD approach. This negative feedback loop is illustrated in Figure 6. The binding of MITF to the non-consensus E-box CAGCTG sequence within the region of the first intron of TFEB was then demonstrated using a luciferase reporter construct and by DNA binding EMSA assay. This sequence element was then further analysed in the ChIP-seq dataset for MITF confirming it as an efficient binding site. The nuclear/cytoplasmic localisation of MITF and TFEB and effect of mTOR were examined by immunostaining and Torin-1 treatment of melanoma cells. This data clearly demonstrates MITF as a major presence in the nuclear fraction, whereas TFEB located in the cytoplasm and nucleus which the can move to the nucleus following inhibition of mTOR by Torin-1. Things to consider: 1\. In the final sentence of the manuscript “… improving our understanding of their role in healthy tissue …” does not consider what the expression status of the MiT-TFE family may be in normal melanocytic cells. The authors may consider if there are any databases examining expression of these genes in different tissues could be consulted e.g. Melanocytes in the skin--comparative whole transcriptome analysis of main skin cell types. Reemann P, Reimann E, Ilmjärv S, Porosaar O, Silm H, Jaks V, Vasar E, Kingo K, Kõks S.PLoS One. 2014 Dec 29;9(12):e115717. This data was extracted from. S1 Table. RPKM values of genes we detected in MC, KC, FB and the whole skin. Melanocytes genes MC_1 MC_2 MC_3 MC_4 ENSG00000187098 199.197 156.316 191.509 167.668 MITF ENSG00000112561 0.400417 1.12329 0.798733 0.35943 TFEB ENSG00000068323 1.13973 1.78558 1.92091 1.13326 TFE3 ENSG00000105967 0.0451891 0.048543 0.0621599 0.0607503 TFEC 2\. The recent paper describing acetylation of MITF, in which the senior author of this manuscript is a co-author, should at least be cited. Some discussion of how MITF has a pool of low affinity binding sites that may keep it in the nucleus and how may this differ or be similar for the other MiT-TFE family, and effects of acetylation on other family members such as TFEB could be included. Tuning Transcription Factor Availability through Acetylation-Mediated Genomic Redistribution. Louphrasitthiphol P, Siddaway R, Loffreda A, Pogenberg V, Friedrichsen H, Schepsky A, Zeng Z, Lu M, Strub T, Freter R, Lisle R, Suer E, Thomas B, Schuster-Böckler B, Filippakopoulos P, Middleton M, Lu X, Patton EE, Davidson I, Lambert JP, Wilmanns M, Steingrímsson E, Mazza D, Goding CR.Mol Cell. 2020 Jun 4:S1097-2765(20)30345-2. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0238546.r002 Author response to Decision Letter 0 6 Jul 2020 Dear editor, We thank PLOS One for the opportunity to resubmit our manuscript after addressing the concerns of reviewers and editor. We also thank the reviewers for their critical comments and analysis. Following are the specifics of the changes we made: 1\. We have made sure that the manuscript fits the PLOS ONE style requirements and renamed the files appropriately. 2\. The uncropped and unadjusted images are provided for all images as supporting information. 3\. The data previously referred to as “data not shown” is now available on Figshare (DOI: 10.6084/m9.figshare.12568646). Reviewer \#2: Comment 1: In the final sentence of the manuscript “… improving our understanding of their role in healthy tissue …” does not consider what the expression status of the MiT-TFE family may be in normal melanocytic cells. The authors may consider if there are any databases examining expression of these genes in different tissues could be consulted e.g. Melanocytes in the skin-- comparative whole transcriptome analysis of main skin cell types. Reemann P, Reimann E, Ilmjärv S, Porosaar O, Silm H, Jaks V, Vasar E, Kingo K, Kõks S.PLoS One. 2014 Dec 29;9(12):e115717. This data was extracted from. S1 Table. RPKM values of genes we detected in MC, KC, FB and the whole skin. Melanocytes genes MC_1 MC_2 MC_3 MC_4 ENSG00000187098 199.197 156.316 191.509 167.668 MITF ENSG00000112561 0.400417 1.12329 0.798733 0.35943 TFEB ENSG00000068323 1.13973 1.78558 1.92091 1.13326 TFE3 ENSG00000105967 0.0451891 0.048543 0.0621599 0.0607503 TFEC We have consulted our own array-based gene expression data where the expression of MITF, TFEB and TFE3 was determined in melanocytes and melanoma cells. We have introduced the following into the results section, page 6, lines 72-76: “Furthermore, analysis of gene expression in 23 human melanoma cell lines as well as in normal human epidermal melanocytes (NHEM) using a microarray platform revealed that the expression of TFEB and TFE3 was roughly 50-fold lower than that of MITF, whereas expression of TFEC mRNA was about 850-fold lower than that of MITF \[21,31\].” We also added the following in the Discussion section, page 17, lines 317-317: “MITF, TFEB and TFE3 are expressed to some extent across melanoma tumors and cell lines, whereas TFEC is not \[21,31\]. Similarly, TFEB and TFE3 mRNA expression can be detected in normal melanocytes, albeit at lower levels than MITF \[31,41\].” Comment 2. The recent paper describing acetylation of MITF, in which the senior author of this manuscript is a co-author, should at least be cited. Some discussion of how MITF has a pool of low affinity binding sites that may keep it in the nucleus and how may this differ or be similar for the other MiT-TFE family, and effects of acetylation on other family members such as TFEB could be included. Tuning Transcription Factor Availability through Acetylation-Mediated Genomic Redistribution. Louphrasitthiphol P, Siddaway R, Loffreda A, Pogenberg V, Friedrichsen H, Schepsky A, Zeng Z, Lu M, Strub T, Freter R, Lisle R, Suer E, Thomas B, Schuster-Böckler B, Filippakopoulos P, Middleton M, Lu X, Patton EE, Davidson I, Lambert JP, Wilmanns M, Steingrímsson E, Mazza D, Goding CR.Mol Cell. 2020 Jun 4:S1097-2765(20)30345-2. We have entered a paragraph discussing this interesting study in the page 20, between lines 374-383, of the Discussion that reads the following: “Recent studies have shown that the acetylation status of MITF impacts genomic occupancy as a means to modulate its transcriptional activity. Non-acetylated high DNA-binding-affinity MITF is able to bind a large pool of DNA loci including non-canonical degenerate motifs. In contrast, K243-acetylated MITF or the acetyl-mimetic K243Q mutant has low DNA-binding-affinity, yet robustly activates expression of melanocyte and melanoma target genes (54). It is possible that acetylation of MITF affects binding to the non-canonical CAGCTG motifs found in TFEB. Furthermore, mTORC1 has been shown to phosphorylate and positively regulate the p300 acetyltransferase (55), which in turn acetylates MITF (49, 54, 56), suggesting that the mTOR pathway might be capable of modulating not only the subcellular localization of the MiT-TFE factors, but also shift their genomic occupancy towards high-affinity sites.” In addition, we reason that this study may be relevant to the observation that although the expression of TFEB is low in melanocytes and in some melanoma tumor and cell lines, the protein levels of this factor may be sufficient for binding and activation for a subet of high-affinity binding sites. This statement can be found in the Discussion section, pages 17-18 between lines 317-320: “Although the expression of TFEB is low at the mRNA level, there may be sufficient protein in the cells to have major effects, especially taking into consideration the biological role of TFEB and recent findings regarding low affinity vs high affinity binding sites across the genome for a given transcription factor \[42\].“ Best regards 10.1371/journal.pone.0238546.r003 Decision Letter 1 Roemer Klaus Academic Editor 2020 Klaus Roemer This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 19 Aug 2020 MITF and TFEB cross-regulation in melanoma cells PONE-D-20-15262R1 Dear Dr. Steingrimsson, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Klaus Roemer Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: 10.1371/journal.pone.0238546.r004 Acceptance letter Roemer Klaus Academic Editor 2020 Klaus Roemer This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 21 Aug 2020 PONE-D-20-15262R1 MITF and TFEB cross-regulation in melanoma cells Dear Dr. Steingrímsson: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Klaus Roemer Academic Editor PLOS ONE [^1]: The authors declare no competing interests. [^2]: Current address: The Buck Institute for Research on Aging, Novato, CA, United States of America [^3]: Current address: Department of Dermatology, Medical University of Vienna, Vienna, Austria [^4]: Current address: Institute of Molecular Life Sciences, University of Zurich, Zurich, Switzerland

# Introduction In recent years, research has suggested animals possess capacities thought to be uniquely human. Perhaps particularly interesting is the observation of such capacities in birds, whose brain anatomy is distinctly different from that of mammals. At least some birds are innovative tool manufacturers and users who can plan for the future and engage in causal reasoning. Others show remarkable long- term memory capacity, inference, episodic-like memory, and striking regularity in how they use past experience to navigate novel present conditions. These avian displays of flexible, future-oriented behavior suggest that at least some elements of complex cognition are fundamental learning processes present in all species. The ability to perceive the passage of time is critical in many complex behaviors. For example, episodic memory is characterised by knowing one’s location in time, and planning for the future requires learning about extended temporal relations in one’s environment and adjusting behavior in accordance with what the present predicts about future events. Thus, the ability to learn about the relevance of time may be a species-general ability that is a fundamental building block of complex, flexible behavior. Understanding how animals learn the temporal structure of an environment, and the relation between time and other events, is therefore critical to understanding how flexible, future-oriented behavior develops. Models of time perception (e.g.,) are particularly useful in this endeavor; they make specific, testable predictions about how temporal learning occurs. Many of these models are based on ‘simple’ associative-learning mechanisms, but offer surprising flexibility in the range of behavior they can explain. One particularly promising model, Learning to Time (LeT), asserts that animals learn to adjust their behavior in accordance with the passage of time when a behavioral state is repeatedly active at the point that a particular behavior produces a reinforcer, and other behavioral states are active at times when the same behavior does not produce a reinforcer. The more a state is associated with a reinforcer, the more it is likely to occasion the associated operant behavior. Associations between states and behaviors are strengthened by reinforcement *and* weakened by the absence of reinforcement, and hence temporal learning is context dependent and flexible. LeT provides a more accurate account of behavior than other models of timing in a number of lab-based timing procedures. One such procedure, the *temporal bisection task*, has been used extensively across different species including pigeons, rats, mice, and humans. In the bisection task, animals learn during training to choose between two visually or spatially distinct *comparison stimuli* on the basis of which *sample duration* was presented immediately beforehand. Exposing animals to two types of trials during training (a *double- bisection task*) allows a particularly nuanced assessment of how animals learn about the relation between different durations and outcomes. In Type 1 training trials, choice of comparison stimulus A is correct following presentation of an *x*-s sample stimulus, and choice of comparison stimulus B is correct following presentation of a longer *y*-s sample stimulus. In Type 2 training trials, choice of comparison stimulus C is correct following presentation of a *y*-s sample stimulus, and choice of comparison stimulus D is correct following presentation of a longer *z*-s sample stimulus. Following training, some element of the task may be altered to assess how learning during training is applied to navigate situations that differ to some degree from those in the animal’s training history–for example, novel sample-stimulus durations (Sample Test trials;), or novel combinations of comparison stimuli. Thus, the temporal- bisection task allows assessment of how learning acquired through extended training is used to solve familiar tasks and novel problems. The double-bisection task is useful not only because it permits a test of different models of timing, but also because it lends itself to measures of behavior (‘signatures’) which are comparable across different species. The *psychometric function* plots the proportion of trials in which one of the two available comparison stimuli was chosen, as a function of the duration of the sample stimulus in that trial. Psychometric functions typically show a sigmoidal shift in preference for the stimulus associated with longer durations as the sample duration increases. The steeper the psychometric function, the more pronounced the change in choice as the sample duration increases, and hence the stronger control by the sample duration. A logistic function may be used to give a quantitative estimate of the slope and other characteristics of the psychometric function. A second measure, the *point of subjective equality* (PSE), reflects the sample duration that occasions equal choices of the two comparison stimuli. The closer the PSE falls to one of the two sample durations from training (i.e., the shorter or longer one), the less the learning from training with that sample duration generalizes to novel sample durations. In the double-bisection task, signatures of performance are similar across different modalities including duration and length, and across species including pigeons, mice, and humans. Extensive work with pigeons has revealed two main patterns. First, functions from trials containing Type-1 and -2 stimuli do not superimpose, but there is individual variation in *how* the slopes of the functions differ. Most individuals show steeper functions in trials containing Type-2 stimuli (associated with longer times overall) than in trials containing Type-1 stimuli, but a minority show the reverse pattern or no difference. These features suggest pigeons learn the *relation* between the two sample-comparison pairs in each trial type, rather than sample-comparison pairs in isolation. Second, the PSE is typically slightly closer to the shorter sample than the longer one, falling at the geometric mean of the training durations for some subjects (as predicted by models of timing; e.g.,), and at even shorter durations for others. Thus, learning from trials with longer samples tends to generalize more widely–perhaps because of increasing error in perception of longer durations. Combining comparison stimuli in a novel way by including one stimulus from Type-1 training trials and another from Type-2 training trials (Comparison and Sample + Comparison Test trials;) allows assessment of how past information from the sample duration is used in the face of an unexpected choice. In such situations, pigeons tend to choose whichever comparison stimulus has previously been ‘correct’ following a similar sample duration over a comparison stimulus never before associated with a similar sample duration, or one that has previously been incorrect following similar sample durations. When faced with a choice between a comparison stimulus that has never been associated with a similar sample duration, and one that has been explicitly incorrect follow a similar sample duration, pigeons tend to avoid the comparison they have learned as being explicitly incorrect. That is, the duration of the sample stimulus functions as a cue that signals both correct *and* incorrect behaviors, and hence learning from training can be applied flexibly to novel situations in which comparison stimuli do not contain an explicitly correct option. Such patterns are generally consistent with LeT’s predictions about what is learned during a temporal discrimination task, and are certainly better explained by LeT than by other mainstream models of temporal learning (e.g., see, for discussion). To provide a clearer picture of the ability of LeT to account for the processes that give rise to temporal learning across species, we examined the behavior of domestic chickens (*Gallus gallus domesticus*) following training on the temporal bisection task. While research has established (perhaps unsurprisingly) that chickens *can* discriminate the passage of time it has not explored *how* they learn to do so. We therefore asked how chickens performed on the double- bisection task–a replication of Machado and Keen’s experiment using a different species of subject. Understanding how chickens–often considered the ‘bird- brains’ of the avian world–learn in the double-bisection task provides a further test of LeT and its strengths and weaknesses. We assessed each individual chicken’s pattern of behavior under a range of novel situations, using a logistic function to provide a quantitative estimate of various aspects of behavior directly comparable to the behavior of pigeons on the same task. A focus on the individual as its own control overcomes many of the constraints of psychological research that have led to replication failures, particularly those which create difficulties for comparative cognition (see for discussion). Our findings add to the literature by demonstrating fundamental similarities in the way different avian species use past experience to solve familiar and novel tasks, and in the strengths and shortcomings of LeT in describing such learning across different species. # Materials and methods ## Subjects Three domestic Barnevelder hens numbered 10.1, 10.3 and 10.6 and three Crossbreed Bantam roosters numbered 10.2, 10.4 and 10.5 (all *Gallus gallus domesticus*) participated in the experiment. The hens all had the same prior experience pecking response keys for food on simple ratio schedules of reinforcement. The roosters had no prior experience pecking response keys for food. All birds were approximately two years of age at the start of the experiment. They were housed individually in wire cages that were approximately 500-mm long x 420-mm high x 500-mm wide in a ventilated room lit on a 12-hr light and 12-hr dark cycle. All birds were maintained at 80% ± 5% of their free- feeding body weight to ensure they were motivated to respond for the wheat reinforcers used in the experiment, maintained by post-session feeding of commercial pellets. All birds had free access to water in their cages, with grit and vitamin supplements provided weekly. The research was approved by the University of Waikato ethics committee (Protocol 894). ## Apparatus An experimental chamber, which measured 615-mm long x 450-mm wide x 580-mm high, was used. The interior of the chamber was white with three keys and a food magazine mounted on the right-hand side of the chamber. The food magazine was located behind an opening (115-mm high x 70-mm wide) and centred 105-mm above the floor and when operated was lit by a 1-W light bulb. Each response key was a frosted transparent Perspex key measuring 30-mm in diameter, positioned 390-mm from the floor and 85-mm apart in a horizontal position and could be lit by either a red, blue, yellow, green or white 28 –V multi-chip LED (light-emitting diode) bulb. Each effective key peck required a force of approximately 0.1 N and produced an audible beep that signalled key activation. When activated, a light above the magazine was illuminated, and the magazine was raised to allow access to wheat. All experimental events were controlled and recorded by a computer running MED-PC IV software. ## Procedure We used a small-N design in which each individual hens experienced all conditions–that is, the individual functioned as its own control. A small-N design allowed us to assess the performance of individuals, and patterns across individuals, permitting a more direct comparison with previous research using different species. Given the observation of two distinct patterns of signature measures (see for example), comparison at the group level would be inappropriate. Further, small-N designs are statistically powerful, and control for many of the factors that cause a failure to replicate. Chickens first learned to choose one of two coloured stimuli according to the duration of a preceding sample stimulus, then were occasionally presented with longer or shorter sample stimuli, and/or with novel combinations of coloured stimuli. Our training and testing procedures followed those used by Machado and Guilhardi as closely as possible in terms of the stimuli and structure of the training and experimental sessions (outlined in detail below). This was to ensure a fair comparison between chicken and pigeon behavior. ## Pretraining Each pretraining session comprised 48 trials in which a sample duration was presented by lighting the center key white for some duration, and then the center key was turned off and two colored comparison stimuli were presented on the side keys. The correct comparison stimulus was always signalled by the duration of the sample stimulus (training;). In Type 1 trials, the two-sample durations were 1-s and 4-s. Pecks to the red key following a 1-s sample, or to the green key following a 4-s sample, resulted in a reinforcer. In Type 2 trials, sample durations were 4-s and 16-s. Pecks to the blue key following a 4-s sample, or to the yellow key following a 16-s sample, produced a reinforcer. In both Type 1 and 2 trials, incorrect comparison choices resulted in the trial being repeated. Trials were separated by a 20-s inter-trial interval (ITI). All birds were first trained on Type 1 trials until all could discriminate between both sample durations with 80% accuracy across repeated trials for ten consecutive sessions. Once this was achieved, all birds were trained in Type 2 trials with the same performance criteria as for Type 1 trials. Following mastery of both Type 1 and 2 trials, all birds received Type 1 and 2 trials across alternate sessions for a period of 8 to 38 days depending on individual accuracy. Finally, both Type 1 and 2 trials were presented in the same session. After all, birds had completed the training and achieved 80% accuracy across ten consecutive days (which took 10 to 20 sessions), the error-correction procedure was removed, so that incorrect color choices resulted in the beginning of the ITI. Following approximately ten sessions of this pretraining, chickens began training. ## Training Training sessions comprised 48 trials that ended in a reinforcer for a correct response, and 24 extinction trials where the correct response ended the trial and initiated the ITI without access to a reinforcer. Type 1 and 2 trials (top panel) occurred in random order within a session. Extinction trials were introduced during training to ensure the absence of reinforcers in testing trials (which also ended without a reinforcer) would not be likely to signal a change in contingencies specific to the test-trial stimuli. Training continued for ten sessions before each type of Test began. In between each set of the Test sessions, the chickens were returned to training for five sessions. In the training sessions preceding the Stimulus-response-generalization tests, and the number of extinction trials increased from 24 to 32 because of the number of test trials required to display each different combination of stimuli in testing. ## Sample tests In Sample test trials, novel sample stimuli of intermediate duration were introduced, and comparison stimulus color combinations were the same as in training (Sample Test;). The sample-stimulus durations were logarithmically spaced: For Type 1 trials, sample stimuli were 1.41 s, 2 s, and 2.83 s long, and for Type 2 trials sample stimuli were 5.66 s, 8 s, and 11.31 s long. The middle duration of the test durations corresponded to the geometric mean of the training stimuli. Responses in test trials were never reinforced. Each test sample stimulus duration occurred four times in a session, and sample stimuli were presented on both left/right key color combinations. Thus, there were 24 test trials in each session. ## Comparison test Comparison test trials used the same sample-stimulus durations as in training, but presented novel combinations of comparison-stimulus colors (Comparison Test; ). These new combinations were Red-Blue, Red-Yellow, Green-Bue, and Green-Yellow (i.e., one comparison stimulus from a Type 1 trial and another from a Type 2 trial). Each of these unique combinations of stimuli occurred twice per session. Each session comprised 56 regular trials and 24 test trials. Novel Comparison testing ran for 20 sessions. ## Sample + Comparison test In Sample + Comparison tests, we presented novel sample-stimulus durations of either. 2- or 8-s, as well as the novel combinations of comparison stimuli used in Novel-Comparison trials (Sample + Comparison Test;). The 8 test trials were presented four times within each session, twice for each left-key/right-key color combination, for 16 consecutive sessions. Due to an intermittent key-light problem caused by a loose wire, the Sample + Comparison test was repeated following an additional ten sessions of baseline. # Results ## Sample tests shows the probability of choosing the comparison stimulus associated with a shorter sample duration (hereafter, the probability of *choosing short*) in Sample test trials, as a function of the duration of the test sample stimulus relative the longer training-sample duration. Filled data points denote data from trials with Type 1 comparison stimuli, and unfilled from Type 2. The solid lines are the best fits of a four-parameter logistic function that provides a quantitative description of how choice changed in response to variations in the sample-stimulus duration (see for a discussion about the utility of logistic functions for describing data): $$P\left( {short} \middle| t \right)\left. = \left( y \right._{0} - a \right)/\left\lbrack 1 + exp\frac{\left( {T - \mu} \right)}{\sigma} \right\rbrack$$ provided an excellent description of choice, accounting for between 98 and 100% of variance in the data. shows the variance accounted for by for each individual, as well as the mean and standard deviation parameters from the fits. In, the probability of choosing short decreased as the sample duration increased. Visual inspection of revealed a clear failure of superposition for all but Chicken 10.1. Fits of to data revealed smaller standard deviations (i.e., steeper functions) in Type 1 trials relative to Type 2 for four of six chickens; the other two showed the opposite pattern. A one-tailed paired-samples t-test did not reveal significant differences in the standard deviation t (5) =.175, *p* =.868 or mean t (5) = 1.308, p =.248, and a Bayesian t test revealed anecdotal evidence for an absence of difference between the standard deviations (BF₁₀ =.378) and between the means (BF₁₀ =.689). The average of the mean from fits of the logistic function fell below the geometric mean of the training stimulus values for Type 1 (0.39, 95% CI \[0.29,0.49\]) and Type 2 trials (0.26, 95% CI \[0.15,0.37\]). ## Sample and Sample + Comparison tests shows the proportion of responses to one of the two comparison stimuli in Sample and Sample + Comparison trials, for each novel combination of comparison stimuli as a function of sample-stimulus duration, for each chicken. Colored ticks and crosses below the x axis denote the sample duration at which a comparison color was correct and incorrect during training. From, patterns of responding in a Test trial were strikingly similar across individual chickens. ## Green + Blue In Green Blue test trials, choice for Green was most extreme following the longest sample duration, which during training had never been associated with Green but had signaled Blue was explicitly incorrect. Choice for Blue was most extreme following the shortest sample duration, which during training had never been associated with Blue but had signaled Green was incorrect. For all chickens except 10.1, choice was approximately indifferent (i.e., Blue and Green were chosen approximately equally) at the 4-s sample duration which had signaled Green and Blue were correct during training, and at the 2-s and 8-s durations which had never been associated with either color during Training. ## Red + Yellow In Red Yellow trials, choice shifted from favoring Red to Yellow as the sample duration increased, with maximal choice for each color following the sample stimulus that had signaled the color as being correct during training (1 s for Red, 16 s for Yellow). For Chickens 10.2 to 10.4 and 10.6, choice shifted progressively from Red to Yellow over the intermediate sample durations which had either never been associated with either color (2 s, 8 s), or had signaled both Red and Yellow as being incorrect during training (4 s). For Chickens 10.1 and 10.5, choice following a 4-s sample stimulus favored Red to a greater extent than it had following the 2-s sample duration. ## Red + Blue In Red Blue test trials, choice functions were v-shaped, with the strongest choice for Blue occurring following the 4-s duration associated with Blue being correct and Red being incorrect in training. Choice shifted progressively toward Red as the sample duration became more different from 4 s–either longer or shorter. For three chickens (10.1, 10.3, and 10.4), choice functions were asymmetrical, with choice for Blue being more extreme at the 1-s sample duration that had never been associated with Blue during training, relative to the 16-s sample duration that had explicitly signaled Blue was incorrect. ## Green + Yellow In Green Yellow test trials, choice functions resembled an asymmetrical inverted v. The most extreme choice for Green occurred following a 4-s sample stimulus, which had signaled Green as being correct, and Yellow as incorrect, in training. The most extreme choice for Yellow following a 16-s sample duration, which had signaled Yellow as being correct, but was never associated with Green, in training. Following a 1-s sample stimulus, which had signaled Green was incorrect during training and had never been associated with Yellow, and a 2-s sample stimulus, which had never explicitly been associated with either color during training, choice was intermediate. Four chickens showed less extreme choice for Green following a 1-s sample than a 2-s sample; the other two showed the opposite pattern. In general, then, shows stimuli that were in training associated with reinforcers after a similar or identical sample duration were more likely to be chosen over those never before associated with the same duration, and those that had explicitly been incorrect at a duration (i.e., were associated with no food). Generally, stimuli never before associated with a particular duration were more likely to be chosen than a stimulus that had been incorrect but associated with a duration during training. # Discussion We examined choices made by chickens on a double-bisection task following novel sample-stimulus durations (Sample Tests), and in the presence of novel combinations of comparison stimuli (Comparison Tests). This is the first study to assess chickens’ performance on the temporal-bisection task. Measures of chicken behavior in all types of Test showed patterns similar to those obtained from pigeons performing the same task. Specifically, we found a failure of superposition of Type 1 and 2 psychometric functions, variation in the way functions failed to superimpose, and flexible application of past experience to solve novel problems in a manner consistent with LeT’s predictions. The similarity between chicken and pigeon behavior on the double-bisection task highlights generality in the way avian species with different evolutionary histories use past experience to navigate novel situations, and also underscores generality in both the strengths and weaknesses of LeT’s approach to explaining learning. What does our chickens’ behavior tell us about how learning in a time-based task occurs? The patterns of behavior that we and others have observed in the double- bisection task are generally consistent with predictions made by the Learning to Time (LeT;) model of temporal learning. Just as pigeons’ psychometric functions on the double-bisection task fail to superimpose, so too did our chickens’ psychometric functions from Type 1 and 2 trials. A failure of superposition reflects differences in the accuracy of judgements about the sample duration in Type 2 and 1 trials, and hence violations of scalar timing. Similarly systematic violations of scalar timing have been reported in other studies. LeT predicts a failure of superposition, but in a specific direction; because the overall reinforcer rate is constant, functions for Type 2 trials should be steeper than those in Type 1 trials. Differences in the standard deviation (an estimate of slope) of each chicken’s Type-1 and -2 psychometric functions were consistent with LeT’s prediction for only two of our six chickens. For the remaining four, Type-1 psychometric functions were steeper than those in Type 2 trials. In pigeons, the direction of the difference is similarly variable across individuals. Our findings demonstrate that this variation is not a specific quirk of pigeon subjects, but is instead a more general outcome of temporal learning. The inability of LeT to account for a bi-directional difference in psychometric functions in the double- bisection task thus highlights a general shortcoming of the model (at least in its present form). Temporal discrimination is typically more accurate and precise for shorter durations, although exceptions–including double-bisection-task performance–exist. These exceptions suggest temporal discrimination depends on more than the duration itself. Longer durations may create more opportunities to engage in temporally regular sequences of behavior which facilitate timing, but such sequences are not a requirement, and their occurrence and nature will thus depend on the individual. Certainly, temporal discrimination is improved in environments that facilitate engagement in other behaviors during the interval to be timed (termed *mediating* behaviors); performance on timing tasks worsens when animals are restrained, when space for movement is restricted, and when usual patterns of behavior are interrupted, and improves when mediating behaviors *must* be performed during the relevant duration, as well as when other behaviors are simply *able to* be performed. Further, individuals who exhibit more behavior during an interval also perform more accurately, and obtain higher numbers of reinforcers. When mediating behaviors are emitted at a different time from their usual occurrence, they tend to occasion incorrect timing responses. So important are these mediating behaviors that humans report–albeit mistakenly–the entire *sequence* is necessary to produce reinforcers. The impact of mediating behaviors on performance in time-based tasks suggests memory for one’s own behavior–episodic-like memory–is a critical component of learning the temporal structure of an environment, perhaps because such sequences act as an additional, enduring cue signalling the most appropriate behavior (see also). Although orderly, sequences of mediating behavior contain considerable variability, in terms of the time taken for each sequence to unfold, the evolution of orderly sequences with experience (e.g.,; see also topographical drift;) and to some extent in the nature of the behaviors that make up each individual’s sequence. Such variation may cause imperfect temporal control, giving rise to systematic differences between the ability of different individuals to perform the same time-based task. Indeed, noted that their pigeons with steeper Type-2-trial functions displayed a different pattern of behaviors before the comparison phase than did those with steeper Type-1 trials. It is reasonable to assume that had our chickens’ behavior been observed before the presentation of the comparison stimuli, we would have seen the same sort of differences in patterns of mediating behaviors according to whether a chicken’s performance different in tests with Type-1 and 2 stimuli. Given the apparent importance of mediating behavior in navigating the temporal structure of the world, it is essential to ask how these mediating behaviors develop, and why some individuals fill time with mediating behavior more efficiently than others. LeT’s conceptual approach to understanding the temporal organisation of behavior might capture the role of mediating behaviors in timing were such mediating behaviors able to be measured with the same rigor as are timing behaviors (e.g., see for discussion). As with psychometric functions from Sample Tests, LeT also makes predictions about functions in Sample + Comparison Tests are generally consistent with but not identical to actual individual psychometric functions. LeT asserts that learning in timing tasks is context-dependent because the associations between states and behaviors are both strengthened by reinforcement *and* weakened by the absence of reinforcement. We tended to observe a lack of systematic variation across intermediate sample-stimulus durations in Sample + Comparison tests when LeT would predict a systematic variation (Green+Blue and Red+Yellow trials), and systematic variation when LeT would not predict it (Green+Yellow trials). noted similar patterns in pigeons’ behavior in Red+Blue and Green+Yellow trials, although their pigeons’ response patterns in Green+Blue and Red+Yellow trials tended to conform more closely to LeT’s predictions than did our chickens’. Nevertheless, the similarity across species and studies in choice *not* predicted by LeT suggests that LeT may capture only some of the various processes underlying timing behavior. One possible shortcoming of LeT is that it attributes all errors to temporal discrimination errors, even though discrimination of the relation the passage of time, behavior, and outcomes requires accurate detection of behaviors and outcomes, as well as the passage of time. Indeed, the double-bisection task cannot be learned without detecting the relation between time and responses to stimuli of a particular color. Time is but one element of any environment; error in discriminating any non-temporal aspect of the environment would also cause weaker control. Such errors cannot be predicted or understood by a model of behavior whose sole focus is time perception. Where non-temporal discrimination errors occur, a model dealing only with temporal control will conflate non- temporal discrimination errors with temporal ones. This conflation will inadvertently creating timing errors. The stream of environmental inputs an animal faces is distributed across both temporal and non-temporal dimensions. Navigating the world requires discrimination of what and where, as well as when–in this sense, even simple operant learning has episodic-like qualities that are not captured by mainstream theories or models of learning. In conclusion, our findings show strong consistency in patterns of behavior in the face of novel situations across individual chickens. These patterns are similar to those observed in other studies with pigeons–that is, findings appear similar across avian species with strikingly different evolutionary histories. These results add to a growing body of data demonstrating similarities in the way humans and non-human animals learn about the relation between the passage of time and other events. Taken together, findings that animals use past experience flexibly to navigate new situations suggest that behavior does not merely fill time; behaviors take place at times when they are the best possible option either because past experience suggests the behavior is likely to produce a valuable outcome, or because it suggests other behaviors are *unlikely* to produce a valuable outcome. Organisms learn what events (stimuli, responses) will produce a particular consequence, and what will not. Such learning occurs in the context of time, and space, and other relevant dimensions, and as such as episodic-like in nature. Learning about time and how it relates to other events bears many similarities to learning about other dimensions (e.g., number, space; see for example). Understanding how the organisation of behavior across time, space, and other relevant dimensions emerges, and why time and space are filled more efficiently in some environments and by some individuals, is key to understanding how simple learning underpins apparently complex behaviors. [^1]: The authors have declared that no competing interests exist.

# Introduction EspB is one of the virulence factors of enterohemorrhagic *E. coli* (EHEC) that is known to be dependent on a type III secretion system (T3SS). EspB is the multifunctional effector with 312 amino acid residues that can bind to various different proteins such as α-catenin, α₁-antitrypsin and myosin from the host cell, and EspA and EspD from the bacterium itself. These interactions of EspB to a range of target proteins are associated with different events in bacterial infection including pore-formation, actin reorganization, and inhibition of phagocytosis. The binding site for α-catenin within EspB is known to be located at the N-terminal region between residues 1 to 98. However, the precise mechanism of α-catenin recognition by EspB is unclear. Previously, we found that binding of EspB to α-catenin induces dissociation of α-catenin from an E-cadherin/β-catenin/α-catenin triple complex formed at the adherence junction on host cell membrane and enhances the intrinsic ability of α-catenin to promote bundling of actin filaments under *in vitro* conditions. Interestingly, EspB in aqueous solution has the characteristics of a partially folded protein that consists of α-helical secondary structures but only small amount of tertiary contacts if any. This conformational property is similar to that of the molten globule state – which is the compact partially folded state usually accumulated at the early stage of folding kinetics of globular proteins. However, unlike the general molten globule states, EspB does not show an increase of fluorescence intensity of 8-anilinonaphthalene-1-sulfonate which usually binds to surface exposed hydrophobic clusters present in the molten globule intermediates and increase the fluorescence intensity around 480 nm. Moreover, sequence-based disorder predictions as well as a multiplicity of experimental data further suggested that many T3SS-dependent pathogens assume entirely or partially unfolded structures when dissociated from their binding targets. Interestingly, it has been clarified that numerous pathogens from infectious viruses also known to possess a characteristic of intrinsically disordered proteins. Thus, intrinsic disorder could be a generic property for certain classes of virulence factors from infectious bacteria and viruses. Recently, solution NMR has successfully elucidated the structural details of highly disordered proteins. However, such an approach is hardly applicable to EspB due to the line broadening of NMR signals caused by the slow chemical exchange in partially folded regions. Application of another modern high resolution technique such as X-ray crystallography is also impossible because of the difficulty in crystallization of EspB containing significantly flexible regions in the unbound state to its target proteins. We therefore employed a hybrid approach combining low resolution structural and thermodynamic data of intact form and the fragments of EspB obtained by circular dichroism (CD) and small angle X-ray scattering (SAXS) in conjunction with various disorder-order prediction algorithms based on amino acid sequences. The functional properties of the EspB fragments were also tested by fluorescence anisotropy changes upon binding to α-catenin. Our results provide a significant insight into the structural basis of the recognition of α-catenin by EspB. # Materials and Methods ## Materials All chemicals used in this study were of analytical grade from Sigma-Aldrich Co. (St. Louis, MO), Nacalai Tesque, Inc. (Kyoto, Japan) or Wako Pure Chemical Industries, Ltd (Osaka, Japan). Synthetic peptides of EspB fragments with an N-terminally attached FITC were purchased from PH Japan (Hiroshima, Japan). EspB was prepared using a combination of Ni²⁺-Sepharose and Resource Q chromatography (GE Healthcare, Milwaukee, WI) as described previously. Recombinant EspB were engineered for expression in *E. coli* BL21(DE3) as N-terminally His₆-tagged fusion proteins using expression vectors derived from pET28a (Merck Chemicals Ltd., Nottingham, UK). The purified proteins were analyzed without removal of the N-terminal His₆-tag sequences. α-Catenin635–906 with an N-terminal His-tag was purified using Chelating Sepharose Fast Flow medium (GE Healthcare) charged with nickel chloride in 20 mM sodium phosphate at pH 7.4. After equilibrating the column in 20 mM sodium phosphate at pH 7.4 the protein solution was applied and unbound material was washed away using the same buffer. Bound protein was subsequently eluted using a 0–0.5 M imidazole gradient. The eluted fractions were dialyzed against 20 mM sodium acetate at pH 4.0 and applied onto a SP-Sepharose column (GE Healthcare) equilibrated in 20 mM sodium acetate at pH 4.0, and then eluted using a 0–2.0 M sodium chloride gradient. Fractions containing α-catenin635–906 were pooled and further purified by gel filtration chromatography using a Sephacryl S-200 column (GE Healthcare) equilibrated in 20 mM Tris-HCl (pH 8.0), 0.1 mM EDTA. The protein concentrations of full length EspB and EspB1–176 were determined from UV absorption at 280 nm using an extinction coefficient of 4470 and 2980 M⁻¹ cm⁻¹, respectively, which were calculated using the ProtParam server (<http://www.expasy.org/tools/protparam.html>) based on their amino acid sequences. The extinction coefficient for EspB177–312 calculated using the ProtParam server (1490 M⁻¹ cm⁻¹) was too small to obtain reliable estimations of protein concentration in dilute solution. We therefore employed Bradford’s method to determine the concentration of EspB177–312 using standard curve obtained by full length EspB. The absorption spectra were obtained using a JASCO UV spectrophotometer, V-550 (Jasco Co., Tokyo, Japan). ## CD Spectra Far-UV CD spectra were monitored using a J-720 spectropolarimeter (Jasco Co.) equipped with a Peltier type thermo-controllable cell holder. A quartz cuvette with a 1 cm pathlength was used for the thermal unfolding experiments and a cell with a 0.1 cm pathlength was used for the analysis of CD spectra of EspB fragments as well as full length EspB. Unless stated otherwise, all data are expressed as mean residue ellipticity. Secondary structure contents of the protein and peptides were analyzed by CDpro using the reference set of SP29. The thermal unfolding transition was monitored by measuring the ellipticity at 222 nm. The temperature was increased from 10 to 90°C at a heating rate of 1°C min⁻¹. The samples cooled to 20°C after the thermal unfolding experiments recovered to the essentially same CD spectra obtained at 20°C before heat treatments, suggesting the unfolding reactions were reversible. The thermal unfolding curves were analyzed by modified method as previously demonstrated assuming the two-state transition between folded and unfolded states based on the following equation,where *T* and *T*₀ are the given and reference temperatures in Kelvin, Δ*G* and Δ*G*₀ are the free energy changes upon unfolding of folded state at give temperature *T* and *T*₀, respectively. Δ*H*₀ and Δ*C*_p is the enthalpy and the heat capacity change upon unfolding reaction, respectively. The ellipticities at 222 nm (\[*θ*\]₂₂₂) are expressed as, where \[*θ*\]_F and \[*θ*\]_U are the baselines of folded and unfolded species that are linearly dependent on temperature, and *f*_U is the fraction of unfolded state. Thus, Δ*G* can be expressed as, where *R* is the gas constant. In practice, *T*₀ was set to 293.15 K ( = 20°C), and all the unfolding curves observed in the presence of different amounts of guanidium hydrochloride (GdnHCl) were simultaneously analyzed by the global-fitting method according to the above equations using the same baselines of \[*θ*\]_F and \[*θ*\]_U, Δ*C*_p and Δ*H*₀. IgorPro ver 6.3 (Wavemetrics, Inc., Portland, Oregon)was used for this fitting analysis. The Δ*G*₀ values were obtained for the individual conditions at different concentrations of GdnHCl. In the absence of GdnHCl, \[*θ*\]₂₂₂ was linearly decreased as in crease in *T*, and this linear dependence of \[*θ*\]₂₂₂ was considered as the baseline, \[*θ*\]_F. This situation prevented us from obtaining the thermodynamic parameters in the absence of GdnHCl based on the above analysis using the data from the thermal unfolding experiments. We therefore estimated Δ*G* in the absence of GdnHCl (Δ*G*_water) from the linear dependence of Δ*G*₀ as a function of the concentration of GdnHCl (\[GdnHCl\]), i.e., which is generally used for the analysis of denaturant-induced unfolding reaction. The *m* value is the measure of cooperativity of denaturant-induced unfolding reaction which is considered to be correlated with the change in solvent accessible surface area upon unfolding reaction. ## SAXS SAXS measurements were carried out at a BL-10C synchrotron beamline in the Photon Factory, Tsukuba, Japan. Scattering intensity was monitored by R-axis VII (Rigaku, Tokyo, Japan) and circularly averaged intensity was used for further analysis. A cell with a 1 mm pathlength was used and the protein solution was prepared by dialysis against 10 mM 3-morpholinopropane-1-sulfonate (MOPS) at pH 7.0. The temperature in the cell was maintained at 20°C by circulating temperature-controlled water. Scattering intensities obtained at various concentrations of protein were subtracted from the intensity of the buffer solution without protein using IgorPro ver 6.1 (Wavemetrics, Inc., Portland, Oregon). These values were used to generate Guinier and Kratky plots. *P*_r functions were calculated by GNOM using the scattering intensity after reduction of noise by singular value decomposition *via* IgorPro. The *P*_r function of phosphotriesterase was calculated using a theoretical scattering curve calculated by CRYSOL based on the atomic coordinates of 1EYW solved by X-ray crystallography. ## Fluorescence Anisotropy Fluorescence spectra were observed using a FP-6500 fluorimeter (Jasco Co., Japan). To observe fluorescence anisotropy, HN32 linear polarizer films (5 cm×5 cm; Polaroid Co., Waltham, MA) were inserted in both excitation and emission lightpaths. The orientations of the polarizers were manually changed to obtain the emission components of *I*_vv, *I*_vh, *I*_hv and *I*_hh. *I*_vv and *I*_vh were the vertical and horizontal emission components obtained for the sample excited with vertically polarized light, and *I*_hv and *I*_hh were the vertical and horizontal emission components obtained for the sample excited with horizontally polarized light. Anisotropy value, *r* is then defined as,where *G* is defined as *I*_hv/*I*_hh. The data were processed by IgorPro ver 6.3 (Wavemetrics, Inc.) to calculate the anisotropy values with different samples and to estimate the dissociation constants, *K*_D by nonlinear curve-fitting. # Results ## Cooperativity of the Thermal Unfolding Transition of EspB Thermal unfolding transition of EspB at pH 7.0 was monitored by the ellipticity at 222 nm. Usually, globular proteins with rigid secondary and tertiary structures stabilized by hydrophobic interactions show highly cooperative thermal unfolding transitions represented by a sigmoidal curve that are consistent with a two-state model from the well-ordered and rigid native state to the highly unfolded state. However, EspB displays a noncooperative transition from a α-helical structure to a unfolded state in which the absolute ellipticity at 222 nm decreased almost linearly with increasing temperature ( and circles in ). The lack of cooperativity in the thermal unfolding reaction suggests the lack of highly ordered tertiary contacts stabilized by hydrophobic interactions in EspB. A similar behavior was previously observed for the thermal unfolding of a partially folded or “molten globule” intermediate state of human α-lactalbumin. Thus, the result is consistent with our previous conclusion suggesting that EspB is a natively partially folded protein. Interestingly, further analysis on the response of EspB against the temperature change by far-UV CD revealed the relatively cooperative cold and heat denaturations in the presence of various amount of GdnHCl. This result indicates the positive heat capacity change for the unfolding reaction (Δ*C*_p) of EspB. We analyzed the transition curves assuming two-state transition by global-fitting procedure according to the method described in Materials and Methods. The estimated parameters of unfolding reactions at various GdnHCl concentrations are summarized in. In this analysis, the estimated Δ*C*_p was 0.34±0.01 kJ mol⁻¹ K⁻¹ or 81.8±1.0 cal mol⁻¹ K⁻¹. According to the BPpred server (<http://www-clarke.ch.cam.ac.uk/BPPred.php>), Δ*C*_p of a globular protein with 332 amino acid residues, i.e. equivalent length of EspB with hexahistidin-tag, should be 5.15 kcal mol⁻¹ K⁻¹. The experimentally determined Δ*C*_p by thermal unfolding experiment of EspB is therefore extremely smaller than that is expected for the globular protein with the same length of EspB. These results from the CD spectroscopy suggest that either the molecule is in the relatively ordered molten globule-like structure or there is a small relatively structured core and other regions with less-ordered regions in EspB. However, the fact that the EspB does not increase the fluorescence intensity of ANS supports the latter possibility. ## SAXS Indicates the Extended Conformation of EspB To gain further insight into the conformational properties of EspB, we performed SAXS experiments for this protein at pH 7.0, 20°C. SAXS provides information about the molecular shape and dimension of a macromolecule under a given solution condition. This particular technique is extremely useful when the protein assumes less-ordered structures such as partially folded or highly unfolded states having conformational heterogeneity or ensemble for which more detailed analysis using X-ray crystallography or solution NMR are nearly impossible. The pair distance distribution (*P*_r) function calculated from the scattering intensity of EspB using GNOM program showed an extremely broad distribution with a maximum distance of 250 Å and the optimum position of *P*_r at 65 Å. For comparison, we calculated *P*_r function of phosphotriesterase (364 amino acids) as a typical model of a globular protein whose molecular weight is similar to that of EspB (332 amino acids including hexahistidine-tag) using its atomic coordinates solved by X-ray crystallography. Unlike the *P*_r function of EspB, the *P*_r function of phosphotriesterase indicates an optimum position of *P*_r at 24 Å and the maximum distance is ∼60 Å. We also determined the radius of gyration (*R*_g) of EspB according to a Guinier plot (ln(*I*(*Q*) vs. *Q*²)) as shown in. The slope of this plot at low *Q* regions corresponds to *R*_g²/3 although the value also depends on the protein concentration. To correct the effect of protein concentration dependency, the *R*_g² values were plotted as a function of protein concentration and the reduced *R*_g value was given by linear extrapolation of this plot to zero protein concentration. This analysis provided the *R*_g value of EspB of 67.8±0.1 Å, which is slightly larger than the expected *R*_g value of urea unfolded state (62.0 Å) according to the empirical equation, where *N* is the number of amino acids, i.e. 332 in the case of our recombinant EspB with N-terminal hexahistidine-tag. Scattering intensity obtained by linear extrapolation of the Guinier plot to zero angle, *I*(0) is linearly correlated with the molecular weights of the proteins. Molecular weight of EspB was estimated to be 32,400 by comparison of *I*(0) of lysozyme with a molecular weight of 14,300. This value is close to the calculated molecular weight of monomeric hexahistidine-tagged EspB (34,800) used in this study. Thus, EspB assumes a highly extended monomeric structure comparable to the ideal random coil. Nonetheless, EspB contains folded regions with α-helical secondary structures according to the far-UV CD spectra. ## α-Helical Structures are Formed at the N-terminal Regions Whereas the C-terminal Half is Highly Unfolded Although the data from SAXS suggested the highly extended structure of EspB, the protein contains a significant amount of α-helical structures. The most α-helical structures may be partially folded according to the noncooperative behavior of thermal unfolding reaction Thus, the overall structure of EspB should consist of both relatively ordered and less well-organized α-helical structures. To establish the regions of EspB that adopt α-helical structures, we performed combinatorial analysis using sequence-based disorder-order prediction algorithms and CD spectra of peptide fragments derived from the EspB sequence. For disorder probability prediction, we used three different algorithms including PONDR – (<http://www.pondr.com/>), IUPred (<http://iupred.enzim.hu/>) and POODLE-L (<http://mbs.cbrc.jp/poodle/poodle.html>). Intriguingly, all three algorithms predicted that the N-terminal half of EspB has a low disorder probability, i.e. relatively high propensity to form ordered structures. By contrast, the C-terminal half of this protein was predicted to have a high propensity to form a disordered structure. To confirm the results of these predictions, we prepared two fragments of EspB corresponding to the amino acid sequences from M1 to T176 (EspB1–176) and from T177 to G312 (EspB177–312) using an *E. coli* expression system and analyzed their ability to form secondary structures by far-UV CD. As expected, the far-UV CD spectrum of EspB1–176 showed the minima around 210 and 222 nm which are indicative of the presence of a α-helical structure, whereas that the spectrum of EspB177–312 was typical of an unfolded protein with a minimum around 200 nm. Furthermore, the sum of the CD spectra of EspB1–176 and EspB177–312 almost completely reproduced the spectrum of full length EspB (broken line). This observation indicates that the α-helical structures in the N-terminal region are almost independent from the C-terminal unstructured region. Hence, the overall shape of the molecule comprises a partially folded N-terminal region with a significant amount of α-helical secondary structure and a largely unstructured C-terminal region. Importantly, EspB1–176 includes the region that was shown to be the α-catenin binding site (from residue number 1 to 98, see also), suggesting that the α-helical region of this protein should be, at least partly, involved in the binding site for α-catenin. ## Structural Properties of α-Catenin Binding Site of EspB by Protein Dissection Analysis To further refine the structural properties of the regions required for α-catenin binding in EspB, we prepared several short fragments of EspB, each 20 amino acids in length, covering the first 100 residues from the N-terminus. To facilitate the estimation of peptide concentration, all the prepared fragments were modified by FITC through the N-terminal α-amino group. This modification was also useful in the study of fluorescence anisotropy as described later. Far-UV CD spectra of the four fragments, covering the region from G41 to E90 (i.e. EspB41–60, EspBN51–70, EspBN61–80, EspBN71–90), indicated the presence of a significant amount of secondary structures. By contrast, the other fragments gave spectra corresponding to highly unfolded structures. According to the ellipticity around 222 nm, both EspB41–60 and EspB51–70 have significantly higher α-helical propensities than those of EspB61–80 and EspB71–90. Interestingly, the CD spectrum of EspB41–60 indicates intensity minima at 222 nm which is greater than that at 208 nm. The ellipticity ratio at 222 vs 208 nm (\[*θ*\]₂₂₂/\[*θ*\]₂₀₈) larger than 1 is possibly an indication of the formation of either the coiled-coils or other assemblies of helices. Since \[*θ*\]_222/\[*θ*\]₂₀₈ of EspB41–60 is 1.8, the fragment may consist of such rather constraint helices. This and the fact that the SAXS for whole EspB indicates the formation of monomeric structure ruled out the possibility for the formation of oligomeric coiled-coil for EspB41–60. Then, we calculated the secondary structure contents from the CD spectra of the EspB fragments using CDpro software package. This analysis predicted that EspB1–20, EspB11–30, EspB21–40, EspB31–50, EspB81–100 and EspB91–110 possess relatively small amount of secondary structures and nearly 20–30% of β-strand and 10–20% of α-helix contents but with 55–60% of other structures including unordered structures and turns. EspB41–60 and EspB51–70 exhibited the highest α-helical contents of 80–90%. Although the spectra for EspB61–80 and EspB71–90 were apparently different from typical α-helical structure and rather similar to β-rich structure, CDpro suggested that the β-strand contents of these peptides are small (∼15%) and α-helical structures are still dominated (∼44%). These results indicates that the α-helical structures of EspB are probably localized around the regions from G41 to E90. For some peptides, the β-strand contents are negative values but the absolute intensities are very small. These values, therefore, can be considered as zero. We also compared the results of CDpro analysis with the α-helical propensities of fragments predicted from the amino acid sequences using AGADIR algorithm. The result of AGADIR prediction against the sequence of whole EspB was consistent with the results from the CD spectra of EspB fragments, in which the region around G41 to E90 should have high α-helical contents. The results of the AGADIR predictions against individual peptides qualitatively agree well but quantitatively less well to the estimated α-helix contents by CDpro. Next, we aimed to identify the fragments that preferentially bind to the C-terminal vinculin homology domain of α-catenin (α-catenin635–906), which includes the target recognition region of intact EspB, using fluorescence anisotropy of the FITC-labeled EspB fragments. This experiment clarified that EspB41–60, EspB51–70 and EspB61–80 bind to α-catenin635–906 although the affinities of EspB41–60 and EspB51–70 were rather higher than that of EspB61–80. Importantly, these two fragments including EspB41–60 and EspB51–70 which indicated the high affinity to α-catenin635–906 showed the far-UV CD spectra indicative of significant α-helical propensities. However, the α-helical propensity of EspB61–80 was lower than that of EspB41–60 and EspB51–70. Interestingly, the EspB fragments which showed the CD spectra characteristic of unfolded structures all displayed negligible affinity to α-catenin635–907. Thus, the data showed a strong relationship between α-helical propensities of the EspB fragments and their binding abilities to α-catenin. The *K*_D values of EspB41–60 and EspB51–70 were 23±2 and 12±3 µM, respectively. These values are five to ten times higher than that of wild type EspB (2.9±0.3 µM), suggesting that neither of the fragments completely mimic the function of intact EspB. The affinity of EspB61–80 was too weak to obtain the reliable *K*_D value. An extremely high concentration of α-catenin635–906 would be required to determine an accurate *K*_D value for this peptide using our experimental conditions. To further refine the regions of EspB that can reproduce the affinity of intact EspB, we prepared longer fragments corresponding to G41 to E70 (EspB41–70). The far-UV CD spectrum of EspB41–70 was consistent with the formation of α-helical structures. The analysis of CDpro indicated that the α-helical content of EspB41–70 becomes 93.9%, i.e. about 28 amino acid residues may be involved in α-helical structures in EspB41–70. AGADIR program also predicted that this fragment possess high helix propensity. Interestingly, EspB41–70 displayed an unique thermal response similar to that of whole EspB in the presence and absence of GdnHCl. The estimated Δ*C*_p value for EspB41–70 was 0.87±0.02 kJ mol⁻¹ K⁻¹. This value is rather similar to that of whole EspB (0.34±0.01 kJ mol⁻¹ K⁻¹). We also estimated *m* values for EspB and EspB41–70 upon unfolding reaction by GdnHCl and the Δ*G*_water at 20°C in the absence of GdnHCl from the linear relationship between Δ*G*_20°C and GdnHCl concentration, \[GdnHCl\], i.e.. Δ*G*_water of EspB (5.5±0.7 kJ mol⁻¹) was smaller than that of EspB41–70 (13.0±0.2 kJ mol⁻¹). Thus, the presence of additional regions at the N- and C-terminal may rather destabilize the relatively rigid α-helical structure formed around G41 to Q70. The *m* values are considered to be correlated with the change in the accessible surface area upon unfolding of a protein. Interestingly, the *m* value of EspB (1.6±0.2 kJ mol⁻¹ M⁻¹) was closely similar to the value of EspB41–70 (2.1±0.1 kJ mol⁻¹ M⁻¹), suggesting that the structural elements responsible for the rather cooperative thermal response of EspB assume the structure similar to that of EspB41–70. These results suggested that the sequence around G41 to Q70 by itself could form the core of rigid α-helical structure formed in EspB. Moreover, the *K*_D value of EspB41–70 for α-catenin was 2.5±0.3 µM, which is very close to the value determined for intact EspB (2.9±0.3 µM). These results suggest that the sequence around G41 to Q70 of EspB must be the core α-helical region which is also important for the recognition of α-catenin. A close analysis of the sequence from G41 to Q70 clarified the possible formation of an amphipathic α-helix in this region if it assumes extended α-helical structure. However, it is unclear if such an extended amphipathic α-helical structure can be formed under normal solution conditions. We then analyzed if any additional secondary structures are induced for EspB upon binding to α-catenin based on far-UV CD spectra. Interestingly, the far-UV CD spectrum for the solutions containing of α-catenin635–906 and EspB in 1 to 1 ratio coincide very well with the sum of the individual spectra obtained for α-catenin635–906 and EspB. This clearly indicates that no additional secondary structures are formed for these proteins upon formation of α-catenin/EspB complex. # Discussion In general, proteins synthesized in biological systems spontaneously folds into well-ordered structures that are unique to their amino acid sequences, thereby conferring distinctive functions. However, our data suggest that EspB possesses highly unique structural properties. Specifically, the N-terminal region around G41 to E70 of EspB has a significant amount of α-helix contents whilst the C-terminal region is largely unstructured. The analysis of EspB fragments further suggested that the α-helical region is directly involved in the binding with α-catenin. A relatively cooperative thermal response including cold denaturation phenomena observed for EspB and its fragment of EspB41–70 suggests the presence of small but distinctively rigid structures around G41 to Q70 that is stabilized by hydrophobic interactions. However, the contribution of such hydrophobic contacts particularly among the sidechains may be very little since the values of Δ*C*_p were extremely small. The length of EspB41–70 is 30 amino acid long and about 28 amino acid residues will be involved in α-helical structures according to the secondary structure prediction by CDpro. If this estimate is correct, the most parts of this fragment assumes α-helical structure and one of the extreme patterns should be an long extended α-helix. Such an extended α-helix may possess only little number of sidechain-sidechain contacts and the contribution of hydrophobic effects to the stabilization should be less high. It is of question if such an extended α-helix structure by a single chain is stabilized by the mechanism same as that of the globular protein even though the helix is sufficiently stable. Interestingly, secondary structure prediction from the CD spectrum of EspB indicated that about 76 amino acid residues should be involved in α-helical structures within 332 amino acid sequence of EspB. Therefore, although the thermal unfolding behavior and α-catenin binding of EspB can be almost successfully explained by the contribution of the sequences from G41 to Q70, some other parts of the protein should be involved in the formation of additional α-helices. These regions may be rather less organized and behave like the partially folded intermediates of globular proteins, namely the “molten globule” states. Interestingly, SAXS data revealed that, although the molten globule states of globular proteins are as compact as the well-defined native states, EspB assumes an extremely extended conformation with the *R*_g value closely similar to the ideal random coil structures. In this sense, the less organized regions of EspB should be in the conformational states similar to the “premolten globule” states, which are only transiently accumulated at the very early stage of folding of globular proteins. Such an extended partially folded structure of EspB should be dominantly stabilized by the local interactions formed between amino acids close to each other in the amino acid sequence rather than the nonlocal interactions formed by the residues apart from each other in the sequence. Taken together, most parts of EspB assume an extended premolten globule-like structures with significant amount of α-helix and the region of G41 to Q70 forms relatively organized α-helical structures that are directly responsible for α-catenin binding. This result is consistent with the observations for various intrinsically disordered proteins that indicate the regions with high propensity to form secondary structures tend to be functionally important. Highly unstructured intrinsically disordered proteins without having any secondary and tertiary structures often show the formation of well-defined structures upon binding to their target molecules in a manner of either “*induced fit*” or “*conformational selection*”. In the former case, the unstructured protein initially binds to its target molecule as an unfolded structure followed by the formation of well-defined three dimensional structure on the target molecule. Whereas, in the latter case, the particular structure that is only transiently formed among the ensemble of the fluctuating structure of the unstructured protein is selectively recognized by the target molecule. The term, “*coupled folding and binding*” is recently used when the secondary structures (usually α-helix) are formed by induced-fit mechanism. No additional secondary structure was induced neither for EspB nor α-catenin635–906 upon formation of their complex structure according to far-UV CD spectra. This result suggests that the binding of EspB to α-catenin proceeds by the recognition of the preformed secondary structure in EspB by α-catenin. This reaction may be classified into an extreme case of “*conformational selection*” or possibly called “*folding before binding*” in which the formation of α-helix structure that is similar to the helix formed in the protein complex. However, the current data cannot rule out the possibility that additional conformational changes like rearrangements of sidechain orientation or helix curvature without changing total α-helical content may take place in a manner of *induced fit* reaction even in the case of EspB. This type of reaction modes that is initially driven by conformational selection followed by induced fit has been proposed for the interaction between transactivation domain of tumor suppressor p53 and the nuclear coactivator binding domain of cyclic-AMP response element binding protein from the result of molecular dynamic simulation. Such a combined mechanism could be the major mechanisms of molecular recognitions particularly for intrinsically disordered proteins having partially folded regions similar to the molten globule or premolten globule states where preformed partially folded structures are involved in binding process. The structural details on the free and complex forms of EspB as well as the kinetic analysis for α-catenin binding are required to clarify the more realistic mechanism for α-catenin binding of EspB. Interestingly, α-catenin binds to β-catenin through its N-terminal vinculin homology domain whereas EspB binds to the C-terminal vinculin homology domain. This fact suggests that the binding of EspB to the C-terminal region of α-catenin should induce the conformational change around the N-terminal region of α-catenin to promote the dissociation of α-catenin from β-catenin. The analysis of 3D structure of EspB/α-catenin complex using e.g. X-ray crystallography will clarify the mechanism of the dissociation of α-catenin from E-cadherin/β-catenin complex by the presence of EspB. In this paper, we identified that only about 30 amino acid residues in EspB play important role for α-catenin binding which probably promotes the rearrangement cellular morphology. Then, what is the role of other regions of EspB? As mentioned earlier in this paper, EspB is a multifunctional protein which is involved in pore-formation, actin reorganization, and inhibition of phagocytosis. The analysis demonstrated here focused on the actin reorganization through binding of α-catenin. Previous analysis on enterophathogenic *E. coli* (EPEC) indicated that the region from I159 to L218 in EPEC EspB is important for binding to myosin. The regions in EHEC EspB corresponding to this region is relatively well conserved and are assigned to V158 to R109. Thus, this region in EHEC EspB can also involved in inhibition of phagocytosis by binding to myosin proteins. The regions responsible for other binding partners of EspB are still unknown, but the regions other than G41–Q70 or V158 to R109 can be also required for recognition of various other proteins. Simultaneous binding of various other proteins using different binding site can also important for all the phenomena that EspB promotes. Then, why EspB has to assume less-organized extended structures with partially folded α-helices? The observations of various other effectors that are secreted through the type III secretion systems of other infectious bacteria suggest that they tend to form less-organized structures as we observed for EspB. This fact implicates that the conformational disorder of T3SS effectors are required for effective secretion of these proteins through the narrow central channel of T3SS needles with inner diameters of ∼20 Å. This size allows the secretion of the globular proteins with \<15 amino acids in length or the hydrodynamic radius (*R*_h) of \<10 Å, according to the empirical correlation of (Å) where *N* is the length of a protein. In conclusion, we identified that the functionally important regions of EspB forms α-helical structures stabilized mainly by local interactions around G41 to Q70 and other regions are less well-organized. As a consequence, the overall structure of EspB is as extended as random coil structures. This result is highly relevant for understanding the mechanism of α-catenin recognition by EspB. Further analysis for the structural property of the complex between EspB and α-catenin will provide further detail of the mechanism of α-catenin binding by EspB. Such information will be also critical for understanding more precise structural details of unbound form of EspB in solution. We thank M. Ozawa (Kagoshima University, Japan) for providing us with a cDNA clone of α-catenin. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: DH MH HK YH MK. Performed the experiments: DH MH KNS HK YH. Analyzed the data: DH HK MK. Contributed reagents/materials/analysis tools: DH MH KNS. Wrote the paper: DH HK MH YH IY IS MK.

# Introduction Social networking sites (SNSs) are virtual communities for building new interpersonal communication, where users can create personal public profiles, interact with friends, and perform activities based on shared interests. SNSs are attractive entertainment platforms that are becoming increasingly popular worldwide. Particularly for adolescents and students, SNSs are the primary means of interaction in their free time. According to reported statistics, Facebook alone, one of the leading SNS today, has more than 2 billion users. Users can present themselves and rebuild their social identities by creating their profiles on SNSs. Meanwhile, because SNSs provide private environments free from parental supervision, adolescents can easily interact with their friends and peers happily anytime and anywhere. These advantages attract youngsters to use social networks frequently. Although using SNSs can bring convenience to individuals and society, it may also have adverse effects, such as the problem of social network addiction (SNA). Previous studies have reported that SNA can lead to harmful psychological problems such as anxiety, depression, loneliness, and low self-esteem. Moreover, it can cause more negative physical problems such as eye diseases, excessive sitting. SNA severely damages the physical and mental health of adolescents and destroys their well-being in life, and in severe cases, it may lead to suicide. The degree of social network addiction is on the rise, which has attracted the continuous attention of scholars. Adolescents are at high risk of social network addiction. Compared with other countries, SNA is the most prominent among Chinese college students, which has also attracted considerable attention from the Chinese government and researchers. Recent studies have shown that SNA among university students worsened during the COVID-19 pandemic; therefore, the factors that affect SNA among university students during COVID-19 pandemic need to be immediately identified. Previous researchers have generally concluded that factors influencing SNA are multifaceted, such as social attachment, lack of self-control, emotions, technology, enjoyment, and stress. Moreover, past studies have reported that stress under certain conditions is one of the driving factors of SNA. As COVID-19 has become health, economic, and social emergency and a unique disaster, great changes have taken place in every aspect of people’s life. These dramatic changes have added to the stress levels of college students in their studies and lives. Numerous studies have confirmed stress to be a significant predictor of SNA \[, –\]. Moreover, previous studies have reported that fear of missing out (FoMO), a negative personal emotion amplifying favorable experiences in social networks, is strongly associated with SNA and drives the growth of social network addictive tendencies. Therefore, this study suggested that the COVID-19 related stress (CRS) and FoMO may directly affect SNA of college students. Meanwhile, recent studies have shown a significant positive correlation between stress and FoMO, and stress significantly positively predicts FoMO. In addition, FoMO is often considered a mediator in studies on negative social network behaviors and has become a crucial variable for research attention. Therefore, this study inferred that FoMO may mediate between CRS and SNA among university students. Furthermore, general and persistent differences exist between genders in perceptions of the external environment (cognitive), personal factors (psychological and personality), and behavioral factors (ability, functioning, and role). Previous related studies that included FoMO, addiction problems, and stress have reported gender as a moderating variable. Therefore, this study explored whether gender plays a moderating role in the effect of CRS on the SNA of college students via FoMO. Although several previous studies exist on the impact of stress on SNA. But in fact, the COVID-19 caused massive school closures; the postponement or cancellation of most activities and events, and a lot of formal and informal interactions to online platforms; the COVID-19 significant changes in academic and life patterns of college students and potentially caused social isolation. Therefore, the COVID-19 related stress is a new phenomenon responding in a specific context, and this stress has been proved to have a very significant impact on the study and life of college students, which needs the attention of researchers. At the same time, the measurement of stress caused by sudden major public health events should be different from the general measurement of stress, and most of the previous studies have adopted the general sense of stress measurement tools for assessment, which may lead to certain limitations in the research results. Therefore, this study adopted a targeted questionnaire developed by Zurlo et al. to specifically evaluate the pressure related to COVID-19 among college students to further expand the previous research. In addition, while previous studies have explored pairwise structural relationships between stress, FoMO, and SNA in general, the relationship between the three has not been discussed in the same hypothetical model, and the effects of gender differences have not been considered on this basis. On the basis of social cognitive theory (SCT), wherein the individual, environment, and behavior interact. This study regarded the CRS as an environmental factor and FoMO as a personal factor to explore their potential influence mechanism on SNA of college students. Moreover, this study took gender differences into account, constructing a moderated mediation model to discuss these comprehensive effects. By trying to expand the social cognitive theory and the above studies through the research results, we can not only enrich our understanding of how closely these variables were associated but also provide a valuable reference for colleges and universities to help students correctly understand SNA and prevent its further deterioration. Making a contribution in these senses is the goal of this study. # Literature review ## Social cognitive theory SCT, proposed by Bandura, believes that the environmental and individual cognition factors have an impact on individual behavior. SCT has been used to study SNA in various countries and groups. Specifically, Chen et al. studied the effect of COVID-19 victimization experience (environmental factor) on mobile phone addiction of college students (behavioral factor) through future anxiety (personal factor) based on SCT. Moreover, the results showed that future anxiety plays a fully mediating role in the relationship between the effect of COVID-19 victimization experience and mobile phone addiction of college students. Wu et al. conducted a study on 277 Chinese Macao youth using the SCT model as a framework considering outcome expectancies as an environmental factor, Internet self-efficacy as a personal factor, and SNA as a behavioral factor; the results revealed that both outcome expectancy and online self-efficacy positively predicted SNA. Yu et al. investigated 395 Chinese people to explore the effects of low optimism and loneliness on SNA of college students based on SCT; the results showed that low optimism is an indirect risk factor for SNA, whereas loneliness is a direct risk factor for SNA. Therefore, on the basis of SCT, this study considered the CRS as an environmental factor, FoMO as an individual factor and SNA as a behavioral factor. In addition, gender was included as a moderating variable to construct a moderated mediation model to explore the effect of CRS on SNA through FoMO and the mediating role of gender. ## CRS and SNA Similar to SARS and H1N1, COVID-19, as a public health emergency with extremely fast transmission speed and high infection rate, has been prevalent all over the world in recent years. Studies have suggested that such public health emergencies can spread stress, lead to psychotic symptoms such as panic and even suicide. The COVID-19 has dramatically changed the lives of college students. Studies have shown that college students are one of the groups that were most affected by COVID-19. The psychological stress caused by COVID-19 among college students is enormous and multifaceted. A recent study on medical undergraduates revealed that remote online examinations adopted during the pandemic were more stressful than on-campus examinations. Moreover, data from a longitudinal study by Hakami et al. supported the findings of Elsalem et al., drastic changes such as social restrictions during the COVID-19 pandemic are putting extra stress on college students. To further explain the stressors caused by COVID-19, Zurlo et al. investigated more than 500 European university students during the Covid-19 pandemic. The results indicated that the stressors of the COVID-19 in university students were “interpersonal and academic life,” “social isolation,” and “fear of contagion,” and a seven-item COVID-19 Student Stress Questionnaire (CSSQ) was developed to evaluate the stress of students related to COVID-19. SNS can be used as a method to relieve stress and is the easiest and convenient method to reduce stress. In response to daily stressors, individuals may use SNS more frequently, leading to SNA. Therefore, stress can be a factor in SNA. Recent studies have indicated that a significant positive correlation between stress and SNA, and higher levels of perceived stress are associated with a greater risk for SNA. Adolescent students who experience stress in a learning environment are most likely to have SNA. Moreover, because the COVID-19 pandemic had put too much pressure on college students in various aspects, it may have led them to overuse social networks, thus leading to SNA. In view of the above discussion, this study speculates that college students in the context of COVID-19 pandemic may become addicted to social networks due to COVID-19 related stress. Therefore, the first aim of this study is to explore whether COVID-19 related stress has a significant positive effect on the SNA of Chinese college students. ## The mediating role of FoMO FoMO is a state of anxiety that one will miss something. It is a fear that others may be having beneficial experiences of which one is not a part, a desire to be continuously connected to others, and a negative emotion caused by unmet social relationship needs. Studies suggest that individuals experience greater levels of FoMO when their social needs are unmet. The decline of individual social activities during the pandemic, may likely to lead to inadequate socialization and thus increase the FoMO. Recent studies have shown a positive association between stress and FoMO. Fabris et al. surveyed 472 European adolescents and found that Sensitivity to Stress associated with negative experience on social media, which is neglected by online peers, was positively associated with FoMO. Yang et al. investigated 2276 Chinese college students and revealed that stress was not only positively associated with FoMO but also positively predicted FoMO, and stress could also further affect problematic mobile phone use through FoMO. Studies have shown that FoMO is strongly associated with SNA. Online social networking creates a platform to socially connect and engage, attracting people with FoMO. Thus, SNSs are often used to satisfy FoMO through online social activities to alleviate negative emotions. Such negative emotions may cause people to dedicate more energy to SNSs at the expense of real life and cause more real-life problems such as interpersonal relationships and work and thus generate new negative emotions in return. In this vicious cycle, the psychological dependence on SNSs is further increased, eventually causing SNA. Findings show that people with high FoMO are more active in responding to messages on SNSs. Fabris et al. suggested that FoMO was a promoting factor for social media addiction. Moore & Craciun further confirmed the significant predictive effect of FoMO on social media addictive tendencies and indicated that FoMO was the most critical driver of growing addictive tendencies in social networks. Previous researchers have found that FoMO often acts as a mediator in studies predicting the behavior of social network use. For instance, Beyens et al. noted that FoMO fully mediated the effect of the need to belong on Facebook. In Liu & Ma’s empirical research model, FoMO mediated the relationship between anxious attachment and SNS addiction. In summary, this study suggestes that CRS may significantly affect SNA through FoMO. The second aim of this study is to investigate whether FoMO has a mediating role in the influence relationship of CRS on SNA of Chinese college students. ## The moderating role of gender Previous researchers have revealed general and persistent differences between genders in the cognitive, psychological, personality, and behavioral aspects of the external environment. First, gender is an essential determinant of psychological stress perception and significant gender differences exist in the perception of stress. Recent studies have shown significant gender differences in perceived stress among different groups such as surgeons, police officers, high school students, and college students. Second, the findings of gender differences in FoMO are controversial. On the one hand, an online study conducted in Germany observed differences in FoMO by age and not by gender. No statistically significant gender differences were observed in the relationship between FoMO and social network use during the COVID-19 pandemic. On the other hand, a cross-sectional study conducted by Li et al. found significant gender differences in FoMO among Chinese college students, and Munawar et al. showed significant gender differences in FoMO in 324 respondents, wherein women were more sensitive to FoMO than men. Third, previous studies have pointed out that addictive behaviors may be associated with gender differences, and significant differences exist in SNA by gender. Specifically, a meta-analysis showed significant gender differences in SNA. A study conducted in Spain with university students found similar results supporting gender differences in SNA. Furthermore, previous studies on stress, FoMO, and addiction have introduced gender differences as moderating variables for discussion. Therefore, this study focuses on gender differences, and the third aim is to explore whether gender has a moderating role on the mediating model of CRS influencing college students’ SNA through FoMO. ## The present study Based on the above theories and empirical studies, this study constructs a moderated mediation model. Compared with simple mediation or moderation models, integrated moderated mediation model provides a deeper understanding of the potential influence mechanism of college students’ SNA. Corresponding to the above research objectives, the following research hypotheses are presented in this study: 1. Hypothesis 1 (H1): CRS has a significant positive effect on Chinese college students’ SNA. 2. Hypothesis 2 (H2): FoMO has a mediating role in the influence relationship of CRS on SNA of Chinese college students. 3. Hypothesis 3 (H3): Gender has a moderating role in the direct and indirect effects of CRS on SNA through FoMO. 4. H3a: Gender has a moderating role between CRS and FoMO. 5. H3b: Gender has a moderating role between FoMO and SNA. 6. H3c: Gender has a moderating role between CRS and SNA. # Method ## Ethics approval This study was approved and ethically reviewed by the Academic Ethics Committee of Hainan Vocational University of Science and Technology (HKD-2022-25). The Declaration of Helsinki and ethical standards were followed. The corresponding author was designated to be responsible for data extraction and analysis. ## Participants and procedure In this study, purposive sampling was used to recruit college students at a university in Southern China. The criteria for recruiting participants were college students who voluntarily wanted to participate should be recruited. First of all, professional training was provided to teachers who would distribute the questionnaire. The questionnaire and items were explained to make them understand the purpose of the study. Second, the participants were informed of the anonymous submission of the questionnaires, the study’s purpose, and the confidentiality agreement. After obtaining consent from participants, questionnaires were distributed through the online questionnaire platform Questionnaire Star ([www.wjx.cn](http://www.wjx.cn)). By scanning a two- dimensional barcode, participants could complete the survey with the help of their teachers. They could refuse or withdraw from the study anytime before submission. According to the formula for calculating the sample size by Israel, the official sample size of this study should not be less than 652. In our case, 777 questionnaires were distributed, 702 valid questionnaires were returned (75 invalid questionnaires were excluded). The effective rate was 90.35%. There were 170 (24.2%) male and 532 (75.8%) female students; 236 (33.6%) only children and 466 (66.4%) non-only children; 317 (45.2%) first year university students, 160 (22.8%) second year students, 152 (21.7%) third year students, and 30 (4.3%) fourth year students; and 43 master’s students (6.1%). The gender imbalance in the sample is because the sampled university was a normal university, including a larger proportion of female students in the natural sample composition. ## Instruments ### CRS This study used the COVID-19 Student Stress Questionnaire (CSSQ),which was developed by Zurlo et al. The questionnaire was tested for good reliability among college students. Three dimensions of the scale were relationships and academic life, isolation and fear of contagion. It was a five-point Likert scale ranging from 1 (not at all stressful) to 5 (extremely stressful), with higher scores indicating more stress from COVID-19. ### SNA This study used the SNA scale for adolescents, which was developed by Wang et al. to measure SNA among college students. It was a five-point Likert unidimensional scale, ranging from 1 (not at all true of me) to 5 (extremely true of me), with higher scores indicating higher levels of SNA. ### FoMO This study used the FoMO scale, which was developed by Przybylski et al. to measure FoMO among college students. It was a five-point Likert unidimensional scale, ranging from 1 (not at all true of me) to 5 (extremely true of me), with higher scores indicating higher levels of FoMO. ## Statistical analysis In this study, first, we first conducted the reliability and validity tests of each measurement instrument. The reliability was reflected by Cronbach’s α, and when Cronbach’s α was greater than 0.7, the reliability was better. Confirmatory factor analysis (CFA) was used to test validity and model fit. Standardized Factor Loading (SFL), Composite Reliability (CR) and Average Variance Extracted (AVE) of each measured model were tested, according to the suggestion of Cheung & Wang, SFL\>0.5, CR\> 0.7 and AVE\>0.5, indicated that the convergent validity of the measurement model is good. Considering the sensitivity of chi-square value to large sample sizes, chi-square value was not reported in this study when the fitness of the measurement model was reported. According to Hu & Bentler, when the sample size was large, the chi-square values tended to reach significance and other fit indexes could be referred to. In this study, the following model fit indexes were reported: RMR\<0.08; GFI\>0.85; CFI\>0.85; NFI\>0.85; TLI\>0.80; IFI\>0.85 and PNFI\>0.5. If the above criteria are satisfied, the measurement model fitness is acceptable. In addition, the discriminant validity of each potential variables needed to be accounted for. The square root of AVE was performed to assess the discriminant validity of each dimension of the measurement model. The criterion was that the square root of the AVE of each dimension should be greater than the correlation coefficient of each dimension. Secondly, since the questionnaire of this study was collected by online self- report, it was necessary to test the Common Method Variance (CMV) before data analysis. Harman’s One-Factor test was used to test the CMV. The Kaiser-Meyer- Olkin (KMO) should be greater than 0.8, with the Bartlett test of sphericity reaching significant (p \< 0.001) and the explanatory power of the first factor no exceeding the critical value of 50%. If the above indexes could meet the criteria simultaneously, indicating that the CMV problem did not affect the findings. Thirdly, descriptive statistics and correlation analysis were performed on the study variables. The descriptive statistics reflected the mean and standard deviation of each variables. Pearson’s correlation coefficient reflected the correlation between the study variables, and the correlation coefficient was less than 0.7, indicating that all variables had no collinearity problem, and regression analysis could be performed. Fourth, Model 4 of PROCESS was used to test the mediation role of FoMO, with the CRS as the independent variable, FoMO as the mediating variable, and SNA as the dependent variable. Furthermore, the mediation models should be classified into four types: the full mediation, the partial mediation (the complementary partial mediation and the competitive partial mediation), only direct effect and no effect. The type of mediation model would be reported in this study. Model 59 of PROCESS added gender as a moderator variable to validate the mediating role of moderation. Meanwhile, the bootstrap confidence interval (CI) was set to 95%, and the number of samples was set to 5,000. The CI cannot contain 0, indicating a significant effect. # Results ## Measurement model ### CRS The CFA results indicated that the last item of the scale had factor loadings less than 0.5 and was removed. Standardized factor loading (SFL) for the retained items ranged from 0.787 to 0.874, both greater than 0.5; Composite Reliability (CR) values were 0.890 and 0.861, both greater than 0.7; and Average Variance Extracted (AVE) values were 0.670 and 0.756, both greater than 0.5. These three indices indicated the ideal convergent validity of the measurement model. The fit indices of the measurement model were as follows: RMR = 0.039, GFI = 0.976, CFI = 0.984, NFI = 0.981, TLI = 0.970, IFI = 0.984, and PNFI = 0.523, indicating a good fit of the measurement model. Cronbach’s α of the total scale was 0.906 (\>0.7). ### SNA The results shown that the SFL ranged from 0.694 to 0.806, all values greater than 0.5. The CR value was 0.919, greater than 0.7. The AVE value was 0.587, greater than 0.5. These three indices indicate the ideal convergent validity of the measurement model. The fit indicators of the measurement model were as follows: RMR = 0.046, GFI = 0.905, CFI = 0.928, NFI = 0.923, TLI = 0.900, IFI = 0.929, and PNFI = 0.660, indicating a good fit of the measurement model. Cronbach’s α of the total scale was 0.919 (\>0.7). ### FoMO The CFA results indicated that three items had factor loading less than 0.5 and were removed. The SFL for the retained items were 0.573–0.858, all values greater than 0.5; the CR value was 0.901, greater than 0.7; and the AVE value was 0.571, greater than 0.5. These three indices indicate the ideal convergent validity of the measurement model. The fit indicators of the measurement model were as follows: RMR = 0.063, GFI = 0.851, CFI = 0.877, NFI = 0.873, TLI = 0.815, IFI = 0.877, and PNFI = 0.582, indicating that the fit of the measurement model to the observed data was acceptable. Cronbach’s α of the total scale was 0.901(\>0.7). ## Discriminant validity To validate the discriminant validity of all dimensions, we adopted the rigorous testing method of the square root of AVE. As shown in, the results demonstrated that the square root of AVE in each dimension was greater than the correlation coefficient in each dimension, indicating that the discriminant validity of each scale was good. ## CMV test To validate CMV, we performed Harman’s one-factor test. Unrotated factor analysis revealed that KMO was 0.925 (\>0.8), and the Bartlett test of sphericity reached significance (p \< 0.001). The analysis revealed three factors, and the explanatory power of the first factor was 42.271%, which did not exceed 50%, indicating that the CMV problem was not significant. ## Descriptive statistics and correlation analysis The descriptive statistics and correlation analysis for the CRS, SNA, FoMO, and gender are shown in. The correlation analysis showed that CRS was positively and significantly correlated with SNA (r = 0.364, p \< 0.001) and FoMO (r = 0.447, p \< 0.001), and FoMO was positively and significantly correlated with SNA (r = 0.579, p \< 0.001). Meanwhile, gender was negatively and significantly correlated with SNA (r = −0.084, p \< 0.05), and no significant correlation existed between gender and CRS and FoMO. The absolute values of the correlation coefficients between any two variables in this study were less than 0.7, indicating no collinearity problem. ## The mediating role of FoMO The mediating effect of FoMO was tested using Model 4 of PROCESS. The results are shown in. After controlling for gender, the CRS significantly and positively predicted SNA in model 1 (B = 0.282, p \< 0.001), and H1 was supported. In addition, it significantly and positively predicted FoMO in model 2 (B = 0.347, p \< 0.001). In model 3, after adding FoMO as a mediating variable, FoMO significantly and positively predicted SNA (B = 0.510, p \< 0.001), and CRS still predicted SNA significantly (B = 0.105, p \< 0.001), indicating that FoMO had a partial mediating effect in the relationship between CRS and SNA, and H2 was supported. The bias-corrected nonparametric percentile bootstrap method was further used to test the mediating effect of FoMO. The indirect effect value was 0.177, and the 95% CI ranged from 0.136 (LLCI) to 0.220 (ULCI), excluding 0, indicating a mediating effect. The direct effect value was 0.105, and the 95% CI ranged from 0.047 (LLCI) to 0.165 (ULCI), excluding 0, again validating the partial mediating effect of FoMO with the mediating effect accounting for 62.766% of the total effect. Meanwhile, the direct (B = 0.105, p \< 0.001) and indirect effects (0.347 × 0.510 = 0.177, p \< 0.001) of CRS on SNA were both positive (the direct and indirect effects pointing in the same direction), indicating a complementary partial mediation. ## The moderating role of gender To test whether gender moderated the direct and indirect relation among the CRS, FoMO, and SNA, Model 59 of PROCESS was used. The results are shown in and. After controlling for gender, a significant positive predictive effect of CRS on FoMO was observed in model 1 (B = 0.358, p \< 0.001), and the interaction term between CRS and gender was not a significant predictor of FoMO (B = −0.036, p \> 0.05), indicating that gender has no moderating effect in the relationship between CRS and FoMO, thus H3a was not supported. In model 2, a significant positive predictive effect of CRS was observed on SNA (B = 0.123, p \< 0.001), and the interaction of CRS with gender was insignificant in predicting SNA (B = −0.064, p \> 0.05), indicating that gender had no moderating effect on the relationship between CRS and SNA, thus H3c was not supported. Meanwhile, a significant positive predictive effect of FoMO on SNA (B = 0.465, p \< 0.001) and the interaction term between FoMO and gender on SNA (B = 0.169, p \< 0.05) was observed, indicating a moderating role of gender between the FoMO and SNA, thus H3b was supported. The moderating effect of gender was further validated using the bias-corrected nonparametric percentile bootstrap method. The results again confirmed that gender had a significant moderating effect only in the second half of the mediation model constructed in this study. Particularly, gender moderated the relationship between FoMO and SNA (B = 0.169, p \< 0.05, 95% CI = 0.025–0.313). Therefore, the indirect effect of CRS on SNA through FoMO was stronger in male college students (B = 0.241, 95% CI = 0.182–0.299) than females (B = 0.173, 95% CI = 0.128–0.221). To visualize the moderating effect of gender, the moderating effect was plotted. From the simple slope analysis, FoMO had a stronger predictive effect on SNA in male college students (simple slope = 0.680, t = 10.323, p \< 0.001) than females (simple slope = 0.490, t = 11.768, p \< 0.001). # Discussion ## Theoretical contributions First, a significant positive effect of COVIID-19 related stress on SNA was observed in Chinese college students and H1 was supported, which is similar to previous studies concluding a significant predictive effect of stress on SNA. During the COVID-19 pandemic, a great deal of offline teaching and learning moved from offline to online, their lifestyle changed dramatically and socialization was restricted and relied on SNSs. Excessive social network use may trigger SNA. In addition, previous studies have indicated that stress under specific conditions is one of the drivers of SNA. Whereas COVID-19 pandemic is a particular social environment, different aspects of stress associated may be faced by college students. Social networks are often used as a channel to relieve the reality of stress, particularly under COVID-19 pandemic. These reasons may contribute to the excessive use of social networks among college students during Covid-19 pandemic and may result in SNA. Second, the findings revealed that FoMO has a complementary partial mediation role between CRS and SNA. Then the H2 was supported. Particularly, CRS not only directly affects SNA but also indirectly influences SNA through FoMO, which is a state of anxiety with unmet social relationship needs is a negative emotional state. The COVID-19 pandemic has brought many restrictions to the life and study of college students, making them unable to get adequate social contact, as well as multi-faceted pressure, which will largely cause such negative emotions (FOMO) among college students. The FoMO, in turn, drove college students to use social networks frequently to satisfy social connections and engagement, increasing their psychological dependence on social networks and leading to excessive use of social networks to the point of SNA. This study further extended the results of the above studies and supported the SCT. Particularly, the mediating role of FoMO was identified, and a complementary mediating model was constructed and verified by considering COVID-19 stress on college students as an environmental factor, FoMO as a personal factor, and SNA as a behavioral factor. It indicates that FoMO is a critical facilitator for SNA among college students and can complement COVID-19 related stress to increase their SNA. Finally, notably, gender has a significant moderating effect only in the second half of the mediation model constructed in this study; the moderating effect in the first half of the path and the direct path did not reach statistical significance. Particularly, gender moderated the relationship between FoMO and SNA among college students, and the effect of FoMO on SNA was stronger among male college students than females. Thus H3b was supported. The findings are consistent with those of Koh & Kim’s study. In addition, in this study, gender had no significant moderating effect on the impact path of COVID-19 related stress on FoMO and SNA, thus H3a and H3c were not supported. The reason may be concluding no significant difference in the perception of stress by gender in a specific context of major public health disasters (e.g., severe acute respiratory syndrome \[SARS\]). Similar to SARS, COVID-19 is also an emergency and major public health disaster. Therefore, under the impact of such emergency and significant public health events, gender perception of pressure is not significantly different. The significant moderating role of gender played in the effect of FoMO on SNA may be because of the following: First, according to gender social role differences, traditional male social roles require them to shoulder more responsibilities for obtaining survival resources. Although the current social situation has changed, such stereotyped thinking still exists in the current social culture, and this traditional concept of social role makes men’s sense of crisis for social competition more intense. This may lead men to worry more than women that others are having rewarding experiences that they themselves are not, and that they therefore have a greater need for constant attention and connection with others. Second, from the perspective of the status characteristics theory, men typically occupy a stronger status position than women; therefore, they seem to socially embody more purpose-oriented characteristics and make more decisions than women. Therefore, men may need more information than women to help them make the right decisions to maintain their dominance. Hence, men may be more afraid than women about missing out on helpful information or experiences, which may make men more susceptible to FoMO. Third, the number of deaths during COVID-19 was higher in men than in women, which may also make men more afraid of missing out on beneficial health-related messages. Fourth, males are more inclined to maintain their individuality socially. SNSs, a private environment free from parental supervision, allow college students to make their identities in an unfettered environment, which may be more attractive to male students. Fifth, males appear to have a more apparent risk-taking attitude than females, wherein males are more likely to post personal information about themselves, such as phone numbers and addresses, on social network profiles, whereas females have greater privacy concerns and avoid disclosing identity information, this may also make men feel more comfortable indulging in social networking in a state of FOMO. In summary, all of these reasons are likely to make the effect of FoMO on social network addiction stronger among male college students than females. This result is an essential contribution of this study, and further enhances the understanding of gender differences among college students between FoMO and SNA. ## Practical contributions This study provides some practical contributions: First, in the future, when facing public health emergencies such as COVID-19, college students should actively make corresponding self-adjustment and actively participate in stress training courses to relieve the pressure brought by major public health disasters. Second, in the case of disasters such as COVID-19 that threaten public health, college students, especially male college students, should actively learn about FoMO, so as to correctly understand the concept and function of FoMO and make appropriate self-adjustment. Third, with the development of The Times, the difference between gender roles in current society has changed dramatically. Women may gradually have the same social power and shoulder the same social responsibilities as men Male college students should actively change their thinking, abandon the traditional thinking of social roles, and adjust themselves to reduce their excessive sense of crisis. # Conclusions Overall, this study focused on some problems due to the COVID-19. We explored the effects of CRS on SNA, the mediating role of the FoMO, and the moderating role of gender. We found that CRS not only significantly and positively predicted SNA of college students in the direct path but also significantly and positively influenced it through the mediator of FoMO in the indirect path. At the same time, this indirect effect was moderated by gender. Particularly, gender moderated the relationship between FoMO and SNA, and FoMO was a stronger predictor of SNA in male college students than in females. These findings support SCT and gender differences and broaden our understanding of the combined effects of COVID-19 pandemic and other major public health disasters on college students. # Limitations and future research directions There are still some limitations in this study. This study only conducted a questionnaire survey among college students in Yunnan Province, China, who were severely affected by the epidemic. Further research should consider sampling additional Chinese provinces. Second, this study was conducted on college students from China. Future studies could be conducted cross-culturally to compare the differences between Chinese and Western college students in the relationship between CRS and SNA or between different groups to expand the general applicability of the findings. Third, this study used a cross-sectional design and causal relationships between variables could not be inferred from the findings. Thus, longitudinal or experimental studies should be considered in future research. # Supporting information Thanks to all the participants in this study. CRS COVID-19 related stress FoMO fear of missing out SNA social network addiction 10.1371/journal.pone.0290577.r001 Decision Letter 0 Vega-Muñoz Alejandro Academic Editor 2023 Alejandro Vega-Muñoz This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 12 May 2023 PONE-D-23-09029The COVID-19 Related Stress and Social Network Addiction among Chinese College Students: A Moderated Mediation ModelPLOS ONE Dear Dr. Li, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Jun 26 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Alejandro Vega-Muñoz, Ph.D. Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at  <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and  <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For more information on unacceptable data access restrictions, please see [http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data- access-restrictions. ](http://journals.plos.org/plosone/s/data-availability#loc- unacceptable-data-access-restrictions. ) In your revised cover letter, please address the following prompts: a\) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially sensitive information, data are owned by a third-party organization, etc.) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. b\) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data-availability#loc-recommended- repositories>. We will update your Data Availability statement on your behalf to reflect the information you provide. 3\. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: [http://journals.plos.org/plosone/s/supporting- information. ](http://journals.plos.org/plosone/s/supporting-information. ) \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Partly \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: N/A \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The authors presented a study on the topic: “The COVID-19 Related Stress and Social Network Addiction among Chinese College Students: A Moderated Mediation Model” which examines the relationship between COVID-19 related stress (CRS) and social network addiction (SNA) among Chinese university students. In this study, the authors adopted a social cognitive theory and gender differences as determining factors in their theoretical framework. Both factors investigate the mediating role of fear of missing out (FoMO) as well as the moderating role of gender. There are few comments I wish to address as follows: 1\. The hypothetical problem is not very clear; I would suggest that the authors outline their hypothesis by identifying the hypothesis and stating them in order. Authors can use bulletins to do that. 2\. Authors should relate their findings according to the stated hypothesis. 3\. The tables in the article should be properly drawn. Some lines are hidden with text hovering over the lines. 4\. According to the gender influence in the study, can the authors provide an explanation or theoretical justification for why males might be more affected by FoMO compared to females. Reviewer \#2: Hypotheses 2 and 3, when speaking of a "mediating or moderating role", it is not clear in what sense. These hypotheses should be stated more concretely. In the method, the steps to be followed for the validation of the instruments and the rest of the analysis should be indicated more clearly and in a logical order. The presentation of results should also follow this order, keeping the concordance between each item of the research. In the practical contributions, the first part emphasizes Covid. A broader context should be given based on the current situation, which is no longer a health emergency. On the other hand, emphasis is placed on the actions to be taken by university teachers, however, the study does not focus on them; in this regard, the practical contributions should be related to the main line of research of this study. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0290577.r002 Author response to Decision Letter 0 23 Jun 2023 Response to Reviewer 1 Comments Dear reviewer 1: Thank you very much for your time involved in reviewing the manuscript and your comments have further improved the quality of the manuscript. We have carefully reviewed the comments and revised the manuscript accordingly. The modified section was already highlighted in yellow. Hope the explanation has fully addressed all of your concerns. Point-by-point response to reviewer are attached below this letter. Please see the attachment. Point 1: The hypothetical problem is not very clear; I would suggest that the authors outline their hypothesis by identifying the hypothesis and stating them in order. Authors can use bulletins to do that. Response 1: Thank you very much for your comments. We have presented the hypotheses in order and use bulletins to present the research hypotheses. Please see the lines 230-244 of our manuscript for detailed revisions. Point 2: Authors should relate their findings according to the stated hypothesis. Response 2: Thank you very much for your advice. We have linked research hypotheses with findings. Please see the lines 394, 399, 420, 423, 427, 450, 463, 483 and 485 of our manuscript for detailed revisions. Point 3: The tables in the article should be properly drawn. Some lines are hidden with text hovering over the lines. Response 3: Thanks so much for your reminding, we have adapted the tables. Please see the lines 369, 388, 409 and 428 of our manuscript for detailed revisions. Point 4: According to the gender influence in the study, can the authors provide an explanation or theoretical justification for why males might be more affected by FoMO compared to females. Response 4: We have added gender social role differences in the discussion section to explain why males were more affected by FoMO than women. Please see the lines 493-504 of the manuscript. Response to Reviewer 2 Comments Dear reviewer 2: Thank you very much for your time involved in reviewing the manuscript and your comments have further improved the quality of the manuscript. We have carefully reviewed the comments and revised the manuscript accordingly. The modified section was already highlighted in yellow. Hope the explanation has fully addressed all of your concerns. Point-by-point response to reviewer are attached below this letter. Please see the attachment. Point 1: Hypotheses 2 and 3, when speaking of a "mediating or moderating role", it is not clear in what sense. These hypotheses should be stated more concretely. Response 1: Thanks for your advice, we have revised the hypothesis to make the hypothesis reflect more concretely. Please see the lines 166-168, 201-203 and 227-244 of our manuscript for specific modifications. Point 2: In the method, the steps to be followed for the validation of the instruments and the rest of the analysis should be indicated more clearly and in a logical order. The presentation of results should also follow this order, keeping the concordance between each item of the research. Response 2: According to your comments, we have modified the data analysis strategy in the "Statistical analysis" section to correspond to the sequence of results presentation. Please see the lines 295-334 of our manuscript for specific revisions. Point 3: In the practical contributions, the first part emphasizes Covid. A broader context should be given based on the current situation, which is no longer a health emergency. On the other hand, emphasis is placed on the actions to be taken by university teachers, however, the study does not focus on them; in this regard, the practical contributions should be related to the main line of research of this study. Response 3: Thank you for your comments. A broader context in the practice contributions has been given based on your recommendations and we shifted our focus from university teachers to the main line of research of this study (college students). Please see the lines 523-533 of the manuscript. 10.1371/journal.pone.0290577.r003 Decision Letter 1 Vega-Muñoz Alejandro Academic Editor 2023 Alejandro Vega-Muñoz This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 11 Aug 2023 The COVID-19 Related Stress and Social Network Addiction among Chinese College Students: A Moderated Mediation Model PONE-D-23-09029R1 Dear Dr. Li, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Alejandro Vega-Muñoz, Ph.D. Academic Editor PLOS ONE Additional Editor Comments (optional): Both reviewers have accepted the manuscript. Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#2: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The publication exhibits a remarkable and sophisticated writing style, showcasing the mastery of English language construction. The study was meticulously conducted, reflecting a commendable commitment to adhering to research ethics and best practices. The authors demonstrated a clear understanding of ethical considerations, ensuring the well-being and rights of participants throughout the research process. The comprehensive nature of this study is evident, with the authors leaving no stone unturned in their pursuit of knowledge and understanding. The methodology employed was sound, fostering confidence in the validity and reliability of the results obtained. The data analysis was both robust and thorough, leading to meaningful insights and valuable contributions to the field of study. It is evident that the authors have taken great care in crafting their publications and presenting their findings in a logical and coherent manner. The clarity of their arguments and the organization of the content make it an engaging read for both experts and non-experts in the subject matter. Given the rigour and high quality of this research, the publication is undoubtedly suitable for inclusion in this esteemed journal. Its significance and relevance to the broader academic community make it a valuable addition to the body of literature in the field. While the publication meets the standards for this journal, it is always beneficial to invite external peer reviews to further validate the research and provide constructive feedback. This process can strengthen the credibility of the study and ensure that potential blind spots or biases are addressed. Additionally, I would like to commend the authors for their adherence to research ethics and responsible conduct. Ethical considerations are of utmost importance in research, and their conscientious approach provides a positive example for others in the academic community. Finally, I suggest that the authors should consider sharing their research data openly, if possible, to promote transparency and facilitate reproducibility. This practice not only benefits the research community but also upholds the principles of scientific integrity. Overall, the publication is a testament to the author's dedication, expertise, and adherence to research ethics, making it a worthy addition to the journal's prestigious collection of scholarly works. Reviewer \#2: The authors have answered the questions of the first round of review, making the required adjustments in the article. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* 10.1371/journal.pone.0290577.r004 Acceptance letter Vega-Muñoz Alejandro Academic Editor 2023 Alejandro Vega-Muñoz This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 16 Aug 2023 PONE-D-23-09029R1 The COVID-19 Related Stress and Social Network Addiction among Chinese College Students: A Moderated Mediation Model Dear Dr. Li: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Alejandro Vega-Muñoz Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Initiation of DNA replication is a critical step in the regulation of cell proliferation. The replicon model proposed 46 years ago has served as a good paradigm for our understanding of the initiation step of DNA replication. According to this model, the origin, and adjacent DNA sequences whose replication depend on it, define an independent unit of replication, or replicon. The initiation step relies on the interaction of trans-acting factors (initiators) with cis-acting DNA sequences (replicators or origins). Based on studies on the single replicon present in E. coli, the role of the initiator protein(s) has been expanded to mark the position of the origin, as well as to serve as a recruitment factor that facilitates the opening of the DNA helix, a step required for the initiation of DNA synthesis. Most of our current understanding about the initiation step in eukaryotic DNA replication is based on the wealth of experimental information obtained in both the budding yeast and frog embryos. In the budding yeast, origins are defined by the presence of a small (10–15 bp) conserved DNA sequence, named autonomous replication consensus sequence (ACS), harboring the autonomous replicating sequence (ARS) motif. Regardless of their source, DNA sequences harboring the ARS motif, can promote DNA replication in yeast. A genome wide functional analysis of the distribution of replication origins in budding yeast has shown significant agreement with a computational analysis based solely on the distribution of ARS-related motifs in the yeast genome. These results strongly indicate that in budding yeast, specific DNA sequences dictate the position of the initiation step of DNA replication. At the opposite end of the spectrum, in frog and fly embryos, DNA replication appears to initiate randomly along the genome. Moreover, any DNA sequence, regardless of its source and composition can replicate in these systems, arguing that no specific DNA sequence is required to initiate DNA replication. In metazoans, the temporal regulation of regional initiation of DNA replication and the identification of defined origins of DNA replication which can function ectopically have been presented as arguments for the occurrence of specific DNA sequences at origins. However, to date a specific DNA sequence has not yet been identified, although some degenerate sequences and motifs have been proposed. There is ample evidence suggesting that the number of potential mammalian origins exceeds what is required to duplicate the whole genome, but the distribution of potential origins along the chromosomes and the manner they are activated are still unclear. In this study we have used a DNA microarray-based nascent strand abundance assay, and high-throughput DNA sequencing to determine the distribution of putative origins of DNA replication along selected regions of human chromosomes covering 1% of the human genome. Data from three different cell lines indicate that potential origins are closely spaced (3–5 kb) and that their positioning is largely conserved. More interestingly, our results indicate that origins are not randomly distributed but that are enriched at the 5′-ends of expressed genes as well as at the locations of intergenic conserved sequences. The association of origin positioning with gene expression was further investigated in MCF-7 cells. We found that origins are preferentially positioned at promoters of highly active genes, and that a statistically significant correlation exists between the positioning of origins and the location of H3K4Me3 and Pol-II binding sites on chromatin. Overall, our results suggest a strong link between the distribution of origins of DNA replication and features of the genome related to gene expression and chromatin organization. # Results ## Overall strategy To study the global distribution of origins of DNA replication in human chromosomes we have followed a strategy which utilizes a DNA microarray hybridization assay to measure the enrichment of short nascent strand DNA obtained from asynchronous proliferating cells. Briefly, nascent DNA strands released from total genomic DNA by heat denaturation, were size fractionated on a 5–30% sucrose gradient. A selected pool of fractions containing DNA in the 0.7–1.5 kb size range, were subjected to digestion with λ-exonuclease, and the resulting DNA constituted our test DNA. Total genomic DNA, obtained from the same cell line and sonicated to a similar size range constituted our reference DNA. In all our preparations, the test DNA showed at least a 20-fold enrichment of origin relative to adjacent non-origin sequences, as determined by a real time PCR-based nascent DNA abundance assay. In contrast, the same assay performed with the reference DNA yielded a ratio close to 1. Thus the enrichment found with our test DNA fulfilled the criterion of at least 10-fold enrichment for a site to be considered an origin of DNA replication. Both test and reference DNAs were then labeled with Cy-5 and Cy-3 dUTP derivatives, respectively and hybridized to a custom made DNA tiling microarray containing 50–60 nt DNA probes staggered in 50–60 bp steps and spanning a total of 33.5 Mb of human DNA (Supporting). Repeat DNA sequences encountered in these regions were masked and excluded in the array. A signal-processing algorithm (see Statistical Methods Supplement) was utilized to analyze the microarray data and to identify peaks indicating the positions on the genome where short nascent DNA strands were enriched. These sites defined the locations of origins of DNA replication. ## Localization of origins of DNA replication in MCF-7 cells Our initial studies were performed with DNA obtained from the breast cancer cell line MCF-7. Prior to the isolation of short nascent DNA, the exponential growth of the culture was verified by fluorescent activated cell sorter (FACS) analysis (a representative FACS profile for an MCF-7 preparation is shown in Supporting, and the quantification of cells in each cell cycle phase for all the preparations used in this study is shown in Supporting). The nascent nature of the DNA pool in the 0.7–1.5 kb size range was confirmed by employing a real time PCR-based enrichment assay focused at a previously reported origin of DNA replication around the human ribulose phosphate epimerase (RPE) gene. We found that the enrichment values (a) were maximal at the fractions containing DNA in the short size range used for array experiments (0.7–1.5 kb); (b) they progressively decreased as the size range of the DNA increased; and (c) this activity was not significantly affected by prior treatment of the nascent DNA preparation with either RNase or λ-exonuclease (Supporting). In addition, in DNA preparations obtained from estrogen-deprived MCF-7 cell (which by FACS analysis showed an arrest of about 80% at G1), origin enrichment was only found upon progression of MCF-7 cells into the S-phase following estradiol addition (Supporting). These observations strongly indicated that the 0.7–1.5 kb pooled DNA contained bona fide short nascent DNA strands arising from actively proliferating cells. Upon hybridization of the 0.7–1.5 kb nascent DNA (test DNA) and similarly size sheared total DNA (reference DNA) to the tiling array, we observed a strong short-range autocorrelation among neighboring probes (Fraction 10–12), which was absent in the input\input hybridization (self-self). These results suggested that the peak signals observed with the short DNA arose from the enrichment of regionally localized DNA sequences in our DNA preparation. If the pattern of peaks and troughs derived from short nascent DNA, as opposed to randomly broken DNA fragments, we predicted that the peak signals would diminish in fractions containing larger DNA fragments. To test this, we examined fractions containing nascent DNA in the ranges of 1.5–3 Kb, and ≥3 kb DNA (Fraction 18 and Fraction 28, respectively). As predicted, their profiles were both quantitative and qualitatively different from that of the fraction 10–12 DNA pool, yielding progressively fewer and broader peaks. We interpreted these results as indicative of enrichment for origins of DNA replication in the 0.7–1.5 kb fraction, that decreased with fragment size, since the ratios of signals emanating from test versus reference DNA were greatest in the short nascent DNA and declined in fractions with larger fragments. To validate the accuracy of our origin mapping method we used two approaches: First, we calculated by real-time PCR the copy number at positions of the array showing 13 peaks and 22 troughs on two contiguous regions of chromosome 17 (for the list of primer sets used see Supporting). As shown in, our real time PCR results paralleled the patterns observed with the microarray assay, thus validating the test/reference DNA ratios deduced from the array hybridizations. Second, we determined the nascent DNA enrichment at four chromosomal regions, embedded into our DNA array (Supporting) that served as internal controls for origins of DNA replication, and positioned around the c-myc, β globin, Lamin B2, and the RPE genes, respectively. We found that nascent DNA peaks detected in the array occurred in proximity or, at each one of these origins, with a mean distance of less than 500 bp between predicted origins and centers of known origin windows (Supporting). Finally, short nascent DNAs from three independent MCF-7 preparations produced the same array profile. To further ascertain that the array profile obtained did not arise from contaminating short double stranded DNA fragments, we obtained two independent nascent DNA preparations from MCF-7 cells (NS71 and NS73) and treated them with λ-exonuclease, following a standard protocol. Upon hybridization of the λ-exonuclease resistant DNA preparation to the array, we observed that the peak profile obtained with these two preparations was almost indistinguishable from that obtained without λ-exonuclease treatment, indicating that the peak profile observed did not arise from contaminating DNA. Finally, to rule out the possibility that a hybridization bias may be responsible for the peak profile obtained, we determined the abundance of DNA fragments present in our short nascent DNA preparation using high throughput DNA sequencing, and compared that profile to the one found through the array method. To this end, we obtained an independent λ-exonuclease-resistant nascent DNA preparation from MCF-7 cells, in the size range of 400–800 nt, which was then converted to double stranded DNA, using DNA polymerase I Klenow fragment and random primers. This DNA was then sequenced using the Illumina Genome Analyzer II. The average of three independent sequencing reads (named NS-seq) were aligned to the UCSC genome browser hg18 build, then converted to hg16 (liftOver, UCSC) for comparison to DNA microarray data (named NS-chip). illustrates the significant correlation between the position of both NS-seq and NS-chip tracks along a 100 kb region of Chr17. This correlation is not confined to regions of high sequence tag abundance but also extends to less abundant regions (Supporting). Altogether, the concordance of the results obtained by these two distinct methods strongly supports the enrichment profiles in nascent DNA observed in our DNA array corresponding to the location of putative origins of DNA replication in MCF-7 cells. It is interesting to note that while there is a good correlation between the positions of peaks observed with both methodologies, a much larger range of peak heights is observed with the sequencing technique. This different peak profile may reflect the higher sensitivity obtained with the sequencing technique. ## Distribution of origins of DNA replication is similar in different cell lines To investigate if the peak profiles varied between cell lines, we compared the MCF-7 profile to that of BT-474 another breast cancer cell line, as well as, to that obtained with H520, a lung cancer cell line. We calculated the number of origins detected in all chromosomal regions contained in the array for peaks with a height \>1 (log2 units), allowing an overlap in independent replicate experiments of at least 750 bp. In MCF-7, we calculated the number of origins detected with both the short nascent DNA (fractions 10–12) as well as with fractions of increasing size (fractions 18 and 28). The short nascent DNA pool yielded 8281 peaks, and as expected, the number of peaks decreased considerably as the average size of the nascent DNA increased in size (3074 peaks and 192 peaks were found in fractions 18, and 28, respectively;). Similarly the spacing between peaks or inter-origin distances, were substantially shorter in the fraction 10–12 pool (about 4 kb) compared to about 10 kb in fraction 18, and 1 Mb for fraction 28, respectively. When we compared the inter-origin distances among BT-474 and H520 cell lines, we found that they fell within the range found in MCF-7 cells (3–5 kb; Supporting ), a spacing similar to that reported in a human lymphoblastoid cell line. Given that the array profiles and spacing were similar in all cell lines, we wished to determine if the distribution of origins was also similar. To this end, we measured the concordance of origin positions across all the chromosomal regions covered in the array. We found a high level of concordance among the three different cell lines. provides an example of this concordance for a 130 kb region of Chr17 in all the cell lines. Overall, in two independent MCF-7 replicates, the concordance of origins was found to be 86%. The comparison with an MCF-7 synchronized sample (see methods section) yielded about 70% concordance. When compared to the other breast cancer cell line BT-474, a 74% concordance with the MCF-7 origins was observed. This high percentage of concordance was also maintained in the lung cancer cell line H520 (79% concordance;). A false discovery rate (FDR) analysis showing a value of \<4% confirmed the statistical significance of our findings. These results strongly suggest that the global distribution of origins is largely similar in all three cancer cell lines studied. ## Origins of DNA replication are enriched at the 5′ ends of expressed genes To examine the relationship of replication initiation with transcription, we initially compared the location of known transcription start sites (TSS) contained in our array to the pattern of origin peaks obtained in MCF-7. Using a window of 500 bp to define the positioning of peak signals, a composite origin profile at the 5′-end of all genes present in the array was generated. We observed a significant enrichment near the transcription start sites of genes covered by the array. This enrichment was even more evident for adjacent genes transcribed in opposite directions. To assess the statistical significance of these findings we analyzed 2000 positions selected at random within the genomic regions covered by the array. A t-test comparison at TSS for the random sample and the origin peaks demonstrated a highly significant difference (p\<10⁻⁴¹). No enrichment was found at the 3′ end of genes (Supporting). However, the TSS enrichment was also observed with synchronized MCF-7, and with the BT-474 and H520 cell lines (Supporting). Next, we investigated the relationship of replication initiation to gene expression level. Using an Affymetrix data set and a cut off of seven units (log2), for highly transcribed genes, we found that origin peaks at TSS were significantly more enriched in highly expressed genes compared to low/unexpressed genes (; t-test, p = 5.8×10⁻⁶). It is important to note that the genome coverage of our array is distributed almost evenly among genic and non-genic regions (Supporting), therefore the observed enrichment of origins at promoters sites does not derive from a gene dense array design. Our results are also consistent with recent reports which point to the association of human and mouse origins with transcriptional initiation. ## Origins of DNA replication correlate with the positioning of non-genic conserved DNA elements Because origin peaks were not confined to genes or their 5′ends, we sought to determine if other features of the genome were significantly related to their localization in intergenic regions. DNA sequence comparison of the human genome with other vertebrates has uncovered significant conservation of non-coding DNA sequences suggesting a functional role for these sequences. Visual inspection of the conserved sequences among the human, chimpanzee, mouse, rat, and chicken genomes (UCSC genome browser hg16 build, table mxPt1 Mm3RnGg_pHMM) along the regions covered by our array suggested a correlation of origin peaks with the position of conserved elements. We therefore developed a composite average conservation score around the highest point of the origin peaks (peak heights with at least 1 or 1.5 log2-fold changes). (green lines) demonstrates an association between the average conservation score with the highest peak enrichment point (solid and dashed green lines for peak/trough ratios of \>1.0, and \>1.5 log-2 fold, respectively). At peak height log-fold \>1.0, the Pearson correlation coefficient was found to be 0.9524, p = 1.19×10⁻³⁰. To further assess the significance of this finding, we selected a similar number of locations at random and calculated the average conservation scores along these locations (red line). No significant correlation was found. In contrast, a t-test performed to compare the average conservation score at origin peaks versus random locations was found to be highly significant (p = 1.1×10⁻¹⁴). To ascertain if the correlation between origin peaks and conserved sequences also held for non-genic regions, we selected for analysis intergenic regions that were separated by at least 1000 bp from the nearest genes on either end of the gene free segment (an illustration of such region is shown in Supporting). Randomly selected sites were subjected to the same criteria. The results shown in indicate that a highly significant correlation still remains (Pearson correlation coefficient = 0.915, p = 9×10⁻²⁴) at these conserved non-genic regions. When compared to the randomly selected sites the t-test p-value (p = 2.95×10⁻⁴) was also found significant. Similar results were found in the other cell lines used in this study (Supporting). An example of the association of origins with evolutionarily conserved regions is illustrated for a 50 kb intergenic segment on chr17 containing several highly conserved sequences (Supporting). These results are consistent with the possibility that evolutionarily conserved elements define functionally active chromatin available as preferred sites of replication initiation. ## Chromatin binding sites for H3K4me3 and PolII correlate with the position of origins of DNA replication in MCF-7 cells To further evaluate the presence of origin enrichment in regions of active promoters in MCF-7 cells, we determined by chromatin immunoprecipitation (ChIP), the positions of H3K4me3 and Pol-II chromatin binding on our array (ChIP on chip). Consistent with previous reports, we found enrichment of H3K4me3 and Pol- II binding at sites of transcription initiation (Supporting). Within 1 kb from the TSS, about 52% of all annotated promoters in our array were found to be enriched for H3K4me3, and 27% were found to be occupied by Pol-II. Interestingly, a composite origin profile around the center of either H3K4me3 (599 sites) or Pol-II (138 sites) binding sites revealed a strong correlation. A t-test of this association versus a sample extracted from the array containing 599 sites chosen at random showed a significant difference (p\<10⁻¹⁰). We also compared the association of origins with genes harboring (N = 90) or lacking (N = 244) Pol-II binding sites. shows that a stronger origin association is found at TSSs of genes harboring Pol-II binding sites (t-test, p = 3×10⁻⁶). These results clearly suggest that the open chromatin structure at these sites may drive the positioning of proteins involved in the initiation of DNA replication. Remarkably, in every nascent DNA preparation tested, origins with strong enrichment were consistently positioned at sites concordant with both Pol-II and H3K4me3 binding (Supporting). # Discussion In the present study through the application of high resolution DNA array and high throughput DNA sequencing technologies, we have considerably expanded the range and sensitivity of a nascent DNA enrichment assay used to determine the position of putative origins for DNA replication in about 1% of the human genome. We have found that the apparent distance between putative origins is about 3–5 kb in all cancer cells lines tested, a much shorter distance than that deduced from single molecule studies of DNA replication. If all these origins were active in a single cell the genome would complete its duplication in a fraction of the duration of the S phase. These results strongly support the current notion, largely based on studies in the budding yeast and embryonic systems, that eukaryotic genomes contain more initiation sites than those required to complete replication, not all of which are used in each cell cycle. In this context, it should be pointed out that since the results obtained by both DNA microarray and DNA sequencing technologies only provide an average profile of origin activation in a population of cells, these data to not define the origin distribution profile in individual cells. Therefore it is likely that while origin spacing appears to be short when averaged across the population, this most likely reflects a stochastic pattern of origin activation at larger intervals in individual cells rather than unique pattern shared in all cells. The pattern of peaks and troughs that we observe in nascent strands could therefore be regarded as defining the local probability of a replication initiation event. To gain a better understanding about origin activation in human cells and how cell lineage or environmental changes disturb replication profiles, it might be necessary to complement genome–wide population studies with single cell analysis. Indeed in one such study in the budding yeast where DNA combing combined with DNA fiber fluorography was used to deduce the replication profile in Chr VI it was found that all yeast VI chromosomes showed different replication profiles when analyzed as single molecules, while recapitulating microarray data when averaged. It would also be of considerable interest to compare origin profiles between isogenic normal and cancer cells to determine if origin selection is altered in transformed cells. Our findings offer a glimpse of the relationship between DNA replication and other aspects of mammalian chromosome function by clearly establishing that origins of replication are non-randomly distributed with respect to genome landmarks. These include the transcription start sites of active genes and conserved elements in intergenic regions. These results may be the consequence of easier access by the DNA replication machinery to specific regions. Transcription start sites must contain relatively open chromatin and are frequently marked by nearby nucleosome free DNase hypersensitive sites. Thus the formation of a replication initiation complex may be favored at these locations. Our findings confirm and extend recent results found with mouse ES cells and human cells regarding mammalian replication origins and their proximity to transcription start sites, as well as, to RNA Pol-II, and histone H3K4Me3 chromatin binding sites. A recent report suggests that DNA over-replication of short DNA fragments around promoter regions may also account for the apparent enrichment of origins around transcription start sites. While it is not clear whether the short fragments observed in the study of Gomez and Antequera elongate into mature replicons, our data is consistent with this novel and provocative finding, and it is possible that our nascent DNA preparation may also contain some of these short over-replicated DNAs. However, given that the genome coverage in our array is almost equally partitioned among genic and non- genic regions, and the fact that we do not observe a strong bias for the localization of origins in the genic regions, other factors must determine the placement of origins in non-genic regions. The function of intergenic conserved elements is largely unknown. Some may function as enhancers for distant genes and might therefore also be accessible to the formation of nucleoprotein complexes. Our results strongly suggest that many of these evolutionarily conserved elements are indeed functionally active in at least one critical process, initiation of DNA replication. We have found that active origins in intergenic regions are strongly associated with conserved sequences. However, this association is not completely explained by H3K4Me3 modification around these sites since nucleosomes containing this epigenetic marker are only slightly enriched at conserved sequences in the intergenic regions represented on our array (Supporting). These results suggest that the increased probability of replication initiation at conserved intergenic sites must be determined by another as yet undescribed property of these regions. Further whole genome investigation, coupled with studies on individual DNA molecules, will be useful to identify DNA elements and their associated chromatin features including, epigenetic modifications, participation in higher order structures, and function in regulating gene expression which enhance the likelihood of forming an active replication initiation complex. This information should provide us with a deeper understanding of the process of replication origin selection in mammalian cells. # Materials and Methods ## (a) Cell lines and FACS analysis Breast cancer cell lines MCF-7 and BT-474, and lung cancer cell line H520 were obtained from the American Type Culture Collection (Manassas, VA). Cells were grown according to recommended specifications, to about 70% confluence. An aliquot of the cell culture, corresponding to about 10⁶cells was set aside for Fluorescent Activated Cell Sorting (FACS) analysis. The aliquot of cells was prepared for FACS analysis using the cellular DNA flow cytometric analysis kit (Roche, IN) following the manufacturer's specifications. The percentage of cells in the S phase served as a good predictor of the amount of nascent strand DNA available in the preparation (see Supporting). ## (b) Isolation of short nascent strand DNA The procedure previously employed to isolate nascent DNA (14) was followed with minor modifications. About 2–5×10⁸cells were washed in PBS and collected by centrifugation. Cells were lysed with SDS in presence of Proteinase K. DNA was extracted with phenol and chloroform, precipitated by centrifugation with ethanol in 0.3 M sodium acetate, and resuspended in TE (10 mM Tris-HCl, pH 8.0; 1 mM EDTA) buffer. The re-suspended DNA was denatured by incubation in boiling water for 12 min, quenched in ice for 6 min, and applied onto a linear 5–30% neutral sucrose gradient. After centrifugation of the gradient in a Beckman SW28 rotor for 24,000 rpm for 20 hrs at 15°C, the gradient was fractionated using an ISCO 185 fractionator. One ml fractions were collected and the linearity of the gradient assessed by measuring the refractive index of every third fraction. Gradients were highly reproducible with regression line R² values larger than 0.99. The reproducibility of the gradients allowed us to identify fractions corresponding to the desired DNA size range, which in our experience falls around a refractive index of 1.35. About 80 µl of every gradient fraction was concentrated 10-fold and analyzed by gel electrophoresis in 1% agarose to confirm the DNA size range in the fractions. Fractions containing DNA in the range of 0.7–1.5 kb in length were pooled and dialyzed against TE buffer. Two independent MCF-7 0.7–1.5 kb DNA pools (NS71 and NS73) were treated with λ- exonuclease following a standard protocol. For comparison purposes an equal amount of total sheared (0.5–1.5 kb) was also digested under the same conditions. The quality and abundance of short nascent strands in the DNA preparations was assessed by real time PCR (Supporting). ## (c) DNA array design and hybridization 60-nt probes spaced by an average length of 50–60 nt were designed to cover about 34 Mb of human DNA and distributed in several chromosomal regions (Supporting). A 5 Mb region of Chr20q12.13 was represented by DNA probes to both strands of this DNA region. This served as an internal control region to assess the reproducibility of signals emanating from the same DNA region. Finally, we eliminated potentially cross-hybridizing DNA probes by checking the uniqueness of each probe in the human genome. The hybridization protocol followed was essentially similar to one used for comparative genomic hybridization (CGH) to NimbleGen arrays (NimbleGen Systems Inc.). The test DNA consisted of selected DNA fractions from the sucrose gradient with or without λ-exonuclease treatment. The reference DNA sample, corresponded to DNA obtained from the same cell line from which the nascent DNA preparation was originated. This DNA was sheared by sonication to yield an equivalent range in size fragments as the test DNA. Each of the DNA samples was independently labeled by random priming with dye-modified dUTPs (e.g. Cy5-or Cy3-), and then combined before hybridization to a NimbleGen array using a MAUI hybridization system at 42°C for 16–20 hrs. The slide containing the array was then removed from the MAUI hybridization chamber while immersed in wash buffer I (1X SSC, 0.05% SDS), placed in a slide rack containing wash buffer I and washed twice in the same buffer for 5 min with agitation. The slide was transferred to wash buffer II (0.1X SSC) and the washing repeated as before. The slide was then removed from the slide rack and dried by centrifugation (1500 rpm for 3 min) prior to scanning. ## (d) Array Data Analysis: Feature Extraction The hybridized microarray was scanned with the Agilent Microarray Scanner (Agilent Technologies, Santa Clara, CA). Two color images were analyzed by NimbleScan software (v2.1, NimbleGen, Madison, WI) and exported with probe intensities from both channels. The data were subsequently converted, without normalization, to log2-ratios in SGR and BED formats, for data visualization in the Affymetrix Integrated Genome Browser (IGB, [www.affymetrix.com](http://www.affymetrix.com)) or as custom tracks in the UCSC genome browser. ## Data Analysis (See Statistical Methods Supplement for additional details.) A peak finder algorithm was developed as follows. Briefly, we first ordered the data according to genomic location, re-sampled the data (log2-ratio) to achieve equal 50 bp spacing and then interpolated to 25 bp spacing in order to meet the requirements of subsequent methods. We then used the Savitzky-Golay convolution smoothing kernel to smooth the data to the degree needed (span  = 7 was the default choice). Peaks were then detected with the first derivative. We determined the minimal detectable peak height by using the error derived from smoothing filtering. We ignored peaks with height less than the minimum detectable peak. After the peak-finder algorithm had identified all of the peaks, the peak height density plot was generated. Self-self hybridization peak heights were mostly less than 1.0 (log2-ratio). By setting a peak height threshold at 1.0, the peak spacing density, which reflects the peak-to-peak distance, was generated by counting only peaks higher than 1.0. ## (e) Real Time PCR Real time PCR analysis was performed on all DNA preparations to ascertain their enrichment for short nascent strands. In addition to origin/non origin sites previously characterized by others around the lamin B2 gene and the β-globin locus, two bona fide sites, around the ribulose phosphate epimerase gene (14), an origin site, STS36.8, at position 211060206211060430 on Chr2q34, and a non- origin site, STS98.4, at position 211121797–211122038, on Chr2q34, were used as markers to determine the enrichment for initiation sites in the fractions containing our short nascent DNA pools. For our real-time PCR assays, fractions around the DNA size range of 0.7–1.5 kb were brought to a concentration of about 10 ng/µl, and 2 µl of these preparations were used for real time PCR assays. As a reference marker, fraction number 25 corresponding to the lower third of the sucrose gradient, was also analyzed. PCR reactions were carried out as previously described. For each primer set used, MCF-7 total DNA which had been sheared by sonication to a size range of 0.5–1.5 kb was diluted to give 20,000, 4,000, 800, 160, and 32 genomic copies per µl respectively. 2 µl of these dilutions were run in triplicate as copy number standards. As a negative control 2 µl triplicate aliquots of water were used. Copy numbers for STS36.8 or STS98.4 were estimated from a standard curve obtained with the samples containing known amounts of genomic equivalents. Ratios of the copy number at STS36.8/STS98.4 larger than 10, were indicative of a good nascent DNA strand preparation. As an internal control, the ratio of STS36.8/STS98.4 in fraction number 25 was always found to yield a value close to 1. Fractions containing the highest ratios of STS36.8/STS98.4 were then pooled and used for hybridization to the DNA tiling microarrays. Once initiation sites along each one of the chromosomal regions represented in the DNA tiling microarray were identified, we also used real-time PCR to validate both initiation and non-initiation regions by selecting STS/primer sets from peak regions and adjacent troughs, and their abundance in short nascent DNA strand preparations determined (for a list of primer sets used see Supporting). ## (f) Cell synchronization Cells were grown in appropriate media until they reach 60% confluence. At this point the cells were placed in a charcoal-treated media and kept in this media for 48 hrs. Estradiol (10 nmoles/ml) was added to the media and at times 0, 2, 4, 8, 12, and 18 hrs after the addition of estradiol, aliquots containing about 10⁶cells were taken and processed for FACS analysis as described above. Once we obtained synchronization of the cells as demonstrated by FACS analysis, nascent strand preparations from each one of the time points were prepared as described above, and the copy number at STS36.8 and STS98.4 on chromosome 2q34 was determined by real time PCR as indicated above. As expected in the time points preceding the shift of the cell culture from G1/G0 to the S phase of the cell cycle, the majority of the cells had been arrested in G1/G0. Accordingly, our real time PCR assays at these time points yielded an STS36.8/STS98.4 copy number ratio that approximated to one, indicating that short nascent DNA strands, corresponding to activated initiation sites, have not yet been produced. As the cells entered into the S phase, the proportion of cells leaving the G1/G0 phase were assessed both by FACS analysis and real time PCR assays. Cultures showing a maximal entry into S phase, around 12–14 hrs after estradiol addition (about 2–4 hrs into the S phase), were selected for analysis. ## (g) Chromatin immunoprecipitation (ChIP) was carried out according to standard protocols using a ChIP-IT kit from Active Motif (Carlsbad, CA), and following the manufacture's instructions with minor modifications. Briefly, MCF-7 cells were crosslinked with 1% formaldehyde at room temperature for 15 minutes. Then the cells were sheared with a VirSonic 100 sonicator for 10 cycles of ten 1-second pulses. After centrifugation, the chromatin contained in the supernatant was collected. Part of it was set aside and served as the input fraction. The rest was immunoprecipitated overnight at 4°C. The antibodies used were: anti-polymerase II antibody (Upstate 05–623) and anti-trimethylated histone H3K4 (Abcam Ab8580). After reversal of crosslinks at 65°C overnight, the ChIP DNA was purified using spin columns provided by the kit. For ChIP-chip, the ChIP DNA was amplified using a ligation mediated-PCR method, as previously described. A second round amplification of 15 cycles was added to increase the yield of DNA. 3 µg of amplified ChIP DNA and Input DNA was labeled with Cy5– dUTP and Cy3-dUTP, respectively, with a BioPrime DNA Labeling System (Invitrogen). The labeled ChIP DNA and Input DNA were then combined and hybridized to the NimbleGen arrays. ## (h) High-throughput DNA sequencing The procedure described above for the isolation of short nascent DNA was utilized on a culture of exponentially growing MCF-7 cells. After λ-exonuclease treatment of the DNA pool in the size range of 400–800 bp, we synthesized a double stranded DNA population required for massively parallel sequencing using the Klenow fragment of DNA polymerase I (Invitrogen, Carlsbed, CA) and random primers (Invitrogen, Carlsbad, CA). Random priming and DNA synthesis were performed according to the manufacturer's protocol except that the samples were incubated for an hour at 37°C. To insure that the resulting population represented bona fide nascent strands and that random priming did not introduce a quantitative bias, real time quantitative PCR was performed before and after second strand synthesis using origin-proximal and origin-distal primers from two regions that contain known replication initiation sites, the human beta globin and lamin B2 loci, as previously described. This DNA was then submitted for DNA sequencing using the Illumina Genome Analyzer II (Illumina, San Diego, CA). Three independent sequencing reads were merged into one single tag-count list, after aligning to hg18 and filtering of multiple occurrences of identical reads. The alignment results were subsequently down-lifted to hg16 (liftOver, UCSC) to compare to other tracks generated from the tiling microarray technology. Before the comparison, counts data from sequencing were further subsampled into 10 bp spacing, and them smoothed with kernel density function with window size of 500 bp with 50 bp interval (similar to the spacing of the tiling microarray, termed as NS-chip). For comparison purposes, we displayed every track in bar charts and omitted ratios less than 1 (log-ratio less than 0) in NS-chip results. ## (i) Data deposition The data discussed in this publication have been deposited in National Center for Biotechnology Information's Gene Expression Omnibus (GEO, ttp://[www.ncbi.nlm.nih.gov/geo/](http://www.ncbi.nlm.nih.gov/geo/)) and are accessible through GEO Series accession number GSE10917. # Supporting Information We would like to thank R. Rajanbabu, and S. Anderson, M. Kirby, and J. Qian for assistance with cell culturing and flow cytometry, respectively. We also thank J. Zhu and L. Long for computational support, and J. Stewart for help with Figure 1. MSV wishes to thank Y. Jiang and the members of the Meltzer lab for assistance and helpful discussions. [^1]: Conceived and designed the experiments: MSV MIA PSM. Performed the experiments: MSV RLW JL MMM FY. Analyzed the data: YC SB SD PSM. Contributed reagents/materials/analysis tools: PPM. Wrote the paper: MSV PSM. [^2]: Current address: Department of Epidemiology and Biostatistics, University of Texas Health Science Center at San Antonio, San Antonio Texas, United States of America [^3]: The authors have declared that no competing interests exist.

# Introduction The job satisfaction of healthcare personnel is a crucial factor in healthcare management, as it has been found to be directly linked to higher quality of care, greater patient adherence to treatments, and higher patient satisfaction. The recent pandemic has escalated the problem of low job satisfaction among healthcare workers, further threatening the sustainability of healthcare systems. The US and the UK are experiencing a staffing crisis due to workers quitting or retiring early, exhausted by the health emergency, and soon other countries might also have to address the consequences of an exodus of healthcare workers triggered by the pandemic. For instance, according to a survey of the Italian Association of Executive Physicians, nearly half of the physicians currently working for the National Health System wish to quit their position in the next two years. Similarly, 4% of Spanish doctors report their intention to leave the profession, while 30% admit considering this option. These challenges, combined with the shortages of healthcare workers already experienced by some countries, might further compromise the quality and safety of patient care. Indeed, the prepandemic estimates of the European Commission indicate a gap in the supply of healthcare human resources of approximately one million workers in 2020, meaning that nearly 15% of the health needs of the EU population was not adequately covered. Recent updates estimate the EU shortage of health workers to increase to approximately 4.1 million units by 2030: 0.6 million physicians, 2.3 million nurses and 1.3 million other healthcare professionals. Therefore, exploring the determinants of healthcare workers’ professional satisfaction and vocation in the aftermath of the pandemic becomes crucial to defining the areas of intervention to support the sustainability of healthcare systems in the long run. We addressed this issue using a unique 50-question survey with 7,681 respondents conducted immediately after the first wave of the COVID-19 outbreak (February- May 2020) in Italy, one of the countries most affected by the pandemic. Among EU countries, Italy was the first to register more than 20,000 deaths, and it reached this threshold between the end of January 2020 and April 14, 2020. Italy was also the country with the second-highest number of deaths (120,053 compared to 128,136 in the UK) by the end of April 2021, approximately a year after the start of the pandemic. During the first wave, the COVID-19 mortality rate and contagiousness were extremely heterogeneous by region, and the outbreak was more severe in the North of the country, with remarkable regional variation. To control the rapid spread of the outbreak, uniform measures were taken at the national level (e.g., case-detection, contact-tracing, isolation, physical distancing). However, each Italian region was responsible for the actual implementation of these interventions within its territory, and the response to the pandemic ultimately differed substantially both in means and timing across the country. In any case, enormous efforts were made to reorganize the available healthcare resources. These include the reallocation of health personnel from ordinary wards to the treatment of COVID-19 patients, the recruitment of additional health personnel, the increase in the number of intensive care units (ICUs) and beds by converting ordinary hospital wards to ICUs and creating temporary hospitals. Overall, during the first wave, healthcare workers faced an unprecedented situation, and their work was undermined by continuous changes in the health procedures and by frequent shortages of protective equipment, which increased their risk of infection. Unlike the general public, they were excluded from the preventive quarantine measures prescribed after having a contact with COVID-19 positives, and they could stop working only in the event of experiencing respiratory symptoms or if they tested positive. Finally, their greater risk of contagion was disregarded for most of the first wave. Our interest lies in understanding the main channels driving the overall level of job satisfaction among healthcare workers in the aftermath of the pandemic. We proxied for job satisfaction with direct questions about it, as well as with questions on the respondents’ willingness to leave the profession or to move to another specialization. This type of analysis is important because, while the literature on the consequences of the pandemic has addressed the mental and psychological hardships suffered by the healthcare workforce due to the COVID-19 emergency, there is scant evidence on the job satisfaction experienced. Moreover, most of this evidence relies on small or selected samples, such as professionals (mostly nurses only) working in a specific hospital/region. Differently, our sample is comparable to the full population of workers, both in terms of the age distribution and gender composition, which are determinant factors in explaining job satisfaction. Finally, to study the channels driving the job satisfaction of healthcare workers, we focused on both groups of drivers identified by the literature: a traditional group of drivers, consisting of personal and contextual factors (e.g., age, wages, and workload) and a COVID-19 group of drivers (e.g., being exposed to the virus). Our survey collected information on both groups. Personal and contextual factors included socioeconomic measures and the characteristics of the workplace (e.g. type of hospital, type of employment contract). The COVID-19 controls included questions on personal experience with the pandemic, such as testing positive for the virus, working with COVID-19 patients, working overtime due to the health emergency, having infected colleagues, or losing colleagues due to the virus. Finally, we also considered administrative data on the COVID-19 first wave mortality rate in the province of work (108 provinces) as an out-of-survey robustness measure. # Methods We conducted an online anonymous survey using the Google Form platform, including 50 short questions—translation available in the companion paper. As described by, answers were collected between June 15 and August 31, 2020. Potential participants received an initial invitation email, followed by two reminders, one and two weeks after the first invitation. The invitation email explained that participation was possible through the use of any electronic device (i.e., PC, tablet, or smartphone) and an internet connection. Potential participants were also informed that the expected completion time was approximately 15 minutes. ## Participants We primarily contacted potential participants through individual email addresses, having recovered their contact information from various sources: provincial boards of physicians and nurses (108 provinces), hospital websites, and representative associations, some of which also agreed to advertise and share our survey on their website, as reported in. Overall, we collected 7,681 answers distributed among 33.2% (2,549) physicians, 59.4% (4,561) nurses, and 7.4% (571) other health workers (e.g., technicians, biologists, safety inspectors, administrative personnel, and researchers). The inclusion of other health workers in addition to physicians and nurses was important to capture the impact of COVID-19 on these professionals, who were often reassigned as contact- tracers, and to account more accurately for the regional disparities in the availability of healthcare personnel. As shown in, our main focus was on the northern areas since they were the most affected, and at the same time, we encountered a general low response rate of workers form southern areas. Specifically, we had 2,797 nurses, 1,657 physicians and 400 other health workers from the northern regions (i.e., Piedmont, Valle d’Aosta, Lombardy, Trentino- Alto Adige, Veneto, Friuli-Venezia Giulia, Liguria, and Emilia-Romagna). We had 1,999 respondents (i.e., 1,206 nurses, 685 physicians and 108 other health workers) from central regions and 828 respondents (i.e., 558 nurses, 207 physicians and 63 other health workers) from southern regions. Note that it is not possible to compute our exact response rate since our survey also circulated through the provincial boards of physicians and nurses, hospital websites, and representative associations, some of which also agreed to advertise and share our survey on their website. In any case, different from most previous studies, our survey targeted all Italian healthcare workers rather than workers working in specific hospitals or geographical areas within the country. In addition, note that having a difference in survey responses across Italian macro areas is quite common in this type of empirical study. compares the distribution of our sample in terms of gender, profession, and region of work (Columns 1 and 3) with respect to the administrative data on the 2019 population of physicians and nurses (Columns 2 and 4). It appears that we achieved good representativeness along the gender composition dimension in the most pandemic-affected areas (i.e., Piedmont, Lombardy, Veneto and Emilia- Romagna), both among physicians and nurses. Indeed, the average percentages of females in the North was equal to 50.1% among physicians and almost 80% among nurses, which are very much in line with the national averages (i.e., 50.8% and 84.5%, respectively). Regarding age composition, our sample was slightly younger as physicians on average were 49 years old and nurses 40, while at the national level, the two groups recorded an average age of 52 and 47 in 2018 (the last available year). ## Ethics The study protocol was approved by the “Comitato di Approvazione per la Ricerca sull’Uomo”, that is, the Ethics Committee of the University of Verona (Prot. N. 0221872—22/06/2020). The protocol was also registered at the AEA-RCT registry (AEARCTR-0007419), while the University of Pavia certified compliance with privacy requirements (Prot. N. 61080—15/06/2020). All participants gave their written informed consent that was embedded on the first page of the questionnaire. After reading a description of the questionnaire, healthcare professionals were asked if they agreed to participate. If they ticked the option “proceed” after the statement “I confirm that I have read the information on the processing of personal data and I agree to participate in this survey” on the electronic form, the survey would begin. Respondents voluntarily participated and could withdraw from the survey at any time. ## Outcomes and covariates The outcome of the study is job satisfaction. This is directly captured by *Satisfaction*_*i* and indirectly proxied by the willingness to change jobs or medical specialization. As described in, *Satisfaction*_*i* is an index that varies from 0 to 8 by summing up 8 dummies (*D*_*ci*) referring to the following aspects: *Profession*, *Job*, *Salary*, *Work-life balance*, *Relationships with colleagues*, *Relationships with the administration*, *Work hours*, and *Career*. Since the aspects along which satisfaction is evaluated are originally measured on a 5-item Likert scale, each related dummy takes value 1 if the respondent stated being satisfied or very satisfied with the aspect recalled by the name of the dummy itself. Note that our measure of job satisfaction is constructed using the questions from the Labor Force Survey (“Rilevazioni Forza Lavoro—RFL”) by the National Institute of Statistics (Istat). For a detailed definition of the outcome variables and the related dummies, see. $$\begin{array}{r} {Satisfaction_{i} = \sum\limits_{c = 1}^{8}D_{ci}\quad\text{where}\quad D_{ci} = \left\{ \begin{matrix} 1 & {\text{sa tisfied}\mspace{360mu}\text{or}\mspace{360mu}\text{very}\mspace{360mu}\text{sati sfied}\mspace{360mu}\text{with}\mspace{360mu}\text{dimension}\mspace{720mu} c} \\ 0 & \text{otherwise} \\ \end{matrix}\operatorname{} \right.} \\ \end{array}$$ The variable *Profession Change* (*Specialization Change*) is instead defined by the answer given to a unique statement: “If I could start over, I would not be in this profession” (“If I could start over, I would choose a different field of specialization”). Then, it is a dummy equal to 1 if the respondent agreed or strongly agreed with the statement (the level of agreement was originally measured on a 5-item Likert scale). Alternative ways to construct the outcomes of interest are discussed in Section. reports the cross-correlations between the outcomes. As expected, *Satisfaction* is strongly and negatively correlated with *Profession Change* and *Specialization Change*, which are in turn positively correlated with one another. The distribution of the outcome variables within professions is shown in. Physicians experience a level of satisfaction that is higher than that of nurses but lower than that of other healthcare professionals. Consistently, nurses are more prone to change both profession and medical specialization than physicians, while other professionals place themselves in between nurses and physicians, although the confidence interval (95%) of the values across professions overlaps. This is consistent with the expectation that since the training for a physician is significantly longer than that for a nurse or for a lab technician, physicians have a higher cost of switching to a different profession. Consistent with the literature, we grouped the covariates into personal, contextual, and COVID-19-related factors as summarized in. ### Personal factors The first group of covariates included the socioeconomic characteristics and basic attributes of healthcare workers, from their gender and health conditions to a proxy for their household wealth as the dimension of their home. Several questions were included to capture the possible higher distress due to the fear of infecting others or having relatives who could become infected (i.e., home dimension, living alone, having health workers in the family). As shown in, female and male nurses almost do not differ along *Satisfaction*, while female physicians are more generally unsatisfied with their working conditions than their male colleagues. Differences also persist in the willingness to change profession and specialization. While physicians are less willing to change profession or specialization regardless of their gender than nurses, male nurses are more willing to change than female nurses. We defined *ad hoc* variables that may capture nuanced differences in the outcomes: the presence of healthcare workers in the family of origin and whether the respondent has always been employed in the facility where she is employed at the time of the survey. Regarding the professional background of the family of origin, the effect may be twofold: on the one hand, sharing the same profession and challenges is a source of support in coping with similar problems; on the other hand, the experience of relatives could serve as a benchmark to evaluate one’s own working conditions. Having changed workplaces proxies for how well the respondent knows her working environment but also indicates the variety of experience she has in terms of different working environments. This could have a positive or a negative impact on job satisfaction. The square footage of the accommodation provides valuable information since it is an indirect measure of wealth that is not necessarily captured by workers’ income (which we control for): an individual earning a low salary could still belong to a wealthy family. We also asked whether survey participants were living alone to account, for example, for unmarried cohabiting couples or non-cohabiting workers. ### Contextual factors The second group of covariates controlled for a set of basic characteristics defining the type of worker. Specifically, participants disclose whether they are nurses and hospital workers, work in a public facility and have a managerial/coordinating role since these characteristics make them among the most exposed to the pressure triggered by the health emergency. The top panel of indicates that professionals with managerial responsibilities tend to report a higher level of satisfaction and a lower willingness to change profession than those with no similar responsibilities. Physicians with managerial duties are additionally less willing to change specialization than physicians without managerial duties. Respondents also provided information on their working conditions because these can impact their satisfaction. These conditions include average working hours, number of years of employment, working in a COVID-19-related specialization (i.e, ICU, anesthesiology, emergency care, cardiology, pulmonary diseases, and infectious diseases), contract with work shifts and monthly salary. From the declared monthly salary, we created a dummy that takes value 1 when the monthly salary is above the median of the distribution in our sample (i.e., above 2,000 euros per month). Within professions, hospital workers are less satisfied than their colleagues working outside hospitals, with physicians being more willing to change profession if working in a hospital (central panel of). Surprisingly, working in a COVID-19-related specialization is not associated with any specific direction of satisfaction or the professionals’ attitudes toward their profession/specialization (bottom panel of). If anything, physicians working in COVID-19-related specializations report a lower satisfaction and a higher willingness to change specialization. Finally, we also considered the quality of the work environment because this is likely to affect job satisfaction. To this end, respondents were asked to report whether they work for a teaching hospital (i.e., a high-quality hospital) and to assess the perceived lack of medical personnel in their province of work and the quality of their employing facilities. The former is a dummy that takes value 1 if the respondent judges that there is a severe or a very severe lack of healthcare personnel in her province of work that might compromise patients’ access to care. The latter is again a binary variable equal to 1 if the respondent defines the facility she works for as being of very good or excellent quality. As shown in the top panel of, health workers who perceived a lack in medical personnel tend to report a lower satisfaction and a higher propensity to declare a profession or specialization change, with physicians driving the effect. Conversely, the bottom panel of graphically describes how a perceived higher quality of the facility is associated with a higher level of satisfaction and a lower propensity to change profession or specialization. The effect is large and significant across all professions. ### COVID-19 related factors To account for the links between the COVID-19 pandemic and the level of job satisfaction, the third and last group of covariates includes both administrative data and the personal experience with the pandemic. As administrative measure of the COVID-19 outbreak, we relied on the COVID-19 mortality rate as computed by the National Institute of Statistics (Istat) together with the Istituto Superiore di Sanitá (Iss) on administrative data. This index, referring to the period January-May 2020, represents the mortality rate due to COVID-19, standardized by the demographic characteristics of the resident population in each province (values expressed per 100,000 inhabitants). As apparent from, the administrative mortality rate at the provincial level is associated with a small and not statistically significant difference in any of our outcomes of interest among both physicians and nurses exposed to different intensities of this measure (i.e., high/low mortality rate). These unexpected results can be interpreted as a first signal that the spread of the pandemic per se might not be significantly correlated with the level of commitment of healthcare workers. As measures of personal COVID-19 experiences in the workplace, we asked respondents to judge the promptness and effectiveness of the policy response to the COVID-19 emergency in the facility where they work. Additionally, we included a set of variables measuring the exposure of the respondents to COVID-19 infection based on their own experience and the experience of their colleagues: whether their colleagues were infected or lost their lives due to COVID-19 and whether respondents themselves were exposed to or tested positive for the disease. Finally, we obtained data on whether respondents directly worked with COVID-19 positives, were reassigned to a specialization or facility devoted to COVID-19 patients, and worked overtime due to the COVID-19 emergency. As shown in Tables, nurses are significantly younger and have a higher prevalence of female workers than physicians. Consistently, nurses are less likely to have children, be married, live in large dwellings, cohabit, and suffer from chronic diseases, while they are more likely to have changed workplace before and to not have Italian citizenship. They are also more likely to work in a hospital (especially in a non-teaching hospital) and in the private sector than physicians and to have shorter tenure. However, nurses are less likely to have managerial or coordination roles. Finally, nurses work approximately 38 hours per week compared to the 44 hours recorded among physicians, with fewer work shifts, and they are less likely to judge their unit as being of high quality than physicians. When analyzing the COVID-19-related factors, physicians had a higher chance of having colleagues who were infected/hospitalized or who died of COVID-19, and they also worked more overtime during the first wave than nurses. However, nurses are more likely to have been reassigned to a different specialization or facility and to have tested positive for COVID-19. ## Statistical analysis For healthcare worker *i* working in region *r*, we estimate the links between each outcome of interest (*Outcome*_*ir*) and the three sets of controls by applying the model in. $$\begin{array}{r} {Outcome_{ir} = \alpha Personal_{i} + \lambda Contextual_{i} + \sigma COVID_{19i} + \beta COVID_{19p} + \tau_{r} + \epsilon_{ir}} \\ \end{array}$$ This model captures the joint impact of workers’ personal characteristics (*Personal*_*i*), contextual conditions (*Contextual*_*i*), and both personal (*COVID*−19_*i*) and administrative (*COVID*−19_*p*, with *p* being the province of work) COVID-19-related factors. In addition, we control for the working region fixed effects *τ*_*r* to account for the time-invariant regional characteristics such as the organization of the regional health system and its performance in regular times, the macro characteristics of the region of work—such as employment or population characteristics—or the cultural factors that might reflect differences in daily life attitudes and behaviors. Standard errors are clustered at the level of the working region of each respondent *i*. # Results shows the share of the explanatory power of personal, contextual and COVID-19-related factors separately as obtained by summing the estimated partial *η*² for each of the related regressors. We observe a large relative importance of contextual factors with respect to personal and COVID-19-related factors in explaining the outcomes. Contextual factors explain 58% of the variation in *Satisfaction*, 42% of the variation in *Profession Change*, and 52% of the variation in *Specialization Change*. The remaining variation in *Satisfaction* is equally explained by personal (21%) and COVID-19 (21%) related, while the latter only account for a small amount of the variation in *Profession Change* (9%) and *Specialization Change* (9%). shows the regression results for the full sample (i.e., all health workers). We observe a U-shaped reduction in *Satisfaction* with respect to age, with the lowest level of *Satisfaction* appearing among middle-aged respondents (i.e., age 40–50). Workers who are married and in good health show a significantly higher level of satisfaction. Among the contextual factors, working in a high- quality facility is the most important determinant of workers’ satisfaction. Those working in a high-quality facility enjoy greater satisfaction by approximately 0.827 percentage points, which is approximately 15.6% at the mean (i.e., 5.3). Perceiving a higher salary and having a managerial/coordination role are positively correlated with *Satisfaction*. By contrast, factors reducing workers’ satisfaction are the hours of work, having an employment contract with work shifts, and working in a hospital and in a province that is perceived to have a lack of medical personnel. Regarding the impact of COVID-19-related factors, when the response to the crisis was considered to be prompt and effective, workers are overall more satisfied—by approximately 0.6 and 0.4 p.p., respectively (corresponding to a magnitude of 11% and 8% at the mean value). However, those workers who worked more overtime or were reassigned to a different specialization or function are significantly less satisfied. Column 2 of shows the coefficients for the willingness to change profession. In line with previous literature showing a strong correlation between risk aversion, age, and gender, younger workers and workers with chronic diseases show higher willingness to change profession, whereas female workers, even if overall less satisfied, are less willing to change. In addition, the lack of medical personnel in the province of work, or working more hours, increases the propensity to change profession. In contrast, working in a high-quality facility, receiving a high salary, or having managerial/coordination responsibilities are all employee retention factors. Surprisingly, the first wave of the COVID-19 pandemic did not threaten workers’ vocation. In particular, workers who had more contact with COVID-19 patients have a significantly lower willingness to change profession (by 0.03 p.p.). Where there was a more effective response to the emergency, workers are more willing to keep their job in healthcare, while the opposite is true only when workers were reassigned to a different specialization or function due to COVID-19. We obtain very similar results when considering *Change Specialization* (Column 3 of). Workers employed in COVID-19-related wards or who had been in contact with COVID-19 patients are less willing to change specialization, further signaling that the first wave of the pandemic did not affect their professional vocation. The disjoint results for physicians (Columns 4, 5, and 6 of) and nurses (Columns 7, 8, and 9 of) are consistent with the full sample estimates. Regarding the U-shaped effect driven by age, physicians report the lowest peak in the age group 50–60, while nurses record it among young to middle-aged professionals (i.e., 30–40). The lack of medical personnel (i.e., a lack of human resources or low number of coworkers) negatively impacts the level of satisfaction of physicians, which decreases by 8.8%, but not that of nurses. Managerial or coordination responsibilities also seem to be more important for physicians’ satisfaction only. However, both types of professionals are less satisfied with longer working hours and work shifts, while being assigned to a different specialization/function during the first wave of the pandemic decreases satisfaction only for nurses. Regardless of the type of profession, both satisfaction and the willingness to change profession or specialization are driven by the perceived quality of the employing facility. Working in a perceived high-quality facility increases the level of satisfaction by +14.9% and +17.3% for physicians and nurses, respectively, while it decreases the willingness to move to another profession (specialization) by -22.6% (-22.3%) for physicians and by -24.3% (-31.1%) for nurses. ## Robustness checks We check the robustness of our results along several dimensions. First, to better understand the impact of the COVID-19-specific factors on our outcomes of interest, we include separately the administrative variable on the provincial COVID-19 mortality rate and the *COVID-19 factors* derived from the survey data. This robustness check clarifies which type of phenomenon better captures the effects of the COVID-19 pandemic—either the recorded mortality in the province of work or the personal experience of the workers. As is apparent from, the inclusion of the COVID-19 mortality rate does not affect the significance of the survey variables describing the workers’ personal experience with the pandemic. This suggests that the administrative index and survey data provide different and complementary information. When only the COVID-19 mortality rate is included, we observe a positive effect on *Satisfaction*. As shown in, in those provinces hit more intensively by the first wave, the level of job satisfaction is higher. This result might be driven by the fact that in the northern provinces, which were more heavily impacted, there is typically a higher level of satisfaction. Moreover, the positive effect of the COVID-19 mortality rate might also reflect both the resilience and the fulfillment that workers experienced during the first wave thanks to emotional support received by the general public. The media often referred to healthcare workers as “heroes,” and many public figures, such as Pope Francis, openly thanked them for their heroic services and praised their dedication, while individuals undertook many private initiatives to express their gratitude (e.g., from individual messages to private donations). Overall, this unexpected public reaction might have given further meaning to the hardships of the exhausting work experienced by the healthcare workforce during the first wave of the pandemic. Second, we verify the stability of the estimates of the baseline specification by including administrative information at the provincial level to capture the objective quality of the healthcare system within which the respondents operate. There are four measures of objective quality. The first measure proxies for workforce availability and coincides with the rate of physicians registered with the provincial board of physicians per 10,000 inhabitants (a correlation of -0.063 with the perceived lack of personnel). The other three measures are the 30-day readmission rate for acute myocardial infarction (AMI), the 30-day readmission rate for stroke and the 30-day readmission rate for chronic obstructive pulmonary disease (COPD) in 2019 as measured by the Ministry of Health in the “National Healthcare Outcomes Program” (correlations of -0.042, -0.091, and -0.073 with perceived quality, respectively). These 30-day readmission rates are computed as the ratio between the number of readmissions for the related disease within 30 days of discharge out of the total number of admissions due to the given disease (e.g., the number of readmissions due to stroke out of the overall admissions due to stroke). Note that the inclusion of administrative proxies for objective quality does not affect the estimates of the self-perceived quality measures, highlighting the relevance of self- perception over objective factors. Third, we consider alternative definitions of *Satisfaction*. Specifically, we work with its discrete version, *Satisfaction 2*, which ranges from 8 to 40, being the sum of the 8 categorical variables related to *Profession*, *Job*, *Salary*, *Work-life balance*, *Relationship with the colleagues*, *Relationships with the administration*, *Work hours*, and *Career*. Each of these variables is measured on a 5-item Likert scale ranging from 1 (very dissatisfied) to 5 (very satisfied). Alternatively, *Satisfaction 3* was computed as the arithmetic mean of the same 8 categorical variables on which we also performed a principal component analysis (PCA) obtaining *Satisfaction PCA*; that is, a continuous outcome varying from -5.42 to 4.69. As reported in, the coding of *Satisfaction* into a binary outcome does not drive our results, as we find no significant difference in the explanatory power of each control group (i.e., personal, contextual and COVID-19-related factors). # Discussion Our results show how contextual factors explain a remarkable amount of the variation in the level of satisfaction (58%), the willingness to change profession (43%) and the willingness to change specialization (52%), while personal factors, however important, matter to a lesser extent (21% of satisfaction, 49% of changing profession, and 39% of changing specialization). In particular, working for a (perceived) high-quality facility and in a province where the worker perceives a lack of medical personnel are the two main components of job satisfaction and of the willingness to change profession or specialization. These findings hold for different types of professionals (e.g., nurses and physicians) and are robust to several checks. Our findings reinforce previous evidence, particularly on the relevance of contextual factors related to the workplace in determining job satisfaction. Numerous studies highlight that, in non-emergency times, healthcare workers’ satisfaction is significantly and negatively associated with workload and working shifts and access to resources but positively associated with economic incentives, quality and having a managerial role and coordination responsibilities. Similarly, these factors also matter during the pandemic. find job satisfaction among Jordanian physicians to be positively associated with age and salaries and negatively associated with working as a general practitioner, as a specialist or in high-load hospitals. The study by shows that the number of office days is an important determinant of job satisfaction and that turnover intentions vary with age among healthcare workers in Bolivia. Similarly, age and workload appear to be the main drivers of job satisfaction and of the willingness to leave the profession among Egyptian nurses; Spanish nurses’ job satisfaction is primarily affected by their workload, access to resources and information. However, in contrast to what is commonly expected, the intensity of the spread of the pandemic did not significantly affect workers’ satisfaction or undermine their vocation, either when it is captured by the administrative measure (i.e., the death rate at the provincial level), or by their personal experience with the virus (e.g., being infected). Rather, job satisfaction is significantly reduced by working overtime and by dealing with infected patients; that is, by changes in the working conditions caused by the health emergency. Moreover, when we compare the determinants of the propensity to change profession or specialization, nurses’ resilience stands out. In particular, nurses who worked in a COVID-19 related ward are significantly less willing to change specialization, and nurses who had direct contact with COVID-19 patients are less willing to change both profession and specialization. However, if nurses had infected colleagues or had been reassigned to a different ward or function, they are more likely to consider a change. Among physicians, none of the COVID-19-related factors influenced their vocation in terms of either profession or specialization. The intensity of the outbreak of the pandemic mainly affected nurses, although the effects are quite small in magnitude with respect to the other controls. In our sample, approximately 50% of physician respondents and 79% of nurse respondents are female, in line with female participation in the public healthcare sector (i.e., 48% of physicians and 78% of nurses). Regarding age, our sample is also consistent with the national trend of physicians being, on average, older than nurses (49 vs. 40). Hence, the present study further contributes to the literature by increasing the representativeness of the recruited sample with respect to the general population of healthcare workers. Overall, our results have important policy implications. On the one hand, they highlight the commitment of healthcare workers whose vocation is not challenged by their own or their colleagues’ struggle with the virus. On the other hand, they indicate that factors that can be affected by policy interventions, such as those related to the workplace, are the main drivers of job satisfaction. Health emergencies, such as the COVID-19 outbreak, undermine workers’ commitment, not necessarily out of fear of contracting the virus but by worsening the working environment. Indeed, job satisfaction and commitment are preserved mainly by guaranteeing good conditions in the working environment, both in normal times and during an emergency. In particular, it appears crucial to foster the quality of facilities and to reduce shortages of medical personnel, since these interventions would improve the provision of care to patients while simultaneously supporting healthcare workforce satisfaction. For a one-standard- deviation increase in healthcare facility quality, the level of job satisfaction increases by 0.40 p.p., which is equal to 7% of the mean value; however, a one- standard-deviation increase in the perceived lack of medical personnel decreases job satisfaction by 0.06 p.p., which is equal to 1.1% of the mean. ## Limitations This study has several limitations. First, our analysis uses a cross-sectional approach rather than a longitudinal perspective. While a panel study would have been ideal to evaluate the change in the level of satisfaction with respect to the period before the pandemic, our approach provides a valid estimation of how job satisfaction varies with both personal and contextual elements and with COVID-19-related factors. In particular, since most of the personal and contextual factors do not vary over time, especially in a short time period (e.g., one year), one can reasonably expect our analysis of the impact of these factors on job satisfaction to be well identified. This is also supported by the fact that our results are well in line with the findings reported in the literature of interest. Moreover, participants were asked to describe the working conditions considering solely the period from the end of February 2020 to the beginning of June 2020. Therefore, the inclusion of the COVID-19-related factors in the regressions allows us to capture the relationship between the working conditions during the emergency situation (e.g., working overtime) and the level of job satisfaction, in addition to the standard working conditions (e.g., average hours worked, or contract with work shifts). However, future studies on this topic seeking to account for the effects of time-varying factors should use a longitudinal approach instead. Second, participants voluntarily participated in our survey; thus, they may be self-selected, and reporting bias cannot be excluded. For instance, if individuals with low job satisfaction were more likely to respond (e.g., to express dissatisfaction), then the estimated satisfaction level may be lower than the worker population average. By contrast, if individuals with below- average satisfaction were less likely to respond (e.g., due to distrust in the system), then the estimated satisfaction level may be higher than the worker population average. Reassuringly, however, our sample is representative of the underlying worker population by gender and age. Third, it is not possible to compute an exact response rate since our survey circulated also through the provincial boards of physicians and nurses, hospital websites, and representative associations. However, this decentralized approach allowed us to target all Italian healthcare workers rather those working in specific hospitals or geographical areas within a country, as most previous studies have done. # Conclusion Immediately after the first wave of the COVID-19 pandemic, we conducted a unique survey of Italian healthcare workers to explore the determinants of their professional satisfaction and vocation focusing on personal, contextual and COVID-19-related factors. In addition to confirming the role of gender, age, good health and chronic diseases among the personal factors, the analysis shows that contextual factors are the strongest determinants of workers’ satisfaction and propensity to change profession or medical specialization. In particular, we find that *working in a high-quality facility* has beneficial effects on workers, increasing work- related satisfaction and willingness to remain in the profession and in the medical specialization. However, *working in a province with a serious shortage of medical personnel* yields the opposite result. Our findings have strong policy implications because the main drivers of professional satisfaction are modifiable. Hence, policymakers should implement effective strategies to improve working conditions in the healthcare sector in general and further support workers along these dimensions in emergency times. This would directly impact the turnover of healthcare workers while indirectly increasing the quality of care for patients. Although our analysis does not offer a one-size-fits-all policy to improve working conditions in the healthcare sector, in the specific case of Italy, policymakers should foster quality of facilities and invest in increasing the number of medical personnel. In examining the intensity of COVID-19 exposure, we find that work accidents, such as being infected or losing colleagues to the virus, do not play a relevant role in affecting the vocation of healthcare workers. Rather, we find they are more affected by changes in working conditions caused by the pandemic, such as having to work overtime or being reassigned to a different ward/function. Healthcare professionals are devoted to helping others, and supporting them through difficult times such that a severe pandemic in the province of work plays a marginal role; more important, it contradicts the common expectation. Indeed, healthcare workers and, especially, nurses are even more satisfied with their job and less prone to change profession or specialization in the most affected provinces following the first wave of the pandemic, further showing the resilience of their vocation. # Supporting information We are grateful for the support of the physicians and nurses associations that promoted the dissemination of our survey, and we thank all the physicians, nurses, biologists, psychologists, obstetricians, and technicians who took the time to complete it. 10.1371/journal.pone.0275334.r001 Decision Letter 0 Taghizadeh-Hesary Farzad Academic Editor 2022 Farzad Taghizadeh-Hesary This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 31 Jan 2022 PONE-D-22-00138Job Satisfaction Among Healthcare Workers in the Aftermath of the COVID-19 Pandemic1PLOS ONE Dear Dr. Grembi, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it requires modifications to meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.  Prior to further steps, it is necessary to organize your manuscript per the PLOS ONE's guidelines (<https://journals.plos.org/plosone/s/submission- guidelines>): Please submit your revised manuscript by Mar 17 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Farzad Taghizadeh-Hesary Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at  <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and  <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Please provide additional details regarding participant consent. In the ethics statement in the Methods and online submission information, please ensure that you have specified (1) whether consent was informed and (2) what type you obtained (for instance, written or verbal, and if verbal, how it was documented and witnessed). If your study included minors, state whether you obtained consent from parents or guardians. If the need for consent was waived by the ethics committee, please include this information. If you are reporting a retrospective study of medical records or archived samples, please ensure that you have discussed whether all data were fully anonymized before you accessed them and/or whether the IRB or ethics committee waived the requirement for informed consent. If patients provided informed written consent to have data from their medical records used in research, please include this information. 3\. During the internal evaluation of the submission we have noted the following statement: "f this research was approved by the Ethics committee of the University of Pavia (Italy) to which Emilia Barili affiliated at the time of the survey" Please could you clarify whether author Emilia Barili is affiliated with the research ethics committee which provided ethical approval. If no, please could you revise this text to avoid confusion. 4\. Please update your submission to use the PLOS LaTeX template. The template and more information on our requirements for LaTeX submissions can be found at <http://journals.plos.org/plosone/s/latex>. 5\. Thank you for stating the following financial disclosure:  "Paola Bertoli is the recipient of a Rita Levi Montalcini Fellowship to promote the moving back in Italian University of Young Italian Scholars based abroad and willing to come back to Italy" Please state what role the funders took in the study.  If the funders had no role, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript."  If this statement is not correct you must amend it as needed.  Please include this amended Role of Funder statement in your cover letter; we will change the online submission form on your behalf. 6\. Please include your full ethics statement in the ‘Methods’ section of your manuscript file. In your statement, please include the full name of the IRB or ethics committee who approved or waived your study, as well as whether or not you obtained informed written or verbal consent. If consent was waived for your study, please include this information in your statement as well.  7\. We note that Figures 1 and 3 in your submission contain copyrighted images. All PLOS content is published under the Creative Commons Attribution License (CC BY 4.0), which means that the manuscript, images, and Supporting Information files will be freely available online, and any third party is permitted to access, download, copy, distribute, and use these materials in any way, even commercially, with proper attribution. For more information, see our copyright guidelines: <http://journals.plos.org/plosone/s/licenses-and-copyright>. We require you to either (1) present written permission from the copyright holder to publish these figures specifically under the CC BY 4.0 license, or (2) remove the figures from your submission: a\. You may seek permission from the original copyright holder of Figures 1 and 3 to publish the content specifically under the CC BY 4.0 license.  We recommend that you contact the original copyright holder with the Content Permission Form (<http://journals.plos.org/plosone/s/file?id=7c09/content- permission-form.pdf>) and the following text: “I request permission for the open-access journal PLOS ONE to publish XXX under the Creative Commons Attribution License (CCAL) CC BY 4.0 (<http://creativecommons.org/licenses/by/4.0/>). Please be aware that this license allows unrestricted use and distribution, even commercially, by third parties. Please reply and provide explicit written permission to publish XXX under a CC BY license and complete the attached form.” Please upload the completed Content Permission Form or other proof of granted permissions as an "Other" file with your submission.  In the figure caption of the copyrighted figure, please include the following text: “Reprinted from \[ref\] under a CC BY license, with permission from \[name of publisher\], original copyright \[original copyright year\].” b\. If you are unable to obtain permission from the original copyright holder to publish these figures under the CC BY 4.0 license or if the copyright holder’s requirements are incompatible with the CC BY 4.0 license, please either i) remove the figure or ii) supply a replacement figure that complies with the CC BY 4.0 license. Please check copyright information on all replacement figures and update the figure caption with source information. If applicable, please specify in the figure caption text when a figure is similar but not identical to the original image and is therefore for illustrative purposes only. Additional Editor Comments: Thank you for submitting the results of your study to the PLOS ONE. Prior to further steps, it is necessary to organize your manuscript per the PLOS ONE's guidelines (<https://journals.plos.org/plosone/s/submission-guidelines>): Manuscripts should be organized as follows: Beginning section The following elements are required, in order: Title page: List title, authors, and affiliations as first page of the manuscript Abstract Introduction Middle section The following elements can be renamed as needed and presented in any order: Materials and Methods Results Discussion Conclusions (optional) Ending section The following elements are required, in order: Acknowledgments References Supporting information captions (if applicable) Other elements Figure captions are inserted immediately after the first paragraph in which the figure is cited. Figure files are uploaded separately. Tables are inserted immediately after the first paragraph in which they are cited. Supporting information files are uploaded separately. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0275334.r002 Author response to Decision Letter 0 16 May 2022 Dear Prof. Taghizadeh-Hesary, Thank you for your letter of January 31st. We reviewed the paper as to make sure it complies with PLOS ONE's style requirements and we adopted the PLOS ONE’s Latex template. Moreover, we now clearly specify in the text that each participant provided written informed consent that was embedded in the first page of the questionnaire and we provide the specific details on the Ethics committee that approved our survey (i.e., the Ethics committee of the University of Verona, called Comitato di Approvazione per la Ricerca sull'Uomo''). We also confirm that the reported funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. In what follows, we answer to each of the specific points raised in your letter detailing how these have been taken into account in the revised version. We hope that this revised version will bring us nearer to a publication in PLON ONE and we again thank you for this opportunity and look forward to hearing from you at your convenience. Best regards, Emilia Barili, Paola Bertoli, Veronica Grembi and Veronica Rattini COMMENTS AND RELATED REPLIES When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> A: The manuscript has been reviewed to ensure compliance with PLOS ONE’s style requirements. Specifically, we adapt the Title formatting and insert the correct authors’ information. We eliminate the Section and Subsection numbering. The main sections are now Introduction, Materials and Methods, Results and Discussion, Conclusion. We took care of Tables and Figures formatting the related captions, as well as of footnotes. 2\. Please provide additional details regarding participant consent. In the ethics statement in the Methods and online submission information, please ensure that you have specified (1) whether consent was informed and (2) what type you obtained (for instance, written or verbal, and if verbal, how it was documented and witnessed). If your study included minors, state whether you obtained consent from parents or guardians. If the need for consent was waived by the ethics committee, please include this information. If you are reporting a retrospective study of medical records or archived samples, please ensure that you have discussed whether all data were fully anonymized before you accessed them and/or whether the IRB or ethics committee waived the requirement for informed consent. If patients provided informed written consent to have data from their medical records used in research, please include this information. A: We obtained written informed consent from each participant at the beginning of the questionnaire. Specifically, the first screen provided participants with information about the content of the survey, its intended use and a privacy statement. After this information, each participant was asked whether she agreed to provide her consent for participation. If participants clicked on the option “Yes, I consent to participate in the survey”, they proceeded with the questionnaire. If they opted for “No, I do not consent to participate in the survey”, they were re-directed to an ending screen. Our study does not include minors and does not use medical records or achieved samples. 3\. During the internal evaluation of the submission we have noted the following statement: "this research was approved by the Ethics committee of the University of Pavia (Italy) to which Emilia Barili affiliated at the time of the survey" Please could you clarify whether author Emilia Barili is affiliated with the research ethics committee which provided ethical approval. If no, please could you revise this text to avoid confusion. A: Emilia Barili was affiliated with the University of Pavia at the time of the survey, but she was not directly affiliated or connected to the Ethics committee of the University of Pavia 4\. Please update your submission to use the PLOS LaTeX template. The template and more information on our requirements for LaTeX submissions can be found at <http://journals.plos.org/plosone/s/latex>. A: We adopted the PLOS LaTex template as requested. 5\. Thank you for stating the following financial disclosure: "Paola Bertoli is the recipient of a Rita Levi Montalcini Fellowship to promote the moving back in Italian University of Young Italian Scholars based abroad and willing to come back to Italy" Please state what role the funders took in the study. If the funders had no role, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." If this statement is not correct you must amend it as needed. Please include this amended Role of Funder statement in your cover letter; we will change the online submission form on your behalf. A: The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. 6\. Please include your full ethics statement in the ‘Methods’ section of your manuscript file. In your statement, please include the full name of the IRB or ethics committee who approved or waived your study, as well as whether or not you obtained informed written or verbal consent. If consent was waived for your study, please include this information in your statement as well. A: We now specify in the “Materials and Methods” section what follows: “The survey was approved by the \`\`Comitato di Approvazione per la Ricerca sull'Uomo'', that is, the Ethics Committee of the University of Verona (Prot. N. 0221872 - 22/06/2020) and registered at AEA-RCT registry (AEARCTR-0007419), while the University of Pavia certified compliance with privacy requirements (Prot. N. 61080 - 15/06/2020). All participants gave their written informed consent that was embedded on the first page of the questionnaire.” 7\. We note that Figures 1 and 3 in your submission contain copyrighted images. All PLOS content is published under the Creative Commons Attribution License (CC BY 4.0), which means that the manuscript, images, and Supporting Information files will be freely available online, and any third party is permitted to access, download, copy, distribute, and use these materials in any way, even commercially, with proper attribution. For more information, see our copyright guidelines: <http://journals.plos.org/plosone/s/licenses-and-copyright>. We require you to either (1) present written permission from the copyright holder to publish these figures specifically under the CC BY 4.0 license, or (2) remove the figures from your submission: a\. You may seek permission from the original copyright holder of Figures 1 and 3 to publish the content specifically under the CC BY 4.0 license. We recommend that you contact the original copyright holder with the Content Permission Form (<http://journals.plos.org/plosone/s/file?id=7c09/content- permission-form.pdf>) and the following text: “I request permission for the open-access journal PLOS ONE to publish XXX under the Creative Commons Attribution License (CCAL) CC BY 4.0 (<http://creativecommons.org/licenses/by/4.0/>). Please be aware that this license allows unrestricted use and distribution, even commercially, by third parties. Please reply and provide explicit written permission to publish XXX under a CC BY license and complete the attached form.” Please upload the completed Content Permission Form or other proof of granted permissions as an "Other" file with your submission. In the figure caption of the copyrighted figure, please include the following text: “Reprinted from \[ref\] under a CC BY license, with permission from \[name of publisher\], original copyright \[original copyright year\].” b\. If you are unable to obtain permission from the original copyright holder to publish these figures under the CC BY 4.0 license or if the copyright holder’s requirements are incompatible with the CC BY 4.0 license, please either i) remove the figure or ii) supply a replacement figure that complies with the CC BY 4.0 license. Please check copyright information on all replacement figures and update the figure caption with source information. If applicable, please specify in the figure caption text when a figure is similar but not identical to the original image and is therefore for illustrative purposes only. A: Figures 1 and 3 do not contain copyrighted images, but rather images we produced. We now specify this in the footnotes to the figures. Additional Editor Comments: Thank you for submitting the results of your study to the PLOS ONE. Prior to further steps, it is necessary to organize your manuscript per the PLOS ONE's guidelines (<https://journals.plos.org/plosone/s/submission-guidelines>): Manuscripts should be organized as follows: Beginning section The following elements are required, in order: Title page: List title, authors, and affiliations as first page of the manuscript Abstract Introduction Middle section The following elements can be renamed as needed and presented in any order: Materials and Methods Results Discussion Conclusions (optional) Ending section The following elements are required, in order: Acknowledgments References Supporting information captions (if applicable) Other elements Figure captions are inserted immediately after the first paragraph in which the figure is cited. Figure files are uploaded separately. Tables are inserted immediately after the first paragraph in which they are cited. Supporting information files are uploaded separately. A: We re-organized the structure of the paper as requested. In particular, the paper now includes Abstract, Introduction, a section providing information on the institutional setting, Materials and Methods, Results and Discussion, Conclusions. We also included the applicable required elements as, among others, acknowledgements and references. 10.1371/journal.pone.0275334.r003 Decision Letter 1 Taghizadeh-Hesary Farzad Academic Editor 2022 Farzad Taghizadeh-Hesary This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 20 Jun 2022 PONE-D-22-00138R1Job Satisfaction Among Healthcare Workers in the Aftermath of the COVID-19 Pandemic1PLOS ONE Dear Dr. Grembi, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. ================================================================= ACADEMIC EDITOR: Please find the reviewers' comments to improve the manuscript. ================================================================= Please submit your revised manuscript by Aug 04 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Farzad Taghizadeh-Hesary Academic Editor PLOS ONE \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#2: All comments have been addressed Reviewer \#3: (No Response) Reviewer \#4: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Partly Reviewer \#4: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes Reviewer \#4: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes Reviewer \#4: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: No Reviewer \#4: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Manuscript presents a current and important theme in the elaboration of human resources policies in the health area, the COVID-19 pandemic was a learning experience, including the need for rapid activation of emergency programs in public health. It presents important and significant results. I add some considerations for the authors' evaluation. 1- How were the other health professionals (571) included in the analysis? Or just in the methodology of the study? 2- The authors mention the nurses' resilience and that they have less intention of changing profession or specialization. Resilience is due only to personal factors or also to the employability of nurses compared to doctors, reducing the intention to change profession or specialization. Reviewer \#2: #P 2 Line 37: after \[18\] There must be some word, it seems to be missing. \#P2 Line 46-53 : These statements should not be included in the introduction part with the discussion, theses should be moved to the discussion part. As well as line 64 onwards till the end of the introduction part. Introduction part should be written in past tense as well the methodology part. \#P5 Research design should the written before the heading of "materials and methods" Reviewer \#3: Dear author, this is an interesting paper about the job satisfaction of healtcare workers after the onset of Covid-19 pandemic. However, the study design doesn't allow to evaluate it's effective impact on those workers, because no data are available about the status of the sample before the pandemic. A longitudinal study could be the best choice to perform the evaluation you intented to do. In addition, to make a stronger evaluation of job satisfaction, you could use a validated tool like The Job Satisfaction Survey (JSS) or others. Other issues with this manuscript are the follow: \- Many studies performed in italy in the same period have not been cited, even if they could help to better understand the context in which the study has been performed (i.e. De Sio et al). \- English language is quite good even if there are many long sentences and others are quite complex to be understood. \- No response rate has been provided. \- The introduction section is too long and, at the end, it deals with results and conclusions that should be put at the end of the paper, not at the beginning. \- The limits of the research have not been disclosed (selection bias etc) \- The authors state that they focused on the respondents from northern Italy, but they didn't indicate how many respondents came from there and how many from other parts of the country \- Many question of the survey have not been justified in relation to the aims of the study (i.e. homes dimensions etc). I think that the paper needs those and more others corrections to be accepted for publication. Reviewer \#4: 1. Introduction: "with more than 7,000 respondents (about 2,500 physicians and 4,500 nurses) conducted". Please designate the exact numbers of participants. 2\. Introduction section is too long. It is recommended the authors summarize the Introduction section (in 1000-1500 words) representing the importance of the study and how the study can fill the literature gap. Many paragraphs-describing the similar articles in other countries- can be discussed in the Discussion section. 3\. Please follow the guidelines of scientific writing. 4\. Please cite the following Covid-19 related article in the Introduction section: \- <https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8184167/> 5\. Introduction: It may not be necessary to describe the first Covid-19 wave in Italy in details. It is suggested the authors summarize this section in a paragraph. 6\. Figs. 1 and 3 can be submitted as supplementary materials. 7\. Please follow the Journal guideline to prepare the manuscript. For example, the Results and Discussion sections must be presented separately. It is recommended the author refer to the PLOS ONE guideline for authors. 8\. Please add a table summarizing the basic characteristics of the participants. 9\. Many figures have been submitted. It is recommended the authors merge Figs. 4-12 to three figures. Figure 13 can be explained in the text and be omitted. Figure 2 can be submitted as a supplementary material. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: **Yes: **Maria Carlota Borba Brum Reviewer \#2: No Reviewer \#3: No Reviewer \#4: No \*\*\*\*\*\*\*\*\*\* \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0275334.r004 Author response to Decision Letter 1 3 Aug 2022 We have attached the file 10.1371/journal.pone.0275334.r005 Decision Letter 2 Taghizadeh-Hesary Farzad Academic Editor 2022 Farzad Taghizadeh-Hesary This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 14 Sep 2022 Job Satisfaction Among Healthcare Workers in the Aftermath of the COVID-19 Pandemic1 PONE-D-22-00138R2 Dear Dr. Grembi, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Farzad Taghizadeh-Hesary Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#3: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#3: All my questions have been addressed. I think that the paper has improved and it is now suitable for puplication. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#3: **Yes: **GIUSEPPE BUOMPRISCO \*\*\*\*\*\*\*\*\*\* 10.1371/journal.pone.0275334.r006 Acceptance letter Taghizadeh-Hesary Farzad Academic Editor 2022 Farzad Taghizadeh-Hesary This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 28 Sep 2022 PONE-D-22-00138R2 Job satisfaction among healthcare workers in the aftermath of the COVID-19 pandemic Dear Dr. Grembi: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Farzad Taghizadeh-Hesary Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction The shells of pearl oysters comprise two layers, termed the prismatic and nacreous layers, formed by secreting organic materials from epithelial cells in the mantle. The prismatic layer is thought to be formed by the secretion of materials from the edge region of the mantle (the mantle edge) and the nacreous layer is formed from the inner part of the mantle (the pallium). The main components of shells are calcium carbonate and organic substances, such as chitin and various proteins. Crystal structures are different between the two layers: calcite is present in the prismatic layer and aragonite is found in the nacreous layer. The accumulation of calcium carbonate as calcite and aragonite crystals was thought to be regulated by proteins secreted from the mantle. Nacrein, the first protein to be isolated from the nacreous layer, exists in both the nacreous and prismatic layers and, thus, is thought to be involved in the formation of the whole shell. MSI60, a matrix protein in the nacreous layer, has several characteristic domains that are thought to be involved in the formation of this layer. Pif, an acidic protein isolated from the nacreous layer, has been reported to regulate the formation of this layer. Although other proteins have been isolated as matrix proteins from the prismatic and nacreous layers, it is not clear how these two layers are formed in molluscan shells. In pearl culture, nuclei are inserted into the gonad gland in mother oysters together with graft tissues of the outer epithelial cells in the mantle from other individuals. The graft tissues then spread over the nuclei to develop a pearl sac that produces the nacreous layer on the nuclei. The pearl sac comprises specialized tissues that form the nacreous layer and produce pearls. To elucidate the molecular mechanisms involved in the formation of the prismatic and nacreous layers, we analyzed expressed sequence tag (EST) sets from the mantle edge, pallium and pearl sac of the pearl oyster *Pinctada fucata*. As a result, we identified more than 70 novel candidate genes that might be involved in the formation of the two layers. RNA interference (RNAi) suppresses the expression of specific genes. This technique has been widely used to investigate functions of uncharacterized genes and has been effective in bivalves. For example, the injection of double stranded RNA (dsRNA) of *Pif* into the adductor muscle resulted in an abnormal appearance of the nacreous layer, as revealed by scanning electron microscopy (SEM). Furthermore, RNAi was useful for screening other candidate genes possibly involved in the formation of the nacreous layer. The objective of the present study was to examine whether or not the candidate genes from the EST analysis would function in the formation of the nacreous layer. First, we determined the full-length cDNA sequences of the genes and then examined their functions in during shell formation by expression and RNAi knockdown analyses. # Materials and Methods ## Novel Gene Candidates for Nacre Formation We screened gene candidates involved in the formation of nacreous and prismatic layers from the EST database of the pearl oyster *P. fucata* as follows. *De novo* assembly using MIRA assembler ver. 2.9.45×1 and the basic local alignment search tool (BLAST) Clust program from NCBI produced 29,682 unique gene sequences that were automatically numbered by the BLAST Clust program as 000001-029682, as described previously. Gene expression was represented as their appearance frequency, which corresponds to the total reads of a given gene against the total reads in the tissue. Among the 29,682 gene sequences, we selected genes with over 200 reads and compared their appearance frequency among shell formation-related tissues, the pearl sac, mantle pallium and mantle edge. As a result, five genes (000081, 000098, 000113, 000118 and 000133) were found to be expressed twice as highly in the mantle pallium as in the mantle edge and pearl sac. Three (000027, 000096 and 000411) were highly expressed in the mantle pallium and pearl sac, whereas six genes (000031, 000058, 000066, 000072, 000104, 000145, 000194 and 000200) were highly expressed in the mantle edge with almost no expression in the pearl sac. BLAST searching identified genes 000113 and 000411 to be *nacrein* and *MSI60*, respectively. ## Molecular Characterization of Candidate Genes We carried out molecular cloning of the candidate genes and predicted the molecular characteristics of their encoded proteins. 5′-rapid amplification of cDNA ends (RACE) was conducted to determine the full-length nucleotide sequences of the genes using total RNA purified from the mantle of live specimens of the pearl oyster harvested in the Mikimoto Pearl Farm, Mie Prefecture, Japan. Gene specific primers for RACE were designed based on the partial nucleotide sequences of the genes obtained from the EST analyses. cDNA templates in 5′-RACE were synthesized using GeneRacer™ kit (Invitrogen, Carlsbad, CA, USA). 5′- RACE products were subcloned into the pGEM-T Easy vector (Promega, Madison, WI, USA) and sequenced with an ABI3100 Genetic Analyzer (Applied Biosystems, Foster City, CA, USA). The sequences obtained were subjected to BLAST analysis with the option of BLASTX or BLASTP, with an expected value of 1E-10, against all sequences registered in the National Center for Biotechnology Information (NCBI) database. The amino acid sequences deduced from the nucleotide sequences were subjected to analyses of their molecular characteristics, using Pfam to predict the motif structure, SignalP v4.0 to predict the presence of signal peptides and their cleavage sites, ProtParam to compute the molecular weight and theoretical p*I*, TMHMM to predict the region of transmembrane helices in proteins, NetOGlyc to predict mucin type GalNAc *O*-glycosylation sites and NetNGlyc to predict *N*-glycosylation sites. All sequences determined in this study were registered in the DDBJ/EMBL/GenBank databases with the accession numbers AB734766- AB734776. ## Tissue Distribution Analysis Adult specimens of the pearl oyster were collected from the Mikimoto Pearl Farm, Mie Prefecture, Japan. Foot muscle, gill, gonad, adductor muscle, mantle and pearl sac tissues were dissected and preserved in RNA*later* (Applied Biosystems). Total RNA was extracted from the above-mentioned tissues with an Isogen kit (Nippon Gene, Tokyo, Japan). Reverse transcription (RT)-PCR was performed with gene specific primers based on the full-length or partial cDNA sequences. *Pif* and the elongation factor-1α gene (*EF-1α*) (DDBJ/EMBL/GenBank accession number, AB205403) were used as the positive reference and internal control, respectively. PCR was conducted as follows: denaturation at 94°C for 5 min; followed by 30 cycles at 94°C for 30 s, 55°C for 30 s and 72°C for 1 min; with a final extension step at 72°C for 7 min. PCR products were then subjected to electrophoresis through a 1% agarose gel. For *in situ* hybridization, mantle tissues containing the pallium and mantle edge were fixed in 4% (w/v) paraformaldehyde overnight, embedded in Optimal Cutting Temperature compound (Funakoshi, Tokyo, Japan), and sectioned at a thickness of 18 µm. cDNA fragments were amplified for RNA probe synthesis by RT-PCR from the mantle using gene- specific primers designed from the 3′-untranslated region (UTR) of each gene. Sense and anti-sense digoxigenin (DIG)-labeled RNA probes were generated with T7 and SP6 RNA polymerases, respectively, from the cDNA fragments, using a DIG RNA Labeling kit (Roche Applied Science, Mannheim, Germany). Hybridization was performed at 58°C. ## RNA Interference Experiments RNAi experiments were performed according to the method reported by Suzuki et al. (2009), with some modifications. cDNA was synthesized with a 3′-Full RACE Core Set (Takara, Otsu, Japan) using the total RNA from mantle tissues as a template. PCR was carried out to obtain the DNA fragments encoding the candidate genes using primers designed based on the sequences of these genes. *Pif* was used as a positive reference gene. The green fluorescence protein (GFP) gene was used as negative reference to verify the RNAi experiments. The T7 promoter sequence was added to the 5′ end of each primer to subject the PCR products to dsRNA synthesis with T7 RNA polymerase. dsRNAs were synthesized using a ScriptMAX™ Thermo T7 Transcription Kit (Toyobo, Osaka, Japan), according to the manual of the kit, using the cDNA clones encoding candidate and reference genes. About 40 µg of dsRNA/100 µl water of MilliQ (Merck Millipore, MA, USA) was injected into adductor muscles in live specimens of two-year-old pearl oysters. The injected individuals were reared in artificial seawater at 23°C for eight days, with feeding once a day. The surface on the inside of shells of the RNAi knockdown individuals was observed with a scanning electron microscope (SEM) S-4000 (Hitachi, Tokyo, Japan), focusing on the prismatic and nacreous layers, and the calcite-aragonite boundary region. The effect of RNAi knockdown was validated as follows. Total RNA was extracted from the mantle of each individual eight days after injection and first-strand cDNA was synthesized as describe above. Real-time quantitative PCR (qPCR) was employed to quantify the expression levels of the target genes, *MSI60*, *nacrein*, and gene 000096. *EF-1α* was used as an internal control, as described previously. qPCR was conducted using the ABI Prism 7300 Sequence Detection System (Applied Biosystems) with an SYBR premix Ex*Taq* II kit (Takara), according to the manufacturer's instructions. The cycling parameters consisted of one cycle of 95°C for 30 s, followed by 40 cycles of 95°C for 5 s, 55°C for 30 s and 72°C for 30 s. Dissociation curves were analyzed to determine the purity of the products and the specificity of amplification. In the control, the expression level of the group injected with MilliQ water was set as 1.0. For the differential gene expression analysis among various samples, statistical analyses were performed using one-way analysis of variance (ANOVA), followed by Tukey's test in Sigma Plot 10.0 (SYSTAT, Chicago, IL, USA). Data were represented as the mean ± standard error (*n* = 3), and the differences were considered significant at *P*\<0.05. # Results ## cDNA Cloning and Molecular Characterization The full-length cDNA sequence of gene 000096 has been determined in our previous study. The 000096 protein contained a N-terminal signal peptide and a galactose binding lectin domain. As described above, genes 000113 and 000411 were identified as known *P. fucarta* nacreous genes, *nacrein* and *MSI60*, respectively. In the present study, we determined the full-length cDNA sequences of genes 000027, 000031, 000058, 000066, 000081, 000118, 000133, 000145, 000194 and 000200, and a partial cDNA sequence of gene 000098. Among them, genes 000027, 000081, 000098 and 000118 showed significant homology (*E* value of 8.00E-71∼2.00E-22) with known gene sequences, as shown in. We tried to obtain cDNA clones of genes 000072 and 000104, without success; therefore, these genes were not subjected to the subsequent RT-PCR, in situ hybridization and RNAi experiments. It was notable that genes 000081 and 000118 showed high homology with glycine rich protein 2 and prism uncharacterized shell protein genes, respectively. The gene encoding glycine rich protein 2 was expressed in the mantle of *P. maxima*, whereas the prism uncharacterized shell protein was found in the shell of *P. margaritifera*, suggesting their direct involvement in shell formation in *Pinctada* species. Genes 000027, 000058 and 000098 showed homology with a ribosomal protein gene of *Sipunculus nudus*, a hypothetical protein gene of *Xenopus tropicalis* and a SCO-spondin like gene of *X. tropicalis*, respectively. The remaining six genes (000031, 000066, 000133, 000145, 000194 and 000200) showed no significant homology with known database sequences. The proteins encoded by genes 000031, 000066, 000194 and 000200 had predicted N-terminal signal peptides. The Pfam program predicted that the deduced amino acid sequence of gene 000058 contained a von Willebrand factor type A (VWA) domain followed by a sequence containing 18 cysteine residues. On the other hand, the translation product of gene 000194 contains a region near the C-terminus that is rich in methionine and glycine, forming the repeat sequences MGG or MGGG. The protein encoded by gene 000200 contains a region where glycine residues were continuously arranged following the signal peptide. The translation product of gene 000133 contains a characteristic region consisting of RVRRI repeats, though its function is unknown. The translation product of gene 000145 comprised 91 amino acids and a predicted molecular weight of 9.8 kDa. No characteristic structure was predicted in the molecule. ## Tissue Distribution of Transcripts RT-PCR and *in situ* hybridization were performed to examine tissue specificity and regional distribution in the mantle of transcripts encoding the target genes. *Pif* is a well known nacreous gene of *P. fucata* and was used as the positive control. RT-PCR confirmed that *Pif* was expressed in the mantle and pearl sac, and its transcripts were mainly detected in the outer epithelial cells of the pallium, supporting the results of our previous report. *Nacrein* (000113) and *MSI60* (000411) are also known nacreous genes. RT-PCR showed that *nacrein* was expressed in the gonad, muscle, mantle and pearl sac, whereas *MSI60* was expressed in the gonad, mantle and pearl sac. Nacrein mRNA was strongly expressed in the outer epithelial cells of both the pallium and the mantle edge, whereas MSI60 mRNA was detected only in the outer epithelial cells of the pallium. These data are consistent with those previously reported. Genes 000081, 000194 and 000200 were detected specifically in shell-forming tissues: the former was detected in both the mantle and pearl sac, whereas the latter two were detected only in the mantle. During RT-PCR of gene 00081, a weak band was also detected in adductor muscle; however, the size of the detected band was larger than those detected in the mantle and pearl sac. Therefore, the band was considered to be a non-specific amplification product. *In situ* hybridization showed that the transcripts of gene 000081 were specifically distributed to the outer epithelial cells of the pallium, whereas those of gene 000194 were observed in the inner epithelial cells of the outer fold in the mantle edge. On the other hand, the transcripts of gene 000200 were detected in the outer epithelial cells in both the pallium and the mantle edge. Genes 000058 and 000118 were expressed in the mantle and some other tissues. *In situ* hybridization localized their transcripts in the outer epithelial cells of the pallium. Genes 000066, 000096 and 000145 were expressed in shell-forming tissues, as well as non-shell forming tissues, such as the muscle, gonad and gill. *In situ* hybridization showed that the transcripts of genes 00096 and 000145 were distributed to the outer epithelial cells in the pallium and mantle edge, whereas the transcripts of gene 000066 were localized in the inner fold of the mantle. Genes 000027, 000031, 000098 and 000133 were expressed in all tissues analyzed in the present study, whereas *in situ* hybridization showed a variety of regional expression patterns in the mantle. The transcripts of gene 000027 were localized in the outer and inner epithelial cells of the pallium, but not of the mantle edge. The transcripts of genes 000031 and 000098 were distributed specifically to the middle fold in the mantle edge. The transcripts of gene 000133 were distributed to the inner epithelial cells of the pallium and mantle edge, and to the outer fold in the mantle edge. ## Effect on Shell Formation of RNAi Knockdown We validated the effect of RNAi knockdown on transcriptional expression by qPCR focusing on known genes and one candidate gene, 000096. As shown in, the transcripts of these genes were successfully decreased to 20–40% of the control value after dsRNA injection. Although the expression levels of all the candidate genes after the RNAi experiments were not analyzed, we considered that the RNAi knockdown was successful in all cases because all individuals injected with dsRNA had abnormal appearances, as described below. RNAi knockdown phenotypes in the shell were observed at three positions: the prismatic layer (position a), the boundary of the prismatic and nacreous layers (position b) and the nacreous layer (position c). The typical prismatic layer of the wild-type pearl oyster comprises calcite tablets in the shape of a polygon separated from each other by the interprismatic walls. The nacreous layer comprises many small aragonite crystals in the shape of hexagons that form striated patterns after they aggregate together, a typical characteristic of the nacreous layer. Injection of GFP dsRNA had no effects on the phenotype of either the prismatic or the nacreous layer. As well as expressional analyses, we used *Pif* as the control gene to verify our experiments. The RNAi knockdown of *Pif* induced an abnormal appearance of the nacreous layer similar to that reported previously. Although the prismatic layer showed calcite tablets with a normal shape, the interprismatic walls that are supposed to form the calcite tablets were slightly disordered. The boundary region between the calcite and aragonite crystals contained growing calcite tablets, the appearance of which was different from that observed after injecting GFP dsRNA. *Nacrein* and *MSI60* are known nacreous genes; however, their RNAi knockdown phenotypes have not been reported yet. *Nacrein* and *MSI60* knockdown oysters had normal patterns of calcite tablets on the prismatic layer, as in the case of the GFP treated oyster. However, the interprismatic walls in a *nacrein* knockdown individual looked spongy, as if some components were not present. Abnormal patterns appeared on the nacreous layer after RNAi knockdown of the two genes, where no aragonite crystals with rectangle shapes were observed, but with many grooves and holes. The calcite-aragonite boundary was occupied with abnormal nacreous layers (See also). As well as *Pif*, *nacrein* and *MSI60*, genes 000027, 000081, 000096 and 000118 were predominantly or specifically expressed in the pallium , suggesting roles in the formation of the nacreous layer. Actually, the effects of RNAi knockdown of genes 000081, 000096 and 000118 on the phenotype were observed only in the nacreous layer, whereas RNAi knockdown of gene 000027 affected both the prismatic and nacreous layers. The details of the phenotypes are as follows. The 000027 knockdown oyster was not able to form the normal prismatic and nacreous layers. Newly produced calcite crystals seemed to be unable to grow correctly into tablets, resulting in the failure to construct the interprismatic walls, with various shapes of calcite crystals on the prismatic layer. Furthermore, aragonite crystals disappeared, with many holes on the nacreous layer (See also). The shape of calcite tablets in the 000081 knockdown oyster appeared to be almost normal, with the interprismatic walls being slightly disordered. In contrast, the nacreous layer showed an abnormal appearance: aragonite crystals were absent and many small holes were observed. The translation product of gene 000081 was predicted to have a signal peptide, three transmembrane helices and two glycosylation sites, suggesting that the 000081 protein is a membrane protein that contributes to transportation of the substances required to form the nacreous layer. RNAi knockdown of gene 000096 had no effect on the prismatic layer, where typical calcite tablets were well arranged, whereas the appearance of the nacreous layer was completely different from the wild-type. Numerous particles with very small sizes were observed on the nacreous layer. The prismatic layer of the 000118 knockdown oyster showed a normal appearance, with calcite tablets tightly arranged with normal interprismatic walls. Growing calcite tablets were observed near the calcite-aragonite boundary, whereas the appearance of the aragonite crystals looked abnormal. No typical crystals were observed in the calcite-aragonite boundary area and striation patterns had completely disappeared. Gene 000133 was expressed twice as much in the pallium as in the mantle edge; however, *in situ* hybridization only detected its expression in the outer fold of the mantle edge. The explanation for this discrepancy between the expression profiles detected by the two methods is unknown. RNAi knockdown of gene 000133 changed the appearance of the nacreous layer completely. Although the shape of the calcite tablets looked quite normal, the interprismatic walls disappeared near the boundary region between the calcite and aragonite crystals. Very abnormal patterns, like fish scales, appeared in the middle region of the shell. These scales were easily chipped away from the surface. Genes 000031, 000066, 000145, 000194 and 000200 were predominantly or specifically expressed in the mantle edge. We predicted initially that these genes would be involved only in the formation of the prismatic layer, because of their regional expression in the mantle. However, RNAi knockdown of these genes affected the formation of both the nacreous and prismatic layers. Details of phenotypes produced by RNAi knockdown of these genes are as follows. In the 000031 and 000066 knockdown oysters, the interprismatic walls disappeared in the prismatic layer and the calcite tablets were fused to each other. The nacreous layer also looked abnormal. The 000031 knockdown individuals showed no growing aragonite crystals and the 000066 knockdown individuals showed no aragonite crystals in the supposed nacreous layer. The 000145 knockdown oysters showed apparently normal prismatic layers , whereas calcite tablets near the boundary were abnormal in shape, and several tablets were fused to each other, resulting in the disappearance of the interprismatic walls. The nacreous layer was also abnormal and showed no granules on their surface. The prismatic layer in the 000194 knockdown oyster showed an abnormal surface. The shapes of the calcite tablets were disordered and the interprismatic walls had disappeared. The appearance of the nacreous layer was also abnormal, showing a mountain ridge shape with many small particles. The prismatic layer of the 000200 knockdown oyster showed a different appearance from typical calcite crystal patterns. There was no interprismatic wall near the calcite-aragonite boundary. Numerous cracks were observed all over the surface of the shell. It was hard to distinguish the nacreous layer from the prismatic layer because of the presence of abnormal crystals with similar appearances. Although genes 000058 and 000098 showed predominant expression in the mantle edge and pallium, respectively, by next generation sequencing, *in situ* hybridization showed the opposite regional expression of these genes in the mantle. RNAi knockdown of these genes caused abnormal phenotypes in both the prismatic and nacreous layers. The prismatic layer of the 000058 knockdown oyster showed abnormal calcite tablets with their interprismatic walls. The nacreous layer was also abnormal, showing no aragonite crystals. On the other hand, RNAi knockdown of gene 000098 resulted in the loss of the interprismatic walls. In comparison with the individual injected with GFP dsRNA, it was difficult to distinguish prismatic and nacreous layers because their boundary region was obscure, with abnormal calcite and aragonite crystals. The nacreous layer looked flat with many holes. # Discussion In the present study, we identified 12 novel genes that are responsible for shell formation, using RNAi methods, RT-PCR and *in situ* hybridization. The fact that the RNAi knockdown of each gene resulted in abnormal shell formation phenotypes confirmed that our previous screening of gene candidates for shell formation based on the expression levels of genes in the EST data was successful. RT-PCR showed that certain genes were expressed in tissues other than those of the mantle and pearl sac. The transcripts of genes 000133, 000098, 000027 and 000031 were observed in the foot muscle, gill, gonad, adductor muscle and mantle. The transcripts encoded by genes *Pif*, *nacrein* (000113), 000118 and 000096 were expressed in the gonad and adductor muscle, in addition to the mantle and pearl sac. These observations imply that tissues other than the mantle and the pearl sac may also be necessary for shell formation. Alternatively, these genes may have different functions in tissues where no shells are formed. It has been reported that shell matrix proteins, such as nacrein from the pearl oyster *P. fucata*, calprismin and mucoperlin from the Mediterranean fan mussel *Pinna nobilis*, dermatopontin from the freshwater snail *Biomphalaria glabrata* and MSP-1 from the scallop *Patinopecten yessoensis*, undergo glycosylation, which is thought to play a key role in shell formation. The present study revealed five of the predicted secretory proteins (000031, 000081, 000098, 0000194 and 000200) have potential glycosylation sites, implying that they are contained in the shell matrix as glycoproteins. Initially, we speculated that genes specifically expressed in the tissues forming the nacreous layer such as the pallium and pearl sac, in our EST data were responsible for nacreous layer formation. This speculation proved correct: RNAi experiments showed that knockdown of these genes specifically expressed in the pallium and pearl sac resulted in an abnormal appearance, mainly in the nacreous layer. We similarly speculated that genes specifically expressed in tissues that form the prismatic layer, such as the mantle edge, would be responsible for the formation of the prismatic layer. Such speculation was also proven to be true by the RNAi experiments. Unexpectedly, these genes also appear to be involved in the formation of the nacreous layer. RNAi knockdown of these genes resulted in severely abnormal appearances in the nacreous layer, raising the possibility that genes responsible for the formation of the prismatic layer might also be involved in the formation of the nacreous layer. Marie et al. (2012) demonstrated that different repertoires of proteins are involved in the formation of the nacreous and prismatic layers, respectively, suggesting that the two layers are not derived from each other. They showed that the composition of shell matrix proteins in the prismatic layer, which may be secreted from the mantle edge, is clearly different from that in the nacreous layer, which may be secreted from the pallium. The abnormality of the appearances of the prismatic and nacreous layers might result from a changed ratio of gene expression caused by gene knockdown. Genes specific to mantle edge, each of which was subjected to RNAi knockdown in the present study, were not expressed in the pearl sac. If the hypothesis described above is true, the pearl sac must have no ability to form the nacreous layer, because genes 000031, 000058, 000066, 000145, 000194 and 000200, considered to be responsible for the nacreous layer formation, are not expressed there. However, in our previous report, we confirmed that the pearl oyster subjected to EST analysis, data from which were used in the present study, produced a beautiful pearl in a pearl sac, with a normal nacreous layer surrounding the nucleus. Why does the pearl sac produce the nacreous layer, even though genes involved in the nacreous layer formation are absent? One possibility is that the shell and pearl sac form the nacreous layer using different molecular mechanisms. The nacreous layer comprises calcium carbonate in the form of aragonite crystals and an organic matrix in which several components have been identified. Our previous investigation showed that gene expression patterns were different between the pallium and the pearl sac, implying that their protein compositions are not identical. This hypothesis has been partly demonstrated by SDS-PAGE of matrix proteins extracted from the nacreous layer of the shell and pearl (unpublished data). Some RNAi knockdown individuals provided a common phenotype of the nacreous layer, as shown in SEM images of. The RNAi knockdown of nacrein (000113), MSI60 (000411), gene 000027, gene 000031 and gene 000081, which all have signal peptides except for 000027, resulted in nacreous layers with similar appearances. This raised the possibility that they participate in the formation of the nacreous layer in cooperation in the same process. As shown in and, BLAST searching showed that gene 000027 is a putative ribosomal protein, which may be involved in normal cell metabolism. RT-PCR showed that it is expressed in all tissues and broadly in mantle by in situ hybridization, suggesting that the product of gene 000027 may be a housekeeping protein. Although knockdown of gene 000027 caused the abnormal appearance of the nacreous layer, gene 000027 may not be involved directly in shell-formation. The RNAi knockdown of Pif, gene 000058 and 000118 also resulted in the abnormal nacreous layer with a common phenotype. These genes are possibly involved in the formation of the layer in a different process from that mentioned above. No other common phenotype was found in the nacreous layer. On the other hand, there were no common phenotypes among the prismatic layers of the RNAi knockdown individuals. For further prediction of the function of candidate genes in this study, we compared the amino acid composition of the gene products with those of shell- formation related proteins reported to date. The protein encoded by 000081 has a glycine-rich sequence and showed clear similarity with MSI60 in the shape of star diagram. Genes 000081 and *MSI60* also showed similar expression patterns, where transcripts of the two genes were found in the pallium and pearl sac. In addition, the RNAi knockdown phenotypes of the two genes showed similar abnormality in the nacreous layer, suggesting a functional relationship in the shell formation. Proteins encoded by genes 000194 and 000200 had similarly shaped star diagram, where glycine is dominant. Their shapes resembled those of prisilkin-39 and shematrin-1. Prisilkin-39 is a matrix protein in the prismatic layer, has been reported to be expressed in the mantle edge and is considered to be involved in the prismatic layer formation. The shematrin gene was cloned from a cDNA library constructed from mRNAs of the mantle by random sequencing. Shematrin is predicted to be secreted from the mantle edge into the prismatic layer to participate in calcification. As stated in the results section, BLAST searching of genes 000194 and 000200 identified no homologous genes in the database including those encoding prisilkin-39 and shermatri-1. However, the similarity in the amino acid compositions of these proteins may reflect their similar functions in shell formation to some extent. The disordered interprismatic walls in the prismatic layer resulting from RNAi knockdown of genes 000194 and 000200 imply that their encoded proteins play important roles in the formation of the framework for calcite tablets, which is coincident with hypothetical function of prisilkin-39 and shematrin-1. Although there is no homology in protein structure, it is also notable that similar phenotypes were observed in the nacreous layer by RNAi knockdown of genes 000200 and *nacrein*. *In situ* hybridization also showed the same distribution of transcripts of the two genes in the mantle, suggesting their cooperative function in nacreous layer formation. It is reasonable to find similarity in the star diagram between nacrein and N66, which was isolated from shell matrix of *P. maxima*, because they are homologous proteins. The Gly-Xaa-Asn repeat domain of nacrein and N66 is thought to contribute to their interaction with calcium to form shells. As described above, we showed that the RNAi knockdown of nacrein produced abnormal nacreous layers. These results provide useful information to understand the function of N66 in shell formation. The proteins encoded by genes 000066 and 000118 are rich in tyrosine and cysteine, which are considered to be involved in polymerization of proteins secreted from mantle, resulting in the formation of matrix sheets in shells. The two proteins are predicted to have signal peptides, indicating that they are accumulated as shell matrix proteins after secretion from the mantle. Pearlin and N16 belong to the same protein family and are known as shell matrix proteins rich in tyrosine and cysteine. The proteins encoded by genes 000066 and 000118 might participate in the formation of protein networks with other secreted proteins, yielding insoluble protein sheets. RNAi knockdown of genes 000066 and 000118 caused disappearance of the interprismatic walls and aragonite crystals, and abnormal aragonite crystal structure, respectively. These phenotypes suggest that they encode components of aragonite crystals and/or interprismatic walls. The two proteins are basic, with p*I* values of 8.90 and 8.67, respectively, whereas those of pearlin and N16 are 6.10 and 4.68, respectively, before glycosylation. The different p*I*s are caused by low aspartic and glutamic acids contents in the two molecules. The proteins encoded by genes 000058 and 000098 are relatively large among proteins encoded by the present novel genes and both are acidic and rich in cysteine. No protein with such properties has been found in those used for comparison. The RNAi knockdown of gene 000058 showed indistinct structure of the interprismatic walls in the prismatic layer and that of gene 000098 showed reduced walls. These data indicate that proteins encoded by the two genes might be involved in constructing the framework of calcite tablets via disulfide bonds formed with other proteins. In the present study, we identified novel genes involved in shell formation, based on the alteration of the surface appearance of the nacreous and prismatic layers induced by RNAi. It should be noted that it is possible that all of these knockdowns have had absolutely no specific effect on protein levels because this has not been measured with antibodies. It is therefore possible that all of the phenotypes we observe are non-specific or general off-target effects. To eliminate those possibilities, further study is needed on the proteins encoded by the genes in shell formation; for example, tissue distribution analysis of proteins with antibodies, and *in vitro* precipitation assays of aragonite crystals in the presence of artificially expressed proteins encoded by the genes could be conducted. # Supporting Information [^1]: The authors have the following interests. Kaoru Maeyama and Kikuhiko Okamoto are employed by Mikimoto Pharmaceutical CO., LTD. Kiyohito Nagai is employed by Pearl Research Institute, Mikimoto CO., LTD. There are no patents, products in development or marketed products to declare. This does not alter their adherence to all the PLOS ONE policies on sharing data and materials, as detailed online in the guide for authors. [^2]: Conceived and designed the experiments: DF SW. Performed the experiments: DF FO S. Kinoshita HK SM AO YO. Analyzed the data: DF FO S. Kinoshita. Contributed reagents/materials/analysis tools: KN KM KO S. Kanoh SA. Wrote the paper: DF S. Kinoshita SW.

# Introduction Myelodysplastic syndromes (MDS) are a heterogeneous group of hematological disorders characterized by ineffective hematopoiesis that results in refractory cytopenia with morphological and functional abnormalities, as well as higher susceptibility to leukemia. However, the molecular basis of ineffective hematopoiesis has yet to be clarified. Peripheral blood cytopenia accompanied by hypercellularity in bone marrow (BM) has been considered to result from increased apoptosis of hematopoietic progenitors. Upregulation of tumor necrotizing factor (TNF)- **α**, a proapoptotic cytokine, has been commonly observed in BM plasma and peripheral mononuclear cells, and is positively correlated with the degree of apoptosis in early stage/low-risk MDS. TNF-**α** is a potent stimulator of nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) that is increasingly recognized as playing a crucial role in malignancy development. Moreover, TNF-**α** itself is also transcribed by NF-κB. Therefore, upregulation of TNF-**α** could play key roles in both the development of ineffective hematopoiesis and the progression of MDS. One of the components of transcription factor activator protein 1 (AP-1), c-Fos, regulates apoptosis and production of inflammatory mediators in neutrophils, and was recently shown to act as a suppressor of LPS-induced production of cytokines including TNF-**α** via physical interaction with p65 protein of NF-κB. The c-Fos transcription factor is encoded by *FOS*, an immediate early gene whose transcripts have a short life span and rapidly increase in response to a wide range of stimuli including the cellular stress that induces cessation of cap- dependent translation. The rapid increase of *FOS* mRNA is attributable to transcriptional regulation and post-transcriptional regulation, which decelerates *FOS* mRNA decay. We recently found that the elevation of c-Fos mRNA following translation arrest was attenuated because of insufficient mRNA stabilization in granulocytes isolated from MDS patients. Since deregulated expression of c-Fos could affect TNF-**α** production, the aim of the present study was to clarify the causes and consequences of the impaired stabilization of *FOS* mRNA in MDS. The life span of *FOS* mRNA is regulated via its 3’UTR, which includes target sequences for multiple microRNAs (miRNAs) and an AU-rich element (ARE) where Hu antigen R (HuR), an mRNA-stabilizing protein, and AUF1, which generally destabilizes its target mRNA. Although *FOS* mRNA stabilization following translation arrest was accompanied by the association of HuR with an ARE in *FOS* mRNA, our previous study showed that the majority of MDS patients expressed similar levels of HuR protein in granulocytes compared to the healthy controls. No involvement of AUF1 in *FOS* mRNA stabilization under translation arrest was detected, and no mutations were found in *FOS* mRNA 3’UTR. Therefore, miRNAs were likely to be the cause of the impaired stabilization of *FOS* mRNA observed in MDS. It has been shown that miRNAs play crucial roles in hematopoiesis. For example, the maturation of BM-derived dendritic cells requires silencing of c-Fos by miR-155 in both human and mice. Additionally, miRNAs are involved in the development of malignancies, and aberrant miRNA expressions have been reported in various malignant diseases, such as the elevation of miR-155 and miR-125b in acute myelogenous leukemia and the decrease of miR-34a in various solid organ cancers. In MDS, miR-378, miR-632, and miR-636 were increased in BM- derived mononuclear cells. CD34⁺ cells isolated from patients classified into subtypes with a high risk of leukemic transformation overexpressed miR-155 and miR-210. In low-risk MDS subtypes, a decrease of let-7a and miR-16 in plasma and an increase of miR-34a in CD 34⁺ cells have been reported. Thus, abnormal miRNA expression possibly affects *FOS* mRNA stabilization under translation arrest in MDS-derived neutrophils. In this study, we identified *FOS*-targeting miRNAs that were overexpressed in granulocytes from patients with early-stage MDS and their effects on expression of c-Fos protein. We also demonstrated that the reduction of c-Fos led to excessive production of TNF-**α** in response to LPS. # Materials and Methods ## Patients and granulocyte isolation Peripheral blood was obtained from 23 patients with MDS consisting, according to the 2008 WHO classification, of four cases of refractory cytopenia with unilineage dysplasia (RCUD), seventeen cases of refractory cytopenia with multilineage dysplasia (RCMD), two cases of refractory anemia with excess blasts-1 (RABE-1), and 17 age-matched healthy controls. All donors provided written informed consent in accordance with the Institutional Human Research Committee and Helsinki Declaration developed by the World Medical Association. The neutrophil fraction was obtained by centrifugation through Lymphoprep (l.077 g/mL, Axis-Shield, Oslo, Norway) followed by hypotonic lysis of erythrocytes, as previously described. The staining of fractionated cells with May-Grünwald and Giemsa solutions revealed that more than 92% of the cells were neutrophilic granulocytes. The hematological and clinical findings of patients are presented in. ## Ethics The present study, and the process of securing written informed consent from the patients and healthy controls, were approved by the Ethics Committee of Fukushima Medical University (approval number: 1077), which is guided by local policy, national laws, and the World Medical Association Declaration of Helsinki. All study participants provided their written informed consent. ## Cells A human promyelocytic leukemia cell line HL60 was purchased from Riken BRC Cell Bank (Tsukuba, Japan) and cultured in RPMI 1640 (Wako, Mie, Japan) supplemented with 10% (v/v) heat-inactivated fetal bovine serum (FBS) (Nichirei Biosciences, Tokyo, Japan). The cell concentration of HL60 cells and granulocytes was adjusted to 0.5 × 10⁶ and 4 × 10⁶ /mL in the medium, respectively, for incubation with or without 1 μM lipopolysaccharides (LPS) (Sigma-Aldrich, St. Louis, MO, USA). ## Transfections Introduction of 50 nM each of human miR-34a-5p (mirVana miRNA mimic, Ambion, Life technologies, Carlsbad, CA, USA), human miR-155-5p (mirVana miRNA mimic, Ambion, Life technologies), and siRNA against *FOS* (Sense: `GAAUCCGAAGGGAAAGGAAtt`, antisense: `UUCCUUUCCCUUCGGAUUCtc`) into 2.4 × 10⁶ HL60 cells suspended in 800 μL Gene Pulser Electroporation Buffer Reagent (Bio-Rad Laboratories, Hercules, USA) was carried out by square-pulse electroporation (280 V, 12 msec) using a Gene Pulser (Bio-Rad). For the controls, corresponding amounts of control miRNA (Life Technologies) or control siRNA (Life Technologies) were introduced. After 40 hours, the cells were subjected to mRNA and protein quantification. The cell viability after electroporation was measured by trypan blue staining (control siRNA: 72.6 ± 8.7%, *FOS* siRNA: 65.7 ± 5.5%, with no significant difference). ## mRNA decay analysis The miRNA-transduced HL60 cells were cultured at a concentration of 0.5 × 10⁶ cells in the presence of 50 μM of 5, 6-Dichlorobenzimidazole 1-β-D-ribofuranoside (DRB) (Sigma) with or without 200 μg/mL emetine (Sigma) for the indicated time. ## Total cellular RNA extraction and reverse transcription Total cellular RNA was extracted using ISOGEN (NIPPON GENE, Toyama, Japan), and treated with DNase I (Takara Bio, Otsu, Japan). The cDNA for mRNA quantification was synthesized as described previously. For miRNA quantification, RNA was subjected to polyadenylation followed by reverse transcription using a Mir-X miRNA First-strand Synthesis Kit (Clontech Laboratories, Inc., Mountain View, CA, USA). ## Real-time PCR The quantification of mRNA was conducted as described previously. Quantified transcripts of *FO*S and *TNFA* were normalized by *ACTB*. The primer sequences were *FOS* forward: `5’-GGGATAGCCTCTCTTACTACCACT-3’`, *FOS* reverse: `5’-CCTCCTGTCATGGTCTTCACAAC-3’`, *TNFA* forward: `5’-CCCAGGGACCTCTCTAATCA-3’`, *TNFA* reverse: `5’-AGCTGCCCCTCAGCTTGAG-3’` *ACTB* forward: `5’- CAAGAGATGGCCACGGCTGCT-3’`, and *ACTBn* reverse: `5’- TCCTTCTGCATCCTGTCGGCA-3’`. The extraction and reverse-transcription of miRNA was conducted according to the manufacturer’s instructions (Clontech), and amplified using an miRNA-specific primer and mRQ 3’ primer (Clontech) and normalized by U6. ## Cell lysate preparation and western blotting Cytoplasmic and nuclear lysates were prepared as previously described. To prepare total cell lysates from granulocytes, the cells were precipitated in 10% trichloroacetic acid (Wako) for 30 min on ice. The TCA-precipitated fraction was treated with a lysis solution containing 9 M urea, 2% Triton X-100, and 1% dithiothreitol, and was disrupted by ultrasonication, followed by an addition of 2% lithium dodecyl sulfate and further ultrasonication. The proteins were separated on a 10% polyacrylamide gel, transferred onto Immobilon-P Transfer Membranes (Millipore, Billerica, MA, USA), and blotted with rabbit polyclonal anti-c-Fos antibody (Santa Cruz Biotechnology, Santa Cruz, CA, USA) or rabbit polyclonal anti-actin (Sigma), followed by incubation with horseradish peroxidase-conjugated anti-rabbit IgG (Santa Cruz). Signals were detected by ECL Western Blotting Detection Reagents (GE Healthcare, Buckinghamshire, UK). ## Flow cytometry Isolated granulocytes stained with phycoerythrin-labeled anti-CD16 (Beckman Coulter, Marseille, France) or isotype IgG (Beckman Coulter) were fixed and perforated using an Intraprep Permeabilization Reagent (Beckman Coulter). After incubation with anti-c-Fos (Santa Cruz) or isotype rabbit IgG (Santa Cruz), goat anti-rabbit IgG labeled with FITC (Abcam, Cambridge, UK) was added, and the expression of c-Fos in CD16⁺ cells was analyzed by FACSCanto II (BD Biosciences, San Jose, CA, USA). ## Enzyme-linked immunosorbent assay (ELISA) The concentration of TNF-**α** in the culture medium was measured using Quantikine HS human ELISA TNF-**α** (R&D Systems, Inc., Minneapolis, MN, USA) according to manufacturer’s instruction. ## CHIP assay Cells were fixed with 37% formaldehyde, lysed by RIPA buffer, and sonicated. After preclearing with agarose beads, the cell lysate was incubated with 5 μg rabbit anti-human NF-κB p65 (Santa Cruz) or isotype IgG (Santa Cruz). The IgG was gathered by sperm DNA-treated agarose beads, followed by treatment with RNase A (Takara) and proteinase K (Roche Applied Science, Mannheim, Germany). The *TNFA* DNA promoter that coprecipitated with p65 was quantified by real-time PCR using a forward primer (`5’-AGTCAGTGGCCCAGAAGACC-3’`) and a reverse primer (`5’-GGCGGGGAAAGAATCATTCAACC-3’`). ## Statistical analyses Comparison between the two groups was performed by a Mann-Whitney *U* test or paired *t* test. Data from three groups were compared using ANOVA (IBM SPSS Statistics, 17.0). *P* values less than 0.05 were considered significant. Regarding miRNA and c-Fos protein levels in individuals, the criteria for a significant increase and reduction were set as higher and lower expressions than two standard deviations from mean values of the healthy controls, respectively. # Results ## Expression of miRNAs that possibly target *FOS* mRNA in MDS We first computationally searched for miRNAs that possibly bind to *FOS* mRNA 3’UTR, using four different databases; microRNA.org, TargetScan, Miranda, and Microcom. We found that twenty individual miRNAs were each listed by two or more databases. We quantified the expression levels of the 20 miRNAs in granulocytes derived from six patients and six healthy controls. Among them, the expression of miR-34a (1.60 ± 0.57-fold, *P* \< 0.05) and miR-155 (1.70 ± 1.10-fold, *P* \< 0.05) was higher in MDS neutrophils than in healthy cells (miR-34a: 1.00 ± 0.16, miR-155: 1.00 ± 0.17). Considering the heterogeneity of MDS, we expanded the measurement of miR-34a and miR-155 to 17 healthy controls and 23 patients. Compared with the controls, the expression of the two miRNAs in the 23 patients was largely diversified. Twelve patients had an overexpression of miR-34a, which was defined as levels higher than the average level plus two standard deviations (SD) of the healthy controls, and ten patients had significantly increased miR-155. Both miRNAs were simultaneously increased in seven patients. ## Effects of overexpression of miR-34a and miR-155 on *FOS* mRNA stability To examine whether overexpression of these miRNAs caused insufficient *FOS* mRNA stabilisation under translation arrest, *FOS* mRNA decay was analysed in the presence of the transcription inhibitor DRB, using cells with exogenously introduced miR-34a, miR-155, or control miRNA. *FOS* mRNA decreased to 29.0 ± 8.6% of the initial level in 30 min in the control cells, which was not altered by overexpression of miR-34a (23.9 ± 5.4%) or miR-155 (21.7 ± 6.4%). However, when translation was stopped by emetine, 74.6 ± 22.7% of *FO*S mRNA was preserved at 30 min in the control cells, while 45.3 ± 16.9% and 36.8 ± 9.5% remained in miR-34a- and miR-155-overexpressing cells, respectively, both of which were significantly lower than those in the controls (*P* \< 0.05, *P* \< 0.05). These results suggested that the impaired *FOS* mRNA stabilization, which we previously found in MDS granulocytes, was attributed to overexpression of miR-34a and miR-155. ## Suppression of c-Fos protein by miR-34a and miR-155 overexpression We next attempted to experimentally examine whether miR-34a inhibited translation of *FOS*. The basal levels of *FOS* mRNA in miR-34-, and miR-155-overexpressing cells were similar to those in the control cells. In contrast, exogenous expression of miR-34a or miR-155 resulted in a decreased levels of c-Fos protein by more than 50% compared with the controls, suggesting that miR-34a as well as miR-155 interfered with c-Fos translation. ## Expression of c-Fos in granulocytes of MDS patients To measure c-Fos protein levels, the neutrophilic granulocytes from MDS patients and the controls were subjected to immunoblotting and flow cytometric analyses. Since the ratios of c-Fos to GAPDH obtained by immunoblotting were well correlated with the rates of c-Fos-positive cells measured by a flow cytometer, c-Fos protein levels were evaluated by flow cytometry, which requires a smaller number of cells than immunoblotting. The percentage of c-Fos-positive cells in total CD16⁺ neutrophilic granulocytes was 88.6 ± 7.8% in the healthy controls. In contrast, eight out of 17 patients tested (Patients \#1, 2, 4, 7, 9, 12, 13, and 18) showed significantly low percentage of c-Fos⁺ cells, while five out of the remaining nine patients (Patients \#3, 6, 8, 15, and 17) showed detectable c-Fos expression in more than 80% of cells. On the other hand, *FOS* mRNA levels in granulocytes were similar in the controls and MDS patients. The ratios of c-Fos⁺ in total CD16⁺ neutrophils were inversely correlated with miR-34a expression levels (r = -0.616, *P* \< 0.05), but not with miR-155 levels (r = -0.135, *P* = 0.630). There were no correlation between ratios of c-Fos⁺ cells and clinical findings including white blood cell counts (r = 0.438, *P* \< 0.079) and neutrophil counts (r = 0.305, *P* = 0.234). None of the medications or karyotypes were specific to the patients with significantly low c-Fos expression. ## LPS-induced TNF-α synthesis in patients with normal and low expression of c-Fos To investigate the effects of c-Fos protein levels on TNF-**α** production in response to LPS, five patients with low c-Fos expression (Group A), five with similar c-Fos levels to the controls (Group B in), and seven healthy controls allowed us to further draw their blood for the subsequent experiments. The miR-34a levels in the patients categorized as Group A were 2.39 (Patient \#2), 1.58 (#9), 4.52 (#12), 1.89 (#13), and 1.87 (#18), when those in the healthy controls were 1.00 ± 0.15. The isolated granulocytes were then cultured in the presence and absence of LPS. shows the *TNFA* mRNA levels in the cells cultured without LPS for 2 and 3 hours. At both time points, no significant differences were observed in *TNFA* mRNA levels among the three groups. When stimulated with 1 μM LPS, all five patients with normal c-Fos expression in Group B showed a similar increase of *TNFA* mRNA (9.4 ± 4.0-fold at 2 hours, 11.5 ± 5.0-fold at 3 hours) to the controls (10.8 ± 4.2-fold at 2 hours, 13.9 ± 2.3-fold at 3 hours). In contrast, the patients in Group A showed a greater response than the controls at both time point. The maximum increase of *TNFA* mRNA was detected at 2 hours in patients 2, 3, and 5 (26.9-fold, 74.1-fold, and 67.4-fold, respectively). At 3 hours, the increase rate of patients 1, 2, 3, and 4 ranged from 19.5 to 47.1-fold. The concentration of TNF-**α** in culture medium was also measured after 2 and 3 hours of LPS stimulation. A similar TNF-**α** concentration was observed in the controls and Group B (2 hours: 115.7 ± 81.2 pg/mL in controls vs. 118.2 ± 115.4 pg/mL in Group B, 3 hours: 150.8 ± 91.5 pg/mL in controls vs. 143.5 ± 65.7 pg/mL in Group B). However, a significantly higher concentration was observed in Group A at both time points (2 hours: 412.5 ± 302.5 pg/mL; *P* \< 0.05, 3 hours: 735.4 ± 237.5 pg/mL; *P* \< 0.05). ## Effects of *FOS* knockdown on *TNFA* transcription in response to LPS To confirm that reduction of c-Fos enhanced TNF-**α** production in response to LPS, c-Fos was reduced by siRNA in HL60 cells that differentiated to a neutrophilic phenotype. In each experiment, c-Fos levels in *FOS* siRNA-treated cells became less than 63% of those in the control siRNA-treated cells. Without stimulation, the basal levels of *TNFA* mRNA did not differ between the control and *FOS* siRNA-treated cells. After stimulation with LPS for 2 hours, a greater increase of *TNFA* mRNA was observed in *FOS* siRNA-treated cells (32.9 ± 26.6-fold, *P* \< 0.05) than in the controls (4.5 ± 2.3-fold). Since *TNFA* is transcribed by NF-κB, we examined whether c-Fos interfered with the binding of NF-κB to the promoter region of TNF-**α** DNA by CHIP assay using anti-NF-κB p65. In the *FOS* siRNA-treated cells, 2.1 ± 1.0-fold more promoter was coprecipitated with p65 in the unstimulated condition. Stimulation with LPS for 2 hours increased the binding 2.7 ± 1.0-fold in the control cells, which was significantly enhanced by reduction of c-Fos (3.6 ± 2.7-fold, *P* \< 0.05). # Discussion To our knowledge, this is the first report to demonstrate (1) overexpression of c-Fos-targeting miR-34a and miR-155, and (2) reduced expression of c-Fos, which led to excessive TNF-**α** production in response to LPS, in MDS-derived granulocytes. To date, aberrant expression of miR-34a and miR-155 has not been described in terminally differentiated granulocytes, although increases of miR-34a and miR-155 have previously been observed in CD34⁺ cells from patients with low- and high-risk MDS, respectively. Interestingly, increases of tumor- inhibiting miR-34a and tumor-promoting miR-155 were simultaneously detected. Our data raised a new question as to what is reflected by the elevated miR-34a and miR-155 expression in granulocytes. Since miR-34a is a p53 target, the increased miR-34a in peripheral neutrophils may indicate the exposure of their progenitors to DNA-damaging stimuli that induce p53 expression. The miR-34a levels may vary according to the accumulation of damage. Some damaged progenitors may have died because of the proapoptotic feature of miR-34a, while some may have survived and differentiated to neutrophils. Thus, miR-34a-induced apoptosis might contribute to dispersion of miR-34a levels in MDS patients. On the other hand, it has been shown that miR-155, which is derived from a non- protein coding gene, B-cell integration cluster, is upregulated by NF-κB. Therefore, elevation of miR-155 could result from exposure to NF-κB-activating conditions, such as inflammation, oxidative stress, and endoplasmic reticulum stress. Increased expression of miR-155 is known to confer proliferative advantages by suppressing SHIP1, a negative regulator of Akt pathway. To generate miR-155-high granulocytes, progenitors with elevated miR-155 need to escape from such a proliferation cycle to differentiate. The clarification as to what increased miR-34a and miR-155 could be a clue in the understanding of the mechanisms for developing aberrant granulopoiesis in MDS. In the current study, overexpression of miR-34a and miR-155 seemed to have caused the insufficient stabilization of *FOS* mRNA in MDS granulocytes under translation arrest, which we previously reported. In miR-34a- and miR-155-overexpressing cells, neither the basal level nor life span of *FOS* mRNA was altered. The mRNA destabilizing effects of these miRNAs became prominent in the presence of the translation inhibitor emetine, which suppressed endogenous mRNA degradation machinery. These phenomena were exactly the same as those which we had previously observed in MDS-derived granulocytes. In the steady state, *FOS* mRNA might be maximally degraded via multiple factors, which could mask the effects of miR-34a and miR-155. In MDS granulocytes, elevation of miR-34a rather than miR-155 seemed to lead to reduction of c-Fos. According to the data from miR-34a-overexpressing cells, miR-34a, as well as miR-155 that had been experimentally shown to target c-Fos, decreased c-Fos protein level. However, an inverse correlation was observed between c-Fos protein and miR-34a but not miR-155. The lack of suppression of basal levels of FOS mRNA was common to HL60 cells that ectopically overexpressed miR-34a and MDS granulocytes. The repression of protein levels without detectable mRNA change is thought to result from incomplete complementarity between miRNA and target mRNA or modest magnitude of mRNA destabilization. Expression levels of c-Fos protein varied among the patients, although all samples were handled in the same way. The variety of c-Fos expression was unlikely to be due to any medications or different karyotypes, because none of them were specific to the patients with significantly low c-Fos levels. Since c-Fos levels were correlated with miR-34a, a target of p53 induced by DNA- damaging stresses, c-Fos levels may have varied with accumulation of DNA- damaging stress which the cells had been exposed to. There might be additional factors that influence c-Fos levels, because c-Fos expression is regulated post- transcriptionally at multiple steps. For example, we already reported that some MDS patients had decreased HuR levels, an mRNA binding protein that stabilizes c-Fos mRNA. The decrease of HuR might affect c-Fos levels without overexpression of miR-34a. When *FOS* mRNA destabilizing proteins, such as AUF1, are increased, c-Fos levels might be further decreased. Exacerbation of ubiquitin proteasome system that degrades c-Fos protein could also influence c-Fos expression. The reduction of c-Fos possibly contributes to development of ineffective hematopoiesis in both TNF-**α**-dependent and independent manners. Firstly, we confirmed that c-Fos inhibited overproduction of TNF-**α** in response to LPS in neutrophilic granulocytes, as previously shown in monocytes. In HL60 differentiated to a neutrophilic phenotype by DMSO, knockdown of c-Fos enhanced the synthesis of *TNFA* mRNA, and the patients with low c-Fos expression secreted greater amounts of TNF-**α** in response to LPS than the healthy controls and the patients with normal c-Fos levels. Accumulation of excessive production of TNF-**α** under inflammatory stimuli may contribute to the formation of TNF-**α**-high condition in BM and plasma, which has been thought to induce apoptosis of hematopoietic cells in low-risk MDS. Secondly, c-Fos itself is known to promote proliferation, and represses transcription of FasL and jun-mediated Fas. Therefore, loss of c-Fos may provide hematopoietic cells with proliferative disadvantage and poor survival. Low c-Fos expression may also affect the prognosis of patients with low-risk MDS via deregulated production of TNF-**α**. Also, c-Fos knockout mice administered LPS showed higher mortality accompanied by greater increase of inflammatory cytokines including TNF-**α** than control mice. Although there have been no reports that compared cytokine secretion under infection of Gram-negative bacteria between MDS patients and controls, excessive production of inflammatory cytokines may deteriorate the prognosis of MDS patients, as observed in *FOS* knockout mice exposed to LPS, which showed higher mortality accompanied by greater TNF-**α** concentration. The enhanced binding of NF-κB to *TNFA* promoter by knockdown of c-Fos suggests that c-Fos interferes with NF-κB p65. This result was consistent with the previous report that showed interference of c-Fos and NF-kB p65. It is likely that cells with reduced c-Fos expression are susceptible to other NF-κB- activating stimuli, such as stresses by oxidants, irradiation, and the accumulation of misfolded proteins. Thus, not only inflammation but also various cellular stresses could be involved in the upregulation of TNF-**α**in BM and plasma via insufficient expression of c-Fos. # Conclusions In the present study, we demonstrated the reduction of c-Fos via overexpression of miR-34a in MDS granulocytes, which to our knowledge has not been reported in hematopoietic diseases before. The low levels of c-Fos resulted in excessive production of TNF-**α**, which is considered to be a contributing factor to the development of ineffective hematopoiesis. Further studies on the mechanisms of aberrant miRNA upregulation and TNF-**α** overproduction would provide insights to unveil the pathophysiology behind ineffective hematopoiesis in MDS. We thank Ms. Michiko Anzai, Mr. Kojun Yamakawa, Ms. Kazuko Kuriyama, Mr. Shun Hiruta, and Ms. Sanae Sato (Fukushima Medical University, Fukushima, Japan) for the technical assistance, Dr. Tsutomu Shichishima (Fukushima Research Institute of Environment and Medicine, Japan) for suggestive discussion. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceived and designed the experiments:** Y. Shikama. **Performed the experiments:** Y. Shikama MC TO XF. **Analyzed the data:** Y. Shikama MC TO XF. **Contributed reagents/materials/analysis tools:** HN HK KO Y. Suzuki KI. **Wrote the paper:** Y. Shikama YT JK. [^3]: Current address: Center for Medical Education and Career Development, Fukushima Medical University, Fukushima, Japan [^4]: Current address: International Research Center for Medical Sciences, Kumamoto University, Kumamoto, Japan

# Introduction With approximately 160,000 deaths annually in the US, lung cancer continues to account for more cancer-related deaths than colon, prostate and breast cancer combined. In 2011, the National Lung Screening Trial (NLST) demonstrated a 20% relative reduction in lung cancer mortality with annual low-dose computed tomography (LDCT). These encouraging results triggered the widespread endorsement of lung cancer screening. However large-scale implementation has been hampered by the high rate of false-positive LDCT studies. In the NLST approximately 40% of individuals randomized to LDCT screening had one or more pulmonary nodules identified during the study period, 96% of which were ultimately proven benign. In addition to lung cancer screening the increasing utilization of diagnostic chest computed tomography (CT) results in an estimated 1.5 million incidentally discovered indeterminate lung nodules in the US annually. With the implementation of LDCT lung cancer screening for the \> 10 million US adults meeting the screening eligibility criteria, this number is estimated to increase substantially. In summary there appears to be a potential emerging global epidemic of newly detected lung nodules. This increased detection of indeterminate pulmonary nodules in the absence of reliable non-invasive strategies to differentiate benign and malignant nodules will almost certainly result in an increase in iatrogenic mortality, treatment related morbidity and health care costs. While unnecessary invasive diagnostic and therapeutic interventions were kept to a minimum in the NLST study, the management of indeterminate pulmonary nodules in clinical practice serving the general population remains a major challenge. Clinical risk calculators have significantly improved the management of indeterminate pulmonary nodules, but additional tools to distinguish benign from malignant nodules are needed, especially for intermediate risk pulmonary nodules, in order to minimize patient anxiety, radiation exposure, health care costs, and procedural morbidity and mortality.\[–\] We have previously demonstrated that quantitative volumetric CT-based nodule characterization effectively risk-stratifies lung nodules of the adenocarcinoma spectrum.\[–\] In addition we have recently reported in a Lung Tissue Research Consortium based case control study that radiological features of the nodule surrounding lung tissue are potentially valuable in distinguishing benign from malignant lung nodules. (manuscript submitted) This approach eliminates the intra- and inter-observer variability and is independent of the training level of the interpreting radiologist. In addition, modern digital CT images include a large amount of valuable high-dimensional data that currently is not fully utilized besides contributing to the overall impression “gestalt” by the radiologist. This invaluable resource can be leveraged by modern quantitative imaging methods. Radiomic approaches to lung nodule analysis consist of extracting reproducible and objective quantitative radiological variables from CT datasets, reducing large volumes of complex data to manageable and clinically relevant information. These quantitative imaging techniques have been proposed to facilitate the development of diagnostic and prognostic models in lung imaging, allowing for example the risk- stratification of lung adenocarcinomas, the classification of screen- or incidentally detected lung nodules and the characterization of lung cancer subtypes and tumor heterogeneity.\[, –\] In this study, we used the NLST dataset to develop and internally validate a radiological multivariate model to distinguish malignant from benign CT-screen detected indeterminate pulmonary nodules. # Methods ## Subject selection The Mayo Clinic and Vanderbilt University Institutional Review Boards approved or exempted this study (IRB numbers: Vanderbilt University 151500 and Mayo Clinic 15–002674). All participants for the present study were selected from the pool of eligible participants in the NLST, and all patient data were fully anonymized. The methods of the NLST have been published elsewhere. Briefly, the NLST was a randomized controlled trial conducted at 33 US centers, approved by the Institutional review boards at all centers. The study recruited asymptomatic high-risk individuals from August 2002 through April 2004, aged 55 to 74 years, with a smoking history of at least 30 pack-years, who quit 15 years or less prior to randomization. Individuals were screened with either annual low-dose CT or chest X-ray for three years and followed through December 31, 2009. 26,722 individuals were randomized to the low-dose CT arm, and over 10,000 nodules (4–30 mm in longest diameter) were detected during the screening rounds. Participants for the present study were selected from the pool of eligible participants in the NLST, who did not withdraw from follow-up, in the CT arm of the study (N = 26,262) and included all screen-detected lung cancer cases: adenocarcinomas, squamous cell carcinomas, large cell carcinomas, small cell carcinomas and carcinoid tumors. Non-lung cancer controls were selected as a stratified random sample from all participants without a diagnosis of lung cancer during the screen or follow-up periods of the NLST. Cases with more than one nodule were excluded. We restricted our analysis to pulmonary nodules with a size defined by a largest diameter between 7 and 30 mm as reported in the NLST database. ## Screening HRCT data All NLST screening scans were low-dose scans with 2.5 mm collimation or less as pre-defined by strict NLST criteria, the details of which have been published elsewhere. The CT datasets were obtained from the Lung Screening Study core laboratory and transferred to a hard drive that was shipped to the investigators. The datasets from the American College of Radiology Imaging Network core laboratory were transferred initially via hard drive, then electronically to the investigators. Information on nodule location was available to the investigators in the NLST database and confirmed by one radiologist (B.J.B.) and two pulmonologists (F.M. and T.P.) using the CT obtained closest in time to the diagnosis of malignant or benign lung nodules. Nodules were electronically tagged for segmentation and analysis. HRCT without visible nodules, nodules with borders indistinguishable from neighboring structures (e.g. mediastinum or pleura) and nodules without related clinical data were excluded. ## Optimization and validation of nodule segmentation The lung nodules were segmented manually using the ANALYZE software (Biomedical Imaging Resource, Mayo Clinic, Rochester, MN). The location and the extent of each nodule was identified visually and a stack of two dimensional borders were traced out along the transverse orientation. A semi-automated region-growing approach based on the operator-specified bounding cube enclosing the nodule and a seed location within the nodule was used for initial segmentation. Manual editing was performed to remove, if needed, intruding structures like vessels and pleura. A parametric feature-based region growing technique based on the texture classification of the voxels within the operator specified bounding cube was used as previously described. ## Radiomic features A comprehensive set of automatically computable, quantitative radiomic metrics was included for the development of a multivariable predictive model to discriminate benign from malignant lung nodules. Based on previous data and preliminary analysis, we considered metrics within the following categories: general characteristics of the nodule (size and location), nodule characteristics (radiodensity, texture and surface characteristics) and features of the nodule-free surrounding lung characteristics, as below: 1. Metrics capturing the spatial Location of the nodule. 2. Nodule Size 3. Bulk metrics based on the global Shape descriptors of the nodule. 4. Radiodensity metrics based on the CT Hounsfield units within the nodule. 5. Nodule Texture/Density metrics based on the texture exemplar distributions within the nodule. 6. Texture/Density nodule surrounding lung metrics based on the parenchymal texture exemplar distributions within a region surrounding the nodule. 7. Metrics capturing the nodule surface descriptors. 8. Metrics capturing the distribution of the nodule surface characteristics exemplars. ## Development of Score Indicative of Lesion/Lung Aggression/Abnormality (SILA) Current literature suggests that no single quantitative metric exists to differentiate benign and malignant nodules. However, multivariate predictive models based on an ensemble of nodule texture/density, surround texture/density, nodule surface and other shape descriptors could improve the discriminability. To facilitate the multivariate analysis we investigated the possibility of replacing our previously developed nodule texture/density and surface categorization using unsupervised stratification into continuous variables that can be thresholded at multiple levels to provide, if needed, the necessary categorization. We developed SILA to map the nine nominal texture/surface exemplar distributions of the nodule onto a continuous scale. The nine nominal exemplar distributions can be ordinated in 362,880 (factorial 9) ways. To identify the unique ordination that correlates with the virulence/malignancy of the nodule, we used qualitative spatial reasoning and multi-dimensional scaling. Based on this, the nine texture exemplars arbitrarily labeled as V,I,B,G,Y,O,R,C, and P were ordinated as V-R-O-I-Y-P-B-G-C identical to that used to represent the distributions via the glyphs (Figure C). The nine surface exemplars were ordinated as unknown-minimal surface-valley-flat-ridge-pit-saddle valley-saddle ridge-peak. SILA was computed as the Cramer-Von Mises Distance of the ordinated exemplar distributions. Using a similar strategy, the seven primal parenchymal exemplars (Normal, Ground Glass, Honeycombing, Reticular, {mild, moderate, severe} lower attenuation areas) were ordinated to compute the SILA for the parenchyma surrounding the nodule (Figure D in shows the operator dependent variations in the SILA mappings for the texture and surface characterization. The 95% C.I for the maximum nodule-specific SILA differences across the 3 operators was 0.217–0.271 and 0.236–0.276 respectively for the texture and surface characterization). ## Multivariate model Quantitative methods were developed to characterize independent radiological variables assessing various radiologic nodule features. Univariate analysis of the discriminatory power of each radiologic variable and receiver operative curve (ROC) analysis were performed for each variable and an area under the curve (AUC) calculated. Statistical significance was calculated and adjusted for multiple comparisons using Bonferroni correction. Spearman rank correlations between all pairs of variables were calculated and displayed via a heat map. Multivariate analysis was performed using least absolute shrinkage and selection operator (LASSO) method for both variable selection and regularization in order to enhance the prediction accuracy and interpretability of the multivariate statistical model. To increase the stability of the modeling, LASSO was run 1,000 times and the variables that were selected by at least 50% of the runs were included into the final multivariate model. The bootstrapping method was then applied for the internal validation, and the optimism-corrected AUC was reported for the final model. # Results ## Study participants We reviewed 649 LDCT of cancers diagnosed in the screening arm of the NLST that included 353 adenocarcinomas, 136 squamous cell carcinomas, 28 large cell carcinomas, 75 non-small cell carcinomas, 49 small cell carcinomas and 5 carcinoid tumors. After exclusion of cases lacking HRCT data, cases with no apparent lesion on last HRCT prior to the cancer diagnosis, cases with nodules invading the mediastinum, cases with missing outcome data, and lesion with size \< 7mm or \>30 mm, 408 LDCT scans with malignant nodules were selected and analyzed. A stratified random sample of non-lung cancer control nodules (size between 7 and 30 mm) was selected on a 1:1 basis, and 318 benign nodules were selected and included in the analysis. The demographic and clinical characteristics of individuals included in the study are summarized in. In order to prevent overfitting of the model, we only considered quantitative imaging variables that were known *a priori* to be potentially associated with the benign or malignant nature of lung nodules (see supplemental material). Quantitative methods were developed to characterize independent radiological variables assessing various radiologic nodule features including 1. Nodule location, 2. Nodule size, 3. Nodule shape, 4. Nodule radiodensity 5. Nodule texture, 6. Texture/radiodensity of the nodule-free surrounding lung, 7. Nodule surface characteristics and 8. Distribution of the nodule surface characteristics exemplars using 726 nodules identified from the NLST dataset (benign, n = 318 and malignant, n = 408). ## Multivariate analysis In order to select the optimal variables, adjust the regression coefficients to optimize the transportability (external validity) of the model and determine the degree of optimism of the model and perform optimism-corrected analysis of the performance of the model by ROC analysis, all 57 quantitative imaging variables were included in the LASSO regression model. Multivariate analysis using LASSO on all features yielded a multivariate model with 8 selected features (selected with frequency \> 50% after introducing bootstrap to reduce variability after 1000 runs) with an AUC estimate of 0.941. These 8 features include: 1. Offset carina centroid_z (Nodule location), 2. Minimum enclosing brick (Nodule shape), 3. Nodule flatness (Nodule shape), 4. SILA nodule texture (Nodule texture), 5. Maximum shape index (Nodule surface Characteristics), 6. Average shape index (Nodule surface Characteristics), 7. Average positive mean curvature (Nodule surface Characteristics) and 8. Minimum mean curvature (Nodule surface Characteristics), all with P\<0.01. To correct overfitting (internal validation) we used the bootstrapping technique to estimate the optimism of the AUC. The optimism-corrected AUC is 0.939. Using Youdan's index, we obtained the optimal cutoff at 0.478 with sensitivity 0.904 and specificity 0.855. A subset analysis of nodules with size between 7 mm and 15 mm yielded an AUC of 0.9477 with an optimism-corrected AUC of 0.9409 (n = 169 nodules). Offset carina centroid_z captures the location of the nodule in the vertical axis in relationship to the carina, the minimal enclosing brick and flatness capture shape and volume, SILA texture is a summary variable capturing the nodule texture, maximum and average shape index capturing the complexity of the nodule surface and average positive mean curvature and minimum mean curvature representing the degree of curvature of the outer surface of the nodule account for the surface characteristics of the nodule. # Discussion In this study, we report the development and the performance of an internally- validated multivariate radiomic model to differentiate malignant and benign screen-identified indeterminate lung nodules. Using a large lung cancer screening dataset of images obtained with a broad spectrum of CT scanners, acquisition protocols and reconstruction kernels, we demonstrate that our automated radiomic approach reliably distinguishes benign from malignant nodules. This approach, if externally validated, could inform management of screen-identified pulmonary nodules and potentially minimize morbidity, mortality, health care costs, radiation exposure and patient anxiety associated with the currently accepted approach for the evaluation and management of indeterminate pulmonary nodules. To eliminate “agnostic” variables with unknown or improbable clinical significance we pre-selected quantitative imaging features with known potentially associations to the benign or malignant nature of lung nodules for our model. In addition to standard nodule descriptors such as size and location we include variables capturing nodule surface characteristics, density and characteristics of the nodule-free surrounding lung. Although a number of these additional features may influence the subjective assessment by trained radiologist, they currently cannot be accurately measured clinically. While predictive in the univariate analysis features of the immediate nodule-free surrounding lung, as determined by quantitative estimates of low-attenuation (emphysema), groundglass and reticular changes within 10 mm of the segmented boundaries of the nodule were not found to be useful predictors after LASSO selection of candidate predictors. Interestingly, nodule size was not one of the eight selected variables. The only potential variable related to size was the minimum enclosing brick. In order to evaluate the performance of our model without the nodule size as a variable, the optimism-corrected AUC was calculated after removing each variable. The AUC for the 7-variable model without minimum enclosing brick was 0.929, suggesting that nodule size did not exert a disproportionate influence on the final model. If externally validated the excellent diagnostic test performance of our multivariate model could significantly advance the management of patients with screen-detected indeterminate pulmonary nodules. The development of this model based on a large and technically heterogeneous screening dataset including a geographically diverse population and various CT scanners and acquisition protocols, strengthen the external validity of our study. In addition, all analyzable nodules from the NLST were included in modeling which used model selection through shrinkage (LASSO) and bootstrap analysis, allowing adjustment for overfitting and validation of the modeling process. One the main limitations to broad implementation of lung cancer screening remains the large number of false positive screening CT. In order to mitigate this problem and decrease unnecessary patient complications, radiation exposure and patient anxiety, the nodule size threshold for a positive screening study was raised to 6 mm.\[–\] This size threshold has accordingly been endorsed by several other societies such as the Fleischner Society. We selected a threshold of 7 mm in our study for its similarity with this threshold, and also for consistency with the DECAMP-1 study we are planning to use for external validation (NCT01785342). While this 6mm threshold is unquestionably an improvement over the NLST criteria, the number of false positive CT remains substantial, and this problem is likely to persist as screening is more broadly implemented and eligible individuals are screened over longer time periods. Another potentially fruitful avenue of research is the applications of longitudinal volumetric assessment of screen-identified lung nodules, which have been associated with a substantial reduction in the incidence of false-positive CT as well. In fact, the recent European position statement on lung cancer screening endorses volumetric analysis for lung nodule assessment. While some blood or bronchoscopy-based biomarkers have been proposed to facilitate nodule classification, they require additional invasive procedures, which may be difficult to generalize at the population level.\[, –\] Leveraging existing and currently unexploited data to refine the sensitivity and specificity of LDCT would therefore be desirable. Our radiomics classifier compares favorably to currently existing clinical, blood or tissue or radiology-based prediction models and focuses specifically on lung nodule variables considered clinically relevant. Rather than replacing current clinical-based assessment of lung nodules based on size or volumetric analysis, we believe that our classifier could represent an adjunct diagnostic tool to inform clinical decisions for intermediate risk indeterminate pulmonary nodules. This radiomics approach would also not require additional expensive imaging such as PET-CT as required by other additive models. There are several limitations to our work. First, our model has not yet been externally validated before it is used clinically. The prevalence of malignancies in our cohort is \> 50%, which is distinctly more than in a typical screening cohort including similar size lesions (12%). Consequently, it is unclear how our model will perform in independent screening cohorts with a more typical nodule prevalence. If our model cannot be validated it may have to be adjusted based on the validation cohort. However, we used an optimal internal validation model (LASSO), which not only surpasses conventional internal validation approaches (split sample and cross validation), but also penalizes the model to avoid overfitting and optimizes the generalizability of the model. Second, the model was developed from a very heterogeneous sample of the NLST CT dataset and we found the selected radiomic features to be robust and stable across CT platforms, acquisition protocols and reconstructions kernels, which we believe strengthens the reproducibility of our model. Third, the semi-automatic segmentation technique used in this study with manual adjustment by the investigators could admittedly introduce operator-driven variability in radiomic analysis. However, we have recently analyzed the reproducibility of radiomic analysis of adenocarcinomas using the same segmentation technique and found excellent Intraclass Correlation Coefficient (0.828 (95% CI 0.76, 0.895) for the Vanderbilt cohort of 50 adenocarcinomas. We believe that these results support the external validity of our work. Fourth, the relatively small number of cases did not allow us to exclude the influence of clinical or demographic variables known to affect lung cancer risk. We did, however, include additional clinical variables known to strongly influence the risk of lung cancer (age and smoking history in pack-years) and found that these variables did not improve the performance of the model. Finally, it is unclear whether our model will extend to other lung nodule cohorts, such as incidentally-detected lung nodules. Future validation of our model in other settings is indeed warranted. Finally, it should be noted that all lung cancer cases suitable for analysis from the NLST were included in our study, some of which were at advanced stage. This could potentially limit the external validity of our model when applied to indeterminate pulmonary nodules. However most of the included cancer cases were stage I which should mitigate this risk. In summary, we present a promising novel radiomics CT-based approach to lung nodule classification, which we believe could revolutionize our approach to screen-detected indeterminate pulmonary nodules and mitigate the risks inherent in lung cancer screening by minimizing unnecessary mortality, morbidity, radiation exposure, patient anxiety and healthcare costs. # Supporting information [^1]: Some of the authors (Fabien Maldonado, Tobias Peikert, Brian Bartholmai, Srinivasan Rajagopalan and Ronald Karwoksi) are co-inventors of a LDCT-based radiomic classifier for lung adenocarcinomas (distinct from the present work) licensed to Imbio, Inc. Royalties per individual have not exceeded \$5,000. This does not alter our adherence to PLOS ONE policies on sharing data and material. None of the other co-authors have any disclosure.

# Introduction Oceans cover most of the planet, but are still poorly known in terms of biological composition and species richness. As climate changes and human pressures are growing, understanding the distribution of marine biodiversity is a crucial step towards an effective monitoring of marine ecosystems. Nonetheless, numerous species of marine invertebrates are still awaiting taxonomic description. The identification and estimation of species diversity based on a single genetic locus often appears the best option available for groups for which taxonomy is poor or inexistent. DNA barcoding proposed by is a method for identifying unknown specimens to taxonomic entities based on sequence similarity of mitochondrial DNA (mtDNA) sequences. Inexact matches are either grouped with taxa already present in the database or identified as new based on whether they fall within a threshold of sequence similarity. Therefore, delimiting species based on barcode requires setting a specific threshold beyond which two sequences will belong to different putative species referred as molecular operational taxonomic units (MOTUs). Several methods have been developed to delimitate species using single locus data such as the barcode gap detection or based on phylogenetic trees. These methods have been successfully used in the exploration of biodiversity in insects, in crustaceans, in polychaetes, in echinoderms. However, the delimitation of species based on a single locus can be problematic due to incomplete lineage sorting, heteroplasmy, introgression, young species showing no variation at COI or because it relies on taxonomic data. Despite its constraints, DNA barcoding has become an important tool in biodiversity investigation leading to an increase amount of barcode data available. Marine crustaceans are notoriously difficult to identify to the species level by traditional approaches due to their enormous morphological diversity and because morphological stasis is frequent in this group. Despite a decade of barcoding, only 7000 crustacean species have been barcoded out of a total of 67,000 described species so far, and it is estimated that there could be as many as 150,000 species world-wide. In crustaceans, barcode gap detection as proposed by has proven to be an efficient tool to discriminate species like marine crustaceans, and more specifically in decapods and amphipods. Amphipods are small crustaceans characterized by a direct development and weak active dispersal capabilities. These characteristics favor cryptic speciation and endemism, and prevent widespread distribution as has been documented in isopods. Despite their key role in the arctic food web, little is known about the biodiversity of marine amphipods. Amphipods dominate the arctic marine fauna in terms of abundance and biomass. Despite the fact that Arctic regions are already impacted by global warming, the Arctic Ocean is one of the less understood region in the world and its marine biodiversity is one of the least characterized. Arctic marine biodiversity has been shaped by the complex history and environment of this region. At least six openings of the Bering Strait since 5.5 million years ago have allowed trans-Arctic exchanges and invasions of the North Atlantic region by the North Pacific species. Moreover, Quaternary glaciations have pushed some taxa out of the Arctic and recolonisation of the Arctic occurred from neighboring oceans during interglacial periods. Several studies have documented these transarctic exchanges in molluscs, in algae, in fishes and in polychaetes. So, over millions of years, marine species have dispersed through the Arctic several times leading to a complex pattern of biodiversity. In this study, we investigated the species delimitation of amphipods and the biodiversity of marine arctic amphipods using publicly available DNA barcodes to which we added new sequences for amphipods from the Canadian Arctic Archipelago. Our aims were 1) to test species delimitation using both distance-based and coalescent methods and 2) to assess large-scale connectivity of marine amphipods across the three Canadian oceans. # Materials and methods ## Sampling and DNA amplification Samples of Arctic amphipods were collected in 2011 in the Canadian Arctic Archipelago during the NCGS Amundsen expedition using vertical and/or horizontal nets with 200, 500 and 750 μm mesh size. Samples from Greenland were provided by the Greenlandic Institute of Natural Resources. Specimens from the southeastern Bering Sea were collected in 2012 during the Bering-Aleutian Salmon International Survey with a 500 μm mesh bongo net. Specimens from Prince William Sound in the Gulf of Alaska were collected in 2013 with a 500 μm mesh ring net. All individuals were preserved in 95% ethanol. Samples were identified to the species level with appropriate taxonomic keys when possible. Description of samples with geographic cordinates can be found under the “CAAB” (Canadian Arctic Amphipods Barcodes) project available in BOLD ([www.boldsystem.org](http://www.boldsystem.org/)). The DNA of 374 samples was extracted using the E.Z.N.A tissue extraction kit (Omega-biotek), or the QuickExtract kit (Omegabiotek) following the manufacturer’s protocols. Individuals with a body size \> 10 mm were extracted using one or two pereopods. Individuals with a size \< 10 mm were used whole for the extraction. A 658 base pair (bp) fragment of the mitochondrial cytochrome *c* oxidase subunit I gene was amplified using the primer pair LCO1490/HCO2198. Polymerase chain reaction (PCR) was conducted as described in: the reaction mix contained 1X PCR buffer, 2.2 mM MgCl₂, 0.5 mM dNTPs, 0.4 μM of each primer, 1.5 U of Taq DNA polymerase (Life Technologies, Mississauga, ON, Canada), DNA template (around 40–80 ng), and water for a final volume of 25 μl. PCR were performed with an initial denaturation step of 3 min at 94 °C, followed by 5 cycles of 45 s at 94 °C, 45s at 46 °C, 45 s at 72 °C and 35 cycles of 45 s at 94 °C, 40 s at 51°C, 45s at 72 °C, and a final elongation step of 5 min at 72 °C. All PCR products were verified on a 1.5% agarose gel and direct-sequenced by Genome Quebec (McGill University, Montreal, Canada). Over the 374 individuals used, 310 individuals were successfully sequenced and their chromatograms were manually checked on MEGA7.The presence of pseudogenes was assessed by translating sequences into amino acids. All sequences were deposited in the the project “CAAB” (“Canadian Arctic Amphipod Barcodes”) available in BOLD and in GenBank database under the accession numbers MH330696—MH331009. ## Dataset and molecular operational taxonomic units construction Sequences found after searching in December 2017 for “Amphipoda” from Canada, Greenland and United States in the public data portal of BOLD conducted were combined to our data. All sequences were aligned using MUSCLE available in the BOLD sequence analysis tools. The final dataset contained 2309 sequences from the three canadian oceans. As reliable MOTU depends on the accuracy of the MOTU retrieved with different methods, we chose to used several gap discovery methods (Barcode Index Number, MOTHUR, ABGD) and coalescent process (bPTP) to find the number of MOTU present in our dataset. ### Barcode Index Number (BIN) The Barcode index number was constructed by first using a 2.2% p-distance threshold for clustering sequences and then each cluster was refined by the examination of the genetic divergence among neighbors. Each cluster was described with a unique and specific identifier (e.g. Barcode Index Number or BIN), already available or newly created if the sequences are clustered in an unknown BIN. Sequences were aligned using Kimura-2 parameters (K2P), with MUSCLE. All analyses were conducted in BOLD with sequence analysis tools. ### MOTHUR We used MOTHUR to cluster our sequences into provisional species (referred as MOTU). In the literature, several thresholds are reported for delimiting species: a 3% threshold commonly used to define species, a 4% threshold proposed by, a 16% threshold for amphipods species proposed by. Uncorrected pairwise distances were first computed for each threshold values and sequences were clustered into MOTU using the nearest neighbor method considering each gap. Briefly, this method allows the clustering of sequences in the same MOTU if they are at most X% distant from the most similar sequence in the MOTU. In order to associate taxonomy to MOTUs, we created a molecular taxonomic database containing all amphipods identified to the species level and its barcode sequence. To do so, we selected all amphipod sequences from BOLD with the following criteria: taxonomic identification at least to the genus level and a COI sequence longer than 500 bp. All sequences were aligned with MAFFT version 7 web server and trimmed to 501 bp. After the construction step, each MOTU was aligned to the reference database with a confidence threshold of 90%. In order to assess the accuracy of our molecular identification, we aligned 305 sequences of 43 taxonomic identified species to our taxonomic database. As no discrepancy was observed, we confirmed the validity for MOTU identification. All analyses were performed with MOTHUR. ## Automatic Barcode Gap Discovery (ABGD) Since intra-specific divergences are smaller than inter-specific ones, a gap in the distribution of all pairwise distances can be identified using the Automatic Barcode Gap Discovery method available at [www.abi.snv.jussieu.fr/public/abgd](http://www.abi.snv.jussieu.fr/public/abgd). This method is described in detail in. Briefly, the data was first partitioned into a number of groups (i.e. species) such that the distance between two sequences taken from distinct groups was always larger than a given threshold distance and then appply recursively this procedure to get a better partitioning of the data into putative species. We used the default value of 0.001 for the minimum intraspecific distance and 0.3 for the maximum intraspecific distance, with 10 steps and K2P distance. We explored the relative gap width (X) for X = 0.5 and X = 1. After the barcode gap discovery, sequences were clustered into MOTU based on the estimated threshold. ### Coalescent approach Poisson Tree Process is a model for delimiting species based on a rooted tree with branching events representing the number of substitutions. As a large dataset is computer challenging, we selected unique sequences to reconstruct the tree and then performed the analysis. To reconstruct the tree, we rooted it with *Pandalus borealis* (Krøyer, 1838) (accession: KY018893.1). We selected the evolutionary model using the Bayesian Information Criterion (BIC) available at W-IQ-TREE. We generated a phylogeny under the General Time Reversible model with empirical frequency, invariable site and under gamma rate (GTR+F+I+G4) using Bayesian inferences in MrBayes 3.2.6 available at [www.phylo.org](http://www.phylo.org/), using two runs for 10,000,000 generations until convergence was observed. Trees were sampled every 10,000 generations and the first 25% of sampled tree were discarded as burn-in. The posterior probabilities (PP) were calculated with the 50% majority-rule consensus tree. We used the web version of bPTP (<http://species.h-its.org/ptp/>) to generate the species delimitation under a coalescent process. Analysis was conducted with 500,000 iterations of MCMC and 25% burn-in. We also removed the outgroup to improve the delimitation. ### Diversity among the three canadian oceans To provide an overview of the similarity in the MOTU composition among the three oceans, a Venn diagram was obtained for each threshold. ## Species threshold identification As the thresholds proposed by or by were based on a single family of amphipods, we used the same method to identify a species threshold for all Amphipoda. To do so, we collected all amphipod sequences from BOLD (January, 2016). Among the 15516 records, all sequences without a taxonomic identification to the species level, with less than 500 bp or associated with pseudogenes were discarded. A total of 8471 sequences corresponding to 89 families were examined further. Families with less than 30 sequences or containing less than 3 species were also discarded from this dataset. The diversity assessments for the amphipods and for the most represented families were analysed from the data set with 3879 sequences from 272 species, 70 genera, and 10 families. After performing a first alignment with MAFFT, all sequences were trimmed to the same length of 501 bp. After this step, Kimura 2 parameters pairwise distances were computed at each taxonomic level intrafamily (F), intragenus (G) and intraspecies (S) in MEGA7 and plotted by family using the boxplot representation available in R and described in. Based on the ABGD, we estimated a general threshold to 7%. Three ranges of thresholds (3%, 7% and 16%) were plotted to see which one best discriminated the different amphipod species. # Results ## Nucleotide diversity Out of a total of 310 sequences obtained, 5 sequences contained stop codon and were removed from the analysis. All sequences were clustered into 28 MOTUs representing 26 BIN of which 3 were uniques. The mean GC content was 32.9%. Intraspecific K2P distance ranged from 0.6 to 18.07% and interspecific distance ranged from 1.3 to 27.3%. The complete dataset consisted in 2309 sequences that were trimmed to the same length of 400 bp. Within the final alignement, 94 conserved sites and 277 parsimony informative sites were detected. The mean GC content was high (GC = 38.01%). Intraspecific K2P distance ranged from zero to 33.33% and interspecific distance ranged from 0.17 to 32.14%. There were 418 sequences from locations within the Pacific Ocean, 998 sequences from the Arctic region and 892 from the Atlantic. ## Gap distance based methods ### BOLD All 2309 sequences were grouped into 285 MOTUs and 263 BINs, of which 83 BINs were unique (and Tables). Among these 285 MOTUs, 113 were from the Arctic, 91 were from the Atlantic and 105 were from the Pacific. Nineteen MOTUs were shared between the Arctic and the Atlantic, three were shared between the Pacific and the Atlantic, and three were shared between the Pacific and the Arctic. A single MOTU was shared among the three oceans. ## Barcode gap detection with ABGD The distribution of K2P genetic distances displayed two modes separated by a gap (‘barcode gap’) between 0.04 and 0.8. The ABGD method split 2309 sequences into 242 groups over a wide range of prior maximum divergence (P = 0.046416—P = 0.100000) after 20 partitions for X = 1 and after 23 partitions for X = 0.5. All analyses produced a single group when P = 0.11. Among the 242 groups, 94 were from the Arctic, 84 from the Atlantic, and 88 from the Pacific. One MOTU was shared between the Arctic and the Pacific, two between the Atlantic and the Pacific, and one among the three oceans. Nineteen MOTUs were shared between the Atlantic and the Arctic. ## MOTHUR The 3% threshold allowed the identification of 261 MOTUs, for which 100 were from the Arctic, 86 from the Atlantic, and 97 from the Pacific. Eighteen MOTUs were shared between the Arctic and the Atlantic. In contrast, three MOTUs were shared between the Arctic and the Pacific and two between the Atlantic and the Pacific, of which one MOTU was common among the three oceans and identified as *Themisto libellula* (Lichtenstein in Mandt, 1822). A large proportion of MOTUs belonged to the Gammaridae, Ischyroceridae, Hyalidae and Aoridae families. We were unable to identify 6 MOTUs to taxonomic level. At the species level, several species were found in the three oceans: *Ampelisca spinipes* (Boeck, 1861), *Aora gracialis* (Spence Bate, 1857), *Parhyale hawaiensis* (Dana, 1853), *Microphasma agassizi* (Woltereck, 1909), *Pontogeneia inermis* (Krøyer, 1838), *Tiron biocellata* (Barnard, 1962), *Weyprechtia pinguis* (Krøyer, 1838). As only one shared MOTU was detected between the three oceans, it suggests that these species consist of distinct MOTUs. Under the 4% threshold, 251 MOTUs were found for which 97 were in the Arctic, 84 in the Atlantic and 94 in the Pacific. Twenty MOTUs were shared between the Arctic and the Atlantic. Three MOTUs were shared between the Arctic and the Pacific, and two MOTUs between the Atlantic and the Pacific among which one MOTU is shared among the three oceans. Six MOTUs were shared among the three oceans and identified as *Ampelisca spinipes* (Boeck, 1861), *Aora gracilis* (Spence Bate, 1857), *Microphasma agassizi* (Woltereck, 1909), *Pontogeneia inermis* (Krøyer, 1838), *Parhyale hawaiensis* (Dana, 1853), *Themisto libellula* (Lichtenstein in Mandt, 1822), *Weyprechtia pinguis* (Krøyer, 1838), *Tiron biocellata* (Barnard, 1962) respectively. Most MOTUs belong to the Gammaridae, Hyalidae and Ischyroceridae families. Six MOTUs were not assigned to a taxonomic level. Under the 16% threshold, 173 MOTUs were found, for which 59 were in the Arctic, 70 in the Atlantic and 75 in the Pacific. Twenty-one MOTUs were shared between the Arctic and the Atlantic. Three MOTUs were shared between the Arctic and the Pacific and nine MOTUs were shared between the Atlantic and the Pacific. Six MOTUs were shared among the three oceans and identified as *Ampelisca spinipes* (Boeck, 1861), *Aora gracilis* (Spence Bate, 1857), *Anonyx nugax* (Phipps, 1774), *Ischyrocerus anguipes* (Krøyer, 1838), *Microphasma agassizi* (Woltereck, 1909), *Pontogeneia inermis* (Krøyer, 1838), *Themisto libellula* (Lichtenstein in Mandt, 1822), respectively. Most MOTUs belong to the Gammaridae family and the Aoridae family. Only 11 MOTUs were not assigned to a taxonomic level. ## Tree based method The tree-based bPTP analysis estimated the number of species between 265 and 287 with a mean of 275 species with high posterior probablilities (\>0.5). ## Canadian Arctic diversity The family identification showed that not all the locations and the families were well sampled across the three Canadian oceans. Eight families (Bathyporeiidae, Gammaracanthidae, Iphimedidae, Pallaseidae, Photidae, Pleustidae, Pseudocrangonyctidae, and Urothidae) were represented by a single MOTU in one location. The Gammaridae family had the highest number of MOTUs. As expected, the number of MOTUs decreased with an increase of the threshold value used for all families. Six families were equally recovered (Bathyporeiidae, Corophidae, Iphimediidae, Pallaseidae, Pleustidae, Pseudocrangonyctidae) regardless of the threshold. ## Species threshold Divergences were estimated for the validated dataset from 3879 sequences representing 272 species, 79 genera, and 10 families. As expected, genetic divergence increased with taxonomic rank: a higher divergence was observed at the family level (K2P from 0 to 0.9), than at the genus level (K2P from 0 to 0.7), and the species level (K2P from 0 to 0.3). However, the threshold used to delimit species varied among families. The threshold proposed by (0.16 substitution/site) discriminated well the Gammaridae species but not the other amphipods species from different genera. The lowest species divergence was observed in the Melitidae (0.001), the lowest genus and family divergences were observed in the Hyperiidae (0.26). The highest species (0.3) and genus (0.7) divergence was observed within the Gammaridae and the highest family divergence (0.9) was observed within the Melitidae. # Discussion Inventories of marine biodiversity are much needed in the context of global changes. A decade of DNA barcoding has resulted in the acceleration of species discovery and has helped to provide partial or complete inventories lists for selected taxa. The ability to assign species identities to DNA sequences depends on the availability of comprehensive DNA reference libraries such as BOLD. The generation of these libraries represents an important task, in particular in some difficult acessible region such as the poles or the abyss where our knowledge on their biodiversity is still lacking. With the increasing use of metabarcoding approach, the need for complete library references will become essential to allow an easier investigation of species richness and to monitor the faith of the biodiversity in an area undergoing global changes such as the Arctic ocean. Our study provides a useful example of how one can retrieve information from BOLD database to produce an exhaustive first step inventory of specific animal groups or specific regions that are otherwise lost among the different publicly available projects. We showed that widespread species are in fact composed of different MOTUs, suggesting the presence of cryptic species. This implies that a careful look at the taxonomic keys is needed for amphipods. Moreover, we found that using a single threshold for species delimitation for different families of amphipods is not always accurate. ## MOTU delimitation The number of MOTU retrieved varied between 173 (MOTHUR-16% threshold) and 285 (BOLD). Excluding the highest threshold used in MOTHUR, ABGD retrieved the lowest number of MOTUs. Some discripancy have been noticed among methods. For example, da Silva *et al*. noticed that ABGD and PTP produce lower estimates of snapper species diversity (based on the COI) than other methods. The relative performance of species delimitation methods has been examined in other studies. The results produced by AGBD vary according to the metrics used and with the number of individuals per species and might be useless in the exploration of species diversity in poorly known groups. Several authors emphasize the importance of combining several methods in species delimitation. Defining species based on genetic data alone might be limiting and additional characters such as life history traits and geographic ditribution are also of interest for species description. Delimiting species based on a single mitochondrial fragment can introduce some bias. First, the DNA barcoding protocol has different steps that can introduce some errors that can lead to misidentification of the specimen. Second, incomplete lineage sorting or homoplasy can contribute to an overlap between the intra- and interspecific distances leading to difficulties in the identification of the barcode gap. “COI like” sequences can contribute to increase the threshold estimate as it has been investigated in crustacean. We also conducted our species clustering analyses using a tree-based method. We found around 270 species which is less than the number of MOTUs found with BOLD (285) and less than the number of MOTUs found with ABGD (242). It is well known that tree based methods like GMYC or PTP are sensitive to the genealogy of a particular locus, wich can be discordant with the true species tree. Moreover, singletons can also represent difficulties. ## Species threshold Delimiting species relies on a threshold over which species belong to the same or to two different species. The 3% of divergence is a commonly used threshold in the literature. Other studies report that a 16% threshold was more suitable for delimiting amphipod species like Gammaridae or Niphargidae. On the contrary, lower value for threshold has been also reported in other family of amphipods like in Talitridae (8% -17%) or in Hyallelidae (4%). By including ten amphipod families, we have showed that the use of 16% is appropriate for the discrimination of Gammaridaea species but is too high to discriminate species from other families. Based on the interspecific distance distribution among amphipod families, we found that this threshold can discriminate most amphipod species but is not always suitable. Based on the automatic barcode gap detection, we estimated a threshold of 7% for discriminating amphipod species, which is intermediate between the previous thresholds proposed. This threshold seems appropriate for the majority of amphipod species but again, based on diverse family distances distribution, this threshold will underestimate the number of Gammaridae or Talitridae species. Instead, we found that for discriminating amphipod species, a threshold specific to each family will be more appropriate. However, we should emphasize that not all amphipod families are equally studied and less than 10 sequences are not enough to estimate intraspecific distances. For example, Gammaridae is one of the best studied amphipod family for which a large amount of molecular data is available but other amphipod families such as Hyperiidae have not received the same attention. Here, we focused on 10 amphipod families that represent less than a quarter of all amphipod families. Nevertheless our analysis provides the first molecular attempt to determine species threshold in amphipods. Additional barcoding data is needed on the other amphipod families for helping to refine our species threshold and to determine factors (e.g. benthic vs pelagic lifestyle, life history traits, speciation rates, effective population size, etc…) responsible for the large variation in sequence divergence among different crustacean families. ## Arctic diversity Our study, based on the analysis of more than 2300 sequences distributed throughout the three Canadian oceans, indicates the presence of at least 250 provisional amphipod species. We found 100 putative Arctic species representing circa 85% of the known amphipods inventory in the Arctic (from the Chukchi Sea). We recovered more putative species in the Atlantic and less in the Pacific. Marine Arctic fauna is mostly derived from recent and repeated colonisations from both Pacific and Atlantic species after Pleistocene glaciation events or multiple Bering Strait openings \[reviewed in and \]. Most studies on trans- arctic interchanges have reported a Pacific origin of the invasion. Regarless of the threshold used, our results indicate a higher similarity between Arctic and Atlantic oceans (\>15 MOTUs shared) than between the Arctic and the Pacific (one MOTU shared). Similarity between Arctic and North Atlantic fauna has also been reported in polychaetes or in bryozoans, suggesting that the Atlantic Ocean contributed significantly to the recolonization of the Arctic. In addition, the Pacific harbors the highest number of MOTUs compared to the number of sequences available (418), corroborating a higher diversity of this ocean compared to others. The limited number of shared MOTUs between the Arctic and Pacific oceans suggests the presence of a barrier restricting exchanges between these oceans. Moreover, the fact that the retrieved Pacific MOTUs were not found elsewhere confirms the isolation of Pacific taxa from colder Arctic waters. In copepods and in amphipods, isolation between Pacific and Arctic populations has been documented. The cold temperature of the Arctic waters has likely impeded the survival and reproduction of North Pacific species. Global warming induced changes in the Arctic ocean leading to less inhospitable barriers for Pacific species and promoting interchanges between Pacific and Atlantic oceans as suggested by recent models. Therefore, further sampling of the Pacific region is needed to confirm this isolation as our results might be biased by differential sampling efforts between the Pacific and the Atlantic oceans. We found a relatively higher proportion of benthic species belonging to the intertidal or infralittoral family of Gammaridae, Hyalidae and Ischyroceridae than pelagic species. In the unique pelagic amphipod family (e.g. Hyperiidae), we recovered 12 MOTUs of which the majority (8 MOTUs) were Arctic. In the Eurasian arctic waters inventory, the presence of eight Hyperiidae species were recorded. This result suggests that pelagic diversity is not well known maybe due to sampling difficulties; further efforts are needed to better characterized ocean diversity. As sampling in the Arctic is quite challenging, it is most likely that rare taxa were not included in our analyses. Although, our study does not include depth information, it will also be interesting to investigate the diversity of MOTUs among oceans according to depth to get a more precise picture of the marine biodiversity. However we were not able to identify all MOTUs to the species level. Several explanations can be considered. First, during the taxonomic reference creation, we discarded more than half of the sequences due to pseudogenes or short sequences. Secondly, the 15 516 sequences of amphipods available in BOLD correspond to 1 514 species which is under the number of amphipods estimates. Despite these constraints, barcoding techniques have provided useful information on the amphipod biodiversity in the three Canadian oceans. This approach can benefit the study of oceanic Arctic region which is one of the least studied \[, \]. Moreover, we also showed that widespread marine amphipods species are composed of different MOTUs, suggesting an underestimated diversity. A large number of studies have revealed the presence of cryptic species complex in marine invertebrates\[, –\], suggesting the utility of combining different types of data (e.g. molecular and morphological) to identify species. # Conclusions Our analyses have contributed to the assessment of marine arctic amphipod biodiversity in revealing potential cryptic species, in showing the sharing of MOTUs between the Arctic and the Atlantic amphipods and in the isolation of Arctic amphipods from those of the Pacific. Moreover, thanks to the increasing barcode data available in amphipods, we were able to show that threshold value for species identification in amphipods needs to be estimated for each family. It has become evident that species definition should not be restricted to a COI sequence but should include additional information such as ecological niche. Pursuing arctic amphipods studies with DNA barcodes will ultimately lead to a better understanding of marine biodiversity and the mechanisms of speciation in marine environments. With arctic waters already showing the presence of temperate invaders, there is an urgency to complete this task. # Supporting information We wish to thank the officers and crew of the CCGS Amundsen during sampling efforts in 2011. We are also grateful to the Greenland Institute of Natural Resources in Nuuk, Greenland for providing samples. We thank Dr. A. Comeau for MOTHUR presentation. (SR) was supported by the Canada Excellence Research Chair program. We also thank Dr. K. Dionne, Dr N. Bierne, Dr. C. Nozais and Dr. A. Derry for comments on a previous version of this manuscript. We wish to thank the Canadian Healthy Ocean Network. We are grateful to Québec Océan and EnviroNord. [^1]: The authors have declared that no competing interests exist.

# Introduction Cancers are a complex set of proliferative diseases whose progression, in most cases, is driven in part by an accumulation of genetic changes, including copy number aberrations (CNAs) of large or small genomic regions, which may for example lead to amplification of oncogenes or loss of tumor suppressor genes. However, cancer progression is also often characterized by increasing genomic instability, potentially generating many “passenger” CNAs that do not confer clonal growth advantage. These processes give rise to a complicated landscape of genomic alterations within an individual tumor and great diversity of these CNAs across tumor samples, making it difficult to identify driver mutations associated with cancer progression. In recent years, array-based comparative genomic hybridization (aCGH), and single nucleotide polymorphism (SNP) arrays have been used to analyze the CNAs of tumor samples at a genomic scale and at progressively higher resolutions. Moreover, numerous large-scale tumor profiling studies have generated copy number data sets for large cohorts of tumors. These large and complex “cancer genome” data sets present difficult statistical challenges. Individual CNAs may be as small as a few adjacent probes or as large as a whole chromosomes and may be difficult to detect above probe-level noise; moreover, it is unclear how to make sense out of diverse CNAs from hundreds of tumors. Typically, two kinds of analyses have been carried out on copy number data sets: 1. clustering of samples by their CNAs, to determine possible tumor subtypes characterized by a common pattern of amplifications and deletions; 2. determining significant genetic aberrations, either gains or losses, that occur frequently in the data set, since these may represent driver mutations important for tumor progression. Almost always, these problems are tackled with a pipeline approach, where aCGH profiles of chromosomes for individual samples are first processed by a segmentation algorithm; individual segments (genomic regions) are “called” as gains or losses, based on their amplitude, using a choice of statistical procedure and significance threshold; and finally the called segments are used as input to a clustering algorithm , or score-based method for determining significant common aberrations. The disadvantage of pipeline approaches, however, is that algorithmic choices and tuning parameters at each step may produce very different results, and mistakes or biases are propagated forward. For the first step, there are numerous segmentation algorithms, that yield significantly different segment boundaries, leading to different calls of gains and losses. The final step of analyzing CNAs across samples depends critically on choices made earlier. As an example, the widely-used GISTIC method for determining frequent aberrations uses as its test statistic, at each locus, the number of samples in which a gain (or loss) is present multiplied by the mean amplitude of the gain (loss). However, both the count and the mean amplitude depend on earlier choices in the pipeline. In this study, we propose a novel and mathematically robust method for finding significant patterns of CNAs in a large copy number data set directly from the probe-level data. By avoiding a pipeline approach involving a segmentation step, our algorithm exploits probe-level correlations in aCGH data to discover subsets of samples that display common CNAs. By applying the approach in a hierarchical fashion to iteratively partition the data set, we discover both large- and small-scale events and can detect statistically significant CNAs occurring on 5% of the samples. In this way, the algorithm addresses both the clustering problem and the frequent aberration problem at the same time. Algorithmically, our approach is related to recent work on maximum-margin clustering, which extends support vector machine-like optimization approaches to the problem of unsupervised clustering. That is, each partition of the data set is achieved by learning a linear classifier of the probe-level aCGH profiles that assigns samples to one group or the other. We also build on ideas developed for supervised classification of aCGH samples, in particular, the use of piece-wise constant and lasso, regularization terms in the optimization problem, which encourages the classifier to make decisions using only a small number of probes in informative contiguous regions. We tested our approach on a large cohort of glioblastoma aCGH samples recently generated by The Cancer Genome Atlas Project (TCGA). We found that the major CNAs detected by our algorithm are largely consistent with the original TCGA study, in that almost all CNAs previously reported were also in our results. However, we found additional significant CNAs missed by the TCGA analysis but supported by earlier studies and/or expression analyses. Moreover, the hierarchical partitioning approach summarizes the set relationships and dependencies between different CNAs, which may be helpful for generating hypotheses about the sequence of CNAs in tumor progression. # Results ## Algorithm overview Our algorithm iteratively partitions a data set of tumor aCGH profiles for a given chromosome to discover subsets of tumors with similar CNAs. Instead of using standard preprocessing techniques like segmentation algorithms, we directly use probe-level data and incorporate prior knowledge about the nature of this data, namely: (1) successive probes are correlated, i.e. are likely to represent the same copy numbers; and (2) a chromosome typically (though not always) harbors few CNAs. At each partitioning step, we learn a linear separator that assigns aCGH profiles to one of two classes, represented geometrically by the two half-spaces (i.e.) on either side of the hyperplane defined by the normal vector and bias term. Here, chromosome profiles and the weight vector are real-valued vectors with dimension equal to the number of probes for the chromosome, and is determined by solving an optimization problem where it is constrained to be piecewise constant (successive probes tend to have the same weights) and sparse (few probes have non-zero weights). Our approach builds on a recently proposed maximum margin clustering algorithm, which brings ideas from large-margin supervised learning techniques like support vector machine classification and support vector regression to the unsupervised clustering problem; the choice of constraints was motivated by recent work on fused lasso regression. Since each linear separator results in a binary partition of samples, we apply our procedure iteratively to separate each group of samples into two new groups in such a way that the new linear separator is orthogonal to the previously determined ones. Therefore, each step will find a new direction of variation in the aCGH data (similar to principal component analysis), and the overall procedure results in a hierarchical partitioning of the data set. ## Large-margin partitioning reveals hierarchy of copy number changes We collected our data set from the Cancer Genome Atlas (TCGA) data portal. It contains 345 glioblastoma tumor samples with copy number changes profiled on Agilent 244K arrays (228K probes). This data set has previously been analyzed to determine major amplification and deletion events using the RAE and GISTIC algorithms. We used the Level 2 data already produced by the previous analysis. This data has already been normalized through the application of a lowess algorithm on the log ratio data, and probes flagged as low-quality (saturated, non-uniform or faint) are excluded. Quality of the arrays was also measured through the proportion of excluded probes and the consistency of values associated with successive probes, and low-quality arrays were removed from the data set. We ran our algorithm separately on every chromosome, with a sparseness coefficient and a piecewise-constantness coefficient. Empirically, we found the following dependence on the choice of these coefficients: if the coefficients were chosen to be too small, it would result in a trivial clustering, with all samples assigned to the same group; if the parameters were too permissive, the clustering obtained would be the same as standard -means. However, between these two extremes, clustering results were not overly sensitive to parameter choice. We expect the suitable range of parameters to depend on the array platform as well as statistical properties of the array profiles in a given data set. We therefore suggest performing a grid search on a subset of the samples and selecting the smallest possible parameters that give a non-trivial clustering on every chromosome. In order to assess the significance of our results, we used a random model where we shuffled the probes of our dataset and compared the distance between the median samples of our two groups to the distribution of 1000 distances of median samples of two random sample groups separated with the same classifier. We verified that the randomized distance distribution follows a normal distribution, and we computed the -value for the distance between the median samples corresponding to the tail of this normal distribution. For each chromosome, we constructed a “clustering tree” by iteratively splitting each group into two if it respected three criteria. The first criterion was that it must contain more than five samples (1.5% of the data set), since it would be difficult to achieve a statistically significant partition of very small subsets. The second criterion was that splitting this group would not make the depth of our tree bigger than 3. The maximal depth was chosen heuristically: after three iterations, we empirically found that the groups were too small or the separation was not significant anymore. The last criterion was that the partition generating this group must satisfy a significance threshold of. While this -value may seem overly permissive, it is important to understand that our estimator (the centroid distance) is not directly optimized by the algorithm; therefore, the empirical -values generated are fairly conservative. gives an example of a “clustering tree” produced by our algorithm for chromosome 19. The first iteration separates the samples into two clusters, one with 17 samples that presents a deletion of a region of the q arm and one of 326 samples, with. The centroid of each cluster is shown in green (leftmost column); in addition, a segmentation of each cluster centroid using a standard tool (circular binary segmentation) is shown to aid visualization of the copy number differences between the two groups. As for this separation and each cluster is bigger than 5 samples, we split each of these subsets into two new groups. The splitting of the group of 17 samples is is not associated with a significant enough median separation and therefore is not split again. On the other hand, the partition of the group of 326 samples produces one group of 250 samples without any apparent significant CNA and a group of 76 samples whose centroid shows an amplification of the whole chromosome. This split has strong significance, and therefore both of these groups are split again. The partition of the group of 250 samples does not achieve significance, and neither of the resulting clusters show any significant CNAs. The group of 76 samples is divided into two new groups of 37 and 39 samples. Each of these groups shows an amplification of the whole chromosome, but the group with 39 samples seems to have a lower amplification of the q arm than of the p arm while the other does not. As we limit ourselves to trees of depth 3, we do not partition either of these groups any further. ## Analysis of glioblastoma aCGH data recovers known CNAs without segmenting samples We applied the iterative procedure to each chromosome independently, as described in the previous section. To call characteristic CNAs of each cluster, we applied circular binary segmentation using default parameters on its centroid, i.e. the median profile of the cluster, and associated the characteristic CNA(s) of this centroid to the cluster. One should understand that the aberrations of the centroid profile may not be shared by every one of the cluster samples, but that it gives a good estimate of these events. We also caution that the size of the partition gives a good idea of the penetrance but is not entirely equivalent. The first iteration of our algorithm found an amplification of the whole chromosome 1, of the whole chromosome 7 and of the whole chromosome 20. It also identified the deletion of the whole 9p arm, as well as a big part of 19q, the whole chromosome 10, the whole chromosome 13, the whole chromosome 14 and the whole chromosome 22. The second iteration of the algorithm found the loss of 6q arm, deletion of the whole chromosome 15, of the whole chromosome 16 and an amplification of the whole chromosome 19. It also demonstrated that some samples that present an amplification of chromosome 7 also contain a focal and very strong amplification event on the 7p arm. The third iteration of the algorithm identified focal amplification events on chromosome 3 and on chromosome 4. It also showed a loss of the whole chromosomes 9 and 21. These results are summarized in, along with the size of the partition in which each CNA was identified in terms of number of samples and percentage of the full data set. An analysis of the same data set using both RAE and GISTIC algorithms has already been published. Both methods agreed on significant large-scale amplification events for the whole chromosomes 7, 19 and 20 and focal amplification events on chromosome 1 and 12; significant large-scale deletion events on chromosomal arms 6q, 9p, 15q, on whole chromosomes 10, 13, 14 and 22; and focal deletion events on chromosome 1. In addition, RAE found significant focal amplification events on chromosome 14, as well as significant focal deletion events on chromosome 11. By contrast, GISTIC found different additional focal amplification events on chromosomes 3 and 4. includes a summary of our results as well as a comparison with the amplification and deletion events found by both of these analysis. As shown in, most of the events found in both RAE and GISTIC analyses are found by the first two iterations of our method, including every large-scale event identified by these methods. Exceptions include a small amplification event on chromosome 12, the events on chromosome 1 (where our method disagrees with the finding of RAE and GISTIC) and an amplification event on chromosome 4, which is found on our third iteration. ## Iterative partitioning reveals novel CNAs supported by independent glioblastoma studies Beyond recovering almost all the CNAs identified by methods like RAE and GISTIC, our iterative partitioning algorithm found a number of significant events that were not discovered by previous analyses of this dataset. These events include an amplification of the whole chromosome 1, a deletion event on the whole chromosomes 9, 15, 16 and 21, as well as a deletion of the 19q arm. Some of these events have been documented in studies of independent copy number data sets, such as the deletion on the 19q arm, and of chromosome 16. The deletion of chromosome 21 has been previously associated with glioblastoma, and it has been proposed that the low incidence of glioblastoma in Down's syndrome patients is linked to the chromosome 21 trisomy that characterizes this genetic condition. Here, we find the chromosome deletion associated with a very small cluster (6 samples), and the low frequency presumably explains why this aberration was missed by previous analyses. The deletion of chromosome 15 actually includes the deletion on the 15q arm found in the previous analyses. The shape of the centroid for this partition shows that the amplitude of the deletion is smaller on the rest of the q arm and on the p arm, and it is possible that full chromosome deletion was not found by RAE or GISTIC due to the smaller amplitude. To identify genes that are well correlated with the CNAs, we performed a significance analysis of microarray (SAM) using the SAMR package. For each cluster, we labeled each sample according to its label (inside or outside the cluster of interest) and looked at the number of genes of the region of the CNA that were significantly differentially underexpressed in the case of a deletion, or significantly overexpressed in the case of an amplification. Calculations were done using the t-statistic, 100 permutations and the Tusher method. Our results, summarized in, show that in most cases a large number of genes had expression levels that are significantly correlated with the assignment of samples to the cluster harboring the CNA. It should be noted that the relationship between expression and copy number is complex, and that the absence of significant correlations does not exclude the presence of the CNA, especially in cases where the low count of genes or samples makes this correlation statistically difficult to prove. The novel CNAs discovered by our analysis are correlated with several important genes. For example, the deletion of the chromosome 16, the 19q13.2–19q13.43 regions, and the chromosome 21 are significantly correlated with underexpression of candidate cancer-suppressor genes, respectively CBFB, or CDH11, TFPT and DSCR1, giving additional evidence in support of these events. ## Several sets of frequent chromosomal aberrations show high correlation One advantage of our method compared to score-based approaches such as RAE and GISTIC is that it gives an assignment of samples to groups – or, more precisely, identifies CNAs by simultaneously finding the groups of samples that harbor them – which makes it easier to identify which samples are affected by which frequent CNAs. We associated each sample to a set of frequent CNAs based on its cluster assignments in the chromosome-based iterative partitioning procedure. We found that co-occurrences of frequent CNAs within a sample were common; indeed, a majority of samples (249 out of 345) contained 2 or more of the frequent CNAs listed in. We further examined co-occurrences of pairs of frequent CNAs, and we found that 31 pairs can be considered correlated (i.e. with an intersection of sample assignment better than expected by background frequencies) with by Fisher's exact test (see Supplementary). A simple analysis of these significant pairs revealed that these correlated CNAs can actually be seen as three groups of co-occurences: 1. The amplification of chromosome 7 and its associated focal amplification event, the deletion on 9p, the deletion of chromosomes 10, 13 and 14 as well as the amplifications on chromosomes 19 and 20 are all highly correlated. 2. The deletion of 6q is well correlated with the focal amplification event on chromosome 7 as well as with the deletion on 9p. 3. The deletion on chromosome 22 is well correlated with the amplification of chromosome 7 (but not with the associated focal event), the deletion of chromosome 10 and the deletion of chromosome 14. # Discussion ## Recovery of CNAs missed by summary statistics Some of the novel glioblastoma CNAs that we found are good examples of how our method improves on summary statistic approaches, such as RAE and GISTIC. For instance, the deletion of chromosome 15 has only been spotted on the q arm by RAE and GISTIC. When we examined the profile of the centroid of a cluster identified by our method, we saw a lower amplitude deletion on the p arm as well. Because of this low amplitude, each probe on its own would not have a significant mean deletion across the data set and would hence be missed by a summary statistic. However, because all of the probes for the chromosome are affected, the deletion should be considered a significant CNA and is readily identified by approach. As a second example, the deletion of the region 19q2–19q13.3 has not been found by other methods applied to the TCGA data set, even though it has been confirmed as a deletion event by previous studies. Here, the problem seems to be the fact that the same region is also present as an amplification event on a larger number of samples, which confounds the detection of this deletion by a summary test statistic. Finally, the deletion of the whole chromosome 21 is presumably missed by other methods because it is presents on only a small number of samples (6 samples or 2%). However, since this event is a deletion of the whole chromosome and therefore supported on many probes, intuitively it should be much more statistically significant that a smaller but similarly infrequent event. Indeed, the importance of this CNA is confirmed by previous studies linking trisomy 21 in Down's syndrome to lower prevalence of glioblastoma as well as by the correlation with the under-expression of a candidate tumor-supressor gene present in this region. ## Recovery of focal events shows that even though the first iteration of our algorithm seems to focus on large aberrations, the following iterations are able to find focal events such as the ones on chromosomes 3 and 4, and that our algorithm is therefore able to find focal events as well as large ones. The only focal event whose presence is agreed on by both RAE and GISTIC and that our method is not able to find is the one on chromosome 12. Looking at the raw data shows us that this event is shared by roughly 40 samples but only affects 2 probes, which makes it a difficult signal to find when looking a multiple probes. However, by restricting our analysis to a small interval centered on the event (300kbp or 40 probes), we were able to identify the common event using our maximum-margin clustering algorithm (see Supplementary), suggesting that our method could perhaps be used in conjunction with a sliding window to improve detection of very small events. ## Analysis of samples with high noise and genomic instability The glioblastoma copy number profiles that we analyzed here have relatively few CNA events and therefore provide a favorable test case for computational analysis. Copy number data sets for other cancers have proven far more problematic. For example, a recent copy number study of lung adenocarcinoma compiled a very large (400 samples) but challenging data set, where the signal to noise varied considerably over samples – potentially due to stromal contamination – and a sizable fraction of samples displayed numerous events. The authors curated the samples into three tiers based on signal quality and restricted analysis to the best tier. Despite the large average number of events per samples, the study identified only a few regions altered in a significant number of samples, with the most common CNA (amplification of chromosome 14q13.3) only present in 12% of the best third (top tier) of their samples. We applied our method to this lung adenocarcinoma data set to see how it would perform in a high noise setting. Since the original assignment of samples to tiers was not readily available, we did a first pass analysis of the entire data set – without attempting to reduce to the cleanest samples – using the same parameters as we used on the TCGA data set. Interestingly, the first iteration of the algorithm partitioned each chromosome into two clusters containing exactly the same samples (with), with one group consisting of samples with a strong but very noisy signal and the other containing samples with a weak signal. This result suggests that our method may be able to automatically distinguish signal quality. The initial choice of parameters did not find any significant aberrations at a -value cutoff of 0.05, possibly due to the different array platform as well as the different statistical properties of the copy number profiles (see Supplementary and Supplementary). However, using our algorithm with a different set of parameters on chromosome 14 allowed us find the amplification of 14q13.3, albeit only in 6 samples (2% of the total count of samples) and with a weak -value. Here, the presence of a large group of very noisy samples in the data set may be responsible for degrading the -value. While we were not able to directly compare to the original analysis on the top tier samples, this quick analysis on the full data set is fairly encouraging, in that we were able to retrieve the main result without an *ad hoc* curation of samples. ## Possible algorithmic extensions The above analysis also underscores the impact of the choice of the two constraint parameters, and, which determine the degree of sparseness and piecewise-constantness, respectively, of our linear classifiers. We chose the parameters for the glioblastoma study through heuristics and recovered most known events as well as several novel and plausible CNAs. However, full exploration of this parameter space could yield additional results; for example, to predispose the algorithm to find focal events, one might try to make the sparsity constraint more stringent. Various strategies might be used to optimize the choice of parameters, including use of a cross-validation loop. To implement this approach, one would have to choose an appropriate method for estimating the quality of the clusters: standard estimators are closely tied to the objective functions optimized by traditional clustering algorithms (such as -means), which do not take into account the properties of copy number profiles (i.e. spatial correlations, sparsity of deletion/amplication events). However, such a cross- validation loop would also entail lengthier computational times. This cost could be greatly reduced if we were able to compute the entire regularization path of the fused lasso in a single pass, as others were able to do with the original lasso and SVM optimization problems. An interesting direction for future research would be to extend this method to incorporate gene expression data in the analysis of copy number profiles. The candidate gene results of show that even a simple analysis is able to find significant correlations between the two types of data. Presumably, CNAs that result in deregulated expression are more likely to be driver mutations. A framework that integrates paired copy number and mRNA expression may yield greater insight into functional gains and losses in cancer. ## Conclusions We have introduced a new mathematically sound method for the identification of frequent alterations in a large cohort of tumor copy number profiles. This method builds on the concept of maximum-margin clustering by extending to more than two groups and incorporating specific properties of copy number data, i.e. the piecewise-constantness and the sparsity of CNAs. We applied this method to a large publicly available glioblastama data set from The Cancer Genome Atlas initiative. Our results include most CNAs already found by previous studies as well as novel CNAs confirmed by other data sets or expression analyses. We showed that we were able to identify large aberrations as well as focal events and found significant correlations between these different CNAs. # Methods Below, we briefly develop the technical background related to our approach and describe the details of our algorithm. We first present the fused lasso classification algorithm and then show how to extend it to an unsupervised setting based on the maximum margin clustering algorithms. Finally, we introduce our iterative partitioning procedure for determining hierarchical clusters characterized by common CNAs. ## Supervised classification We first consider the supervised learning problems for aCGH profiles. Here we are given a training set of aCGH samples of dimension, where is the number of probes; each example has an associated label or an explanatory variable, where the labels can be discrete (classification) or real-valued (regression). Given our labeled set of samples, the goal of linear supervised classification or regression is to build a linear function that will be able to predict the correct explanatory variable for a new sample. We use a general formulation of supervised learning as an optimization problem:where is a loss function that penalizes the error between the predictions and the real explanatory variables, is a regularization function, and the value of the constraint, to be adjusted to find a suitable compromise between minimizing of the error term and regularizing (avoiding overfitting) the model. Problem (1) describes a whole family of algorithms that includes (i) the support vector machine (SVM), when is the set of binary labels, is the hinge loss, and is the Euclidean norm; (ii) the L-SVM, when is the set of binary labels, is the hinge loss, and the L-norm; or (iii) lasso regression, when, is the squared error, and is the L-norm; among many others. ## Maximum margin clustering Recently Xu et al. proposed to generalize this optimization framework to the unsupervised clustering problem, i.e. trying to find the best linear separator between (latent) classes of samples when the labels are not known. The general optimization problem described in (1) then becomes However, in the case of binary classification, i.e., Problem (2) becomes a mixed integer problem (MIP), which is not easily solvable using standard optimization techniques. Instead, Zhang et al. proposed an algorithm similar to conjugate descent to solve this problem, alternating between (a) training the linear separator given current label assignments and (b) updating the label assignment based on the linear separator. They found that a standard support vector machine (SVM) converges quickly in this alternating procedure to a fixed set of labels without finding more favorable cluster assignments. Therefore, they proposed using support vector regression (SVR) for the linear separator. SVR is more often used in the case of regression, i.e., than in binary classification but performs well for the clustering problem. ## Incorporating prior knowledge In choosing the regularization function to use in training a linear separator, we want to take into account two different properties of copy number profiles: 1. Successive probes on the same chromosomes are likely to represent the same copy number and should therefore tend to be attributed similar weights in the linear function. 2. There are usually only a small number of CNAs in a given sample, often (but not always) occupying relatively small genomic regions, and therefore only a small number of probes should have non-zero weights in the linear function. Tibshirani and Saunders introduced a fused lasso method for regression and classification that gives a sparse and piecewise-constant linear function by imposing two separate constraints ; the regression formulation takes the form:where is the least squares loss function. Here, the first constraint is the lasso regularizer, which induces sparsity, i.e. few components in the solution vector are non-zero; the second constraint enforces piecewise constantness, i.e. adjacent probes tend to be assigned the same weight. In the case of high-density copy number profiles, another issue is the non- uniform distribution of the distances between successive probes. Older low resolution aCGH technologies used probe sets designed to have relatively uniform inter-probe distances, or at least, these distances varied within an order of magnitude. New higher resolution technologies have higher disparities in inter- probe distances. To take these into account, we modify the constraints to include a coefficient that normalizes for inter-probe distances:where if and refer to succesive positions on the same chromosomal arm and is the weight of the corresponding relation. In the case of aCGH profiles, we define aswhere is the genomic distance between probes and. Incorporating these modifications, we obtain the following quadratic problem under linear constraints: Using this quadratic problem, we propose an algorithm similar to the maximum margin clustering algorithm : ### Algorithm 1 *Iterative fused lasso*. 1. Initialize the labels, for example with standard -means. 2. Calculate the linear separator obtained by solving Problem (4). 3. Assign the labels using the linear separator:. 4. Repeat steps 2–3 until convergence. ## Iterative partitioning One limitation of the method proposed in Problem (4) is that it only achieves a binary partition of the data, while in fact there may be more than two distinct subgroups defined by common CNAs. In order to overcome this limitation, we use the following iterative partitioning algorithm: ### Algorithm 2 *Iterative partitioning*. 1. Initialize the partition of the data with Algorithm 1. 2. Partition each of the groups of the partition into two new groups. 3. Repeat steps 2 until the size of a new group or the significance of the partition falls below threshold. In order to guarantee that the newly discovered groups at each step will explore different directions of variation, we make each classifier orthogonal to the preceding ones. This can be done by the following equation, assuming that we know the classifiers, we can then learn a new classifier (and associated partitioning), written as to simplify notation:where is the number of samples that we want to separate. Using the same method as in the last section, Problem (7) can be transformed into a quadratic problem under linear constraints. ## Implementation The method has been implemented under Matlab using the commercial Tomlab/CPLEX package. Both this implementation and another one using the free SeDuMi package are freely available. # Supporting Information We thank Barry Taylor for helpful discussions and assistance with the TCGA data set. [^1]: Performed the experiments: FR. Analyzed the data: FR. Wrote the paper: FR CL. Supervised the research: CL. [^2]: The authors have declared that no competing interests exist.

# Introduction Heart failure develops in a substantial number of patients with acute myocardial infarction (AMI), despite early reperfusion therapy. Stem cell-based therapies and chemokine- or cytokine-based therapies are promising procedures for myocardium regeneration and the attenuation of ventricular remodeling, fibrosis or dysfunction. The stromal cell-derived factor (SDF)-1/CXCR4 (the receptor of SDF-1) axis has been demonstrated to be critical for the recruitment of stem cells to the ischemic area. A higher level of SDF-1 in the injured myocardium improves cardiac repair. Many studies have shown that SDF-1 is upregulated or released spontaneously in the infarct and peri-infarct regions in AMI. The autologous SDF-1 level peaks 24–72 h after AMI and becomes normalized seven days after AMI. However, this natural modulation is not sufficient for the induction of functional recovery in many AMI models or patients. SDF-1 delivery by a virus vector, plasmid or direct injection into the injured tissue early after AMI has been proven to be more effective, but these methods are not convenient or feasible in clinical therapy. The development of an oral medicine that could be an alternative to these methods may allow a markedly easier and safer application. Resveratrol (RSV) is a component of the polyphenol extract of grape skin/seed, red wine and the root of *Polygonum cuspidatum*. There are seven members in the mammalian sirtuin (SIRT) protein family, the many beneficial effects of RSV are mediated by SIRT proteins. It should be said that most researches are focus on SIRT1, which is an essential deacetylase and can regulate the acetylation of various proteins, such as peroxisome proliferator-activated receptor gamma co- activator-1 and p53, a transcriptional activator of apoptosis. The inhibition of p53 may increase the concentration of local vascular endothelial growth factor. In oncology research, p53 deletion causes the upregulation of SDF-1 and accelerates the migration of CXCR4⁺ tumor stem cells or tumor development. We attempted to administer atorvastatin and RSV to a murine model of AMI and found that short-term RSV administration can increase cardiac SDF-1 expression and improve the mobilization of very small embryonic-like stem cells (over 80% of which are CXCR4⁺) to the injured heart. However, the role of SIRT1 in RSV induced SDF-1 regulation has not been fully elucidated. In the present study, we aimed to investigate how RSV increases cardiac SDF-1 in AMI and determined that p53 inactivation through its deacetylation by RSV mediates SDF-1 regulation. # Methods and Materials ## Animals and materials The study was approved by the Institutional Animal Care and Use Committee of Fuwai Hospital and Cardiovascular Institute. All of the research protocols conformed to the guiding principles for animal experimentation articulated by the Ethics Committee of the Chinese Academy of Medical Sciences and Peking Union Medical College. The authors of this manuscript have certified that they comply with the ARRIVE guidelines, and all efforts were made to minimize suffering. Male C57BL/6 mice, 8–10 weeks of age, were housed in an accredited institute facility in accordance with the institutional animal care policies. For the *in vivo* experiments used to evaluate the effect of SIRT, the mice were randomly divided into four groups: 1) a sham operation group, 2) a myocardial infarction (MI) group, 3) an MI+RSV (25 mg/kg/day in drinking water, from five days before AMI to two days after AMI as described previously) group, and 4) an MI+RSV+nicotinamide (NAM, a SIRT inhibitor, 75 mg/kg/day i.p. by osmotic pump, from five days before AMI to two days after AMI) group. The dose of NAM administered was selected according to a previous report. The mice were sacrificed two days after AMI or four weeks after AMI. The osmotic pump used was purchased from Durect Co. (Alzet 1007D, Cupertino, CA, USA). RSV and NAM were purchased from Sigma (St. Louis, MO, USA). ## AMI model The mice were anesthetized with an i.p. of 2% chloral hydrate (2 ml/100 g), endotracheally intubated by tracheotomy, and mechanically ventilated using the Inspira Advanced Safety Ventilator (ASV, NP 55–7059, Harvard Corp., Evansville, WI, USA), which supplied 0.75 ml of room air/oxygen 110 times per minute. The heart of each mouse was exposed via a left thoracotomy. After removing the pericardium, the left anterior descending coronary artery (2 mm below the tip of the left auricle) was occluded with an 8.0 Prolene suture (ETHICON, Inc., Somerville, NJ, USA). Occlusion was confirmed by observing the LV pallor immediately. ## *In vitro* experiments H9C2 cells were purchased from ATCC (Manassas, VA, USA) and incubated in DMEM (Hyclone Laboratories, Logan, UT, USA) with 10% vol/vol FBS and 1% vol/vol antibiotics at 37°C in a humidified atmosphere of 5% CO₂. To determine whether p53 activity affects the cardiac SDF-1 level and investigate the specific effect of SIRT1, p53 and SIRT1 were silenced. p53 and SIRT1 siRNAs were obtained from Santa Cruz Biotechnology (Santa Cruz, CA, USA) and transfected into the cells using the corresponding agents (Santa Cruz Biotechnology) according to the manufacturer’s protocol. The control cells (treated with the transfection reagents only) and the cells transfected with siRNA were treated in serum-free medium and then incubated in a sealed hypoxic GENbox jar fitted with a catalyst (Bio-Me´rieux, Marcy l'Etoile, France) to scavenge free oxygen for 24 h. The oxygen tension in the medium was measured using an air indicator (Bio-Me´rieux). After hypoxia, the cells were collected and subjected to western blotting. ## Western blotting After the mice were sacrificed two days after AMI via cervical dislocation, the cells from the bone marrow (BM, at least three samples in each group) and the LVs (n = 5–7 in each group) were isolated after drawing blood. The protein in the total cell lysates, including H9C2 cardiomyocytes, was extracted by pipetting the cells or homogenizing the tissues in RIPA buffer with a proteinase inhibitor cocktail (Roche Diagnostics, Basel, Switzerland). Thirty to forty micrograms of protein were separated on SDS-polyacrylamide gels, transferred to PVDF membranes, and incubated with primary antibodies overnight at 4°C. Anti- SDF-1 (Santa Cruz Biotechnology) antibody was used for evaluating the SDF-1 level. The nuclear protein fractions were isolated using commercially available kits (BioVision, Mountain View, CA, USA) and immunoblotted with anti-p53 (Cell Signaling Technology), anti-acetyl-p53 K379 (K382 human protein, Cell Signaling Technology) (tissue samples), anti-acetyl-p53 K370 (K373 human protein, Millipore, Billerica, MA, USA) (tissue samples), and anti-SIRT1 (Cell Signaling Technology, Inc., Beverly, MA, USA for tissue samples, Abcam Inc., Cambridge, MA, USA for cell samples) antibodies. After washing and incubation with secondary antibodies, the signals were visualized with the ECL substrate (Pierce Biotechnology), quantified with the Image Scan software (Scion Image; Scion Co.), and standardized to the expression of β-actin (Sigma) and nucleophosmin (NPM, Sigma, for nuclear proteins). ## ELISA SDF-1α is an important isoform of SDF-1. To inspect the changes in SDF-1α, the LVs were isolated and maintained at -80°C until use. The tissue SDF-1α was measured with a mouse ELISA kit (RayBiotech Inc., Norcross, GA, USA) according to the manufacturer’s instructions as described previously. ## Echocardiographic and pathological evaluations 1. **Echocardiographic assessment.** Transthoracic echocardiography was performed (n = 4–9 in each group) at the end of the fourth week. The mice were anesthetized with an i.p. of 2% chloral hydrate (2 ml/100 g), maintained in the decubitus position and allowed to breath spontaneously during the procedure. Transthoracic echocardiography was performed with a 35-MHz phased-array ultrasound system (VisualSonics Inc., Toronto, Canada). M-mode tracings of the LVs were recorded at the papillary muscle level to measure the interventricular septal dimension (IVS), LV end-diastolic dimension (LVEDD) and LV end-systolic dimension (LVESD). The ejection fraction (EF) was calculated automatically. 2. **Pathological studies at the end of the fourth week.** After the mice were sacrificed via cervical dislocation after echocardiography, the LVs were obtained. At least three LVs per group were fixed with 4% paraformaldehyde, embedded in paraffin and cut into 3-μm-thick sections (three sections of each LV at the papillary muscle level). Masson-trichrome (MT) staining was performed to quantify the extent of fibrosis in the LVs. The fibrotic area and total area of the LV on each image were measured using the Image-Pro-Plus software (Media Cybernetics), and the fibrotic area was calculated as a percentage to the total LV area. ## Statistical analysis All of the values are expressed as the means ± SEMs. Unpaired Student's *t* test was used for comparisons between two groups. For multiple comparisons, ANOVA followed by Scheffé’s method was used. A value of *p* \< 0.05 was considered significant. # Results ## Effect of p53 and SIRT1 silencing/RSV on SDF-1 regulation in cardiomyocytes under hypoxic conditions H9C2 cardiomyocytes were transfected with p53 or SIRT1 siRNA, and successful transfection caused a significant downregulation of p53 or SIRT1 (Fig). The cells were then treated with hypoxia and serum deprivation. The SDF-1 level after hypoxia in the cells transfected with p53 siRNA (1.87±0.07) was higher than that observed in the cells without siRNA transfection (1.00±0.09). The SDF-1 levels in the cells administered RSV (15 μM) and the cells transfected with p53 siRNA were highest among all of the cells after hypoxia (1.59±0.03 and 1.52±0.03 respectively). SIRT1 siRNA transfection did not inhibit SDF-1 expression when compared with control (hypoxia only), but inhibited the effect of RSV. RSV did not further enhance the SDF-1 expression after p53 silencing. ## Drug loading affected SIRT1 expression in AMI The amount of nuclear SIRT1 was investigated by western blotting. The levels of nuclear SIRT1 decreased during acute injury (MI: 0.62±0.04; sham: 1.00±0.03). RSV loading induced a recovery in the nuclear SIRT1 level compared with MI (0.90±0.06), whereas NAM inhibited the effect of RSV (0.60±0.03). ## SIRT status mediated the modulation of p53 The amount of total p53 increased after AMI, but a significant difference was not found among the groups. Mouse p53 acetylation was evaluated at K379 and K370. p53 was acetylated at K379 early after AMI compared with the sham group. The acetylation in the MI+RSV mice (1.12±0.03) was less than that in the MI mice (1.59±0.07), and the effect of RSV could be eliminated by NAM administration (K379 acetylation in MI+RSV+NAM mice: 1.61±0.20). We did not find significant acetylation at K370. ## The increase in cardiac SDF-1 was dependent on SIRT We did not detect an obvious difference in BM SDF-1 among the groups. In contrast, compared with the sham mice, the cardiac SDF-1 expression level was increased in the MI mice. The MI+RSV group exhibited the highest cardiac SDF-1 expression level, and this level in the MI+RSV+NAM group was reduced to the level observed in the MI group. To confirm these results, we evaluated the level of SDF-1α in each group by ELISA. The ELISA results (sham: 191.67±6.26 pg SDF-1α/mg cardiac tissue; MI: 243.07±9.15 pg/mg; MI+RSV: 297.12±13.11 pg/mg; MI+RSV+NAM: 233.64±18.53 pg/mg) were consistent with the western blotting results. ## SIRT status was parallel to the cardiac function At the end of the fourth week, the IVS, LVEDD and LVESD in the MI group were significantly different from the corresponding values in the sham group (*p*\<0.01 for each). RSV improved LVESD (*p*\<0.01 vs. MI), but the improvements in IVS and LVEDD (*p*\>0.05 vs. MI) at the end of the fourth week were not significant. The sham mice had an average LVEF of 75.10%. The MI mice had an average LVEF of 39.30%, and this level was elevated to 57.91% by RSV. The LVEF in the MI+RSV+NAM mice (43.98%) was similar to that of the MI mice (*p*\>0.05 vs. MI and *p*\<0.05 vs. MI+RSV). The heart rate during the echocardiographic examination showed no difference among the groups. Furthermore, the MI mice presented a larger infarct size or severe fibrosis (fibrotic area/the total LV area×100% = 16.74±1.78%). RSV attenuated this effect to some extent (8.39±1.10%), and NAM abolished the effects of RSV (14.77±1.20%, Fig). # Discussion In the present study, we showed that RSV can promote an increase in the SDF-1 level in infarcted LV early after AMI through SIRT1 upregulation/p53 inactivation. To the best of our knowledge, this study provides the first demonstration of the interaction of RSV/p53 with SDF-1 modulation in AMI. First, we silenced p53 in cardiomyocytes and found that p53 is upstream of SDF-1 under hypoxia/serum deprivation, which is similar to the results obtained in tumor cells. Second, we found that cardiac SIRT1 was reduced early after AMI and that SIRT1 activation by RSV enhanced the expression of SIRT1, similarly to the results of previous studies. Third, p53 activity is regulated at both the transcriptional and post-translational levels, e.g., by acetylation. Acetylation enhances p53 binding to target genes and causes transcription of the target. As a mark of p53 activation, the K379 and K370 acetylation of p53 have been well investigated in various cells or models. Although we used the nuclear protein fractions of the whole LVs for the evaluation of p53, which may result in similar cardiac p53 levels in sham and MI mice, we clearly showed that p53 is highly acetylated (K379) early after AMI, and this hyperacetylation was reversed by RSV. Furthermore, NAM attenuated the effect of RSV on the post-translational modulation of p53 in the MI model. These results indicate that SIRT1 is upstream of p53 and that the signal transduction affected the SDF-1 level in the injured myocardium. The activation of p53 contributes to apoptosis, and p21 (CIP1/WAF1) is one of the targets of p53. However, previous reports have shown that p53 hyperacetylation is not accompanied by increased p21 under ionizing radiation and that p21 inhibits the integration of STAT with the STAT-binding site within the SDF-1 promoter and directly obstructs the expression of SDF-1 during arterial wound repair. Therefore, p21 appeared to not be involved in signal transduction. The mechanism through which p53 inactivation causes SDF-1 translation or expression in AMI requires further research. In a cutaneous wound model, p53 silencing enhances the secretion of chemokines, including SDF-1, and improves wound healing. The results support the data obtained in the current study, suggesting that the activation of p53 limits the increase in SDF-1 during injury and that targeting p53 is not the only way to inhibit apoptosis but also may cause regeneration. In parallel, we found that greater levels of the myocardium were present in scar tissue four weeks after AMI in the RSV group. In the current study, ischemia was the primary reason responsible for the increase in the expression of SDF-1. We surmise that hypoxia-induced factor 1 (HIF-1) induced by MI may be one of the factors involved in the regulation of SDF-1. However, HIF-1 increases p53 expression, and p53 downregulates SDF-1 transcription. Further investigations are required to elucidate the mechanism. One of the limitations of the current study was that we did not obtain data to support or exclude the possibility whether other SIRT families are involved or not in the pathway. We also did not evaluate the SIRT1 activity. Moreover, RSV has multiple protective effects related to cardiovascular diseases. SDF-1 also contributes to cardiac repair by augmenting other cytokines, chemokines or hormones. Therefore, we cannot conclude that the function improvement observed in the RSV-loaded mice was completely due to SDF-1 upregulation or SDF-1/CXCR4 interaction. We should use both SDF-1 blockade and SDF-1/CXCR4 interferent or label the stem cells to evaluate the role of stem cell recruitment in future research. In conclusion, RSV activates SIRT1, decreases the activity of p53 via deacetylation, and increases the SDF-1 gradient at the site of cardiac injury. These observations may have clinical importance because they imply another beneficial biological effect of RSV for repair of the infarcted myocardium. We gratefully acknowledge the technical assistance provided by Qing Xu (echocardiography; Capital Medical University, Beijing, China). This study was supported by grants from the National Natural Science Foundation of China (81170129 to Y.J. Yang), the Health and Medical Development Foundation of China (2011-H25 to Y.J. Yang), and the Basic Clinical Research Collaboration Fund of Capital Medical University (14JL61 to H. Wang). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: WH ST WSD QHY YYJ. Performed the experiments: WH ZQ HJF WJ. Analyzed the data: WH ZQ JC. Wrote the paper: WH ST YYJ.

# Introduction Ovarian cancer (OC) is the second most common gynecological malignancy and the first cause of death. Cancer metastasis, rather than primary tumors, are responsible for most cancer deaths. Owing to a lack of definitive early symptoms and appropriate markers for OC diagnosis at an early stage, the majority of patients are diagnosed with late-stage OC accompanied with metastasis, which typically has a 5-year survival rate of \< 30%. It is critical to understand the molecular mechanisms involved in OC metastasis, and to determine efficient, specific, and sensitive molecular targets that can be applied to metastasis diagnosis, prognosis, and individual treatment. Double minute chromosomes (DMs) are cytogenetic hallmarks of gene amplification. DMs appear in various kinds of human cancer cells, but not in normal cells. As extra-chromosomal elements carrying amplifications of genomic DNA sequences, DMs contribute to cancer formation and progression, oncogenes are frequently present in the amplified sequences and the proteins they encode are often over- expressed. Examples of genes amplified on DMs include *MYC* in colon cancer, *MYCN* in neuroblastoma, *EGFR* in gliomas, and *EIF5A2* in ovarian cancer. As DMs are vehicles of amplified genes including many oncogenes, functional studies of genes that are amplified on DMs is a good way to explore candidate oncogenes. Our team previously identified 3q26.2 as an origin of DMs in the human ovarian cancer cell line UACC-1598, and a series of genes were co-amplified on the same ovarian DMs, including *MYCN*, *EIF5A2*, and *RPL22L1*. Both *MYCN* and *EIF5A2* play important roles in cancer progression. However, the relationship between *RPL22L1* and cancer is not known. In this study, we showed that *RPL22L1* is commonly over-expressed in clinical OC individuals and its expression level is strongly related to tumor invasion and metastasis. An *in vivo* experiment showed that forced expression of *RPL22L1* promotes intraperitoneal xenograft tumor development in nude mice, and enhances cell migration and invasion *in vitro*. Furthermore, knocking it down with small interfering RNA (siRNA) inhibits migration and invasion *in vitro*. During this process, *RPL22L1* over-expression resulted in elevated expression of mesenchymal markers such as vimentin and α-SMA, and decreased expression of epithelial markers, such as E-cadherin, α-catenin, and β-catenin, indicating that the induction of epithelial-to-mesenchymal transition (EMT) may explain the observed increases in motility and invasion ability for metastasis. Our data showed that *RPL22L1* plays an important role in the process of OC metastasis. # Materials and Methods ## Ethic statement This study were approved by the Ethics Committee of Harbin Medical University with the following reference number, HMUIRB20150023. The ovarian cancer tissue microarrays (TMAs) for immunohistochemistry were purchased from US BIOMAX (ov951, ov1912, ov6161; Rockville, MD, USA) and Xin Chao (HOva-Can90PT-01; Shanghai, China). Both companies provided ethical statements to confirm that the local ethics committees approved their consent procedures, all participants provided their written informed consents and all efforts had been made to protect patient privacy and anonymity. The ethical statements provided by companies and the protocol of experiment had been checked carefully and approved by Ethics Committee of Harbin Medical University (HMUIRB20150023). Four-week-old female BALB/c mice (specific-pathogen-free) were purchased from SLAC (Shanghai, China) and housed in the Harbin Medical University Animal Laboratory. Mice were housed under standardized light-controlled conditions at room temperature (24°C) and 50% humidity, with free access to food and water. Animal experiments were performed in strict accordance with the recommendations in the Guidelines of Laboratory Animal Usage of Harbin Medical University. The protocol was approved by the Ethics Committee of Harbin Medical University (HMUIRB20150023) and all efforts were made to minimize suffering. ## Cell lines and cell culture Human ovarian cancer cell line UACC-1598, SKOV3, HO-8910, and HO-8910PM were purchased from the Type Culture Collection of the Chinese Academy of Sciences (Shanghai, China). All cells were cultured following the methods described by the ATCC (Manassas, VA, USA), and were authenticated in 2012 at the Micro-read Genetics Company (Beijing, China) using a short tandem repeat analysis. ## Preparation of metaphase spreads and fluorescence *in situ* hybridization (FISH) analysis Cells were harvested for metaphase spread preparation according to the methods described in previous studies and were stained with Giemsa. Two BAC clones, GFP- RP11-355H10, which specifically covers the *MYCN* (amplified in UACC-1598 and located on the DMs), and Cy3-RP11-726H11 for *RPL22L1*, were selected as DNA probes and hybridized to metaphase spreads of cells as described previously. Chromosomes were counterstained with DAPI (4, 6-diamidino-2-phenylindole). Images were captured using a Leica DM5000 B fluorescence microscope (Wetzlar, Germany), and analyzed using the MetaMorph Imaging System (Universal Imaging Corporation, West Chester, NY, USA). ## DNA copy number and gene expression analysis The Oncomine DNA Copy Number Datasets (<https://www.oncomine.org/resource/main.html>) includes data deposited in The Cancer Genome Atlas (TCGA Ovarian 2 Dataset, <http://tcga- data.nci.nih.gov/tcga/>) was used to determine the differences in DNA copy number between OC and normal blood / ovarian tissues. 607 ovarian serous cystadenocarcinoma, 431 normal blood, and 130 normal ovary tissue samples were analyzed. Data were obtained through the following steps: type gene name: *RPL22L1* in the search box and in the browse tree select Primary Filters \> Datasets type \> DNA Copy Number Datasets. The result of TCGA Ovarian 2 is directly in the first, and graph was on the left. Click on the histogram button on the upper left corner of the graph title to get a histogram graph. On the group filters above the graph title, click on the triangle symbol, in the menu selected "Cancer and Normal Type" to make the analysis grouped by cancer and normal samples. Data access requires an academic email account. Authors should be contacted for login information if necessary. Two independent microarray gene expression datasets were used. Both datasets were generated using the Affymetrix Human Genome U133 Plus 2.0 Array platform (Santa Clara, CA, USA). The raw.CEL files for the two datasets were downloaded from the Gene Expression Omnibus (GEO) website (GSE27651 and GSE28450), and normalized using the RMA (robust multi-array average) algorithm. ## qRT-PCR Total RNA for each cell line was extracted using the High Pure RNA Isolation Kit (Roche, Basel-Stadt, Switzerland) following the manufacturer’s instructions. PrimeScript ^® RT Reagent Kit Perfect Real Time (Takara, Dalian, China) was used to reverse transcribed the total RNA. The cDNA was quantified using LightCycler ^® 480 SYBR Green I Master (Roche) in a LightCycler^™ 480 II Real Time System (Roche). The specific primers for human *RPL22L1* and *GAPDH* were as follows: *RPL22L1* forward primer, `5′-AGAAGGTTAAAGTCAATGG-3′` and reverse primer, `5′-ATCACGAAGATTGTTCTTC-3′`; *GAPDH* forward primer, `5′-ATCACTGCCACCCAGAAGAC-3′` and reverse primer, `5′- TTTCTAGACGGCAGGTCAGG-3′`. Expression of *RPL22L1* in samples was normalized to that of *GAPDH* and the fold-change in expression was calculated using the 2^-ΔΔCt method. ## Western blot analysis Cells were lysed at 4°C in RIPA (8990, Thermo Fisher Scientific, Waltham, MA, USA). Cell lysates containing 40μg of total protein from each sample were loaded onto 12% sodium dodecyl polyacrylamide gels and transferred to a polyvinylidene fluoride membrane (Millipore, Billerica, MA, USA). After blocking in 10% blocking solution (Roche), membranes were incubated overnight at 4°C with primary antibodies followed by 1-h incubation with anti-mouse/anti-rabbit secondary antibody (200-332-263/610-132-007, Rockland Immunochemicals, Limerick, PA, USA) at room temperature. Images were obtained using the Odyssey Infrared Imaging System (LI-COR Biosciences, Lincoln, NE, USA) and cropped using Adobe Photoshop CS5 (Adobe Systems Inc., San Jose, CA, USA), representative of five independent experiments. GAPDH was used as a control. The details of the primary antibodies are provided in the supplementary materials. ## Tissue microarray The OC tissue microarrays (TMAs) were purchased from US BIOMAX (ov951, ov1912, ov6161; Rockville, MD, USA) and Xin Chao (HOva-Can90PT-01; Shanghai, China). Approval for this study was obtained before it was initiated from the local regional ethics committee. Written informed consent was obtained from all the participants. ## Immunohistochemistry Immunohistochemistry (IHC) studies were performed using PowerVision^™ Two-Step Histostaining Reagent (Zhongshan Golden Bridge, Beijing, China) according to the manufacturer’s instructions. In brief, TMA sections were deparaffinized and rehydrated. For antigen retrieval, TMA slides were microwave- treated in 10mM citrate buffer (pH 6.0) for 8 min. Endogenous peroxidase activity was blocked with 3% hydrogen peroxide for 10 min. The TMA slides were incubated with a 1:50 dilution of monoclonal against human RPL22L1 (1:50) overnight at 4°C in a moist chamber. The slides were then sequentially incubated with goat anti-rabbit IgG antibody-horseradish peroxidase conjugates for 30 min at 37°C, and 3′–3′ diaminobenzidine was used as the chromogen substrate. Finally, all slides were counterstained for nuclei with hematoxylin, dehydrated, and mounted. For the negative control, the primary antibody was replaced with normal rabbit IgG. RPL22L1 immunoreactivity in TMA samples was evaluated by three pathologists. The intensity of immunoreactivity on TMAs was graded on a scale from 0 to 3 based on a consensus of the three investigators. ## Vectors *RPL22L1*-ORF was PCR amplified and cloned into the pcDNA3.1 (+) expression vector (Invitrogen, Carlsbad, CA, USA). The siRNA (h) and control siRNA were purchased from RiboBio (Q000200916-1-B, Guangzhou, China). The luciferase reporter plasmid pGL4.17 was kindly provided by Dr. ZH Zhong (Department of Microbiology, Harbin Medical University, Harbin, China). ## Transfection All plasmids and siRNAs were transfected into cells using Lipofectamine 2000 (Invitrogen) according to the manufacturer’s instructions. ## Nude mouse tumor xenograft model The animal experiments were approved by the Ethics Committee of Harbin Medical University (HMUIRB20150023) and performed according to the Guidelines of Laboratory Animal Usage of Harbin Medical University. Four-week-old BALB/c nude female mice were randomization assigned in each group, five mice per group. Each mouse was injected with 1 × 10⁶ cells (in 200μl of phosphate-buffered saline \[PBS\]). Tumor growth was measured twice a week via bioluminescence imaging. Mice were injected intraperitoneally with 150μg / g D-luciferin (Biosynth, Naperville, IL) in PBS and anesthetized with 2.5% isoflurane. Photons emitted from the mice were recorded by IVIS Lumina (Advanced Molecular Vision, Lincolnshire, UK) and presented as pseudo-color images overlaid on a gray-scale body image. All mice were euthanized by CO₂ inhalation 30 days after injected. To ensure death following CO₂ asphyxiation, cervical dislocation was performed. ## Cell migration and invasion assays Transwell cell migration and invasion assays were performed using Corning 8.0-mm Transwell inserts (8-μm pore size, 24-well plate) and BD BioCoat^™ Matrigel^™ Invasion Chambers (Corning Incorporated Life Sciences, Tewksbury, MA, USA) according to the manufacturer’s instructions. ## Wound healing assay The cell monolayer was scratched with a 10-μl pipette tip (in a 6-well plate). Photographs (magnification, ×40) were taken immediately and at each 24-h post- wounding until the wounds were obviously healed. For each assay, the experiments were performed at nine different positions and the whole assay was repeated three times. ## Immunofluorescence staining Cells were fixed in methanol and blocked for 1-h with 10% normal goat serum, 0.3% bovine serum albumin, 0.05% saponin, and 0.3% Triton X-100 in PBS. The primary antibodies were added and incubated at 4°C overnight. Cells were later washed with PBS and incubated with fluorescence-labeled secondary antibody. A fluorescence microscope (Leica) was used to capture the images. ## Statistics The RPL22L1 protein expression levels of TMAs were analyzed by Wilcoxon’s signed-rank tests. The χ² test was used to examine associations between gene expression and clinical parameters. Other data were expressed as means ± SD, and the statistical significance of differences between two groups was evaluated using two-tailed independent Student’s *t*-tests. Statistical significance was declared if *P* \< 0.05. Calculations were carried out using SPSS 13.0 (IBM; Armonk, NY, USA). # Results ## *RPL22L1* was amplified via DMs and over-expressed in UACC-1598 cells The OC cell line UACC-1598 contained amplified genes in the form of DMs. Oncogene amplification on DMs is representative of tumorigenesis, but does not occur in normal cells. Using a FISH analysis, we detected amplified regions of *RPL22L1* that co-localized with *MYCN* on DMs. Furthermore, we detected the amplification of *RPL22L1* in normal ovarian tissues, human lymphocytes, and UACC-1598 cells using PCR. We also examined the amplification of *RPL22L1* in another three ovarian cancer cell lines. RT-PCR confirmed the expression of *RPL22L1* in different samples. These results indicated that *RPL22L1* was amplified via DMs in UACC-1598 cells. ## DNA copy number and expression level of *RPL22L1* in clinical OC samples To determine the amplification of *RPL22L1* in clinical, we used the Oncomine database (<https://www.oncomine.org>) to analyze the DNA copy number profiles of the *RPL22L1* in a TCGA ovarian dataset (TCGA Ovarian 2, <http://tcga- data.nci.nih.gov/tcga/>). This dataset clearly indicated a significant increase in the DNA copy number of *RPL22L1* in OC compared to normal blood and ovarian tissues, threshold by *P* \< 10⁻⁴. Further, we obtained publically available GEO expression datasets to analyze the expression of *RPL22L1* in OC samples. Two GEO datasets based on the Affymetrix Human Genome U133 Plus 2.0 Array were used to evaluate *RPL22L1* mRNA expression levels in OC and normal ovarian tissues. In each of the two datasets, *RPL22L1* expression was significantly higher in OC than normal ovarian tissues (*P* \< 0.05, Student’s t-test). To further examine the protein expression level of *RPL22L1* in clinical specimens, four OC TMAs were used for IHC analyses. Staining intensity was estimated using an index of 0–3 in the cell nucleus and cytoplasm. RPL22L1 was clearly expressed more highly in OC cytoplasm than in that of the adjacent normal tissues (*P* = 0.008, Wilcoxon’s signed-rank test). We examined the associations between RPL22L1 protein expression levels and clinicopathologic variables. RPL22L1 expression in cytoplasm of OC cells was strongly associated with stage, invasion, and lymph node metastasis (*P* \< 0.05, Pearson’s χ² test). These results indicated that the expression of *RPL22L1* is commonly higher in clinical OC specimens, and its expression level is associated with tumor progression, especially with invasion and lymph node metastasis. ## *RPL22L1* enhanced intraperitoneal xenograft tumor development *in vivo* To detect the role of high level *RPL22L1* in tumor progression, we selected an OC cell line with low RPL22L1 expression, SKOV3, for further functional studies. SKOV3 cells were stable transfection with luciferase previously and then stable transfection with *RPL22L1*. To explore the role of *RPL22L1* in tumor progression *in vivo*, we intraperitoneally injected SKOV3-RPL22L1 and control cells into nude mice. The xenografted tumors were measured using bioluminescence at the 30^th day after injection, area of xenograft tumor in mice with SKOV3-RPL22L1 cells injected was larger than that of control cells injected. The result suggested that high level of *RPL22L1* contribute to tumor development. ## *RPL22L1* promoted OC cell migration and invasion To confirm the role of *RPL22L1* in OC cells, UACC-1598 cells were treated with three specific siRNAs against *RPL22L1* (siRNA-1, siRNA-2, and siRNA-3). Since siRNA-2 exhibited the most efficient knockdown of endogenous *RPL22L1*, it was chosen for subsequent analyses. Except SKOV3-RPL22L1 and the control cells, we used another two OC cell lines HO-8910 and HO-8910PM with lower *RPL22L1* expression to transient transfected with *RPL22L1* for further functional study. The effect of *RPL22L1* on cell motility was characterized by wound-healing, transwell migration, and Matrigel invasion assays. Knockdown of *RPL22L1* in UACC-1598 cells (1598-siRPL22L1) caused clear ression of cells migration and invasion. Over-expression of *RPL22L1* in SKOV3-RPL22L1, HO-8910 (8910-RPL22L1), and HO-8910PM (8910PM-RPL22L1) cells could significantly enhance cells migration and invasion (*P* \< 0.05). Cell growth rate was analyzed by an MTA assay, *RPL22L1* had no influence on cell proliferation (data not shown). These results suggested that high level of *RPL22L1* enhances OC cells migration and invasion. ## The expression level of *RPL22L1* influences OC cell line EMT It has become increasingly clear that EMT is an integral component of the progression of epithelial-derived tumors. We found that SKOV3-RPL22L1 cells exhibited a spindle-shaped and fibroblastic morphology whereas 1598-siRPL22L1 cells exhibited shrinkage shapes and increased cell–cell adhesion. Therefore, we examined whether *RPL22L1* induced EMT could account for the *RPL22L1*-mediated changes in cells motility and invasion. Biochemical hallmarks of EMT include the loss of expression of epithelial marker proteins and concurrent increase in mesenchymal marker expression. Western blots and immunofluorescence analyses were used to evaluate the expression of epithelial and mesenchymal markers. After ectopic over-expression of *RPL22L1* in SKOV3 cells, the expression of epithelial makers, such as E-cadherin, β-catenin, and α-catenin decreased, whereas the expression of vimentin, α-SMA, and fibronectin increased. In 1598-siRPL22L1 cells, the expression levels of the mesenchymal markers vimentin and N-cadherin were degraded. These results suggested that the expression level of *RPL22L1* influences EMT in OC cells. # Discussion DMs are malignant cytogenetic markers and are closely correlated with tumor progression. Previously, we determined that *RPL22L1* amplified on DMs, but its function is unclear. Many oncogenes are amplified via DMs in malignant tumor cells \[34.38.39\], such as *EIF5A2* and *MYCN*, both of which are located at the same locus as *RPL22L1* on DMs. We inferred that *RPL22L1* may be involved in OC progression, and verified this in a series of *in vitro* and *in vivo* assays. Using publically available TCGA and GEO expression datasets we found both DNA copy number and mRNA expression of *RPL22L1* was significantly higher in OC tissues than in normal ovarian tissues. Protein expression of *RPL22L1* was examined by IHC using four OC TMAs. We found that RPL22L1 expression was frequently higher in OC tissues compared with normal adjacent ovarian tissue (*P* = 0.008, Wilcoxon’s signed-rank test). Further, we found the expression level was significantly correlated with disease stage (81.7% in stage 2–4 versus 64.5% in stage 1) invasion depth (81.9% in T2–T4 versus 64.5% in T1), and lymph node metastasis (86.6% with versus 70.3% without metastasis), suggesting that the high level of RPL22L1 in OC cells may facilitate the invasive/metastatic phenotype. These findings underscore a potentially important role of *RPL22L1* as an underlying biological mechanism in the development and progression of OC. The results of the IHC analysis indicated that *RPL22L1* may be involved in the pathogenesis of OC progression especially in metastasis. To confirm this hypothesis, nude mouse tumor xenograft, wound-healing, transwell migration, and Matrigel invasion assays were carried out to investigate the role of *RPL22L1* in regulating OC cells motility, invasion. Over-expression of *RPL22L1* obviously promoted tumor development *in vivo* and increased the migration and invasion ability of three OC cell lines *in vitro*. Moreover, the knockdown of endogenous *RPL22L1* by siRNA in UACC-1598 cells inhibited migration and invasion *in vitro*. Migration and invasion are two important stages in tumor metastasis, and these functional assays strongly suggested that *RPL22L1* plays an important role in promoting tumor metastasis via enhancing cell migration and invasion. Many biological processes are associated with migration and invasion including EMT, a key event in tumor invasion and metastasis. Over-expression of *RPL22L1* reduced the expression of epithelial marker proteins (E-cadherin, β-catenin, and α-catenin) and increased the expression of mesenchymal markers (vimentin, α-SMA, and fibronectin), which are biochemical hallmarks of EMT. However, after knockdown of *RPL22L1* in UACC-1598 cells, the mesenchymal markers vimentin and N-cadherin were obviously down-regulated. EMT plays a critical role in promoting metastasis and during this process cells obtain migration and invasion capabilities, which promote metastasis. Accordingly, we inferred that the over-expression of *RPL22L1* likely induced EMT to promote cell invasion and metastasis. RPL22L1 is a paralog of RPL22, which is a RNA-binding protein component of the 60S ribosomal subunit. Ribosomal proteins are major components of ribosomes where cellular proteins are synthesized. To date, approximately 80 ribosomal proteins have been identified. In addition to their key roles in protein synthesis, some ribosomal proteins are involved in extra-ribosomal functions, such as DNA repair, apoptosis, transcription regulation, and translation regulation. An increasing number of reports have suggested that ribosomal proteins have oncogenic potential. Recently, *RPL22* has been found to be mutated or down-regulated in various cancers, including T-acute lymphoblastic leukemia, invasive breast carcinoma, and lung adenocarcinoma. Additionally, *RPL22* directly represses the expression of *RPL22L1* mRNA, and a lack of RPL22 causes a compensatory increase in RPL22L1. These studies ort our conjecture regarding the role of *RPL22L1* in cancer progression. Taken together, our results suggested that *RPL22L1* plays an important role in OC progression by enhancing cell invasion and metastasis via inducing EMT. As *RPL22L1* is a DM carried amplified gene, our data could help explain the function of DMs in cancer. *RPL22L1* is a novel gene that is correlated with metastasis, and studies aiming to elucidate its function and modes of action may improve our understanding of the mechanisms underlying metastasis in OC. # Supporting Information The statistics analysis was guided by Dr. Yan Liu (Department of statistics, Harbin Medical University). We thank Mr. Yu-Zhen Zhao for the guidance in cell culture. We thank Mrs. Zhe Wang and Mrs. Xi-Lin Cui for the help for preparation of reagents and experiment consumables. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: YJ SF. Performed the experiments: NW JW YW JY YQ BP. Analyzed the data: NW CZ FC DS HS. Contributed reagents/materials/analysis tools: DT JB JG YY. Wrote the paper: NW JW KL WS XM.

# Introduction Melanogenesis is the physiological process by which melanin is synthesized in melanocytes located in the basal layer of the epidermis to protect the skin from UV irradiation. UVB-exposed keratinocytes secrete cytokines and growth factors, including endothelin (EDN) 1, that stimulate cellular functions, especially proliferation and melanization, of adjacent melanocytes in the epidermis. The corresponding specific receptors are constitutively expressed by human melanocytes and the binding of cytokines and growth factors to their receptors transduces intracellular signals to initiate melanogenesis through specific signaling cascades. On the other hand, UVB radiation directly induces the generation of reactive oxygen species (ROS) in epidermal keratinocytes and melanocytes and stimulates stress activated protein kinases, such as p38, c-jun N-terminal kinase (JNK) or extracellular regulated protein kinase (ERK). In UVB- exposed human melanocytes, the p38 pathway predominantly contributes to the increased expression of microphthalmia-associated transcription factor (MITF), a master regulator of melanocyte functions, including differentiation, proliferation, survival and melanogenesis. MITF regulates the expression of many melanogenic enzymes, melanosome structural proteins, transporters and receptors, such as tyrosinase, tyrosinase-related protein 1 (TYRP1), dopachrome tautomerase (DCT), melanosomal protein 17 (PMEL17), melanoma antigen recognized by T-cells 1 (MART1) and endothelin B-receptor (EDNRB). EDN-EDNRB binding is one of the key paracrine interactions between keratinocytes and melanocytes that regulates skin pigmentation \[, –\]. EDN1 is a vasoconstrictor peptide originally isolated from porcine endothelial cells. We first reported that human keratinocytes produce a prepro-EDN1 and then convert it by metallo-proteinases including EDN-converting enzyme α, sequentially to big-EDN1 and EDN1, which is the final secretable form. UVB-exposed human keratinocytes distinctly enhance the secretion of EDN1, which triggers adjacent melanocytes in the epidermis via EDNRB to stimulate melanin synthesis. EDNRB is a seven-transmembrane receptor coupled with G-protein that interacts equally with all forms of EDN, EDN-1, EDN-2 and EDN-3. Mutations of those genes causes Waardenburg Syndrome Type IV, which is an auditory-pigmentary syndrome characterized by hearing loss, abnormal pigmentation of skin, hair and eyes in association with Hirschsprung disease, which is a disorder that causes blockage of the intestine. The role of the EDN-EDNRB interaction was reported to induce mitogenesis and melanogenesis in melanocytes. Although EDN secretion from keratinocytes stimulated by UVB has been well investigated, little is known about EDNRB expression in UVB-exposed melanocytes. EDNRB expression has been shown to increase in the epidermis when human skin is exposed to solar-stimulated radiation or UVB radiation and in skin with lentigo senilis. The finding that a dominant-negative mutant of MITF reduces the expression of EDNRB in cultured melanocytes strongly suggested that EDNRB expression is predominantly regulated by MITF. Skin pigmentation is a major factor that prevents the skin from UV-induced damage. Pigmented skin is unwanted by people who desire a lighter skin color, and many natural products have been utilized historically for cosmetic purposes in order to obtain a lighter skin appearance. As depicted in, the French maritime pine (Pinus maritima) bark extract (PBE) is a complex mixture of flavonoids, which contains 72.5% polyphenol (determined by Folin Denis method) including 5% procyanidin B1, 2.98% catechin, 0.23% epicatechin and about 60% (including the percentage of dimer) oligomeric proanthocyanidin (OPC). PBE has been used as a traditional medicine for scurvy by maritime Indians. PBE has potent antioxidant properties and oral administration of PBE has protective effects on age-related diseases, such as cardiovascular dysfunction, diabetes and arthritis. It was also reported that PBE by itself is highly effective in protecting the skin from UV irradiation. Kim et al. demonstrated that PBE inhibits melanogenesis not via inhibition of tyrosinase but rather by inhibiting the autoxidation of melanin due to its antioxidant activity. In a clinical study, oral administration of PBE at 40 or 100 mg daily for 12 weeks reduced the pigmentation of age spots. Here we show that the expression of EDNRB is accentuated in UVB-exposed human melanocytes via activation of the p38/MSK1/CREB/MITF pathway where MSK1 activation is essentially responsible for CREB activation. Post-irradiation treatment with PBE does not affect p38 activation but can directly interrupt the UVB-induced activation of MSK1, which leads to abrogation of the UVB-induced up- regulation of melanocyte-specific proteins such as EDNRB. Thus, it is anticipated that PBE can serve as an anti-pigmenting agent in a ROS depletion independent manner. # Materials and Methods ## Materials Anti-MITF (C5), anti-EDNRB (EPR7013), anti-CREB (48H2), anti-phospho-CREB (87G3), anti-β-actin (AC-15), anti-rabbit IgG HRP-conjugated and H89 dihydrochloride were purchased from Abcam (Cambridge, MA). Anti-mouse IgG HRP- conjugated was purchased from Jackson ImmunoResearch (West Grove, PA). Antibodies for MAPK and phosphorylated MAPK, the MAPK family sampler kit and the phospho-MAPK family sampler kit were purchased from Cell Signaling Technology (Beverly, MA). Antibodies for MSK1 (C27B2) and phosphorylated (S376 and T581) MSK1 were purchased from Cell Signaling Technology. For Real-time RT-PCR, primers for β-actin (Hs_ACTB_1\_SG Quantitect Primer Assay; QT00095431), EDNRB (Hs_EDNRB_1\_SG Quantitect Primer Assay; QT00014343) and MITF (Hs_MITF_1\_SG Quantitect Primer Assay; QT00037737) were purchased from Qiagen (Hilden, Germany). PBE (Flavangenol) which obtained by hot water extraction method from French maritime pine (*Pinus pinaster*) bark was supplied by Toyo Shinyaku (Saga, Japan). ## Melanocyte culture Primary normal human epidermal melanocytes (NHMs) pooled from 250 individual human foreskins were purchased from Cell Systems (Kirkland, WA) and were maintained in Dermalife Ma culture medium (Lifeline Cell Technology, Walkersville, MD) supplemented with all of the supplements from the manufacturer. ## UVB source The UVB source employed in this study was a Phillips TL20W/12RS lamp (Phillips, Eindhoven, Holland). The energy exposed was measured using a UVX radiometer with a UVX-31 sensor (UVP Inc., San Gabriel, CA). ## UVB irradiation and PBE treatment NHMs were plated in 6-well plates at a density of 1×10⁵ cells per well in complete medium. Twenty-four h later, NHMs were washed with warmed phosphate buffered saline (PBS) once and irradiated once with 60 mJ/cm² UVB in a thin layer of warmed PBS, with the lid removed. Complete medium with or without the indicated concentration of PBE was added to the well immediately after the UVB irradiation and the plates were then cultured for the indicated periods. Non-irradiated NHMs were subjected to the identical procedure but without UVB irradiation. H89 treatment was carried out instead of PBE at the indicated concentration. ## NHM viability NHMs were plated in 96-well plates at a density of 1×10⁴ cells per well in complete medium. Twenty-four h later, the medium was removed and NHMs were washed with warmed PBS once and irradiated once with the indicated energies of UVB with the lid removed. Complete medium with or without the indicated concentration of PBE was added to the well immediately after the UVB irradiation and the plates were then cultured for 24 h. Viable NHMs were determined by a colorimetric assay with a Cell counting kit 8 (Dojin Chemical, Kumamoto, Japan), according to the manufacturer’s protocol. ## Real-time RT-PCR Total RNAs from NHMs cultured for the indicated times were prepared using an RNeasy mini kit (Qiagen, Hilden, Germany) according to the manufacturer’s protocol. Reverse transcription and Real-time PCR reaction were used with a QuantiTect Reverse Transcription kit and a Rotor-Gene SYBR PCR kit with the gene specific primer of β-actin as a reference and the gene of interest described in Materials section according to the manufacturer’s protocol. The Real-time PCR reaction and the signal detection were carried out with Rotor-Gene Q (Qiagen, Hilden, Germany) and data analyses were carried out with Rotor-Gene Q Series Software (Qiagen, Hilden, Germany). ## Western blotting analysis At the end of the culture, NHMs were washed twice with ice cold PBS and were lysed in RIPA buffer with the Halt Protease Inhibitor Cocktail and Halt Phosphatase Inhibitor Cocktail (Thermo Scientific, Rockford, IL). Amounts of total protein were quantitated using BCA protein reagent (Thermo Scientifc, Rockford, IL). Total proteins (5 μg/lane) were denatured by heating at 95°C in Laemmli sample buffer (BioRad, Richmond, CA) for 5 min and loaded onto 10% sodium dodecyl sulfate (SDS)-polyacrylamide gels (BioRad, Richmond, CA). After electrophoresis, proteins were transferred onto Polyvinylidene difluoride (PVDF) membranes and were immunoblotted with appropriate primary and secondary antibodies. Immunoblotted proteins were visualized using an ECL substrate (BioRad, Richmond, CA) and were detected and analyzed by ChemiDoc XR+ System and Image Lab software (BioRad, Richmond, CA). ## Statistical Analysis All data are expressed as means ± SD (n = 3) unless noted otherwise. For pairwise comparisons, either Student’s t-test or Welch’s t-test was applied. For multiple comparisons, data were tested by one-way ANOVA, and subsequently using the Tukey or Dunnett multiple comparison test. P values less than 0.05 are considered statistically significant. # Results ## Effect of PBE and UVB on the viability of NHMs We examined the effect of UVB irradiation and/or PBE treatment on the viability of NHMs. While treatment with PBE slightly enhanced the cell viability at concentrations of 10–30 μg/ml, it did not decrease the cell viability at concentrations less than 60 μg/ml. UVB irradiation had no affect on the viability of NHMs at energy doses less than 60 mJ/cm², but had a distinct effect on cell viability at a dose of 120 mJ/cm². The addition of PBE to UVB-exposed NHMs at a concentration of 30 μg/ml had no substantial influence on the viability of NHMs at energy doses of less than 60 mJ/cm². ## UVB stimulates the expression of EDNRB in NHMs, which is abrogated by post-irradiation treatment with PBE We have already reported that EDNRB expression is increased by the exposure of human skin to UVB. However, little was known about the biological mechanism(s) by which UVB irradiation stimulates EDNRB expression in the epidermis in vivo. Hence, we examined the effects of UVB irradiation at a dose of 60 mJ/cm² on the expression of EDNRB in NHMs. Real-time RT-PCR analysis revealed that the mRNA expression level of EDNRB was significantly increased by UVB irradiation (60 mJ/cm²) at 24 h but not at 6 or 12 h post- irradiation. When added immediately after UVB irradiation at a dose of 60 mJ/cm², the enhanced expression of EDNRB mRNAs at 24 h post- irradiation was significantly abrogated by PBE at concentrations of 10, 20 and 30 μg/ml. Western blotting analysis demonstrated that EDNRB protein levels were significantly increased at 24 h post-UVB irradiation (60 mJ/cm²), which was significantly abrogated by post-irradiation treatment with PBE at a concentration of 30 μg/ml. ## UVB stimulates MITF expression in NHMs, which is abrogated by post-irradiation treatment with PBE Sato-Jin et al. reported that transfection of the dominant negative form of MITF suppressed the expression of EDNRB mRNA and suggested that EDNRB gene expression occurs downstream of MITF. Therefore, we examined MITF expression when NHMs were exposed to UVB irradiation and then were treated with PBE. Real-time RT-PCR analysis revealed that, when exposed to UVB at 60 mJ/cm², the expression level of MITF mRNA was increased at 6, 12 and 24 h post-irradiation with a peak at 6 h post-irradiation. When treated post-irradiation with PBE at a concentration of 30 μg/ml, the increased expression levels of MITF mRNA were significantly abrogated by PBE at 6 and 24 h post-irradiation. Further, the enhanced expression of MITF mRNA at 6 h post-irradiation was significantly abrogated by PBE at a concentration of 20 and 30 μg/ml in a fashion similar to those observed for the EDNRB mRNA expression. Western blotting analysis demonstrated that the MITF protein level was significantly increased by UVB irradiation with 60 mJ/cm² at 12 h post-irradiation, which was significantly abrogated by post-irradiation treatment with PBE at a concentration of 30 μg/ml. ## CREB phosphorylation is attenuated by PBE Cyclic AMP response element-binding protein (CREB) is a transcription factor regulating MITF gene expression, which is activated by phosphorylation in response to various signaling molecules. Since the expression of MITF and its downstream target gene EDNRB was up-regulated by UVB irradiation at the transcriptional and translational levels, and that could be abrogated by post- irradiation treatment with PBE, we next determined if the phosphorylation of CREB is increased by UVB irradiation and/or whether the post-irradiation treatment with PBE can abrogate the CREB activation. Western blotting analysis using an antibody to phosphorylated CREB revealed that the phosphorylation level of CREB was significantly increased by UVB irradiation with 60 mJ/cm² at 15 min post-irradiation, which was significantly abrogated by the post- irradiation treatment with PBE at a concentration of 30 μg/ml. These results indicate that the up-regulation of MITF protein level is mediated via CREB activation in UVB-exposed NHMs and the post-irradiation treatment with PBE abrogates the CREB activation. ## PBE interrupts the phosphorylation of MSK1 but not ERK, JNK and p38 We have already reported that in UVB-exposed NHMs, the generated ROS triggers p38 and JNK but not ERK activation, leading to their downstream target CREB activation predominantly via p38 activation. Based on this evidence, we next determined which signaling molecule(s) upstream of CREB are attributable to the interruption of CREB phosphorylation by post-irradiation treatment with PBE. As expected, while UVB irradiation stimulated the phosphorylation of p38 and JNK but not of ERK, the post-irradiation treatment with PBE did not abrogate the increased phosphorylation of p38 and JNK, which suggests that the interruption of CREB phosphorylation by PBE is not attributable to its effect on p38 activation. In UVB-exposed human primary keratinocytes, the activated p38 is known to stimulate nuclear kinase mitogen-and stress activated kinase (MSK)1 which phosphorylates CREB and NFkBp65 in the nucleus during the NFkB signaling pathway. Therefore, we next determined if UVB radiation stimulates the phosphorylation of MSK1 in NHMs and/or if PBE can serve as an inactivator for MSK1 even when treated post-irradiation. Western blotting analyses revealed that the phosphorylation of Ser376 and Thr581 residues of MSK1 was significantly increased 15 min following UVB irradiation, which was significantly abrogated by PBE when treated post-irradiation at 30 μg/ml. This suggests that the interruption of CREB phosphorylation by PBE is attributable to its abrogating effect on MSK1 activation. We next asked if the inhibition of MSK1 activation results in the down-regulated MITF and EDNRB expression in UVB-exposed NHMs. When the MSK1 inhibitor H89 was added to NHMs immediately after UVB irradiation, the increased expression level of MITF and EDNRB mRNA elicited by UVB irradiation was significantly abrogated by H89. This suggests that the abrogation of UVB-stimulated expression of EDNRB via MITF transcription by the post-irradiation treatment with PBE is mediated via the interruption of MSK1 activation. # Discussion In this study, we found that a single exposure of NHMs by UVB stimulates EDNRB expression. Since the increased levels of EDNRB seem to respond to EDN1 to a greater extent than in unexposed NHMs, that finding suggests that UVB causes NHMs to become highly responsive to environmental stimuli such as EDN1 via an increased expression of the corresponding receptor, EDNRB. Consistently, we have already found that KIT ligand (KITL) up-regulates the expression of EDNRB in NHMs where the binding of ¹²⁵I-labeled EDN1 to EDNRB increases significantly 2 days after incubation with KITL. Similarly, we reported that a single exposure of NHMs with UVB stimulates expression of the KIT receptor, whose function was assessed by an increased phosphorylation following KITL stimulation. Thus, it is likely that in addition to the increased production of melanogenic cytokines by UVB-exposed keratinocytes, EDNRB also plays a coordinated role in UVB-induced pigmentation by augmenting EDN1/EDNRB signaling through the accentuated function of EDNRB. In support of this, in UVB-exposed human skin where pigmentation is being stimulated, there is a significantly up- regulated expression level of EDNRB mRNA. However, little is known about intracellular signaling mechanisms involved in the stimulation of EDNRB expression in UVB-exposed NHMs. Anti-pigmentation agents have been developed as a target for various redox- sensitive biomolecules, including tyrosine hydroxylase or intracellular signaling intermediates during the melanogenesis cascade. Compounds including phytochemical agents or botanical extracts are adequate candidates for this purpose due to their distinct anti-oxidant properties. In this study, we found that a French maritime PBE containing rich flavonoids including OPC distinctly abrogates the increased expression of EDNRB at the transcriptional and translational levels following UVB radiation even when treated post-irradiation. PBE has a distinct antioxidant activity stronger than vitamin C and vitamin E as measured by lipid peroxidation of bovine retinal tissue. Additionally, PBE possesses a potent scavenging activity for peroxynitrite (ONOO-), superoxide (•O₂) and nitric oxide (NO•), which play a central role in inhibiting the generation of these ROS. Further, PBE up-regulates the reduced- glutathione/oxidized-glutathione ratio. Owing to these strong antioxidant properties, it was anticipated that PBE has a potential to inhibit pigmentation by preventing the autoxidation of melanin. Since PBE can behave as an antioxidant and a scavenger for ROS generated by UVB irradiation, its possible inhibitory effect on the increased expression of EDNRB could be accounted for by the depletion of generated ROS if treated pre-irradiation. However, our observation that post-irradiation treatment with PBE can also abrogate the increased EDNRB expression strongly suggests that PBE abrogates the up- regulation of EDNRB expression via an unknown novel signaling mechanism(s) in a ROS depletion-independent manner because the ROS lifetime is very short (e.g. lifetime of •O₂ is 4 μs), not sufficient to deplete the generated ROS when treated immediately after UVB radiation. UVB exposure of human keratinocytes was reported to activate NFκB signaling by stimulating IKK kinase which phosphorylates IkB, causing NFκBp65 to transduce toward translocation into the nucleus during the signaling pathway downstream of the preceding p38 or JNK activation. In contrast, UVB exposure of human melanocytes induces little or no activation of the NFκB pathway compared to the distinct activation of their upstream pathways such as p38 and JNK. In melanocytes and melanoma, UVB has been shown to induce phosphorylation of the p38 and JNK/stress-activated protein kinase pathways, whereas NFkB remains at a constantly high expression level. The activation of p38 or JNK following UVB radiation is mediated by initial stress-activated protein kinases, which are activated by ROS via redox-interfering mechanisms involved in protein kinases as well as their conjugated protein phosphatases. Owing to these mechanisms, many antioxidants can suppress UVB-induced cellular events by scavenging generated ROS when treated pre-irradiation. This evidence indicates that the hitherto reported inhibitory effects of antioxidants on the UVB-induced activation of IKKinase leading to the diminished nuclear translocation of NFκB may occur via the abolishing effect on the activation of p38 or JNK due to the preceding ROS depletion by pretreatment with antioxidants. Therefore, it is of considerable importance to determine the signaling mechanism(s) by which the post-irradiation treatment with PBE can abrogate the increased EDNRB expression. We have already reported that EDNRB gene expression occurs downstream of the melanocyte-master transcription factor MITF. Consistently, in this study, the gene and protein expression levels of MITF are significantly up-regulated by UVB radiation, and can be significantly abrogated by the post-irradiation treatment with PBE. This suggests that the up-regulated EDNRB expression by UVB radiation is mainly associated with the increased protein expression level of MITF and the abrogating effect of PDE on the increased expression of EDNRB is mainly attributed to the down-regulated level of MITF protein. In NHMs, at the terminal point of the EDN1-triggered intracellular signaling cascade, the gene expression levels of melanocyte-specific proteins including EDNRB are strictly associated with the steady state levels of MITF protein. The MITF gene expression level is positively regulated by the levels of activated (phosphorylated) CREB in association with other transcription factors including SOX10, PAX3, lymphoid-enhancing factor-1 (LEF-1) and T cell factor (TCF). Therefore, the abrogating effect of PBE on the up-regulated expression of MITF led us to determine whether CREB phosphorylation is stimulated by UVB radiation and whether PBE abrogates this stimulation. As expected, UVB exposure of human melanocytes distinctly stimulates the phosphorylation of CREB, which is abolished by the post-irradiation treatment with PBE. This suggests that the abrogating effect of PBE on the up-regulated protein expression of MITF is mainly attributed to the interruption of CREB activation. Therefore, we next determined how the CREB is activated by UVB radiation in human melanocytes. In our previous similar study focusing on KIT receptor expression in UVB-exposed human melanocytes, the inhibition of p38 activation by its inhibitor SB203580 results in the complete abrogation of both the up-regulated phosphorylation of CREB and the increased gene expression levels of MITF up to the non-stimulated control levels. This suggests that the increased phosphorylation of CREB by UVB irradiation is mediated predominantly via the activation of p38 but not the cyclic AMP/PKA pathway. In this study, in agreement with our results and another study, UVB exposure of human melanocytes significantly stimulates the phosphorylation of p38 and JNK but not of ERK, whereas the increased phosphorylation of p38 and JNK is not abrogated by the post-irradiation treatment with PBE. Since p38 cannot directly phosphorylate CREB, these findings strongly suggest that the post-irradiation treatment with PBE affects signaling intermediates capable of phosphorylating CREB, which occur downstream of p38 activation. There are at least four protein kinases that have a distinct ability to phosphorylate CREB, protein kinase A (PKA), p90 ribosomal protein S6 kinase (p90RSK), MAPK-activated protein kinase-2 (MK2) and MSK1. MSK1 has a Km value much lower than the other 3 kinases, all of which are distinctly activated by p38 or ERK. Therefore, we next determined whether MSK1 is activated by UVB radiation in human melanocytes and whether the post-irradiation treatment with PBE can abrogate the MSK1 activation. MSK1 is generally expressed in epidermal keratinocytes and, as shown in , is activated by p38 MAPK or the ERK p44/42 MAPKs through phosphorylation of either Thr581 or Ser360. The phosphorylation of Ser360 is an essential requirement for MSK1 activation. Further, Ser376 is auto-phosphorylated as a result of the phosphorylation at Ser360 and Thr581 by either ERK1/2 or p38 MAPK activation. However, little is known about the role of MSK1 in the signaling pathways leading to melanogenesis in human melanocytes. Western blotting analysis of MSK1 activation revealed that the phosphorylation of MSK1 at Thr581 and Ser376 was distinctly accentuated by UVB radiation, and could be significantly abrogated by the post-irradiation treatment with PBE. Since in this study ERK is not activated by UVB irradiation and PBE has no affect on ERK phosphorylation, the above findings strongly suggest that MSK1 is a signaling target of PBE, leading to the abrogation of CREB activation in UVB-exposed human melanocytes when treated post-irradiation. In this study, we also corroborated that the MSK1 inhibitor H89 significantly abrogates the increased gene expression level of MITF and EDNRB even when treated post-irradiation. This strongly indicates that MSK1 inhibition leads to the attenuated expression of MITF and EDNRB. Although the abrogated expression of MITF and EDNRB may also be attributable to the inhibition of cAMP-dependent PKA by H89, this possibility can be ruled out by the fact that the activation of CREB in UVB-exposed human melanocytes is mediated predominantly via the activation of p38 but not the cyclic AMP/PKA pathway, an indication that H89 treatment could not abrogate the UVB-stimulated expression of MITF and EDNRB via an interruption of the cAMP/PKA pathway. This is the first report showing that the MSK1 activation is essentially involved in the CREB activation in UVB-exposed human melanocytes and an antioxidant can directly interrupt UVB-induced MSK1 activation, which leads to the abrogation of UVB-induced up-regulation of melanocyte-specific proteins such as EDNRB. In conclusion, as shown in, our findings indicate for the first time that the increased expression of MITF leading to the up-regulation of the melanocyte- specific protein EDNRB in UVB-exposed human melanocytes is mediated via the activation of the p38/MSK1/CREB pathway but not of the ERK/RSK/CREB pathway. The mode of action by PBE demonstrates that the interruption of MSK1 activation is a new target for antioxidants including PBE which can serve as anti-pigmenting agents in UVB-melanosis. This study provides a deep insight into understanding of signaling mechanisms involved in UVB-accentuated expression of melanocyte- specific proteins as well as the regulatory role of redox-sensitive MSK1 in the UVB-activated melanogenic signaling pathway in human melanocytes. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: HT GI. Performed the experiments: HT AM SK MT. Analyzed the data: HT AM SK. Contributed reagents/materials/analysis tools: KY KT. Wrote the paper: HT GI.

# Introduction It is now widely accepted that a network of brain regions, distributed across the frontal and parietal cortices, form the components of an adaptable global system for the deliberate and intentional control of thought and action. This global ‘executive’ system underlies the flexibility of human behaviour, by enabling us to deliberately and selectively focus our attention on those items that are currently of relevance to the task at hand. Whilst the existence of this network is no longer controversial, the contributions made by the anatomically distinct components from which it is comprised remain poorly defined. For example, to date, there have been several influential models proposing a dorsal-ventral axis across the lateral portion of the prefrontal cortex. These include the suggestion that the dorsolateral prefrontal cortex (DLPFC) and the ventrolateral prefrontal cortex (VLPFC) are differentially involved in exogenous vs. endogenous attentional orienting, first order vs. higher order executive functions , and the active maintenance vs. the controlled manipulation of items in working memory. Much of the current confusion regarding the precise nature of frontoparietal organisation results from the use of complex and cognitively heterogeneous task manipulations when attempting to functionally dissociate frontoparietal sub-regions. Hence, functional dissociations are often hard to interpret, with the (sometimes rather specific) cognitive functions that the tasks seek to examine typically being confounded with other more global parameters such as the general level of difficulty, and the overall level of engagement. Target detection paradigms, in which the individual monitors a sequence of distractor objects for a learnt target stimulus, allow the effects of the relevance of the attended stimulus to the current task set to be examined whilst minimising variations in the complexity of required task parameters from one condition to another. Previously, we have examined the way in which different sub-regions of the frontoparietal network tune to respond selectively to the presentation of a frequently redefined target object whilst undertaking a simple event-related fMRI task. We reported that whilst regions across the frontoparietal network were sensitive to the presentation of current target objects, their response was not homogeneous. The VLPFC, particularly in the right hemisphere, responded with a high degree of specificity to the current target object. Whilst our previous results clearly identify the VLPFC as being particularly sensitive to the presentation of current targets, a number of questions regarding the precise nature of that sensitivity remain unresolved. Most importantly, in our previous task, the target selective response could be explained in terms of two popular hypotheses, which are commonly confounded. The first hypothesis - derived primarily from the findings of non-human primate single unit recording - relates to the type of information that is represented within the frontoparietal network. More specifically, it has been suggested that the frontoparietal network rapidly adapts to code for those items that are relevant to the currently intended goal. In this case it would be predicted that the target selective response should resemble a simple similarity function, with the level of response related directly to the level of congruence between the currently attended object, and the recently defined target. An alternative hypothesis, however, derived primarily from neuroimaging research, refers to the type of cognitive demands under which the frontoparietal network is typically recruited. It has been observed that the BOLD signal within this network increases when the level of difficulty is parametrically varied across a wide variety of different task contexts. On this basis, it has been proposed that the frontoparietal network forms a global system for attention that is engaged whenever the general level of difficulty increases and effortful executive control is exercised. If the difficulty hypothesis is correct, then the BOLD response of frontoparietal regions should be predicted by the proximity of the currently presented object to the 50% target/distractor decision boundary, where the response decision is at its most ambiguous. Another pertinent question relates to the fact that much of our current understanding of the nature of the information represented within the lateral prefrontal cortex comes from the non-human primate electrophysiology literature. The tasks used in these studies are almost invariably extensively pre-trained to ensure good task performance. Herein lies a question regarding the relevance of results from these studies when seeking to understand how the frontoparietal network contributes to normal human behaviour. The frontoparietal cortex has often been proposed to play a particularly important role in *novel situations* by exerting deliberate ‘top-down’ or ‘executive’ control over those systems that would otherwise be governed by more habitual/learnt responses. This top-down executive influence from the frontoparietal network thereby facilitates flexible/adaptable behaviour. A further question, therefore, concerns whether the target selective response within frontoparietal sub-regions varies as a function of increasing task familiarity, and if so, in what way? Here, we addressed these questions using a modified version of our original task design. Volunteers monitored sequences of visually displayed objects for the presentation of a current target item. Distractor stimuli were morphed at varying degrees of similarity to the current target object, and the BOLD response could therefore be measured at each of these degrees of similarity. As the 50% decision boundary and the target object were at different positions on this similarity scale, it was possible to examine whether functions corresponding to the probability of positive identification (similarity) and distance from the 50% decision boundary (ambiguity) played significant roles in predicting the BOLD response. Furthermore, because volunteers undertook three identical blocks of experimental acquisition, it was possible to examine how the selective tuning functions varied as the task became increasingly familiar. # Results ## Behavioural results Twenty volunteers monitored sequences of visually displayed objects for the presentation of a current target item. At the beginning of each sequence a new target item was presented with the word ‘target’, subsequent to which presentation of objects began. Responses, however, were made only when cued at the end of the sequence. Responses were cued by the question ‘was the last stimulus the target?’ and referred only to the last object. In this way, all critical events were kept free from overt motor activity. The lengths of the sequences were varied unpredictably, and within a given sequence, the current target could appear at any or multiple points to ensure attention throughout. To allow the target selective BOLD response to be examined in detail, monitored sequences were comprised of objects at six degrees of similarity to the current target, these being; the current target object (target), morphs one through three, distractors from the same category as the target (same type), and distractors from a different category to the target (other type). In the behavioural analysis, the proportions of positive responses were examined in an ANOVA in which the conditions were similarity (target, morph 1, morph 2, morph 3, same type, other type)\*experimental acquisition block (blocks 1 through 3). The analysis showed a significant interaction of block\*similarity (F_(1,19) = 4.98 p\<0.05), a significant main effect of similarity (F_(1,19) = 525.41 p\<0.001), and a main effect of block (F_(1,19) = 16.11 p\<0.001). The pair-wise comparisons between block 1 and 3 confirmed this result with significantly lower probabilities of positive response for morph 2 and same type distractors in the final block (target t = 0.25 p = 0.80; morph 1 t = −1.52. p = 0.14; morph 2 t = −2.90. p\<0.01; morph 3 t = −1.00 p = 0.32; same type t = −3.94 p\<0.001; other type t = 0.04 p = 0.97). In general the behavioural data reveal a small but significant trend towards increased selectivity across the three blocks. ## Plotting the target selective tuning functions in sub-regions of the frontoparietal network Our first analysis examined responses to the different possible stimulus types in order to examine the question of whether target selective tuning functions varied between different frontoparietal sub-regions, and also to examine whether they varied within those sub-regions across the three blocks of experimental acquisition. Group level analyses were carried out using focused regions of interest (ROIs) representing the DLPFC, the VLPFC, and the posterior parietal cortex (PPC). Data extracted from the frontal and parietal ROIs for the presentation of targets, morphs, and distractors, were examined using repeated measures analyses of variance (ANOVA). The first ANOVA examined the effects of similarity to the current target object averaged over the three blocks of experimental acquisition. The conditions were ROI (VLPFC, DLPFC, PPC)\*hemisphere (left, right)\*similarity to the target (target, morph 1, morph 2, morph 3, same type, other type). The within subject effects revealed a significant interaction of hemisphere\*similarity (F_(1,19) = 17.34 p\<0.001), and a significant interaction of ROI\*similarity (F_(1,19) = 8.07 p\<0.001). There were also significant main effects of similarity (F_(1,19) = 12.42 p\<0.001), and hemisphere (F_(1,19) = 20.96 p\<0.001), and ROI (F_(1,19) = 6.51 p\<0.005). The interactions indicated that different ROIs followed different selective tuning functions, and they were therefore examined separately in a series of one way ANOVAs in which the condition was similarity (target, morph1, morph 2, morph 3, same type, other type). There were strong main effects of similarity in the ventrolateral prefrontal cortex, particularly in the right hemisphere (VLPFC left F_(1,19) = 12.31 p\<0.001, VLPFC right F_(1,19) = 38.36 p\<0.001). There were also significant main effects of similarity in the right DLPFC and the right PPC (DLPFC left F_(1,19) = 1.63 p = 0.16, DLPFC right F_(1,19) = 10.38 p\<0.001; PPC left F_(1,19) = 1.98 p = 0.09, PPC right F_(1,19) = 10.79 p\<0.001). Overall, the results revealed that within the frontoparietal network, the target selective response was greatest in the VLPFC with a general lateralisation effect favouring the right hemisphere. ## Examining the effects of task familiarity on the target selective tuning functions in the frontoparietal network The data were then examined for the effects of task familiarity. The ROIs were examined separately in a series of two way ANOVAs, in which the conditions were similarity (target, morph1, morph 2, morph 3, same type, other type)\*experimental acquisition block (block 1, block 2, block 3). A significant interaction of similarity\*acquisition block was observed in the right VLPFC (left F_(1,19) = 2.22 p = 0.15; right F_(1,19) = 5.88 p\<0.05). The other ROIs displayed no significant familiarity\*similarity interactions. The nature of this learning effect would appear to be a shifting of the peak of the tuning function from morph 1 in the first acquisition block, towards the target in the third acquisition block. This shifting of the tuning function peak was explored further by calculating the average peak position on a voxel by voxel basis across the right lateral prefrontal cortex. A general shift in the peak position from morph 1 to the target was apparent across the lateral prefrontal cortex. This change in the peak position of the tuning function could be accounted for by a general shifting in frontoparietal resources away from resolving the ambiguous target-distractor decision at morph 1 and towards recognition of the target as the task becomes more familiar. For block 1, a direct contrast between morph 1 distractors and the target generated no significant results. To investigate whether this effect was more reliable when the task was at its most novel, the data were remodelled for the first half of session 1 only. Contrasting morph 1 distractors vs. the target using FDR correction for the whole brain mass at p = 0.05 revealed a significantly greater BOLD response in the right IFG (x = 46 y = 10 z = 26 and x = 34 y = 26 z = −4), in the left IFG (x = −36 y = 18 z = −2), and in the right PPC (x = 30 y = −58 z = −56). This finding confirms that the resolution of ambiguous target-distractor decisions recruited frontoparietal resources to a particularly large extent when the task was novel. ## Ambiguity and similarity as predictor functions A further analysis was carried out to test whether the BOLD response in frontal and parietal sub-regions was best accounted for in terms of a) perceptual similarity to the current target object, b) the difficulty of the current target/distractor discrimination, or c) a combination of these two cognitive factors. To address this issue, we examined the extent to which functions derived from the behavioural data representing the probability of positive identification (similarity to the target -), and proximity to the 50% decision boundary (degree of ambiguity –), could predict the BOLD response within the same frontoparietal sub-regions. Group level analyses were carried out using the focused ROIs representing the DLPFC, the VLPFC, and the PPC. In each case, regressors were formed by weighting the onsets and durations of stimulus presentation with the behavioural similarity and ambiguity functions prior to convolution with the canonical haemodynamic response function. In the group level analysis, we first examined each frontoparietal ROI for the positive effects of ambiguity averaged across the three acquisition blocks. Ambiguity played a significant role in predicting the BOLD response in both the right VLPFC, and right PPC (right VLPFC t = 2.45 p\<0.01; right PPC t = 2.65 p\<0.005). Whole brain analysis (FDR corrected for the whole brain mass at p = 0.05) confirmed the results from the ROI analysis, with significant BOLD activation in the ventrolateral prefrontal cortex bilaterally. It should be noted that the peak VLPFC co-ordinates for the ambiguity regressor were located posterior and medial to our ROI between BA 44/BA 47 and the anterior insula, and the activation spread across the anterior insula and the inferior operculum. There were also significant activation peaks in the right DLPFC, right pre-motor cortex, right PPC and right occipital cortex. The frontoparietal ROIs were then examined for significant positive effects of similarity to the target averaged across the three acquisition blocks. There were large significant effects of similarity in the VLPFC bilaterally, the right DLPFC, and the right PPC (left VLPFC t = 4.01 p\<0.001; right VLPFC t = 6.37 p\<0.001; right DLPFC t = 3.52 p\<0.001; right PPC t = 3.76 p\<0.001). Whole brain analysis confirmed the results of the ROI analysis, with significant BOLD activation throughout much of the frontoparietal network for the positive effect of similarity, including the VLPFC bilaterally, the PPC bilaterally, and the right DLPFC. In addition, a network of other brain regions was activated, including visual cortex, temporal cortex, the anterior insula, pre-motor cortex, the anterior cingulate, the pre-SMA, and areas within the striatum. Overall, therefore, the response within the frontoparietal network, particularly within the right VLPFC, was best predicted by a combination of both the ambiguity and the similarity functions, with similarity especially important. Examination of the ROI data separately for each acquisition block indicated that there was a general trend towards increased weighting on the similarity regressor, and decreased weighting on the ambiguity regressor across the three acquisition blocks. We examined the significance of this trend in a full factorial model in SPM 5 in which the factors were predictor (ambiguity or similarity)\*acquisition block (block 1, block 2, block 3). Our results revealed a significant interaction of acquisition block\*predictor function in the VLPFC bilaterally (left F = 5.09, p\<0.01; right F = 14.77, p\<0.001), in the right DLPFC (left F = 0.92, p = 0.40; right F = 3.07, p = 0.05), and in the right PPC (left F = 2.31, p = 0.1; right F = 8.94, p\<0.001), indicating that with practice, similarity becomes relatively more important than ambiguity in predicting the BOLD response across the frontoparietal network. The whole brain analysis did not reveal any peak activation foci for the block\*predictor interaction at the corrected threshold. # Discussion The advantage of using a simple target detection paradigm to investigate frontoparietal function is that it enables the selectivity of the BOLD response to be examined whilst minimising differences in the complexity of the current task parameters. In this tightly controlled context, any observed results must be driven by the similarity of the currently attended stimulus to the object that is at the focus of currently intended behaviour (i.e. the target). Here, the use of target-distractor morphs has allowed us to examine the target selective BOLD response in the human frontoparietal network at a higher degree of acuity than has previously been possible. Our results reveal that a broad swathe of cortex rapidly adapts to respond selectively to the current target object. In line with models that posit a global/adaptive system for working memory and attention, this ability appears to be generalised across different stimulus categories. The target selective network includes a large swathe of frontal and parietal cortex, including the PPC, the DLPFC, and the VLPFC. The selective tuning functions are not homogeneous throughout the frontoparietal network, however, with distinct sub-regions displaying significantly greater sensitivity to the current target object. Previously, we have reported that the ventral portion of the lateral prefrontal cortex is particularly sensitive to the presentation of target objects. On this basis, we have suggested a degree of specialisation within the frontoparietal network, with the more ventral and posterior portion of the lateral prefrontal cortex tuning to respond to those items that are at the current focus of intended action with a particularly high degree of selectivity. This specialisation is replicated here, with heightened activity in the VLPFC compared to other regions of the frontoparietal cortex, including the anatomically adjacent DLPFC. With the increased power afforded by the current design, however, it is clear that this apparent specialisation is quantitative as opposed to absolute, with other frontal and parietal regions following similar shaped tuning functions, but to a lesser extent. Previously, we have also reported a lateralisation effect favouring the right hemisphere during target detection. This finding is replicated again here, with heightened target related activation in the right hemisphere throughout the frontoparietal network. This lateralisation effect is most prominent in the DLPFC, with the left DLPFC appearing to be almost completely insensitive to the presentation of the current target object. Whilst it is now clear that frontal and parietal regions are consistently more activated in the right hemisphere during target detection, the question still remains whether this lateralisation effect is due to the right hemisphere being more involved in the detection of targets, or to the type of stimuli used. One way of testing the possibility that the lateralisation effect is due to the type of stimuli would be to replicate the current task design, but with words instead of objects. One might predict that, in such a situation, the lateralisation effect could be reversed to favour the left hemisphere. It is also important to note that the left DLPFC may play a less transient role in target detection, a hypothesis that cannot be tested here due to the rapid event related design not allowing the estimation of a resting state baseline. In our previous study the right DLPFC was observed to respond at a more categorical level than the VLPFC, with similar increases observed in the BOLD signal during the presentation of both targets and distractors from the same category. Here, the previous findings were only partially replicated, with the particular sensitivity of the right VLPFC to target objects appearing to be robust across experiments, but the wider tuning of the DLPFC appearing to be more sensitive to the exact task parameters. This lack of replication when task demands are changed is a running theme in studies that seek to functionally dissociate the DLPFC and the VLPFC. Hence, whilst dissociations have been reported, subsequent studies that use similar task manipulations often report that the VLPFC and DLPFC follow a similar activation profile. One relatively constant factor, however, is that when these anatomically distinct sub-regions of the lateral prefrontal cortex *are* functionally dissociated, the VLPFC tends to be implicated in simple executive functions, for example the maintenance of items in working memory, , whereas the DLPFC tends to be implicated in more complex, although not necessarily more difficult task demands such as manipulating, monitoring, and structuring items in working memory. It seems sensible to propose, therefore, that differences between the DLPFC and the VLPFC are statistical as opposed to absolute, with both brain regions capable of supporting similar cognitive processes. Under certain conditions, however, the roles played by these two brain regions may dissociate and when they do, they dissociate in a hierarchical manner. This study was designed not only to replicate our previous findings, but also to address two key questions using the higher degree of acuity afforded by the use of morphed distractors. 1) Which cognitive factors can predict the target selective response in the frontoparietal network? 2) Are the target selective tuning functions static, or do they change as a function of task familiarity? We addressed the first of these questions using two cognitive predictor functions. The first, similarity, represented the probability of positive vs. negative response at each of the six degrees of similarity to the current target item. This similarity function relates most closely to findings from the electrophysiology literature in which frontal neurons have been observed to respond selectively to a broad range of task-relevant information, for example responses, rewards, and learnt target stimuli. Based on the electrophysiology findings it seems sensible to predict that the better the currently attended object matches the item that the currently intended action plan is programmed around, the more it will activate the frontoparietal network, which is assumed to represent the currently relevant objects, actions, and task criteria. Another popular hypothesis posits that the frontoparietal network forms a highly adaptable system that is recruited whenever the general level of cognitive demand increases. This latter hypothesis is repeatedly supported by the neuroimaging literature, which tends to reveal increased BOLD signal in the frontoparietal network when a wide variety of cognitive demands are parametrically increased. The second cognitive predictor function was based, therefore, on how close the probability of target identification was to 50%: this was highest, when the target/distractor decision was at its most ambiguous, and resolution of this ambiguity required maximal processing. It is important to note that these two hypotheses are not mutually exclusive and it was our aim to disentangle them in order to test whether either or both played a significant role in predicting the BOLD response. Our results demonstrate that both similarity to the target item and degree of ambiguity in the target/distractor decision play a significant role in predicting the target selective BOLD response. Our results did not simply show an activation profile sharply peaked for the target, as predicted by the similarity regressor. Neither did they show a profile sharply peaked for the most ambiguous stimulus morph 1, as predicted by the ambiguity regressor. Instead the balance of activity between target and morph 1 varied over regions and stages of practice. It is clear from the results that the selective tuning functions are not static, particularly in the right VLPFC, where the BOLD response becomes increasingly selective as the task becomes more familiar. This increased selectivity can be interpreted in terms of a redistribution of cognitive resources. Hence, the ambiguity of the target-distractor decision places a particularly high demand on frontoparietal resources at the earliest stages of the task, when the task parameters and stimulus set are novel. Conversely, the similarity of the attended stimulus to the current target object plays a larger role in predicting the BOLD response in the later stages of the task, when the task parameters and stimulus set are familiar. Herein lies a question over the relevance of findings from much of the current electrophysiology literature when attempting to understand the contribution of the frontoparietal network to normal human behaviour. The results of selective frontal lobe lesions have often been used to suggest that the frontoparietal network plays a particularly important role when dealing with novel problems. However, the vast majority of electrophysiology experiments use extensive pre-training, and it seems sensible to suggest, therefore, that the findings from those studies relate to the way in which neurons within this network maintain attention to, and solve, routine, habitual problems. Our results would suggest that, with practice, frontoparietal processing related to ambiguity/cognitive demand is at best minimised, and may therefore appear to be less significant in heavily pre-trained studies. By contrast, the extent of adaptive tuning to the currently relevant objects increases with learning, and would therefore seem to be more representative of the frontoparietal role in attention in studies that employ over-learnt tasks. It remains to be answered whether the current effects of familiarity on the target selective response relate to learning of the stimulus set, or a lowering of general engagement as the task becomes increasingly familiar. In either case, our data confirm the importance of task familiarity in frontoparietal function. An important aside is the relevance of the current findings to those studies of inhibitory control that have reported activation throughout a very similar network. Of particular relevance to the current findings is the increased BOLD response in the right VLPFC during the suppression of a routine motor output following an infrequent stop cue. During Go/NoGo tasks, the maintained task program is to look for an infrequent and previously learnt cue to stop, and on receiving that cue to interrupt a routine motor response. It is plausible to suggest that a large component of the ‘inhibition’ condition in the Go/NoGo task is recognition of the cue to stop, a process that is very similar to identifying a learnt target stimulus. The process of subsequently stopping the routine response is probably facilitated by the ‘top-down’ biasing signals that are widely held to be the primary mechanism by which control is exercised by the executive system. Whilst this process could be described as inhibition, it could also be described as the *implementation of the currently maintained task program*. In that respect, it should be noted that this manipulation differs from inhibitory control in the more classical sense of an *effortful change in the current task program,* which usually occurs as a consequence of previously rewarded responses leading to sub-optimal feedback from the environment. Inhibitory control in this more classic sense is known to rely on additional frontal lobe circuitry, most particularly sub-regions of the orbitofrontal cortices. Finally, we have presented here a working proof that tuning in simple target detection is a useful scale for measuring the degree of attentional focus in brain activity. Here this scale has been used to compare tuning functions across distinct frontoparietal regions of interest, and across varying levels of task familiarity. The same method may well be useful for testing a variety of hypotheses, for example, differences in attentional selectivity across clinical populations, and under varying cognitive and pharmacological conditions. # Materials and Methods ## Experimental design Volunteers were instructed to look for a visually displayed target object within sequences of distractor objects. At the beginning of each sequence a new target item was presented with the word ‘target’ for 3400 ms. Subsequent to the target stimulus being defined, presentation of the sequence of targets and distractors began. Each item of the sequence was displayed for 1500 ms and was followed by an inter-stimulus-interval of 400 ms. Sequences were predefined and pseudo- randomised. The sequence length was varied unpredictably from 1 to 8 items, and within a given sequence, the current target could appear at any or multiple points. At the end of each sequence a probe stimulus consisting of the question ‘Was the last stimulus the target?’ was presented on the screen for 3400 ms, and volunteers were required to respond yes or no, using a button box with the first two digits of their right hand. The words ‘yes’ and ‘no’ also appeared below the probe, randomly assigned to the left and the right of the display, indicating which buttons to press for the positive and negative response. Critical contrasts were therefore kept free of overt motor activity, whilst attention was ensured throughout the monitored sequence. Twenty healthy right-handed volunteers between the ages of 20 and 40 undertook the fMRI task, which consisted of 3\*12 minute blocks of scanning acquisition, each containing 40 stimulus sequences. Targets, same category distractors, and other category distractors, were drawn from the same fixed set of stimuli, consisting of five standard objects from each of four distinct categories: faces, rooms, line figures, and abstract shapes. To allow the target selective BOLD response to be examined in detail, morphs were generated between all standard objects of the same category, at three physically equidistant degrees of similarity (for example stimuli see). The monitored sequence consisted, therefore, of objects at six degrees of similarity to the current target, these being; the current target object (target), morphs one through three, distractors from the same category as the target (same type), and distractors from a different category to the target (other type). Within each block of scanning acquisition, volunteers monitored 22 targets, 22 morphs from each of the three degrees of similarity, 22 same type distractors, and 74 other type distractors. Each standard object was used as the target twice in a block of scanning acquisition, and multiple times as a distractor. The presentation of targets, morphs, and distractors was balanced across the experimental block so that the relative probabilities were equivalent across all eight positions in the stimulus sequence, in this way averaging out any effects due to reconfiguration to a new target object, or the expectancy of an impending probe. The sequences were identical across the three experimental blocks to ensure maximum cross block comparability when examining the effects of learning. ## Scanning acquisition Scanning was carried out at the MRC Cognition and Brain Sciences Unit using a 3 Tesla Siemens Tim Trio. 32\*3 mm slices (1 mm inter-slice gap, descending slice order) were acquired in 2 seconds for each image (in-plane resolution 3×3 mm). 360 T2-weighted echo-planar images depicting BOLD contrast were acquired per block of scanning acquisition, with the first 10 discarded to avoid T1 equilibrium effects. The experiment was programmed in Visual Basic 6 and the display projected onto a screen, visible from the scanner via a mirror, with stimuli subtending a visual angle of approximately 6.5 degrees. Images were pre-processed and analysed using the Statistical Parametric Mapping 5 software (SPM5, Wellcome Department of Cognitive Neurology). Prior to analysis, images were slice time corrected, reoriented to correct for subject motion, spatially normalised to the standard Montreal Neurological Institute template, smoothed with an 8 mm full-width at half-maximum Gaussian kernel, and high-pass filtered prior to analysis (cut-off period 180 s). ## Event modelling Two separate fixed effects analyses were carried out on each volunteer's data using general linear models. The first design examined how the BOLD response varied when the participant was presented with objects at different degrees of similarity to the target. This model was used to examine the question of whether the selective tuning functions varied between the different frontoparietal sub- regions, and also to examine whether they varied within those sub-regions across the three blocks of experimental acquisition. 20 regressors were included in this model, with the onset and duration of each picture presentation event described according to three orthogonal parameters. For each event, the first descriptor was similarity to the target object, with 6 levels: target, a morph at one of the three degrees of similarity to the target, a ‘same type’ distractor, or an ‘other type’ distractor. The second descriptor was object category with 4 levels: faces, rooms, abstract line figures, and abstract shapes. The third descriptor was temporal position in the monitored sequence (positions 1 through 8). The target definition stage was included as a further regressor, and the final regressor was formed from the onsets and durations for the probes at the end of the sequences with the corresponding motor responses. Regressors were created by convolving these timing functions with a basis function representing the canonical haemodynamic response. Group level analyses were carried out using focused regions of interest (ROIs) representing different sub-regions of the frontoparietal network. 10 mm radius spherical ROIs were defined bilaterally in the DLPFC, the VLPFC, and the PPC, based upon averaged coordinates taken from a previous analysis of common frontal and parietal activity associated with diverse cognitive demands. The centre points of these regions were located at ±38, 30, 22 for the DLPFC, ±39, 20, 2 for the VLPFC, and ±31, −51, 40 for the PPC. For each participant, the level of response to each of the six degrees of similarity to the target (targets, morph 1, morph 2, morph 3, same type, other type) was estimated using fixed effects analysis. These data were averaged across voxels within each of the ROIs using the MARSBAR toolbox, and the mean values were exported for analysis using SPSS. The second linear model was identical to the first, except that the regressors corresponding to the six degrees of similarity to the target item were replaced with two new regressors, weighted according to two predictor functions, similarity and ambiguity. The behavioural data, averaged across all participants, from the responses to probes was used to plot the predictor functions, separately for each of the three acquisition blocks. The similarity function was defined by calculating the probability of a positive vs. a negative response across the six degrees of similarity to the target object. The ambiguity function was estimated from the response data by taking the un-signed result of 0.5 minus the probability of positive response, and then subtracting this value from 0.5. This renders a function that is maximal for a positive decision probability of 0.5, and at zero for a probability of 0 or 1. The onsets and durations for monitored objects were then weighted according to the two predictor functions, to form the two new regressors. These regressors were convolved with the canonical haemodynamic response function and, to control for scaling, were normalised by dividing by the root mean square value of the entire regressor before entry into the design matrix. In the group level analysis, each volunteer contributed six whole brain images, containing the parameter estimates for the similarity and ambiguity regressors, separately for each block of scanning acquisition. [^1]: Conceived and designed the experiments: JD AO AH RT. Performed the experiments: AH. Analyzed the data: AH. Contributed reagents/materials/analysis tools: AH. Wrote the paper: AH. [^2]: The authors have declared that no competing interests exist.

# Introduction Malaria is a deadly infectious disease caused by protozoan parasites of the genus *Plasmodium*. The recently released World Malaria Report estimated that malarial parasites infected over 200 million people worldwide causing 627,000 deaths in 2012. The increasing incidence of drug resistance, absence of an effective vaccine and lack of diversity amongst current compounds in development renders this ancient disease an ongoing global health problem. Novel anti- malaria therapeutic approaches are urgently required to confront these challenges. The blood stage of *Plasmodium* infection is the major cause of the clinical symptoms of malaria and the mechanism of erythrocyte invasion is highly conserved in all apicomplexan parasites. Therefore, proteins involved in this process have been actively pursued as targets for both vaccine and drug development. Apical membrane antigen 1 (AMA1), an integral membrane protein that is highly conserved throughout the phylum Apicomplexa, represents one of these protein targets. The initiation of merozoite invasion is marked by formation of the moving junction (MJ), a ring-like protein structure, between the merozoite and the erythrocyte. In our current understanding of the structure and function of the MJ, AMA1 presents a conserved hydrophobic cleft that interacts with rhoptry neck protein 2 (RON2). This interaction is essential to the formation of the junction, which commits the parasite to invade. Both AMA1 and RON2 are provided by the parasite to enable an active invasion mechanism. AMA1 is initially stored in the parasite micronemes and subsequently translocated to the merozoite surface before invasion, while RON2 is secreted from the parasite rhoptry and transferred to the erythrocyte surface prior to invasion. The essential role of AMA1 in host cell invasion has been questioned recently by genetic studies, which showed AMA1-depleted parasites can still form a functional MJ. As such, the specific role of AMA1 in host cell invasion remains a matter of debate, , but it is clear that inhibition of the AMA1-RON2 interaction by various agents effectively disrupts invasion and validates AMA1 as a viable therapeutic target. Specifically, antibodies raised against AMA1 can inhibit invasion by binding to the hydrophobic cleft, although the inhibition is usually strain-specific. Consistent with these observations, AMA1 evolves under strong selective pressure from the host immune system, and loops surrounding the hydrophobic cleft are polymorphic. Nonetheless, the AMA1-RON2 interaction is highly conserved. In addition, the interaction between AMA1 and RON2 can be inhibited by peptides. One such peptide, R1, was identified from a random peptide library using phage-display. R1 showed a high binding affinity for 3D7 *Pf*AMA1 (*K*_D∼0.08 µM) and spans the full-length of the hydrophobic cleft. Comparison with the structure of a complex between AMA1 and a peptide derived from RON2 reveals that the two peptides occupy the same region of AMA1 and exhibit structural mimicry. Consistent with these structural studies, R1 can effectively inhibit erythrocyte invasion by malaria parasites *in vitro*. Although the inhibition is strain-specific, it has been demonstrated that *N*-methyl modification of R1 broadened its strain specificity. It is evident from the current data that effective targeting of AMA1 from multiple strains requires inhibitors whose interaction is mediated by conserved residues within the hydrophobic cleft, which bind AMA1 without making extensive contact with polymorphic residues. It is likely that this goal will be more easily realized by using smaller molecules as inhibitors. We and others have recently reported the identification of small molecules that bind to AMA1, with the goal of developing these molecules into therapeutically useful antimalarials. A common problem faced in small molecule inhibitor design is difficulty in improving the binding affinity and specificity of screening “hits”. Identification of binding “hot spots”, *i.e.* the subset of residues at the binding interface that contribute most of the free energy to high affinity binding, provides important information to guide the design of high-affinity ligands. This is especially critical for targeting protein-protein interactions (PPIs). As R1 has high binding affinity and makes extensive interactions with the hydrophobic cleft of AMA1, characterization of the AMA1-R1 interaction provides valuable insights into the key interactions that contribute to binding. Indeed, there are many examples showing that small molecule inhibitors can be designed that mimic the interaction of a peptide with a protein target. In the current study we have undertaken a detailed biophysical characterization of the interaction of R1 with AMA1 and used computational solvent mapping to identify hot spots at the binding interface. Collectively our data provide a rational basis for designing high-affinity inhibitors of AMA1-RON2 interaction. # Materials and Methods ## Expression and purification of AMA1 Domain I+II of 3D7 *Pf*AMA1 (residue 104–442) was expressed, purified and refolded as described. The folding of the purified protein was assessed by monitoring its binding affinity and stoichiometry to R1 using surface plasmon resonance and recording a 1D ¹H spectrum, which is characterized in the correctly-folded material by the presence of several upfield-shifted methyl protons (Figure S1). Randomly fractionally deuterated (f-²H) AMA1 was prepared by growth of expression cultures in 100% ²H₂O/M9 minimal medium supplemented with ¹⁴NH₄Cl (1 g/L) and protonated ¹²C-D-glucose (10 g/L). The high-cell-density method was implemented to achieve high protein yield as described in. The hexahistidine (His₆) tag of AMA1 was cleaved by tobacco etch virus (TEV) protease in a ratio of 0.02 mg TEV per mg fusion protein in phosphate buffer, pH 8.0 at 4°C for 24 h. The resultant protein was purified on a linear gradient of 0–500 mM NaCl using HiTrap QFF column chromatography (GE healthcare) and dialyzed against 20 mM ammonium bicarbonate solution at 4°C over 2 days before it was lyophilized. ## DNA manipulation, expression and purification of R1 R1 peptide was produced recombinantly as an enterokinase-cleavable fusion to thioredoxin. An insert encoding DDDDKVFAEFLPLFSKFGSRMHILK was ligated into pET32a (Novagen) at KpnI/NcoI and transformed into *Escherichia coli* BL21 (DE3). The f-²H, u-¹³C, ¹⁵N-labelled R1 fusion was expressed in 100% ²H₂O/M9 minimal medium supplemented with ¹⁵NH₄Cl (1 g/L) and protonated ¹³C₆-glucose (4 g/L) using the high-cell-density method as described in. The cells were harvested by centrifugation at 5,000 g for 20 min and resuspended in lysis/wash buffer (20 mM Tris-HCl pH 8, 20 mM imidazole, 200 mM NaCl). The cells were lysed by sonication and the supernatants were recovered by centrifugation at 12,000 g for 30 min at 4°C. The His₆-tagged R1 fusion in the soluble fraction was purified on a linear gradient of 45–500 mM imidazole by HisTrap column chromatography (GE healthcare). Fractions were analyzed by SDS-PAGE and those containing a band consistent with the expected size of the R1 fusion (∼20 kDa) were pooled and dialyzed against enterokinase cleavage buffer (20 mM Tris-HCl pH 7.4, 50 mM NaCl, 2 mM CaCl₂, 1 mM EDTA) overnight at 4°C. The fusion protein was then incubated with recombinant enterokinase (Novagen) in a ratio of 0.5 units enterokinase per mg fusion protein at room temperature for 21 h. The sample was then filtered through a 0.22 µm membrane (Millipore, Merck) and purified using HiTrap QFF column chromatography using a gradient of 0–500 mM NaCl in a buffer of 20 mM Tris-HCl pH 8. R1 peptide was finally purified by prep-RP-HPLC using a Phenomenex Luna 5 u C18 column (100×10 mm). The identity and purity ( \>95%) were confirmed by liquid chromatography mass spectrometry (LCMS) (Figure S2). About 1 mg of f-²H, u-¹⁵N, ¹³C-labelled R1 was produced from 0.7 L of minimal medium. ²H incorporation was ∼72%, ¹⁵N incorporation was ∼90%, and ¹³C incorporation was ∼95%. ## Synthetic R1 analogues Truncated and mutant R1 peptides used in the SPR study were synthesized by Mimotopes (Melbourne, Australia) with purity \>90% and all were *N*-terminally acetylated and *C*-terminally amidated. ## NMR sample preparation NMR samples were prepared in a buffer consisting of 20 mM sodium phosphate pH 7, 1 mM EDTA, 0.01% (w/v) sodium azide, 0.2% (w/v) Complete protease inhibitor cocktail (Roche), 50 mM Arg, 50 mM Glu and 6% (v/v) ²H₂O unless noted otherwise. For the NMR study of free R1, two samples of u-¹³C, ¹⁵N-labelled R1 at a concentration of 0.4 mM were prepared at pH 5 and pH 7, respectively. To study the AMA1-R1 complex, lyophilized f-²H-labelled AMA1 was added to a sample of f-²H, u-¹³C, ¹⁵N-labelled R1 to give final concentrations of AMA1 and R1 of 320 and 300 µM, respectively. Based on the measured *K*_D of R1 for AMA1, \>90% of the peptide should be bound to AMA1 under these conditions. ## NMR spectroscopy NMR experiments for free R1 were performed at 5°C or 40°C at a ¹H frequency of either 500 MHz or 600 MHz on Bruker Avance spectrometers equipped with a TXI-cryoprobe. Chemical shift assignments were made using the following experiments: 2D ¹H-¹⁵N-HSQC, ¹H-¹³C-HSQC and 3D triple-resonance experiments including HNCACB, CBCA(CO)NH, HBHA(CO)NH and HCCH-TOCSY. All spectra were processed using NMRPipe and analyzed with CARA. All NMR experiments for the AMA1-R1 complex were performed at 40°C in a 5 mm Shigemi tube. The backbone H^N, C^α, and N resonances of f-²H, u-¹³C, ¹⁵N-R1 bound to f-²H-AMA1_104–442 were assigned using 2D ¹H-¹⁵N-TROSY HSQC/conventional ¹H-¹⁵N-HSQC, 3D TROSY-HNCA and TROSY-HN(CO)CA. The 3D TROSY-HNCA was acquired on a Bruker DRX-900 spectrometer equipped with a cryoprobe. Non-uniform sampling was utilized during acquisition, with sampling points chosen randomly from a probability distribution matching the signal decay, as described previously. The spectra were re-constructed using the maximum entropy method with automated parameter selection using the Rowland NMR toolkit. A ¹³C(F₂)-¹H(F₃) plane of the 3D TROSY-HN(CO)CA was acquired on a Bruker Avance 600 MHz spectrometer. The data were processed using NMRPipe or Topspin 3.0 (Bruker-Biospin) and analyzed with CARA. Chemical shifts are reported relative to sodium 2,2-dimethyl-2-silapentane-5-sulfonate (DSS). ## Surface plasmon resonance binding analysis A Biacore T200 biosensor instrument was used to measure the affinity of the interaction of peptides with 3D7 *Pf*AMA1_104–442. AMA1 was immobilized onto a CM5 chip as described. Surface plasmon resonance (SPR) experiments were performed at 25°C using HBS-EP (10 mM HEPES, 150 mM NaCl, 3.4 mM EDTA, and 0.05% surfactant P20, pH 7.4) as the running buffer either with (alanine scanning mutagenesis study) or without (truncation study) 1% DMSO. All peptide samples were prepared in the appropriate running buffer. To generate the peptide binding data, peptide at concentrations ranging from 10 nM to 10 µM was injected over immobilized AMA1 at a constant flow rate of 60 µL/min for 1.5 min; peptide dissociation was monitored by flowing running buffer at 60 µL/min for 5 min. The surface was regenerated after each cycle by injecting glycine/HCl at pH 2.0. Sensorgrams were first zeroed on the y-axis and then x-aligned at the start of the injection. Bulk refractive index changes were eliminated by subtracting the reference flow cell responses. For kinetic analysis, *k_a* and *k_d* were determined from the processed data sets by globally fitting to a 1∶1 binding model. For rapidly associating/dissociating truncated peptides, *K*_D was determined by fitting to a steady-state affinity model using a fixed R_max that was calculated based on the response of R1_5–16. ## Analytical size exclusion chromatography (SEC) Analytical SEC was performed on a Superdex 75 HR 10/30 column (column dimension 1.0×30 cm, column volume 23.6 mL) at room temperature. Samples (100 µl) containing AMA1 (200 µM) with or without R1 peptide (250 µM) were injected onto the column, which was pre-equilibrated with 20 mM sodium phosphate pH 7. Samples were prepared in NMR buffer (20 mM sodium phosphate pH 7, 1 mM EDTA, 0.01% (w/v) sodium azide, 0.2% (w/v) Complete protease inhibitor cocktail (Roche), 50 mM Arg and 50 mM Glu). The flow rate was maintained at 0.5 mL/min and the elution was monitored by measurement of UV absorbance at 280 nM (A₂₈₀). ## Small angle X-ray scattering (SAXS) SAXS measurements were made at the SAXS-WAXS beamline of the Australian Synchrotron, Melbourne, Australia. For each SAXS measurement, 10×1 s exposures were measured and averaged together after verifying that there was no evidence of radiation damage (systematic change in the shape of the scattering curves as a function of exposure time). During data collection the sample was flowed through a 1.5 mm quartz capillary at a rate of 4 µl/sec to further control for radiation damage. Measurements were performed on a dilution series of AMA1 alone from 3.3 to 0.14 mg/ml in NMR buffer and AMA1+R1 (ratio of 1∶1.15) from 3.0 to 0.19 mg/ml in the same buffer. Some concentration-dependent aggregation was observed at protein concentrations above 1 mg/ml as evidenced by increases in Rg and disproportionate increases in I(0) (data not shown). The SAXS data used in this study were from protein at 0.5 mg/ml for AMA1 and 0.75 mg/ml for AMA1:R1. Dilution of the protein below these concentrations did not result in changes to the shape of the scattering curve and calculated molecular weights at these concentrations were consistent with monomeric protein. The molecular weights of the scattering species were estimated from the total forward scatter of the SAXS measurements that were normalised by comparison to water scatter and with reference to the measured protein concentrations. Partial specific volume and scattering length density were calculated using the program MULCh. The monomeric state of the protein was inferred by comparison of the theoretical molecular weight of the protein sequence with the calculated molecular weight from the SAXS experiment. A 1.6 m camera was used with an X-ray energy of 11 keV giving a Q range from 0.01 to 0.5 Å⁻¹. Data were collected on a Pilatus 1M detector (Dektris) and averaging of images, subtraction of blanks and radial integration was performed using the beamline control software ScatterBrain (Australian Synchrotron). Measurements were made at 25°C. Calculation of scattering intensities from molecular models was done using CRYSOL. Radius of gyration (Rg), total forward scatter (I(0)) and P(r) functions were derived using the automated functions in PRIMUS and without manual intervention. ## Computational mapping of binding hot spots FTMAP was employed to map the binding hot spots of AMA1 (<http://ftmap.bu.edu/>) using the AMA1 structures with PDB ID 3SRJ and 2Z8V, which were downloaded from the Protein Data Bank. All ligands and water molecules were removed before mapping. FTMAP searched the global surface of AMA1 with a library of 16 small organic molecules (ethanol, isopropanol, isobutanol, acetone, acetaldehyde, dimethyl ether, cyclohexane, ethane, acetonitrile, urea, methylamine, phenol, benzaldehyde, benzene, acetamide and *N,N*-dimethylformamide). The small molecule probes have different hydrophobicity and hydrogen bonding capability. FTMAP employs a fast Fourier transform correlation approach to efficiently sample billions of protein-probe complexes. The 2000 most favourable docked positions of each probe were energy-minimized and clustered. The six clusters with the lowest average free energy were selected for each probe. The clusters of different probes were further clustered into consensus sites (CSs) based on the distance between the cluster centres. The details of the FTMAP algorithm are described in. ## Accession Numbers Chemical shift assignments for free R1 (pH 5, 40°C) and AMA1-bound R1 (pH 7, 40°C) have been deposited in BMRB under accession codes 19864 and 25134, respectively. # Results and Discussion ## Truncation of the R1 peptide We sought to identify key residues in the interaction of R1 with AMA1. Firstly, in order to define the minimal R1 construct that retains high binding affinity for 3D7 *Pf*AMA1, a series of truncated R1 analogues was synthesized and screened by SPR. Kinetic analysis of data generated for native R1 binding to AMA1 produced a *K*_D of 0.11 µM by globally fitting to a 1∶1 binding model, which is consistent with the reported value (∼0.08 µM). Our previous mutagenesis studies had shown that Phe5, Pro7, Leu8 and Phe9 of R1 were essential for high affinity binding of R1 to AMA1. This conclusion was supported by the current data, in which the truncated R1_11–20 showed no binding to AMA1 up to 10 µM. Interestingly, R1_1–11 containing the Phe5-Phe9 segment also displayed no detectable binding to AMA1 up to 10 µM, implying that the residues Phe5-Phe9 are necessary but not sufficient for interaction with AMA1, and that other key residues are required to facilitate high affinity binding. To test this hypothesis, R1_4–17 and R1_5–16 peptides were synthesized and their binding affinities were measured by SPR. Since both of these truncated mutants showed fast association and dissociation kinetics, a steady-state affinity model was used to fit the data, producing *K*_D values of 0.88 µM for R1_4–17 and 0.99 µM for R1_5–16. Although the binding affinity of the peptides was reduced nearly 10-fold relative to native R1, the fact that both peptides retain *K*_D\<1 µM suggests that Val1-Glu4 and His17-Lys20 do not contribute substantially to high affinity binding with AMA1. Further truncation to the 11-residue peptide R1_5–15 resulted in a further ∼5-fold reduction in *K*_D to 4.6 µM. However truncation of this peptide by deletion of Arg15 to generate the 10-residue R1_5–14 completely abolished measurable binding, indicating that Arg15 is essential for high-affinity binding of R1 to AMA1. This result is consistent with the co-crystal structure of AMA1 bound to R1, in which Arg15 of R1 is bound in a pocket at one end of the hydrophobic cleft of AMA1, where it forms four hydrogen bonds and is the residue that contributes the largest proportion to the buried surface in the interface. Therefore, R1_5–16 was determined to be the minimal construct that retained relatively high binding affinity (∼1 µM) to AMA1. This segment of R1 displays remarkable structural similarity to the Ala2031-Met2042 segment of *Pf*RON2, with an RMSD of 1.2 Å over the twelve C^α positions in their respective structures, implying that the high affinity of R1_5–16 originates from direct mimicry of the natural ligand RON2 as previously suggested. ## Alanine-scanning mutagenesis of the R1 peptide To identify the key interacting residues of R1_5–16, alanine-scanning mutagenesis was performed and the binding affinities of the mutants were determined by SPR. It was necessary to include 1% DMSO (v/v) in the running buffer for this SPR study to maintain the solubility of all of the peptides. This resulted in a small drop in the affinity of the interaction with R1_5–16. Previous ELISA assays on four single-point mutants of R1 had demonstrated that mutation of Pro7 abrogated R1 binding, while mutation of Phe5, Leu8 and Phe9 each resulted in 7.5-, 86- and \>140-fold reductions in affinity relative to the full-length peptide, respectively. In the current SPR study, substitution of Pro7 to Ala resulted in a 35-fold reduction in affinity for AMA1 relative to R1_5–16, indicating that Pro7 is one of the residues that are crucial for high affinity binding. In the crystal structure of the AMA1-R1 complex, Pro7 does not make any direct contact with AMA1, suggesting that it may play a structural role to maintain the adjacent residues in an appropriate conformation for binding. In addition, substitution of Leu6 to Ala caused a 33-fold reduction in affinity for AMA1 relative to R1_5–16. Leu6 makes interactions with a cluster of five Tyr residues in AMA1 (Tyr142, Tyr 175, Tyr234, Tyr 236 and Tyr 251). Importantly, Tyr 251 is highly conserved in *Plasmodium* species and has been shown to be essential for AMA1-RON2 interactions. Combining current and previous data, every residue in the hydrophobic sequence Phe5-Phe9 contributes significantly to AMA1 binding. In the crystal structure of the AMA1-R1 complex, Phe5-Phe9 interacts with a well- defined pocket on one end of the hydrophobic cleft. All of the above suggest that the pocket is a binding hot spot on AMA1 and potentially an attractive target site. A substantial drop in affinity (48-fold relative to R1_5–16) was observed for Ala mutation at Phe12. In the crystal structure, the aromatic ring of Phe12 interacts with two of the key Tyr residues Tyr236 and Tyr251 in the hot spot. In addition, it interacts with Phe183, which was previously identified as a key residue for *Pf*AMA1-*Pf*RON2 interaction. Ala mutation at Phe2038 of RON2 (equivalent to Phe12 of R1) abolished the binding of RON2 to AMA1. A 15-fold reduction in affinity relative to R1_5–16 was observed for the Lys11Ala mutant. This may be caused by disruption of the H-bonds that are observed in the structure between the Lys side chain and Asp227 of AMA1. Mutation of Gly13 resulted in a 21-fold reduction in binding affinity relative to R1_5–16. Since Gly13 interacts with AMA1 through backbone residues only, this loss in affinity may be the result of conformational changes or steric clashes introduced by the mutation. In contrast to the residues discussed above, individual replacements of Ser10, Ser14 and Met16 with Ala resulted in less than 3-fold reductions in affinity, implying that these residues do not contribute significantly to the binding affinity for AMA1. In the crystal structure of the complex, the side chains of these residues are pointing away from the hydrophobic cleft of AMA1 such that mutation to Ala can be accommodated. Consistent with both the truncation studies and the crystal structure, substitution of Arg15 to Ala resulted in largest reduction in affinity ( \>60-fold relative to R1_5–16). The importance of the Arg residue at this position is similar to the case with a peptide derived from RON2, where substitution of Arg2041 of RON2 (equivalent to Arg15 of R1) to Ala abolished the binding to AMA1. In the structures of their complexes, Arg2041 of *Pf*RON2 interacts with the same pocket of AMA1 as Arg15 of R1 and is the residue that contributes most of the buried surface in the interface. In addition to R1 and RON2, antibodies IF9 and IgNAR, which bind with high affinity to AMA1, also have either Arg or Lys residues that fit into the same pocket in the hydrophobic cleft in their respective structures. Together, these data confirm that this “Arg pocket” is a binding hot spot on AMA1, which may serve as a pivotal anchor point for RON2 binding and an attractive site for inhibiting the AMA1-RON2 interaction. ## Backbone resonance assignments of the AMA1-bound R1 peptide The crystal structure of the AMA1-R1 complex revealed a somewhat unexpected 2∶1 binding stoichiometry, which contrasted with the 1∶1 binding observed previously by SPR and ITC studies. To resolve this apparent anomaly, we investigated the AMA1-R1 interaction by solution NMR spectroscopy. A recombinant protein expression system was established to produce uniformly (u-) ¹³C, ¹⁵N-labelled R1 peptide (Figure S2–4). Backbone resonance assignments for free u-¹³C, ¹⁵N-labelled R1 were obtained at pH 7 and 5°C using standard triple-resonance experiments. For the free peptide it was necessary to record the spectrum at a lower temperature as several peaks were not observed at 40°C (which was found to be the optimum temperature for recording spectra of the complex), presumably due to their rapid exchange with water. To enable comparison of the free and bound states, amide chemical shifts for free R1 were extrapolated to 40°C by recording a series of ¹H-¹⁵N-HSQC spectra at increasing temperatures and calculating the temperature dependence of the amide resonances (Table S1 in). A sample of fractionally deuterated (f-²H), u-¹³C, ¹⁵N-labelled R1 with excess f-²H-labelled AMA1 was prepared for backbone assignment of bound R1. To ensure that all the R1 peptide was in the bound form, samples with different ratios of the R1:AMA1 were also prepared. It was found that when the R1:AMA1 ratio was \>1, the ¹H-¹⁵N-HSQC spectrum contained two sets of peaks corresponding to free R1 and bound R1, respectively (Figure S5 in). This indicates that R1 is in slow exchange with AMA1, which is consistent with its high binding affinity. If two peptides were bound to one AMA1 molecule as shown in the crystal structure, this would either give rise to a second set of bound signals in the spectrum or lead to perturbation of the chemical shifts of free R1 in the spectrum recorded with a sub-stoichiometric amount of AMA1; however, no additional peaks or chemical shift perturbations corresponding to a second bound state of R1 were observed. Thus the NMR result supports the 1∶1 binding stoichiometry indicated by our SPR data and previous ITC data. R1 is a 20-residue peptide containing a single proline and has a free *N*-terminal amine. Therefore, a total of 18 peaks were expected in the ¹H-¹⁵N-HSQC spectrum of bound R1. Of these, 17 were observed for the bound R1 peptide at pH 7 and 40°C, although the peak intensities were non-uniform across the spectrum. Both analytical size-exclusion chromatography and small-angle X-ray scattering (SAXS) data indicate that AMA1 interacts with R1 as a monomer, with no evidence for higher order oligomers of protein. The monomeric state is inferred from the SAXS data both from the goodness of fit to the monomeric crystal structures and from the molecular weight calculated from total forward scattering (37 kDa for apo AMA1 and 41 kDa for AMA1+R1. These values compare to theoretical molecular weights of 41.3 and 43.7 kDa respectively). This suggests that the poor sensitivity of certain residues in the NMR spectra is most likely caused by local conformational exchange in the complex that results in significant broadening for the peaks of affected residues. This effect also resulted in poor sensitivity in 3D experiments and hindered full backbone assignment. Through careful analysis of both TROSY-HNCA and TROSY-HN(CO)CA spectra (Figure S6,7), 12 out of 18 expected amide resonances and 15 out of 20 expected C^α resonances were assigned (Table S2). ## Structural analysis of the AMA1-bound R1 peptide Free R1 displayed narrow chemical shift dispersion in the proton dimension (7.7 ppm–8.5 ppm) of the ¹H-¹⁵N-HSQC, consistent with the largely disordered structure in solution that has been observed previously ( and Figure S3–4). Upon binding to AMA1 the ¹H-¹⁵N-HSQC spectrum of the peptide showed broader chemical shift dispersion in the proton dimension (7.0–9.5 ppm), consistent with the peptide assuming a more ordered conformation. The crystal structure of R1 bound to AMA1 identified two R1 peptides bound to AMA1, which were described as the “major” and “minor” states. However, only one set of amide peaks was observed in the ¹H-¹⁵N-HSQC for bound R1. As R1 “minor binder” makes several contacts with R1 “major binder” in the crystal structure, we sought to evaluate the possible structural changes of bound R1 for the 1∶1 binding stoichiometry that is observed in solution. Due to the poor sensitivity in 3D experiments, we were not able to solve the solution structure of bound R1 and make direct comparison with the crystal structure. Instead, we probed the secondary structure of bound R1 based on a limited set of assigned C^α chemical shifts and compared that with the secondary structure of bound R1 in the crystal structure. The deviation of the C^α chemical shifts in R1 relative to their random coil values (secondary shifts, Δδ) was calculated as these are correlated with the polypeptide backbone torsion angles φ and ψ and can be used to predict the secondary structure of AMA1-bound R1 in solution. The secondary shifts of both free and bound R1 are plotted in. The secondary shifts for free R1 are close to zero. In contrast, bound R1 showed larger secondary shifts. Although the C^α chemical shift of Ser14 remained unassigned, Phe12 and Gly13 showed reasonably strong negative secondary shifts (Phe12 and Gly13 \<−1), which is consistent with the presence of extended β-structure in Phe12-Gly13-Ser14 as revealed by the crystal structure of major R1 bound to AMA1. The *C*-terminal residues His17-Lys20 of bound R1 showed nearly identical C^α secondary shifts to those of free R1, which is consistent with this region being flexible in solution and suggests that these residues may not be directly involved in the interaction with AMA1. To further evaluate secondary structure similarity between bound R1 in solution and in the crystal, a comparison was made of C^α chemical shifts, which were determined experimentally for bound R1 in solution and predicted for the major form of R1 in the crystal structure. The predictions for R1 in the crystal structure (Chain C, PDB ID: 3SRJ) were performed using SHIFTX2. The predicted results are plotted as secondary shifts in. Although the C^α chemical shifts of some residues were unassigned or missing, the secondary C^α shifts of the experimental data and the crystal structure prediction were strikingly similar. C^α chemical shifts were also predicted for the minor R1 in the crystal structure, but the correlation between the predicted data and experimental data was much poorer (correlation coefficient for Glu4-Leu8 of minor R1 = 0.85; correlation coefficient for Glu4-Leu8 of major R1 = 0.99; Figure S9 in). The distinction between the two binding modes was equally unambiguous when chemical shifts were predicted by SPARTA+ instead of SHIFTX2 (Figure S10). This suggests that the secondary structure of AMA1-bound R1 in solution is similar to that of the major R1 in the crystal structure. Taken together, the stoichiometry observed in the SPR data we report here, the previous ITC data and comparison of the experimental and predicted NMR data suggest that the minor R1 conformation was most likely an artifact due to the high concentration of R1 peptide used in the crystallographic study. ## Computational solvent mapping of AMA1 Using truncation and mutagenesis of R1 peptide, we have identified two binding hot spots on the AMA1 surface that contribute to high affinity of R1 binding. To further assess the capacity of these hot spots to effectively bind small organic molecules, we employed FTMAP, a fragment-based computational solvent mapping algorithm. FTMAP searched the global surface of the AMA1 with a library of 16 small molecule probes that vary in hydrophobicity and hydrogen bonding capability. The regions that bind to probe clusters are designated consensus sites (CS) in FTMAP and the site that binds the highest number of probe clusters is identified as the most druggable. We performed the initial mapping on the structure of 3D7 *Pf*AMA1 co-crystallized with R1 (PDB ID: 3SRJ). Prior to mapping, the bound ligands and water were removed. The five largest consensus sites were located in the same pocket, which was also identified from SPR analysis as a hot spot for binding Phe5-Phe9 of R1, implicating it as a prospective pocket for small molecule targeting. The most notable features of the pocket are its hydrophobicity and conservation. The pocket is flanked at one end by a cluster of five Tyr residues (Tyr142, Tyr175, Tyr234, Tyr236 and Tyr251) and at the other end by Leu176, Ala254, Met273 and Phe274. Tyr251 is highly conserved across *Plasmodium* species and essential for AMA1-RON2 interactions. All the other residues that form the pocket, except Tyr175, are also highly conserved in all known *P. falciparum* sequences. Of the probe clusters identified, CS2 (magenta, 17 probe clusters) overlaps well with Leu6 of R1 and partially with Phe9 of R1; CS5 (grey, 9 probe clusters) overlaps well with Phe5 of R1. Probes in CS2 favour hydrogen bonding to phenol hydroxyl groups of Tyr234, Tyr236 and Tyr251. More importantly, the largest consensus site CS1 (cyan, 21 probe clusters) and CS3 (yellow, 13 probe clusters), which are located on the base of the pocket, revealed the key interactions that are in addition to those formed by the Phe5-Phe9 segment of R1. Probes in CS1 and CS3 make additional interactions with Leu131, Arg143, Leu144, Pro145, Ala253 and Gln255. CS4 (salmon, 10 probe clusters) extends one end of the pocket by interacting with Val129, Gln256 and Gln349. Probes in CS4 favour hydrogen bonding to the amide group of Gln349. The mapping results presented here suggest that this hot spot, which interacts with Phe5-Phe9 of R1, is druggable and effectively binds various small organic molecules. Identification of additional key interactions in the hot spot is potentially useful in the development of small molecule inhibitors. The Phe5-Phe9-interacting hot spot is partially protected by the domain II loop, which is displaced by the binding of R1 and RON2. Part of this hot spot was identified previously as a pocket for small molecule targeting, although it has been suggested that the domain II loop may limit small molecule binding at this site. We have shown that ∼420 ų solvent-accessible volume of this pocket is still available for small molecule binding when the domain II loop is not displaced. To further address this issue, we performed mapping on the structure of 3D7 *Pf*AMA1 co-crystallized with antibody IgNAR (PDB ID: 2Z8V). This is the only AMA1 structure that has a complete description of the domain II loop. IgNAR binds to a region distant from the hot spot and does not induce any significant changes in the structure of AMA1 (C^α RMSD of 0.34 Å between IgNAR-bound and apo AMA1). Our mapping for 2Z8V results showed that two consensus sites CS3 (yellow, 14 probe clusters) and CS5 (grey, 9 probe clusters) are located in the domain II loop-protected pocket. Importantly, several of the key residues, which bind Phe5-Phe9 of R1 or small molecule probes when the domain II loop is displaced, are still involved in the formation of the loop- protected pocket and remain accessible to small organic molecules (Phe181, Tyr234, Tyr236, Tyr251, Ile252, Ala253, Ala254 and Phe274;). All these residues are highly conserved in *P. falciparum* and their side chain conformations remain almost unchanged when the domain II loop is displaced. Notably, no consensus sites were found in the Arg pocket in either AMA1 structure. One possible explanation could be that the Arg pocket is relatively small and has a polar surface area. Although the interactions mediated in the Arg pocket are absolutely crucial for R1/RON2 binding to AMA1, it may be more difficult to identify suitable small molecules to access and interact with this site in isolation. # Conclusions Using SPR and NMR spectroscopy we have validated that R1 binds to AMA1 in solution with 1∶1 stoichiometry, as suggested by previous ITC data, and adopts a secondary structure consistent with the major form of R1 observed in the crystal structure of the complex. The minor form of R1 in the crystal structure was not observed in solution and is likely to be a crystallographic artifact. The truncation and mutational studies for R1 presented here have identified several key AMA1-interacting residues scattered along the peptide. Amongst these key residues, the hydrophobic segment Phe5-Leu6-Pro7-Leu8-Phe9, residues Phe12 and Arg15 are those that contribute most to the AMA1 binding affinity. They interact with two distinct binding hot spots, which are located at the two ends of the hydrophobic cleft of AMA1. Both of the pockets are highly conserved across the *P. falciparum* strains and likely to be suitable for designing broad-spectrum AMA1 inhibitors. The “Arg pocket” at one end of the cleft mediates key interactions of several known inhibitory agents, although fragment-based computational solvent mapping on AMA1 suggests that it may be a difficult site to target with small organic molecules because of its small surface area and polar nature. Mimicking the Arg side chain using peptidomimetics based on R1 or RON2 might be a more productive approach to target this important pocket. In contrast, mapping results showed that the Phe5-Phe9-interacting hot spot is druggable and identified key AMA1 residues for small molecule targeting. Our results provide a basis for designing novel high affinity inhibitors of AMA1-RON2 interaction that are effective against the majority of *Pf*AMA1 genotypes. # Supporting Information The authors thank San Sui Lim, Dr. Martin L. Williams and Dr. Mark D. Mulcair for their kind assistance in this project. The Queensland NMR Network (QNN) is acknowledged for providing access to the 900 MHz spectrometer at University of Queensland. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: GW CAM RFA JSS SM RSN MJS. Performed the experiments: GW CAM BM MM NPC SM. Analyzed the data: GW CAM BM MM NPC SM. Contributed reagents/materials/analysis tools: MM NPC. Wrote the paper: GW CAM BM MM NPC RFA JSS SM RSN MJS.

# Introduction *Dalbergia odorifera* T. Chen, also named yellow flower pear, is a species that belongs to Family Leguminosae. This tree can reach 10–15 m in plant height. *D*. *odorifera* is a fragrant rosewood species under the second-grade state- protection of the Chinese government. This species is naturally distributed in the tropical regions of China, especially in Hainan Province. *D*. *odorifera* has abundant secondary plant metabolites that have antibacterial, anti-inflammatory, antioxidative, antithrombotic, antiosteoporosis, antiangiogenesis, antiosteosarcoma, and antiplatelet activities. Therefore, *D*. *odorifera* species is widely used as a medicinal drug for the treatment of various illnesses, such as cancer, cardiovascular diseases, blood disorders, necrosis, diabetes, and rheumatic pain. *D*. *odorifera* was introduced in subtropical areas in China and has been widely cultivated in recent decades due to its high medicinal and economic value. However, *D*. *odorifera* seedlings cultivated in typical tropical red soil exhibit poor plant growth and root architecture due to the soil’s low-porosity and pH value, and high ferrum and aluminum ionic contents. This condition results in low survival rates when seedlings are transplanted. Hence, the growth and root system development of *D*. *odorifera* seedlings must be improved through the use of appropriate culture matrixes. In recent years solid wastes, such as cottonseed hulls, sawdust, carbonized rice coirs, sugar slag, sludge, almond shells, and pineal compost, are increasingly being utilized as seedling culture substrates to promote plant growth. Coconut (*Cocos nucifera* Linn.) coir and the fallen leaves of *Ficus elastica* Roxb. ex Hornem, are very common and typical solid wastes in Hainan Province. Coconut coir may improve the growth of *Lycopersicum esculentum* and *Capsicum annuum* when used as a substrate. The fallen leaves of sycamore and aspen may enhance soil structure, soil organic matter content and fertility status thus plant growth may be promoted by deciduous matrixes. Several studies have demonstrated that mixed matrixes are better for plant development because they improve soil aeration, water retention capacity, and soil nutrients than any single matrix. For example, the growth behavior of tomato and marrow squash seedlings is directly related to the composition of seedling substrate. Decomposed vinasse, mushroom residue, and cattle manure can improve the total primary nutrient condition of cultivation matrixes and further enhance the stem diameter, dry weight accumulation, and health index of the tested seedlings. The mixture of perlite and rice coir biochar as a hydroponic substrate for growing cabbage, dill and red lettuce is better than a single substrate treatment, and has the advantage of increased yield. In addition, waste wood, pine bark, sand and volcanic rock have been used as single or compound substrates for the cultivation of commercial vegetable seedlings. Few studies have reported that temperature affects the germination rates of *D*. *odorifera* seeds, and lighting spectra and exogenous substances influence the growth and nutrient absorption capacity of its seedlings. Intraspecific geographic variations in the optimal germination temperature have been noted in *D*. *odorifera* seeds collected from four geographic sites in southern China; the optimal germination temperature for the seeds is reportedly within the range of 25 °C–30 °C. The exogenous application of chitosan oligosaccharide (1/800, v/v) under 280 W light-emitting diode panels for a 15 h daily photoperiod can efficiently promote the synthetic quality and nutrient utilization of *D*. *odorifera* seedlings. However, few studies have been conducted on the culture substrates of *D*. *odorifera* seedlings, and the influence of matrix on seedling growth and root development has been rarely explored. Root systems are the important bridge between soil and plant, and play a vital role in the life cycle of plants, especially for plants that need to be transplanted. Different substrate combinations and proportions from coconut coir powder, deciduous *F*. *elastica* leaf powder, local river sand, and red soil were used as culture substrates to determine their effects on the growth, root system development, and physiological characteristics of *D*. *odorifera*. This study aimed to answer the following questions: (1) Can solid wastes such as coconut coir powder and deciduous *F*. *elastica* leaf powder, be used as substrate for *D*. *odorifera* seedlings? (2) How will seedling growth and root systems of *D*. *odorifera* respond under various substrate combinations in different proportions? (3) Which substrate combinations and in what proportions are suitable for plant growth, root system and root nodule development, leaf pigment synthesis, and root physiobiochemical characteristics? We hypothesized that the growth, root system development, and physiological characteristics of *D*. *odorifera* differentially respond to the same optimal culture substances. # Materials and methods ## Plant material and experimental setup Mature and healthy *D*. *odorifera* seeds were collected from the plantation forest at Ledong County, Hainan Province (18° 42′ 57.91″ N, 108° 52′ 18.65″ E) on Februrary 10, 2017. Although *D*. *odorifera* belongs to protected plant species, the trade of its commercialized seeds and seedlings is permitted and legal in China. Therefore, we confirmed that no specific permissions for the seed collection in these locations were required by Forestry Bureau of Hainan Province, China. The seeds were stored at 4 °C in a refrigerator. The seeds were placed in an artificial climate incubator for germination at 27 °C and 85% humidity in light for 12 h and 24 °C and 75% humidity in the dark for 12 h on September 11, 2018. After germinating and growing, the seedlings with the same size were transplanted in pots (20 cm diameter and 15 cm height) on October 10, 2018. Two seedlings were planted in each pot. The pots were filled with equal volumes of culture substrates from coconut coir powder, deciduous *F*. *elastica* leaf powder, river sand and red soil according to the following experimental designs. Powder size was less than 5 mm. The pots with transplanted seedlings were placed in the greenhouse with natural light source in Hainan University, Haikou, Hainan Province (20° 03′ 26.77″ N, 110° 19′ 0.90″ E). Average air temperature was 22 °C– 28 °C, average air humidity was 42%– 85%, average light intensity was 3200–6500 Lux in daytime. A random block design was employed. The substrate combinations were prepared as follow. C1 group had three subgroups with coconut coir powder and red soil (C1_-1: 3/1; C1_-2: 2/2; C1_-3: 1/3, v/v), C2 group had three subgroups with deciduous leaf powder and red soil (C2_-1: 3/1; C2_-2: 2/2; C2_-3: 1/3, v/v), C3 group had three subgroups with red soil and sand (C3_-1: 3/1; C3_-2: 2/2; C3_-3: 1/3, v/v); C4 group had three subgroups with coconut coir powder and sand (C4_-1: 3/1; C4_-2: 2/2; C4_-3: 1/3, v/v); C5 group contained coconut coir powder, deciduous leaf powder and red soil (1/1/1, v/v); C6 group contained coconut coir powder, red soil, and sand (1/1/1, v/v); C7, C8, C9, and C10 groups were entirely constituted of coconut coir powder, red soil, deciduous leaf powder, and sand, respectively. The pots were watered until 100% field capacity every day, and excess water in the dish placed under the pot was re-watered to the corresponding pot to prevent nutrition loss. Six replicates with six seedlings each were used for each group or subgroup. Plant height, leaf area, and number of leaf were determined after 60 days growth. Fresh leaves and root samples for physiobiochemical analysis were collected and immediately frozen in liquid N. Fresh leaves, shoots, and roots were individually harvested for biomass and image analyses. ## Determination of plant growth, dry matter accumulation and allocation, and root system development The height of *D*. *odorifera* seedlings was determined with a ruler with a precision of 0.1 cm. Leaf area was measured by an LI-3000C portable leaf area meter. The total number of leaves was counted. All seedlings were harvested at the end of the experiment, and divided into leaves, shoots, and roots. The roots were washed, dried in the shade, and scanned with a Perfection V700 photo color image scanner. Total root length, entire surface area, total volume, average diameter, number of root tips, and number of branches were analyzed via the Winrhizo system. Biomass samples were dried (at 70 °C for 48 h) until constant weight, and specific leaf area (SLA) was then calculated by dividing the leaf area by the leaf dry weight according to Xu et al.. Dry matter allocation was calculated according to the ratio of root to the sum of leaves and shoots. ## Determination of leaf relative electrolyte leakage and pigment content In brief, five freshly harvested leaf disks (0.5 cm in diameter) were placed in tubes containing 10 mL of deionized water and incubated at 25 °C on a shaker for 6 h. Then, the initial electrical conductivity was determined using a conductivity meter (Mettler-Toledo Instruments Co., Ltd, Shanghai, China). The final conductivity was measured after boiling at 100 °C for 30 min using previous samples. Relative electrolyte leakage (REL) was determined as the ratio of the initial conductivity to final conductivity. Chlorophyll pigments were extracted in 80% (v/v) chilled acetone and quantified using a spectrometer (MPDA-1800, Shanghai, China). The absorption spectra of the samples were recorded at 663, 647, and 470 nm wavelength for chlorophyll a, chlorophyll b, and carotenoids, respectively. The absorbance values were converted to concentrations and/or contents according to the experimental equations described by Lichtenthaler. ## Estimation of root malondialdehyde, superoxide dismutase, and soluble sugar contents Fresh root samples (0.3 g) were homogenized in 8 mL 5% trichloroacetic acid (TCA) solution and centrifuged at 10,000 rpm for 10 min, thiobarbituric acid (2 mL. 0.6%) in 10% TCA was added to 2 mL of the supernatant. The mixture was boiled at 100 °C for 30 min. Then, the absorbances of the supernatant at 450 (A₄₅₀), 532 (A₅₃₂) and 600 nm (A₆₀₀) was determined with a MAPDA spectrometer. Malondialdehyde (MDA) concentration was calculated through the following formula: C (μmol/L) = 6.45 × (A₅₃₂ − A₆₀₀)– 0.56 × A₄₅₀. Soluble sugars were extracted and spectrophotometrically estimated using an anthrone sulphuric acid reagent according to the method of Xiao et al. and Renaut et al.. Superoxide dismutase (SOD) was extracted with sodium phosphate buffer containing 50 mM sodium phosphate buffer (pH 7.8), 1 mM ethylenediamine tetraacetic acid, 15% glycerin, 1 mM ascorbic acid, 1 mM dithiothreitol, 1 mM glutathione, 5 mM MgCl₂, and 1% (w/v) polyvinylpolypyrrolidone. SOD activity was measured spectrophotometrically at 560 nm by monitoring the inhibition of the photochemical reduction of nitroblue tetrazolium according to the method of Xiao et al.. ## Comprehensive evaluation based on the multidimensional space mathematical model of Euclid The responses of *D*. *odorifera* to different culture substrate compositions based on a single and independent parameter are difficult to evaluate. However, the multidimensional space (Euclid, Eⁿ) mathematical model is a useful tool to comprehensively assess the effects of various culture substrate compositions on the above- and below-ground growth and development, and physiological responses of *D*. *odorifera*. This model can be calculated by using following equation as reported by Wang et al.: $$\sum\text{P}_{i}{}^{2} = \sum{(1 - \text{a}_{ij})}^{2},$$ where P_*i* stands for the distance from the *i*-th treatment to the standard point, and the performance of treatment is better when the value of P_*i* is smaller; *i* represents treatment; *j* stands for the trait, and a_*ij* represents the absolute value of the ratio of *j* trait of the *i*-th treatment to the maximum value of the *j*-th trait (the minimized value is selected under negative correlated condition). ## Statistical analysis Results were expressed as means ± standard errors (n = 6). SPSS 13.0 software package was used for statistical analysis. One-way ANOVA followed by Duncan's multiple range test at P \< 0.05 was employed to assess the statistical significant difference between treatments. # Results ## Morphological variations The composition of different substrates had remarkable effects on the above- ground and below-ground growth and development of *D*. *odorifera*, as shown in the Figs and, respectively. In general, seedlings from the C1_-2, C2_-2, C3_-1, and C6 groups had a larger crown width, more leaves and branches, and dark green in color compared with those in the other groups. Seedlings from the C7, C9, and C10 groups were remarkable shorter, smaller, and yellower than the seedlings described above. The root architecture of *D*. *odorifera* from the C1_-2, C2_-2, C3_-1, C6, and C8 groups were well developed, and several fine root hairs and branches or long roots were observed. The root system from the C7 and C9 groups had the worst performance in comparison with the others. Interestingly, several visible root nodules could be found in the C6 group but not in other groups. Thus, culture substances consisting of sand, coconut coir powder, and red soil could improve the whole performance of *D*. *odorifera* seedlings. ## Effect of substrate combinations and proportions on the seedling growth of *D*. *odorifera* Significant changes (P \< 0.05) in plant height, leaf area, number of leaves, and total leaf area were observed among groups from different substrate combinations and proportions. In general, the mixed substrates might efficiently promote the growth of *D*. *odorifera* seedlings than any single substrate. Seedlings from the C6 group filled with equal volumes of coconut coir powder, red soil, and sand achieved the highest plant height and largest average leaf area compared with those in other groups. Seedlings from the C3 group filled with red soil and sand had the most leaf numbers, the largest total leaf area, and a considerable higher plant height than those in other groups. Different substrate proportions from different subgroups caused significant differences (P \< 0.05) in plant height, leaf area, number of leaves, and total leaf area within the same group with the same substrate composition. Overall, the C1_-2, C2_-2, and C3_-1 subgroups had better performances than the other subgroups within the same group. Interestingly, the single red soil had a balanced and better performance in plant height, leaf area, number of leaves, and total leaf area even when compared with some groups filled with mixed substrates. The single substrate of coconut coir, deciduous leaf powder, and sand (especially deciduous leaf powder and coconut coir) remarkable inhibited the growth and development of *D*. *odorifera* seedlings. Thus, culture substances consisting of sand, coconut coir powder, and red soil could improve the growth of *D*. *odorifera* seedlings. The C1_-2 and C3_-2 subgroups had the best effects on plant growth. ## Effect of substrate combinations and proportions on the root architecture of *D*. *odorifera* Total root length, total root surface area, total root volume, and number of root tips and branches showed significant differences (P \< 0.05) among groups from different substrate combinations and proportions, but no significant difference (P \> 0.05) in the average diameter of roots was observed. In general, the mixed substrates may efficiently promote the root growth and development of *D*. *odorifera* than any single substrate. Seedlings from the C6 group had the most excellent performance in total root length, total root surface area, and the total root volume, whereas seedlings from the C3 group had the most number of root tips and branches and the largest total root length compared with those in other groups. Different substrate proportions from different subgroups caused considerable difference in total root length, total root surface area, total root volume, and number of root tips and branches within the same group with same substrate composition. Broadly, the C1_-2, C2_-2, and C3_-1 subgroups had better overall performance in all the above parameters than the other subgroups within the same group and even better than the C6 group. Interestingly, the single substrate of red soil had a balanced and relatively better performance in total root length, total root surface area, and total root volume but poor performance in terms of number of root tips and branches. The single substrate of coconut coir powder, deciduous leaf powder, and sand substantially inhibited the root system and the growth and development of *D*. *odorifera*. Thus, culture substances consisting of sand, coconut coir powder, and red soil could improve the root system and the growth and development of *D*. *odorifera*. Subgroups C1_-2, C2_-2, and C3_-2 could remarkably promote development of root systems. ## Dry matter accumulation and allocation in *D*. *odorifera* affected by various composite substrates Different substrate combinations and proportions significantly (P \< 0.05) affected the root dry weight (RDW), leaf dry weight (LDW), shoot dry weight (SDW), total dry weight (TDW), SLA, and dry matter allocation (DMA). Among the 10 groups filled with different substrate combinations, the C6 group had the highest RDW and SDW, the C3 group had the best LDW and TDW, and no statistically significant effects (P \> 0.05) on RDW, SDW, LDW, and TDW were found among the C1, C2, C3 and C6 groups. The C9 group accumulated the least biomass. Different substrate proportions from different subgroups caused substantial effects on RDW, SDW, LDW, and TDW within the same group with the same substrate composition. For example, the C1_-2, C2_-2, and C3_-2 subgroups had a substantially higher biomass in the above parameters than the other subgroups within the same group, and even better than the C6 group. Interestingly, the single substrate of red soil had a balanced and relatively better performance in biomass accumulation. In addition, the C5 group achieved the maximum values in SLA and DMA, whereas the C9 group obtained the minimum values. The groups with good performance in dry matter accumulation, such as the C1, C2, C3 and C6 groups, had intermediate levels of SLA and DMA, but the effects were not significant. Thus, culture substances consisting of sand, coconut coir powder, and red soil could improve the dry matter accumulation of *D*. *odorifera* seedlings, and subgroups of C1_-2 and C3_-2 had the best effects on biomass accumulation. ## Effects of substrate combinations and proportions on the leaf REL and pigment content Various substrate combinations and proportions significantly (P \< 0.05) changed the leaf REL and the contents of chlorophyll a, chlorophyll b, carotenoids, and total chlorophyll in *D*. *odorifera*. Among the 10 groups filled with different substrate combinations, the C1, C2, C3, and C6 groups had relatively lower REL and higher levels of chlorophyll a, chlorophyll b, carotenoids, and total chlorophyll. The C1_-2, C2_-2, and C3_-2 subgroups had a substantially lower level of leaf REL but had higher levels of chlorophyll contents than the other subgroups within the same groups and even better than the C6 group. The C2_-2 subgroup and the C7 group had the lowest and highest levels of leaf REL, respectively. The C2_-2 and C3_-1 subgroups had the highest levels, whereas the C9 group had the lowest levels of chlorophyll contents. In addition, the single red soil substrate had balanced and intermediate performance in leaf REL and photosynthetic pigment contents. Thus, culture substances consisting of sand, coconut coir powder, and red soil could increase the pigment content of *D*. *odorifera* leaves. ## Effects of substrate combinations and proportions on the MDA content, SOD activity, and soluble sugar contents of *D*. *odorifera* root systems Various substrate combinations and proportions significantly (P \< 0.05) changed the MDA, SOD, and sugar contents of *D*. *odorifera* root systems. Among the 10 groups filled with different substrate combinations, the C9 group had the highest levels of MDA content and SOD activity. The C1, C2, C3, and C6 groups had relatively lower levels of MDA content and SOD activity. The C6 group had the lowest levels in these parameters. The C1_-2, C2_-2, and C3_-2 subgroups had a remarkable lower levels of MDA content and SOD activity than the other subgroups within the same groups but almost the as those of the C6 group. In general, the C3_-1 subgroup had the highest level and the C5 group had the lowest level of sugar contents. The C1_-2 and C2_-2 and the C6 groups had remarkably higher sugar contents than the other groups, whereas the C5, C7, and C9 groups had substantially lower sugar contents. In addition, the single red soil substrate had balanced and intermediate effects on the MDA content, SOD activity, and sugar content of *D*. *odorifera* root systems. Thus, culture substances consisting of sand, coconut coir powder, and red soil, particularly C3_-1 subgroup, could improve the physiological traits of *D*. *odorifera* roots. ## Comprehensive evaluation of *D*. *odorifera* seedlings affected by various substrate combinations and proportions A smaller value of ∑P_*i*² indicates a better effect in Euclid's model. In general, the multidimensional space analysis showed that the groups containing red soil had better performance than the groups without red soil, and appropriate proportions of coconut coir powder, deciduous leaf powder, and sand could improve the performance of *D*. *odorifera* in red soil. The comprehensive analysis showed that the C1_-2, C2_-2, C3_-1, and C3_-2 subgroups and the C6 group are the top five groups in terms of performance in most of the given parameters. The C1_-2 subgroup filled with equal volumes of coconut coir and red soil had the best overall performance in the comprehensive evaluation, but the gaps were small among the top five groups. The C8 group filled with red soil only ranked seventh in the comprehensive evaluation because of its balanced performance in most parameters but was inferior to the top five groups in terms of overall performence. The C4 group, including the three subgroups filled with coconut coir and sand had a worse performance in the comprehensive evaluation compared with the other groups containing the red soil component. The C7 and C9 groups filled solely with coconut coir powder and deciduous leaf powder, respectively, had the worst performance in the comprehensive evaluation (especially the single deciduous leaf powder substrate). # Discussion Some solid wastes, such as rice coir biochar, almond shells, pineal compost, waste wood, and pine bark, can be used as culture materials to improve germination rate, seedling growth and root development by improve the soil pH and aeration, increasing cation exchange capacity and water-retention capacity, and even providing nutrients. In the present study, the basic physical and chemical properties of culture substances evidently varied when single red soil was mixed with coconut coir powder and sand. For example, the single red soil used in this study had pH 4.69, 0.61 g/kg total nitrogen, 0.18 g/kg total phosphorus, 7.11 g/kg organic matter, and 1.65 g/kg organic carbon. However, the culture substance consisting of equal volumes of coconut coir powder, red soil, and sand had pH 6.23, 1.77 g/kg total nitrogen, 0.64 g/kg total phosphorus, 58.01 g/kg organic matter, 33.65 g/kg organic carbon. Coconut coir and the fallen leaves of *F*. *elastica* possess a large amount of fibers, abundant nutrients, and intense physical elasticity, which can be used to improve the structure, porosity, organic matter content, and fertility of soil and thus can improve the environment for plant growth and development. Therefore, the results of the present study provide the answer to the first research question that coconut coir and the fallen leaves of *F*. *elastica* (typical solid wastes in Hainan Province) could be used as culture substrates to improve the growth and development of *D*. *odorifera* in red soil. The rational utilization of coconut coir and fallen leaves is also favorable for environmental protection. Several coordinated morphological, physiological, and biochemical responses occur when a plant is exposed to stress. In this study, nutrition deficiency and poor ventilation status were the two main abiotic stresses caused by the different combinations and proportions of culture substrates for *D*. *odorifera* seedlings. Poor performances under stressful environments were evaluated on the basis of the appearance of yellow leaves, observation of poor seedling growth and development, and root architecture, and attainment of low levels of chlorophyll and sugar contents, low ratios of dry matter accumulation and dry matter allocation, and high levels of leaf REL and root MDA. Conversely, the excellent performances under desirable growth environments were evaluated on the basis of opposite trends in these parameters. The above- and below-ground growth and development of seedlings depend not only on the physical properties of culture matrixes such as pH values and porosity, but also on soil fertility. The substrate combinations caused remarkable differences in all the parameters analyzed. The poor growth and development of *D*. *odorifera* in the C5, C7, and C9 groups might have resulted from nutrients deficiency due to the absence of red soil. According to Liebig’s law of the minimum, the factor with the minimum amount is the key limiting factor. In this case, nutrients were the key factors limiting the growth and development of *D*. *odorifera* in the absence of red soil. Membrane permeability (REL), membrane lipid peroxidation (MDA), and antioxidant enzymatic systems would positively respond when plants are exposed to biotic or abiotic stresses. The higher levels of leaf REL, root MDA contents, and root SOD activities in *D*. *odorifera* from the groups without red soil suggested that the seedlings experienced nutrient deficiency, which resulted in poor seedling growth and root architecture and low dry matter accumulation and pigment content. Seedlings in the group C8 cultured solely with red soil had balanced performance in most of the traits analyzed but were inferior to seedlings in groups cultured with mixtures containing red soil. This phenomenon suggested that the physical properties of red soil should be improved for the establishment of *D*. *odorifera* seedlings in spite of its abundant nutrients. Thus, the answer to the second research question was that the seedlings growth and root systems of *D*. *odorifera* under single substrates were inferior to those under mixed substrates. Some exogenous solid wastes can improve soil environment; However, the proper combinations and proportions of cultures are the key factors to promote plant growth and development. For example, the mixture of perlite and rice coir biochar is superior to single cultivation substrate in increasing the development and yield of cabbage, dill, and red lettuce. Coconut coir powder, deciduous leaf powder, and sand may improve the red soil’s pH and porosity, which in turn enhance the soil’s breathing ability. However, appropriate combinations and proportions among them are essential to improve the growth and development of *D*. *odorifera* seedlings. Among the 10 groups filled with different substrates, seedlings from the C6 group cultured by equal volumes of coconut coir powder, red soil, and sand had the best performance in the above- and below-ground morphological growth and development attributes. Moreover the dry matter accumulation and allocation, pigment content, and root physiological traits, especially development of root nodules, were also enhanced. The substrate proportions from different subgroups within the C1, C2, and C3 groups remarkably improved the overall performance of *D*. *odorifera* seedlings. The performances of seedlings from the C1, C2, and C3 groups were inferior to those of the C6 group because the lower performances of some subgroups within the same group affected their total mean values. In fact, the comprehensive performance and evaluation of seedlings from the C1_-2 (cultured by equal volumes of coconut coir and red soil), C3_-1 (cultured by threefold volumes red soil and sand), and C2_-2 (cultured by equal volumes of deciduous leaf powder and red soil) subgroups were remarkably superior to that of the other subgroups cultured by various proportions of the same substrates within the same group and were even superior to the performance of seedlings from the C6 group. The excellent performance of these seedlings mainly resulted from the appropriate composition of different substrates. Therefore, the answer to the third research question was that the ideal combinations and proportions of culture substrates came from the C1-₂, C3-₁, C2-₂, and C3-₂ subgroups and C6 group, which are suitable for *D*. *odorifera* seedlings in terms of plant growth, root system and root nodule development, leaf pigment synthesis, and root physiobiochemical characteristics. # Conclusion In general, an appropriate proportional amount of coconut coir, deciduous leaf powder, and sand may improve the overall performance of *D*. *odorifera* seedlings in red soil. The ideal substrate composite for *D*. *odorifera* seedlings came from the C1_-2, C3_-1, C2_-2, and C3-₂ subgroups and C6 group. The results confirmed our hypothesis that the growth, root system development, and physiological characteristics of *D*. *odorifera* differentially respond to the same optimal culture substances. Subgroups C1_-2 (coconut coir/red soil = 2/2, v/v) and C3_-2 (red soil/sand = 2/2) exhibited the best effects on plant growth and biomass accumulation. Subgroups C1_-2, C2_-2 (deciduous leaf powder/red soil = 2/2), and C3_-2 could substantially improve root system development. The C6 group (coconut coir/red soil/sand = 1/1/1) remarkably promoted root nodule development. The C3_-1 subgroup (red soil/sand = 3/1) showed the best effects on physiological characteristics. On the basis of the comprehensive evaluation of Euclid’s multidimensional space mathematical model, the suitable substrate combinations were C1_-2, followed by C3_-1, and then C2_-2. This study provides a scientific basis for the production of healthy seedling cultures of *D*. *odorifera* and a rational utilization of solid wastes such as coconut coir and the deciduous leaves of *F*. *elastica*. [^1]: The authors have declared that no competing interests exist.

# Introduction Modern radiation oncology will require a synergy between high-precision radiotherapy protocols and innovative approaches for biological optimization of radiation effect. From a clinical perspective, new insights into molecular radiobiology will provide a unique opportunity for combining systemic targeted therapeutics with radiotherapy. One example is the use of histone deacetylase (HDAC) inhibitors as potentially radiosensitizing drugs. Inhibition of HDAC enzymes leads to acetylation of histone and non-histone proteins, and the resultant changes in gene transcription cause alterations in key molecules that orchestrate a wide range of cellular functions, including cell cycle progression, DNA damage signaling and repair, and cell death by apoptosis and autophagy. Following the demonstration that HDAC inhibitors enhanced radiation-induced clonogenic suppression of experimental *in vitro* and *in vivo* colorectal carcinoma models, but independently of the actual histone acetylation level at the time of radiation exposure, we conducted the Pelvic Radiation and Vorinostat (PRAVO) phase 1 study. This trial, undertaken in sequential patient cohorts exposed to escalating dose levels of the HDAC inhibitor vorinostat combined with pelvic palliative radiotherapy for advanced gastrointestinal malignancy, was the first to report on the therapeutic use of an HDAC inhibitor in clinical radiotherapy. It was designed to demonstrate a number of key questions; whether the investigational agent reached the specific target (detection of tumor histone acetylation), the applicability of non-invasive tumor response assessment (using functional imaging), and importantly, that the combination of an HDAC inhibitor and radiation was safe and tolerable. The ultimate goal of a first-in-human therapy trial is to conclude with a recommended treatment dose for follow-up expanded trials, and in achieving this, a phase 1 study typically is designed to determine treatment toxicity and tolerability (in terms of dose-limiting toxicity and maximum-tolerated dose (MTD), respectively). For molecularly targeted agents, the dose that results in a relevant level of target modulation may differ greatly from the MTD, and generally, we do not have a good understanding of the relationship between the MTD and the dose required to achieve the desired therapeutic effect. An optimum biological dose may be the dose that is associated with pharmacodynamic biomarkers reflecting the mechanism of drug action. In the setting of fractionated radiotherapy, this would ideally represent a radiosensitizing molecular event occurring at each radiation fraction, or in other words, a biological indicator with a transient and periodic expression profile. Importantly, tumor specimens for this particular purpose cannot be sampled after the patient has commenced the radiation treatment. Any signaling activity in on- treatment tumor samples would reflect the combined effect of radiation and the systemic drug, and the contribution of the latter would probably be indistinguishable from the effect of the actual accumulated radiation dose. Instead, the study can be designed to collect non-irradiated surrogate tissue both before the commencement of study treatment and on-treatment at time points reflecting the timing of administration of the systemic drug with regard to the fractionated radiotherapy protocol. In addition, as a general rule, biomarkers that have been previously established for single-agent therapy will require reevaluation in a first-in-human clinical trial combining a molecularly targeted compound with radiotherapy. Within this context, *i.e.,* that the possible mechanism of radiosensitizing action of the molecularly targeted agent should be regarded a main objective in a combined-modality study with radiotherapy, the present study reports on a correlative analytical strategy for identifying possible biomarkers of HDAC inhibitor activity, using peripheral blood mononuclear cells (PBMC) from the PRAVO phase 1 study patients receiving pelvic palliative radiotherapy as an easily accessible surrogate tissue for vorinostat exposure. Gene expression array analysis identified PBMC genes that from experimental models are known to be implicated in biological processes governed by HDAC inhibitors, and might be further developed as pharmacodynamic biomarkers of vorinostat activity in the setting of fractionated radiotherapy. # Materials and Methods ## Ethics Statement Both of the protocols for the PRAVO study (ClinicalTrials ID NCT00455351) and the phase 2, non-randomized study for patients with locally advanced rectal cancer (LARC) given neoadjuvant chemoradiotherapy (ClinicalTrials ID NCT00278694) were approved by the Institutional Review Board and the Regional Committee for Medical and Health Research Ethics South-East Norway (REC South- East, Permit Number S-06289 and S-05059, respectively), and were performed in accordance with the Declaration of Helsinki. Written informed consent was required for participation. Housing and all procedures involving animals were performed according to protocols approved by the Animal Care and Use Committee at Department of Comparative Medicine, Oslo University Hospital (Permit Number 885–2616–2919–2928–3688), in compliance with the Norwegian National Committee for Animal Experiments' guidelines on animal welfare. ## PRAVO Study Patients and Objectives The patient population was enrolled between February 2007 and May 2009. The principal eligibility criterion was histologically confirmed pelvic carcinoma scheduled to receive palliative radiation to 30 Gy in 3-Gy daily fractions. Other details on eligibility are given in the initial report. This phase 1 dose- escalation study adopted a standard 3+3 expansion cohort design, where patients with advanced gastrointestinal carcinoma were enrolled onto four sequential dose levels of vorinostat (Merck & Co., Inc., Whitehouse Station, NJ, USA), starting at 100 mg daily with dose escalation in increments of 100 mg. The primary objective was to determine tolerability of vorinostat, defined by dose-limiting toxicity and MTD, when administered concomitantly with palliative radiation to pelvic target volumes. Secondary objectives were to assess the biological activity of vorinostat, including the identification of possible biomarkers of HDAC inhibitor activity, and to monitor radiological response when given with pelvic radiotherapy. The study data on patient treatment tolerability, tumor histone acetylation following vorinostat administration, and treatment-induced changes in tumor volume and apparent distribution coefficient, as assessed by magnetic resonance imaging, have been reported in detail previously. ## Patient Blood Sampling and RNA Isolation As depicted by, peripheral blood, drawn on PAXgene Blood RNA Tubes (Qiagen Norge, Oslo, Norway), was collected at baseline (before commencement of study treatment; termed T0) and on-treatment the third treatment day, two and 24 hours after the patient had received the preceding daily dose of vorinostat (termed T2 and T24), respectively. A full set of three samples (T0, T2, and T24) was obtained from 14 of the 16 evaluable study patients. The tubes were stored at −70°C until analysis. Total PBMC RNA was isolated using PAXgene Blood RNA Kit (Qiagen), following the manufacturer's protocol. RNA concentration and quality were assessed using NanoDrop 1000 and Agilent 2100 Bioanalyzer (Thermo Fisher Scientific Norway, Oslo, Norway), respectively. ## Gene Expression Array Analysis This analysis was performed by the Norwegian Genomics Consortium (Oslo, Norway). Briefly, cRNA synthesis, amplification, and hybridization to Illumina Human WG-6 v3 Expression BeadChip arrays (Illumina, Inc., San Diego, CA, USA), containing 48,000 probes, were carried out as per manufacturer's instructions. Signal intensities were extracted by the BeadArray Reader Software (Illumina), and raw data were imported into the GenomeStudio v2010.1 Software, Gene Expression module v1.6.0 (Illumina). The primary array data are available in the Gene Expression Omnibus data repository (GEO accession number GSE46703). ## Statistical and Functional Annotation Analyses of Array Data Analysis was performed using Bioconductor vR2.11.1 and the Bioconductor packages lumi 1.14.0, linear models for microarray data (limma) 3.4.4, and illuminaHumanv3BeadID.db 1.6.0 ([www.bioconductor.org](http://www.bioconductor.org)). Following quality control and pre-processing, the data were log₂-transformed, and differential gene expression between the sample groups T0, T2, and T24 was determined by applying a Benjamin and Hochberg false discovery rate-adjusted *P*-value cut-off of 0.05. The total number of probes that were identified as differentially expressed was analyzed using the Database for Annotation, Visualization and Integrated Discovery, DAVID v6.7. Enriched biological processes and pathways were identified using the GOTERM_BP_FAT and KEGG_PATHWAY algorithms, applying a *P*-value cut-off of 0.01. Differential expression analysis of the array data was also performed using a *P*-value of 0.01 and a log₂-fold change cut-off of 1.0 in order to identify genes whose expression changes could have potentially high biological significance. ## Experimental Human Colorectal Carcinoma Models The HCT116 and SW620 colorectal carcinoma cell lines were originally purchased from American Type Culture Collection (Manassas, VA, USA), and the identities of our laboratory's versions were confirmed by short tandem repeat analysis. The LoVo-92 colorectal carcinoma cell line was kindly provided by Dr. Paul Noordhuis (VU Medical Centre, Amsterdam, The Netherlands). The cell lines were cultured as previously described. Xenografts were established by subcutaneous injections of HCT116 or SW620 cell suspensions (2×10⁶ cells) bilaterally on the flanks of locally bred female BALB/c nude (nu/nu) or Athymic Nude- Foxn1^nu mice, 6–8 weeks of age. Vorinostat (Cayman Chemical, Ann Arbor, MI, USA; 100 mg/kg, dissolved in dimethyl sulfoxide to a concentration of 100 mg/ml immediately before use) or vehicle was given by intraperitoneal injection 13 days (HCT116) or 20 days (SW620) after establishment of xenografts. Three and 12 hours after administration, the tumors were extirpated, snap-frozen in liquid nitrogen, and stored at −70°C. The xenografts were sectioned using a cryostat microtome prior to RNA extraction using TRIzol® Reagent (Invitrogen Dynal AS, Oslo, Norway). RNA concentration was assessed using the RNA/DNA calculator Gene Quent II (Pharmacia Biotech, Piscataway, NJ, USA). ## Tumor Samples from LARC Patients Primary tumor biopsies were sampled at the time of diagnosis from LARC patients enrolled onto a phase 2 study on neoadjuvant chemoradiotherapy. The biopsy samples were snap-frozen in liquid nitrogen and stored at −70°C, and sectioned on the cryostat microtome, essentially as previously reported, before RNA was extracted. ## Reverse Transcriptase Quantitative Polymerase Chain Reaction (RT-qPCR) Analysis cDNA was synthesized from total RNA using the qScript™ cDNA Synthesis Kit (Quanta BioSciences, Inc., Gaithersburg, MD, USA). The qPCR was run in Perfecta qPCR Supermix (Quanta), on iCycler (Bio-Rad Laboratories Norway, Oslo, Norway) and with all reactions in parallel. Primers were designed using ProbeFinder Assay Design Software ([www.roche-applied- science.com/sis/rtpcr/upl/ezhome.html](http://www.roche-applied- science.com/sis/rtpcr/upl/ezhome.html)), and were obtained from the Universal ProbeLibrary collection (Roche Applied Sciences, Oslo, Norway). Primer sequences are listed in. Amplified cDNA generated from the reference cell line (LoVo-92) was included on all PCR plates for relative quantification purposes (correction of plate-to-plate variation). Data were normalized to the expression levels of two reference genes; *YARS*, encoding tyrosyl-tRNA synthetase, and *TBP*, encoding the TATA box-binding protein. When tested in the patient samples, the reference genes had equal expression per ng of cDNA, independent of patient treatment (vorinostat dose and time after administration). The data were analyzed using the GeneExpression Analysis for iCycler iQ® Real-Time PCR Detection System Software (BioRad), and were calculated relative to the level in the reference cell line and subsequently log₂-transformed. ## Statistical Analysis of qPCR Data Analysis was performed using Predictive Analytics SoftWare Statistics version 19.0 (SPSS Inc., Chicago, IL, USA). Q-Q plots were applied to test whether the data were normally distributed or not, before differences between groups were analyzed using two-sided Student *t*-test for the PBMC samples and Mann-Whitney *U* test for xenograft samples. *P*-values less than 0.05 were considered statistically significant. # Results ## PBMC Transcriptional Response to Vorinostat – Biological Processes and Pathways gives study patient baseline characteristics; the full study data on treatment tolerability and response have been reported previously. Of the 14 patients that provided a full set of PBMC samples (T0, T2, and T24), one patient was treated at vorinostat 100 mg once daily and three patients at the 200 mg dose level, whereas four and six patients received the medication at 300 or 400 mg once daily, respectively. Importantly, as vorinostat-induced tumor histone acetylation had been observed at all dose levels, the array data from all patient samples at each time point (T0, T2, and T24) were pooled, irrespective of the vorinostat dose administered to the patients. This was done to increase the statistical power of the testing on analysis of differential gene expression between the individual time points. As shown by, approximately 2,100 probes were differentially expressed both at two hours of vorinostat exposure (T2 *versus* T0) and on the T24 *versus* T2 comparison when applying the *P*-value cut-off of 0.05. Of these, 1,602 transcripts were found to be altered in both comparisons, and furthermore, no significantly differential expression was observed when comparing the T0 and T24 groups. Hence, all of the 1,602 mutual probes that were identified had a transient change in expression level from T0, with approximately one half found to be up-regulated and thus, the other half down- regulated at T2, followed by the opposite directional change to baseline expression at T24 (data not shown). Functional annotation analysis of the differentially expressed genes in patients' PBMC identified several enriched biological processes. Comparison of the baseline PBMC transcription profile with that obtained two hours after vorinostat administration (T2 *versus* T0) showed that 69 biological processes were over-represented, whereas the corresponding comparison of T24 *versus* T2 transcriptional profiles identified 106 processes. As seen from, displaying the top-ten Gene Ontology terms for each of the two comparisons, seven out of the ten biological processes were present in both, with transcription being the most significant. In addition, the analysis identified enrichment of genes involved in catabolic processes, the cell cycle, RNA processing, chromatin modification, and chromosome organization. The top-three pathway networks for each of the two comparisons, in common for both, comprised signaling factors of the cell cycle, including the p53 pathway. ## Vorinostat Activity in PBMC – Verification of Selected Biomarkers Next, by introducing a log₂-fold change cut-off of 1.0 while decreasing the *P*-value to 0.01 in order to identify gene expression changes with presumably high biological significance, the list of differentially expressed probes, all with a biphasic pattern of regulation from T0 through T2 and T24, was reduced to 38 candidates. Within this panel, two genes had duplicate array probes, whereas no reference sequence could be identified for three other probes, leaving 33 known genes as transcriptionally regulated by vorinostat following this stringent statistical analysis of the array data. Selection of genes for verification analysis by RT-qPCR was based on both the relevance in the DNA damage response, which is recognized as a significant mechanism contributing to clinical radiation sensitivity, and previous indication of regulation by HDAC inhibitors. Five of the 33 genes were found to fulfill both criteria: *MYC*, among the ten genes repressed at T2 and correspondingly, *GADD45B*, *MSH6*, *BARD1*, and *DDIT3*, among the 23 induced genes; mean PBMC expression levels at T0 relative to reference cell line expression are given in. These genes were present within the enriched biological processes and pathways identified by the functional annotation analysis of the differentially expressed genes, and the biphasic pattern of regulation in PBMC through T2 and T24 was confirmed with significant time-dependent changes (*P*\<0.01) for all of the five genes. ## Vorinostat Activity in Experimental Tumors – Validation of Selected Biomarkers We have previously shown histone hyperacetylation in vorinostat-treated human colorectal carcinoma xenograft models (HCT116 and SW620), peaking three hours after vorinostat administration and with restored baseline levels of histone acetylation three to six hours later, without accumulative effect following repeat daily administration. Hence, expression of the five selected genes was further assessed by RT-qPCR in HCT116 and SW620 xenografts, three and 12 hours after administering vorinostat to tumor-bearing mice; median control expression levels relative to reference cell line expression are given in. In the HCT116 model, a significant change (*P*\<0.05) in vorinostat-induced expression was found for *MYC* only. A similar transient *MYC* repression, but without statistically significant differences in expression levels through the time points, was seen in the SW620 tumors. ## LARC – Primary Tumor *MYC* Expression On identifying *MYC* repression as a possible biomarker of HDAC inhibitor activity from the strategy of analyzing, firstly, PRAVO study patients' PBMC, and secondly, vorinostat-treated colorectal carcinoma xenografts, and additionally recognizing this drug as a rational approach for biological optimization of radiation effect in pelvic gastrointestinal carcinoma, we investigated whether *MYC* might be expressed in the target tissue of a well- established pelvic radiotherapy protocol. In 27 LARC patients receiving neoadjuvant chemoradiotherapy, *MYC* expression was detected in all primary tumor samples, though at highly variable levels (median expression value was 0.47 (range 0.020–4.9) relative to reference cell line expression), but was essentially not associated with patient characteristics or treatment outcome in this small cohort. # Discussion Within the design of the PRAVO phase 1 study, combining the HDAC inhibitor vorinostat with fractionated radiation to pelvic targets volumes for determination of treatment tolerability and response, gene expression array analysis was performed of study patients' PBMC, sampled at baseline (T0) and on- treatment two and 24 hours (T2 and T24) after the patient had received the daily dose of vorinostat, in order to identify possible biomarkers of HDAC inhibitor activity. This strategy revealed 1,600 array probes with biphasic pattern of expression from T0 through T2 and T24 across all of the study patients. A significant number of these genes were found implicated in processes comprising gene regulation, the cell cycle, and chromatin biology. Applying stringent criteria for array data analysis, five genes were recognized both as players in the DNA damage response and targets for regulation by HDAC inhibitors, and were therefore selected for validation of expression pattern both in study patients' PBMC and in human colorectal carcinoma xenograft models. Of these, only *MYC* consistently showed rapid and transient repression in all conditions that were tested. In the setting of fractionated radiotherapy, a synergistic drug should preferably elicit a radiosensitizing molecular event at each radiation fraction; hence, a pharmacodynamic biomarker should reflect the timing of drug administration with regard to radiation exposure in a periodic manner. Importantly, in a prior preclinical *in vivo* study combining vorinostat and fractionated radiation, we observed that tumor histone acetylation, considered a biomarker of vorinostat activity in the radiotherapy target tissue, reached a maximum three hours after intraperitoneal vorinostat injection into the experimental animals and was restored to baseline acetylation level three to six hours later, but with a repetitive, transient induction of acetylation following repeat injections. Of note, tumor growth inhibition after fractionated radiation, representing a long-term phenotypic outcome of the experimental manipulations, was significantly enhanced both when radiation was delivered at peak and restored histone acetylation levels. Consequently, tumor histone hyperacetylation did not seem to be required at the time of radiation exposure, leaving the question of the optimum temporal relationship between administration of the radiosensitizing drug and radiation delivery unaddressed. In the PRAVO study, one patient at each vorinostat dose level had both baseline (before commencement of study treatment) and repeat tumor biopsy two-and-a-half hours after administration of vorinostat (on day 3 of the treatment protocol). Histone hyperacetylation was observed in all on-treatment biopsy samples, confirming the presence of vorinostat in the target at the time of the daily radiation exposure. However, given that one of the objectives of the study was to determine mechanisms of the presumed radiosensitizing action of vorinostat that were not simultaneously manifesting molecular perturbations elicited by the radiation itself, non-irradiated surrogate tissue was collected for the purpose of identifying new biomarkers. Several investigators have demonstrated PBMC histone hyperacetylation on HDAC inhibitor treatment,. With these aspects in mind, PBMC were deemed to represent a relevant surrogate tissue for studying radiosensitizing effects of vorinostat in the context of this clinical trial. Interestingly, using the study patients' PBMC as surrogate tissue for vorinostat exposure, all of the 1,600 probes that were found to be common for the comparisons T2 *versus* T0 and T24 *versus* T2 in principle represented pharmacodynamic biomarkers of the chosen timing of vorinostat administration in the fractionated radiotherapy protocol. The genes showed rapid and transient induction or repression, thus mirroring the kinetics of the histone acetylation response. This observation implies that the design of the PRAVO study, undertaken in patients with advanced gastrointestinal cancer, may not have provided the optimum context for detailed capture of molecular effects of vorinostat. Thus, ethical concerns may challenge the structure required within a clinical trial setting for evaluating novel biomarker endpoints. Nevertheless, in the PRAVO study, functional annotation analysis of the panel of 1,600 probes identified biological processes and pathways comprising gene regulation (transcription, RNA processing), cell cycle progression (including p53 signaling, commonly involved in the DNA damage response), and chromatin biology. These findings are consistent with well-known cellular perturbations following exposure of experimental tumor models to HDAC inhibitors. Investigation of biomarkers of HDAC inhibitor activity has been undertaken in a number of clinical therapy trials. These include the demonstration of increased histone acetylation in patients' PBMC in the early trials, and the more recent confirmation of changes in tumor expression of acetylated histone and non- histone proteins, the HDAC2 enzyme and HR23B protein, the latter been proposed as predictive biomarker, and of tumor proliferation index. Plasma protein profiling has been done in glioblastoma patients receiving vorinostat in combination with an established cytotoxic regimen. Furthermore, tumor gene expression array analysis has been performed in a study with the HDAC inhibitor panobinostat as single agent and in one trial each of combining either vorinostat or valproate with other biologic agents (in non-small cell lung carcinoma and acute myeloid leukemia, respectively). To our knowledge, the present study is the first to report on gene expression array analysis as an attempt to identify pharmacodynamic biomarker(s) reflecting timing of HDAC inhibitor administration with regard to an established cytotoxic regimen. The criteria for selecting genes for validation were both their presumed relevance in the DNA damage response and previous indications of regulation by an HDAC inhibitor, and additionally, in order to find ‘tumor-specific’ markers, omitting genes that typically might be associated with leukocyte biology. Four of the selected genes were induced by vorinostat in the study patients' PBMC but did not show a similar response in the experimental tumor models. *BARD1* encodes a nuclear factor with tumor suppressor activity, the stress response effectors encoded by *GADD45B* and *DDIT3* are implicated in cell cycle arrest, DNA repair, and apoptosis, and *MSH6* encodes a DNA mismatch repair protein. To date, only three studies seem to have been published on their potential use as biomarkers of therapy response. In contrast, the confirmation of *MYC* as the only one of the selected genes with rapid and transient change in expression in all tested conditions (*i.e.,* both in the study patients' PBMC and experimental tumor models) may point to a particular importance of myc in the therapeutic setting with fractionated radiation. Future investigations of vorinostat as possible radiosensitizing agent might be within a long-term curative radiotherapy protocol, for example as an additional component of neoadjuvant chemoradiotherapy for LARC. The confirmed presence of *MYC* expression in the intended radiotherapy target tissue (primary rectal tumors) in LARC patients encourages future exploration of this proto-oncogene as a novel biomarker endpoint. The myc protein acts both as transcriptional activator and repressor, regulating a myriad of genes that collectively conduct cell cycle progression, apoptosis, angiogenesis, and genetic instability. Specifically, it has been suggested that myc activates DNA damage repair genes, and interestingly, that myc in hypoxic tumors acts synergistically with the transcription factor hypoxia-inducible factor type 1α, HIF-1α. Recent evidence indicates that HDAC inhibition suppresses HIF-1α activity. Consequently, mitigation of DNA damage repair capacity through suppression of myc/HIF-1α synergy in hypoxic tumors, typically being resistant to radiation, provides an appealing explanation for the radiosensitizing effect of HDAC inhibitors. However, conflicting data have been presented as to how HDAC inhibition may influence the myc protein itself. Whereas inhibition of various HDAC enzymes has been shown to cause myc repression in a range of human cancer cell lines,, which corresponds well with the data in the present study, specific nuclear induction of myc to mediate HDAC inhibitor-induced apoptosis in glioblastoma cell lines has also been demonstrated. Interestingly, in nasopharyngeal carcinoma cells that were resistant to radiation, myc was found to be essential through the transcriptional activation of cell cycle checkpoint kinases, which are signaling factors implicated in DNA damage repair, thereby facilitating tumor cell survival following radiation exposure. On the contrary, although radiosensitization was conferred by HDAC inhibition both in hypoxic and normoxic hepatocellular carcinoma cells, a lower level of myc expression was associated with the hypoxic and more radioresistant condition. Of particular note, in the present study, the vorinostat-induced repression of *MYC* was found both in study patients' PBMC, clearly representing normoxic tissue, and experimental tumors that also were tested under normoxic conditions. In conclusion, integral in the PRAVO study design was the collection of non- irradiated surrogate tissue for the identification of biomarker(s) of vorinostat activity to reflect the timing of administration and also suggest the mechanism of action of the HDAC inhibitor. This objective was achieved by gene expression array analysis of study patients' PBMC and as a consequence, the identification of genes that from experimental models are known to be implicated in biological processes and pathways governed by HDAC inhibitors. Importantly, all of the identified genes showed rapid and transient induction or repression and therefore, in principle, fulfilled the requirement of being pharmacodynamic biomarkers for this radiosensitizing drug in fractionated radiotherapy. Among the identified candidate genes, *MYC* repression was found in all patient samples and tested experimental conditions, possibly underscoring the impact of the myc proto-oncogene in this particular therapeutic setting. # Supporting Information The authors thank Dr. Siri Tveito and Ms. Tove Øyjord for valuable assistance with laboratory procedures and Prof. Rob G. Bristow for helpful discussion. [^1]: This work was supported by MSD (Norge) AS, the Norway branch of Merck & Co., Inc., directed at overheads associated with the study. Moreover, this does not alter our adherence to all the PLoS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: AHR SD KF. Performed the experiments: AHR MGS EK IHGØ KS KR TWA KF. Analyzed the data: AHR MGS EK KR KF. Contributed reagents/materials/analysis tools: AHR MGS EK IHGØ KS KR TWA KF. Wrote the paper: AHR. Managed patients, databases, and tissue banking: AHR MGS SD KF.

# Introduction Deep-sea and open ocean waters are the largest and yet least understood environments on Earth. They are characterized by distinctive habitats and organisms and support an important part of the world’s biodiversity. Moreover, these ecosystems provide valuable direct and indirect goods and services, such as food provision and climate regulation. In the last decades, the human pressure on these systems has sharply increased threatening their health, biodiversity and resilience. In fact, the decrease of natural and mineral resources on land and in shallow waters, coupled with a rapid technological development which now allows the exploitation of formerly inaccessible areas, has caused a constant expansion of human-related activities toward deeper and more distant areas. Therefore, appropriated forms of governance and management of deep and open ocean ecosystems are essential to preserve their structures, processes and the services they provide. One of the major challenges in designing and implementing governance and management strategies in these environments is the fact that deep seas and open oceans frequently fall in areas beyond national jurisdiction (ABNJ). Therefore, international commitments are necessary to undertake effective conservation actions. In the last twenty years, several global meetings on biological conservation and sustainable development have proposed to set aside for protection 10–30% of all the marine biomes (including deep sea and open ocean realms) by the year 2012. The failure in meeting these objectives has been internationally recognized with just 1.17% of the world’s oceans currently included in marine protected areas (MPAs) mostly located in coastal waters. Revised biodiversity targets aiming at preserving 10% of all the marine biomes by 2020 were agreed at the 10th Convention of the Parties (COP) to the Convention on Biological Diversity (CBD). In order to speed up deep sea and open ocean conservation and achieve the proposed targets, the Parties to the CBD have adopted in 2008 seven scientific criteria for identifying ecologically or biologically significant areas (EBSA) in need of protection in open-ocean waters and deep-sea habitats (COP decision IX/20 paragraph 14). These criteria are: uniqueness or rarity; special importance for life-history stages of species; importance for threatened, endangered or declining species and/or habitats; vulnerability, fragility, sensitivity or slow recovery; biological productivity; biological diversity; and naturalness. The application of the CBD EBSA criteria should ultimately allow the establishment of representative marine protected area networks in the high seas and help the implementation of ecosystem based managements. These networks should cover a full range of examples across biogeographic regions as defined, for example, in the Global Open Ocean and Deep Sea (GOODS) lower bathyal biogeographic classification. CBD also defined five criteria for the definition of representative networks of MPAs: identification of ecologically or biologically significant areas, representivity, connectivity, selection of replicated ecological features and selection of viable and adequate sites. The identification of ecologically and biologically significant seamount areas and the selection of the seamounts more suitable for conservation may represent an important first step in the creation of such networks. However, the present framework is intended to be applied to individual seamount features and not for the identification of networks of such sites. The methodologies suggested by the CBD should not be restricted to ABNJ and could also be adopted and implemented within areas of national jurisdiction. Pilot studies have identified several potential EBSAs in different marine regions ([www.gobi.org](http://www.gobi.org)). However, the patchy nature of biological and ecological data regarding deep and open ocean ecosystems hinders a systematic application of these criteria and implies a wide reliance on global models and remote sensed data. Moreover, areas of critical importance in the water column tend to shift in time and space, making the location of pelagic EBSAs even more difficult. Dynamic marine protected areas have been suggested as tools for conserving open ocean biodiversity. However, there has been some debate on their workability and utility questioning the possibility of a rapid implementation of pelagic MPAs in real world conservation actions –. Thus, non-dynamic features such seamounts and ridges may represent good starting points for a systematic implementation of offshore marine reserves, since they have been demonstrated to be easier to conserve, map, survey, and enforce than ephemeral areas. At the same time conservation of seamount ecosystems seems to be beneficial both for benthic and pelagic organisms. Seamounts are prominent and ubiquitous features of the world’s underwater topography, and constitute one of the largest biomes of the deep-sea. Several authors have illustrated their importance for the benthic and pelagic realms. For example, Samadi et al. found an increased species richness and abundance of galatheid crabs on seamounts and proposed that benthic invertebrates are more abundant and attain higher diversity on submarine reliefs compared to other deep-sea habitats (‘oasis hypothesis’). The high densities of filter feeders, especially corals and sponges, that can be encountered on seamounts seem to confirm the oasis hypothesis, even though no robust quantitative estimates are currently available. The interaction of seamounts with vertically migrating organisms and passing oceanic flows appears to facilitate trophic exchanges toward top pelagic predators. Therefore, seamounts seem to be important hotspots for pelagic biodiversity and visitor organisms and play an important role in enhancing fishery catches of some pelagic species,. In Morato et al., in particular, the aggregating behavior of large pelagic fish, both visitors and not, was showed to be diffuse throughout Southwest Pacific seamounts and to occur within 30–40 km of seamount summits. However, seamounts are very heterogeneous habitats and the above mentioned properties may not be common to all submarine features. In fact, seamounts are generally characterized by diverse geophysical properties, which in turn are likely to affect the biological diversity and production of resident and associated organisms,. As a consequence, the protection of different seamounts may ultimately result in very different outcomes. The use of the CBD EBSA criteria can help to identify seamounts more likely to be suitable for protection. Although the CBD EBSA criteria suite represents a powerful tool in identifying areas of particular ecological or biological importance, parallel socio-economic and governance analysis are needed if marine policies are to find a balance among multiple ecological, socio-economic and other governance objectives. An important part of this process is represented by a correct definition and measurement of the major human activities occurring in these areas. Fishing is considered one of the major threats to seamount ecosystems, having long-term impacts on different habitats, such as coral and sponge aggregations e.g. – and on vulnerable, long-lived fish stocks e.g.. The detrimental effect of fishing on seamounts has been stressed in the FAO guidelines for sustainable fishing, where submarine elevations are listed as an example of vulnerable marine ecosystems. Besides fisheries, deep-sea mining is emerging as an important issue in seamount management. Different types of metal-rich deposits can be found on seamounts, of which Fe-Mn crusts and massive polymetallic sulphide are of highest commercial interest. Even though no substantive exploitation has started, with the exception of few exploratory surveys, mining activities on submarine features are likely to pose a serious threat to seamount ecosystems in the near future. In this study we propose a framework for applying the CBD EBSA criteria to locate potential ecologically or biologically significant seamount areas, based on the best information currently available. In particular, this work developed methods for applying the EBSA criteria to individual seamounts and methods to assess the impact of different fishing gears and mining activities on the various components of individual seamounts such as pelagic, benthopelagic and benthic environments. This framework will allow managers to identify EBSAs and to prioritize their choices or policies in terms of protecting undisturbed areas, protecting disturbed areas for recovery of habitats and species, or both. CDB prioritize areas having low levels of disturbance relative to their surroundings. However, where no natural areas remain, areas with high possibilities of recovery after the cease of anthropogenic related activities should be considered. Thus, measuring major human activities is of paramount importance in the seamount conservation process. The application of the present framework to seamounts and the possibility to redesign it for other habitats (e.g., hydrothermal vents, pelagic fronts, etc.) could strongly enhance a systematic approach to deep sea and open ocean management. The outcomes of these evaluations should serve as a powerful tool for identifying sites of particular importance for conservation which can then be integrated in MPA networks following the set of principles and criteria guiding design and implementation of MPA networks e.g.. # Methods The framework proposed in this study for assessing seamount EBSAs was developed within the conservation part of the Seamount Ecosystem Evaluation Framework, SEEF – and consists of both a semi-quantitative scoring of the biological value individual seamounts have with respect to the EBSA criteria and an evaluation of the main threats posed to each seamount. The overall EBSA and threats scores can then be used to visualize all seamounts on a scale ranging from low to high likelihood of being an EBSA and low to high threats allowing for the comparison of different features. Different definitions of the term seamount are available in literature : here, seamounts are considered as topographically distinct seafloor features greater than 100 m in height but which do not break the sea surface. ## Application of the EBSA Criteria to Seamount Ecosystems The relevance of individual seamounts with respect to the different EBSA criteria were assessed based on the presence of particular habitats, communities and species (i.e. indicators) that capture the most relevant pelagic, benthopelagic and benthic components of seamount ecosystems. The presence of such indicators can be assessed by both using real data coming from sampled features or global models which could complement data deficient sites. ### Uniqueness or rarity (C1) The presence of hydrothermal vents on seamounts was used as a proxy for uniqueness. In fact, vents host a number of special communities and organisms that are found nowhere else in the marine environment e.g.. A second factor used to assess seamounts uniqueness was the presence of macrophytes. Benthic primary producers are extremely rare in the open oceans and therefore represent a valid proxy for rarity. This indicator has already been proposed to candidate the Saya de Malha Banks as an EBSA (<http://www.gobi.org/candidate-ebsas>). The presence of endemic organisms to assess the level of faunal uniqueness has to be considered with extremely caution, since seamount endemicity has been recently questioned,. In fact, the exploration of new seamount areas is generally followed by the description of several endemic species e.g., which tend to keep their endemic status only until new studies with wider and more detailed spatial and taxonomic coverage are performed e.g.. The number of endemisms may thus be directly related to sampling effort. Therefore, considering the difficulty in discriminating between true and apparent endemism, it was decided to consider this indicator in the uniqueness or rarity criteria but its implementation will not be done until more clues on the seamount endemicity hypotheses are revealed. ### Special importance for life-history stages of species (C2) Areas containing breeding or spawning grounds, juvenile habitat and important habitats for migratory species are considered good examples of this criterion. Seamounts represent an important feeding and/or spawning ground for a number of different seamount-associated fishes. These species are known in literature as “aggregating deep sea fishes” and are described and listed in Koslow and Morato et al.. Furthermore, seamounts play an important role for large visiting pelagic species (i.e., tunas, billfishes and large pelagic sharks) e.g., and air- breathing visitors (i.e., marine mammals, marine turtles and seabirds) e.g.. Therefore, the presence of aggregating deep sea fishes, large visiting pelagics and air-breathing visitors was used to identify areas with special importance for life-history stages of species. ### Importance for threatened, endangered or declining species and/or habitats (C3) The IUCN red list provides a comprehensive list of threatened, endangered or declining species. Air-breathing visitors, large visiting pelagic and bottom fish and shark species which are included in the red list as critically endangered, endangered, vulnerable or nearly threatened were used to assess this criterion. Thus, when such species are reported for a seamount, this location is considered important for threatened species or habitats. In addition, the presence of habitat-forming cold water corals and sponge aggregation which represent declining habitats e.g. was also reckoned to be relevant to this criterion and included in the present analysis. It is important to notice that the cold water corals used here as a proxy for criteria 3, 4 and 6 are exclusively those forming deep-water reefs and gardens. A list of structure- forming corals was adapted from Roberts et al. and is provided as Supporting Information. ### Vulnerability, fragility, sensitivity, or slow recovery (C4) This criterion values the degree of risk that will be incurred from human activities or natural events. Coral gardens and reefs, sponge aggregations and vent communities are highly sensible to human disturbance and are listed in international guidelines as examples of vulnerable marine ecosystems deserving particular protection. Besides these groups, aggregating deep-sea fishes *sensu* were considered in the assessment of this criterion since these species are long-lived and slow-reproducing organisms extremely vulnerable to human disturbance,. ### Biological productivity (C5) The dynamics of marine production in the open and deep oceans are poorly understood. To date, the depth of submarine features represents the best indicator of seamount productivity. In fact, shallow seamounts may intercept the diel vertical migration of zooplankton and micronekton, trapping these vertically migrating organisms and/or aggregate small zooplanktonic animals horizontally advected. This could result in an increased prey availability, which may benefit resident and visiting animals enhancing secondary production and aggregating behaviors. The lower range of these migrations is thought to range between 400 and 800 m. Therefore, seamounts shallower than 800 m were assumed to have a higher productivity than deeper features. In addition, metazoan meiofauna, macrofauna and megafauna abundance and biomass tend to decrease sharply with depth as a consequence of limited nutrient availability in deeper water. Therefore, deeper seamounts are likely to be less productive than shallower ones. Benthic primary producers (macrophytes) and hydrothermal vent communities were also regarded as indices of high biological production e.g.. ### Biological diversity (C6) Reliable estimates of biodiversity for seamounts are difficult to obtain with the data currently available in the scientific literature. However, the presence of structural species may increase local biological diversity and be used as a proxy for this criterion e.g.. Therefore, the presence of seamount habitats dominated by cold water corals, sponges or macrophytes was used to assess this criterion. ### Naturalness (C7) These are areas with a comparatively higher degree of naturalness as a result of the lack or low level of human disturbance. The naturalness of individual seamounts depends on the typology and intensity of the anthropogenic activities over time. In the present work we considered a seamount to have a high degree of naturalness when no fishing or mining activities were known to occur. It is well recognized that different fishing activities have very different impacts in the ecosystem. This distinction was taken in consideration when quantifying the human-induced disturbances on seamount ecosystems. ## Seamount EBSA Scoring Procedure Ten indicators were used to identify seamount EBSAs: four benthic (hydrothermal vents, macrophytes meadows, cold water corals, sponge aggregations), two benthopelagic (aggregating deep-sea fishes and threatened bottom fish/sharks), two pelagic (large visiting pelagic and air-breathing visitors), one historical (naturalness) and one geological (depth). The different proportion of indicators adopted for each seamount component reflects the relative importance benthic, benthopelagic, pelagic, geological and historical indicators had in the individuation of potential seamount EBSAs. Individual indicators were weighted based on the relevance they have in the assessment of the EBSA criteria. Factors used to verify more than one criterion had a higher weight in the analysis. For example, the presence of habitat- forming cold water corals was used as a proxy for criteria 3, 4 and 6 and thus was weighted three times higher than depth, which was considered only for criterion 5. Visiting pelagic and air-breathing visitors were generally used as a proxy for criterion 2. However, if the large pelagic or air-breathing species present were listed in the IUCN red list as critically endangered, endangered, vulnerable or nearly threatened they became relevant both for criterion 2 and 3 and their weight was doubled in the scoring process. The final score represents an index of the likelihood of having ecologically or biologically significant seamount areas on a particular seamount and was calculated based on the proportion of indicators present and on their weight in the analysis. It can range from 1 when no indicator is present to 5 when all indicators are present at a specific seamount. The final outcomes were presented as two nominal categories, from here on referred as “seamount EBSA likelihood score”, indicating the chance of having EBSAs on the assessed seamounts. These categories were: low, for total scores ≤3 and high, for total scores \>3. ## Quantify Human-induced Threats to Seamount Ecosystems The major human activity currently impacting seamounts is fishing, and its effects are highly dependent on the type of fishing gear used. The main fisheries occurring on seamounts use trawls, longlines, gillnets and pots and traps but other methods such as hook and line are also present in some small- scale seamount fisheries. In addition to the type of fishing gear used, fishing effort (duration and frequency of fishing events) and catch data (landings, bycatch and discards) are essential to determine the actual impact of any fishery. However, considering the lack of specific data regarding seamount fisheries, it is generally not possible to consider catch data and fishing effort. Therefore, the evaluation of fishing impacts on individual features was here exclusively based on the types of fisheries present. Mineral exploitation is likely to pose a serious threat to seamount ecosystems in the near future, and was therefore included in the evaluation as a potential threat factor. Finally, climate change will probably constitute the greatest threat to aquatic ecosystems. Meanwhile, considering our poor understanding of the repercussions it will have on deep-sea organisms and habitats and the consequent difficulties in quantifying the effects of phenomena such as ocean acidification and hypoxia, it was decided not to implement climate change in the present analysis. A revision of the present framework should be considered as soon as new studies will clarify the consequences climate change will have on seamount ecosystems. ## Scoring Procedure for Human-induced Threats to Individual Seamounts The effects of fishing and mining on individual seamounts were quantified using an expert knowledge system. This system was adapted from two recent reviews where experts were asked to rate the impact of several fishing gears on different taxonomic groups and habitats using a scale ranging from very low to very high. The set of ratings, threats and ecological groups considered in these studies were revised in order to obtain categories and scores more meaningful for seamount ecosystems. A total of nine types of threats (1 mining and 8 fishing activities) were regarded as particularly relevant for seamounts and included in the framework: bottom gillnet, hook and line, bottom and pelagic longline, pots and traps, purse seine, midwater and bottom trawl, seafloor mineral extraction. Five potentially threatened components of seamount ecosystems were identified: two benthic (physical habitat and habitat-forming corals and sponges), one benthopelagic (groundfish) and two pelagic (large pelagic fish and air-breathing visitors, i.e. marine mammals, marine turtles and seabirds). The impact of each fishing gear and mining activity on each ecological group was rated on a scale from 1 -very low- to 5 -very high-. These ratings ultimately resulted in a weighting system for each human activity where, for example, bottom trawling poses a different degree of threat to seamount ecosystems than pelagic longlining. The “threats score” of individual seamounts was determined by the anthropogenic activities occurring on that features. In fact, benthic, benthopelagic and pelagic seamount components will experience different levels of disturbance depending on the set of human related activities present. Among the set of fishing and mining practices present on a specific feature, it is possible to identify a subset of activities likely to pose the highest impacts to the different components considered (i.e. the ones posing the greatest threat to physical habitat, the ones posing the greatest threat to habitat-forming corals and sponges, etc). The final threat score was thus calculated as the average of the maximum impacts posed to the different ecological groups and therefore it takes into account the benthic, benthopelagic and pelagic statuses of all the evaluated seamounts. The threats score range from 1 when no activity is present to 5 when the activities present potentially pose very high impacts on all the considered ecological groups. The final “threats score” (TS) were presented as three nominal categories: 1) none (TS = 1), seamounts with no anthropogenic impacts; 2) low (1\<TS≤3), anthropogenic activities do not have severe impacts on any components of the seamount ecosystem, have moderate impacts on the seamount ecosystem, or impact severely only one component; 3) high (TS\>3), anthropogenic activities are impacting several components of the seamount ecosystem and more than one is severely affected or have severe impacts on all the considered components of the seamount ecosystem. ## Data Uncertainty Index To account for data uncertainty, data quality issues and the varying degree of knowledge regarding different seamounts and geographical area, a data uncertainty index similar to the one elaborated in Wallace et al. was developed. This index is evidence-based and serves as a measure of our confidence about EBSA likelihood and threats score assigned to individual seamounts. This index is calculated independently for EBSA likelihood and threats score. Two measures were incorporated in the index: a data quality index (DQ) and a data deficiency index (DD). Data quality reflects origin and nature of the collected data and was divided into three categories: low (scored as 1), medium (scored as 0.5), and high (scored as 0) data quality. Considering the wide nature of seamount studies, the definition of these categories was not very strict but was based on general guidelines. The high data quality category was designed to include information mainly derived from rigorous scientific surveys. Even though data included in this category should be predominantly quantitative, qualitative data may also be considered as high quality data if in great detail. Medium quality data are incomplete quantitative information or qualitative descriptions of EBSA indicators and human impacts present on individual seamounts. These data should always be specific to a particular feature and validated in the literature. Low quality data include undisclosed data regarding wide geographic areas which do not specifically address seamounts, information inferred from models or from different seamount properties or data not properly referenced. The detailed scoring standards used to assign the data quality scores to all EBSA and threat indicators are described in the supplementary information. Data deficiency (DD) was defined as the proportion of threats or EBSA indicators lacking data. Data deficiency could range from 0 (information available for all threats and indicators) to 1 (no information available). DQ and DD are combined into a data uncertainty index associated with each final EBSA likelihood and impact score. Data uncertainty is the sum of the average DQ score and the DD score. The data uncertainty index has a minimum value of 0 (all factors scored with high data quality) and a maximum value approaching 2 (few factors scored with low data quality). The data uncertainty index is visualized as error bars in plots of the seamount EBSA likelihood scores versus human threats scores. Minimum and maximum values of each final seamount EBSA likelihood and human threats score are calculated by subtracting and adding the data uncertainty index. The error bars can therefore potentially range up to two units above and below the original score. In this manner, a seamount lacking data and/or with low data quality is shown as possibly belonging to different EBSA likelihood and threat categories, reflecting the uncertainty of the outcome. ## Seamount EBSA Portfolio Protecting all seamounts is neither particularly rewarding nor practically feasible considering the high variation that exists in terms of their ecology, geophysics and potential human impacts and the large number of seamounts present in the world’s oceans. Therefore, approaches that systematically highlight conservation priority areas for seamount ecosystems can constitute a valuable tool for marine management purposes. We hereby propose an approach that combines the likelihood of a seamount constituting an EBSA and the level of human impact posed to a submarine feature to locate priority areas for seamount conservation at global, regional and local scales. This methodology allows the classification of individual seamounts into four main conservation categories, which can help in optimizing management efforts toward the protection of the most suitable areas. The portfolio categories are: Low EBSA likelihood-Low threats; Low EBSA likelihood-High threats; High EBSA likelihood-Low threats; High EBSA likelihood- High threats. EBSA likelihood and threats for individual seamounts can be easily summarized and graphically compared. Additionally this approach is designed in a way that helps in visualizing what parts of the ecosystems (e.g., benthic, benthopelagic or pelagic) are contributing to the EBSA score or being threatened by human induced activities. This is of paramount importance in identifying seamounts that may be ecologically or biologically significant for both the benthic and pelagic components of the ecosystem and in complementing preexisting conservation strategies with the protection of underrepresented seamount components. ## Case Studies and Methodology Test In order to test the framework developed here, we have randomly assigned the presence or absence of the ten indicators developed in the EBSA scoring procedure and of the nine types of threats considered in the threats scoring procedure to 1000 dummy seamounts (i.e. hypothetical seamounts having randomly assigned configurations of EBSA and threat indicators). In this way it was possible to assess the ability of our framework to assign real world seamounts to the different portfolio categories considered. In addition, a set of eight seamounts from different geographical areas were selected as case studies for applying the present framework, six located in the Atlantic Ocean (Sedlo, Condor, Anton Dohrn, Rosemary and Josephine seamounts) and two located in the Gulf of Alaska (Cobb and Bowie seamounts). Data for the evaluation process were obtained by reviewing the existing literature and were presented in detail. # Results ## Testing the Methodology The dummy seamounts were assigned to all 4 portfolio categories with only 5.1% seamounts considered as highly likely to be identified as EBSA with low degree of threat and 3.6% seamounts with low likelihood of being identified as EBSA with low degree of threat. Most of the dummy seamounts fall in the category low likelihood of being identified as EBSA with high degree of threat (36.2%) or high likelihood of being identified as EBSA with high degree of threat (55.1%). These results indicate that the framework is adequate to assign seamounts to different portfolio categories. In the outcomes of the framework can be visualized and seamounts compared allowing managers to prioritize their choices or policies in terms of protecting undisturbed areas, protecting disturbed areas for recovery of habitats and species, or both. ## Case Studies For the 8 case study seamounts considered, seamount EBSA likelihood scores ranged from low for Rosemary seamount (2.5±0.59) to high for all other seamounts. Sedlo and Gorringe presented the highest EBSA likelihood scores (3.86±0.25 and 3.86±0.58, respectively). The uncertainty around these estimates is high for seamounts with low data quality and less attributes scored. For example, Josephine was identified as having a high likelihood of being a seamount EBSA but its score ranged from 2.59 to 3.99, i.e. from low to high likelihood of being an EBSA. Seamount threats scores ranged from low for Bowie (2.4±0.5), Sedlo (2.6±0.28) and Cobb (2.6±0.54), to high for Condor and Anton Dohrn seamounts (3.6±0.50 and 3.6±0.61, respectively) and were highest for Gorringe, Josephine and Rosemary seamounts (4.6±1.34, 5.0±1.16 and 5.0±0.61, respectively). Note the very high uncertainty around the threats estimates demonstrating the low data availability and quality. Overall, the eight seamounts evaluated were allocated to three different portfolio categories of EBSA likelihood and threat exposure: high likelihood of being an EBSA-high threat exposure (Condor, Anton Dohrn, Gorringe and Josephine), high likelihood of being an EBSA-low threat exposure (Sedlo, Bowie and Cobb), and low likelihood of being an EBSA-high threat exposure (Rosemary). This framework also allow for the exploration of the parts of the ecosystem contributing for the definition of an EBSA or under major threat. For example, Gorringe seamount has a high likelihood of being an EBSA mainly because of its benthic and benthopelagic environments. The main threats posed to this seamount are also on deep-water corals and groundfish. On the other side, Sedlo seamount may be rich on benthopelagic communities but its groundfish or deep-water corals are not being impacted by human activities considered in this study. In fact, fishing on Sedlo seamount is exclusively based on longliners. The presence of pelagic longline fisheries on a seamount will results in a final threat score of 2.6 because of the high level of bycatch related to pelagic longlining. However, considering the apparently low levels of bycatch in the Azorean fisheries e.g., this classification might overestimate the threats to Sedlo and other Azorean seamounts. Of the case studies considered, Sedlo and Gorringe are those with the highest EBSA likelihood scores and, therefore, they may represent the most suitable areas where to adopt conservation measures. These two seamounts are experiencing very different levels of human pressure, with Sedlo presenting lower chances of detrimental effects caused by human activities than Gorringe. However, given the large uncertainty associated with the Gorringe threat score, a further evaluation of the activities should be undertaken. Sedlo has already been proposed as a suitable site for a marine protected area. The preliminary outcomes of the present study seem to support this proposal highlighting a possible management strategy whose goals would be to protect a biologically and ecologically valuable area and, at the same time, to limit eventual conflicts between conservation and socio-economic interests. On the contrary, at the moment no conservation measure is scheduled for Gorringe seamount. In planning future actions on these two seamounts it should also be considered that Gorringe is the only one hosting benthic primary producers (macrophytes) among the Atlantic seamounts considered in this analysis, while Sedlo seamount seems to have a higher relevance for pelagic organisms and is the only submarine feature, among those considered, having high naturalness. # Discussion In order to achieve the conservation goals established under the Convention on Biological Diversity, scientists from different fields were asked to define and apply criteria which can highlight marine areas of particular interest. In this context, identification and management of submarine mountains suitable for protection may represent an important step toward the systematic preservation of deep sea habitats and open ocean waters. In fact, even though many aspects of seamount ecosystems persist unknown, mounting evidence shows that they might play a key role in sustaining the pelagic and benthic production and biodiversity of deep seas and open oceans. The framework proposed here was designed to set priorities in seamount conservation and to help developing spatially explicit seamount management policies. In order to avoid the location of protected areas in places that contribute little to preserve ecosystem structures and processes, the biological or ecological value of specific areas should always be considered as a primary criterion for the identification of conservation priorities. However, this criterion alone is not sufficient. The implementation of protected areas has to be included in a wider management context which integrates bio-ecological, economic and social goals to be successful. The present framework, by considering both the conservation value of different seamounts with respect to the EBSA criteria and the importance of specific areas to human activities, represents one of the few attempts to implement an ecosystem approach to management in the deep sea. It allows, in fact, the definition of different strategies based on the governance objectives. For example, if there is the intention of restoring damaged ecosystems, seamounts having both high EBSA likelihood and high human threats scores will be chosen for conservation (uppermost right part of). An alternative management policy similar to what was approved by CBD might focus toward the preservation of pristine areas with low levels of fishing and mining, where the likelihood of interactions between human activities and seamount EBSAs is low (uppermost left section of). The case studies Sedlo and Gorringe represent an example of these two strategies. The final choice of the most appropriate strategy should always be site-specific and should depend upon socio-economic (e.g., economic importance; economic replaceability of the site; etc.) and ecological factors (e.g., likelihood of recovery after the cessation of the human activities; etc.) and on the specific goals managers have. Furthermore, management-related criteria (i.e., criteria measuring how feasible is to effectively manage a site to achieve conservation goals) should also be considered in the final selection of the most suitable sites for conservation. Three aspects were central in the practical definition of the proposed methodology. Our first concern was to develop a system which could provide solid measures of the relative value and threat status of individual seamounts. The choice of the selected seamount EBSA indicators and the definition of the most relevant human activities to seamount ecosystems were based on an extensive review of the existing literature and through intense consultations with seamount experts. This approach constitutes, therefore, a complete synthesis of what is presently known regarding seamount ecosystems. Our second concern was to design a system compatible with the data currently available. The major constraint faced by this kind of analysis is generally the scarcity of information readily available. In fact, while a very small portion of submarine relief has a fairly detailed ecological and biological description and an accurate report of ongoing anthropogenic activities, the large majority of seamounts have either never been explored or even charted through direct scientific measurements or only partially described. The development of a methodology which could evaluate the relative importance and the threat status of individual features based on presence/absence data and deal with data deficiencies represented an attempt to overcome these limitations. The use of global models or questionnaires addressed to seamount experts may furnish information regarding the human activities currently present on individual features and allow the application of the present framework in area where little information is available, while a data quality index shows the confidence we have about any outcome provided. Finally, particular care was paid to keep the results simple to visualize and understand in order to facilitate their implementation in future management actions. The outcomes of the dummy seamounts may serve as an indication of the robustness of our methodology. Since our analysis was based only on presence/absence data, conservative outcomes regarding the threat status of individual seamounts should be expected. This is reflected in the outcomes of the analyses, where the highest proportion of seamounts fell within the high threat score category. Moreover, the consistent allocation of the dummy seamounts into high and low EBSA likelihood categories, as shown in, is important to effectively highlight seamount areas of particular importance for conservation and to speed up management actions. Finally the combination of EBSA and threat scores identified seamounts belonging to all the four main portfolio categories indicating that the framework is adequate to assign seamounts to different portfolio categories. Another important characteristic of this framework is that allows the identification of seamount EBSAs and threats considering different ecological groups in the pelagic and benthic realms (see example). This is a major step forward in the integration of these often segregated parts of the ecosystem and may allow managers to complement pre-existent conservation measures and to selectively mitigate the negative effects of human activities on particularly relevant seamount components. This framework will also allow the identification of seamounts with high data uncertainty and thus in urgent need of research. The methodology proposed here may constitute an important step forward in the implementation of conservation measures in deep see habitats and open ocean waters and help to fulfill the international commitments signed under the CBD. The simplicity of its scoring procedure and the nature of the data required to perform the analysis make it easy to understand and implement in actual conservation actions. Its systematic application at local scales and in different biogeographic provinces may enhance the conservation status of marine areas difficult to manage and ensure the protection of a wide range of habitats and organisms (both benthic and pelagic). Future improvements to the methodology may consider the inclusion of additional threats such as climate change and pollution and tailor the threats score on a regional basis. Additionally, better quantitative assessment of the threats posed to a seamount could be implemented by including, for example, year when activities started, a measure of fishing effort and prospective mining areas. Currently, we assume that the presence of a particular threat would always have the same effects on the considered ecological groups which likely is a simplification. Future improvements to the framework should also take into account spatial and temporal patterns. In fact, both seamount EBSA indicators and human activities are not constant in time and space. However, detailed knowledge of seamount ecosystems and long time series data with high spatial and temporal resolution are required. At the moment, with a few exceptions, these conditions cannot be met. Serious doubts regarding the sustainability of seamount trawl fishing and mining have been raised several times in the scientific community e.g.. Increasing human pressure and poor knowledge of seamount ecosystems leave us little room to effectively direct conservation actions. Therefore, this framework which attempts to synthesize the best information currently available and guide conservation actions on the basis of ecological and economic values may represent an important tool to mitigate the biodiversity loss of one of the most representative deep water ecosystems. Its capability of highlighting seamount areas of particular importance, coupled with the spatial assessment of two key activities such as mining and fishing, may constitute the first step toward the implementation of representative and viable networks of marine protected areas. # Supporting Information We thank Ricardo Serrão Santos for critical advice and valuable comments on the manuscript. We acknowledge many colleagues around the world by sharing their vision and data on seamount ecosystems. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: GHT KØK TJP TM. Performed the experiments: GHT KØK. Analyzed the data: GHT TM. Contributed reagents/materials/analysis tools: GHT KØK TJP TM. Wrote the paper: GHT KØK TJP TM.

# Introduction The current state of academia is sometimes referred to as a system affected by hyper-competition. This goes hand in hand with strong emphasis on quantitative assessment of scientific output through journal impact factors, citation analyses and the *H*-index. The number of publications, citations and grants determine to a large extent the status and recognition of academic researchers. Consequently these indicators influence the recruitment, promotion and tenured appointments of researchers. This may in turn induce a high level of perceived publication pressure. In line with Woolf, we define perceived publication pressure as the subjective pressure resulting from the feeling that one *has* to publish. In line with work stress literature, strong perceptions of pressure could provoke stress, but need not to when one has many resources available to manage the pressure. Applied to publication pressure: Publication demands and attitude towards the current publication climate determine the perceived pressure, yet pressure can be alleviated by resources like helpful co-authors, involved colleagues or supervisors, and a sense of academic competence. Some degree of publication pressure can be an incentive to produce high quality scientific work. Yet, too much publication pressure may have detrimental effects on the scientific enterprise in general and on individual researchers in particular. Excessive publication pressure is associated with poor quality research (and teaching), a decreased willingness to share raw data, less involvement from researchers in public and policy issues, and less academic creativity. The perceived hypercompetition is thought to lead to less rigorous (“rushing into print”) and less reliable science. Publication pressure is associated with a greater likelihood to engage in research misbehaviours. Lastly, publication pressure is associated with a disproportionate focus on positive and specular findings. Publication pressure may also have detrimental effects on individual researchers. It is linked to a poor research climate and may render academic researchers emotionally exhausted. Previous research on publication pressure found junior researchers to experience more publication pressure compared to their senior counterparts. Studies investigating publication pressure thus far have mainly included academic researchers from particular disciplines like biomedicine, management and population studies, and included only a subset of academic ranks. This limits the generalizability of the degree to which researchers perceive publication pressure. The current study aims to assess whether researchers from all academic ranks (including PhD students) and all disciplinary fields perceive publication pressure. This is important, as differences between academic ranks could signal the need for tailored interventions. Besides, comparing different disciplinary fields may enable us to determine fields that perceive less publication pressure. This may generate new insights in the nature of publication pressure and possible protective factors. Our research question was: What is the level of perceived publication pressure in the four academic institutions in Amsterdam and does the pressure to publish differ between academic ranks and disciplinary fields? # Materials and methods ## Ethical statement Our study was ethically reviewed and approved by the Scientific and Ethical Review board of the Faculty of Behavioural and Movement Sciences (Vrije Universiteit Amsterdam). ## Participants All academic researchers in Amsterdam employed in research for at least one day per week at one of the four academic institutions (Vrije Universiteit Amsterdam, University of Amsterdam and the two Amsterdam University Medical Centers) were eligible to participate. This included PhD students, as in The Netherlands PhD students are employees. ## Procedure First, we set up a data sharing agreement with all participating institutions to safely obtain the e-mail addresses of their researchers. Second, we sent an information letter inviting all academic researchers in Amsterdam (*n* = 7465) to take part in our study. The information letter contained links to the study protocol and the study’s privacy policy. In addition, we included a link to a short non-response questionnaire where we asked researchers to report their academic rank, gender and enquired whether the reason for declining participation resulted from a sense that their data were not protected. For the full non-response questionnaire, see. A week later, researchers were invited to complete an online survey. The survey started with an informed consent statement followed by the inclusion check (“Are you currently employed in research for at least one day per week?”) and ended with the demographic items about participants’ academic rank (PhD student, postdoc, assistant professor, associate professor and full professor) and major disciplinary field: biomedicine (consisting of life and medical sciences), natural sciences, social sciences (included both social and behavioural sciences) and humanities (consisting of humanities, language, communication, law and arts). We used Qualtrics (Qualtrics, Provo, UT, USA) to create and distribute the survey, that took approximately 15 minutes to complete. We sent three reminders, each 10 days apart. ## Instruments We used the revised Publication Pressure Questionnaire (PPQr) to measure publication pressure. The PPQr is a valid and reliable instrument to measure publication pressure and consists of 3 subscales scored on a 5-points Likert scale (‘Totally agree’ = 5, ‘Totally disagree’ = 1). The Publication Stress subscale (6 items—Cronbach’s α =.804) regards the stress a researcher experiences due to the feeling she/he has to publish and includes items such as “I feel forced to spend time on my publications outside office hours”. The Publication Attitude (6 items—Cronbach’s α =.777) subscale reflects researchers’ attitudes towards publication pressure, for example: “Publication pressure harms science”. Finally, the Publication Resources subscale (6 items—Cronbach’s α =.754) consist of factors that can help prevent publication pressure (e.g. feeling of competence, freedom to choose topics of scientific investigation; involved colleagues). A typical item would be: “When working on a publication, I feel supported by my co-authors.”. The full PPQr questionnaire can be found in. PPQr subscale scores are computed by taking the average of all items in the subscale. A higher score on all subscales means the researcher perceives publication stress, has a negative attitude towards the publication climate and perceives little publication resources to alleviate publication stress. The survey contained two other instruments (Survey of Organizational Research Climate and 60 major and minor misbehaviours), but those analyses will be part of another report see and. The interrelations between these concepts will be reported in a separate future paper. ## Statistical analyses We preregistered our analyses on the Open Science Framework, see [osf.io/w4t7u](https://osf.io/x6t2q/register/565fb3678c5e4a66b5582f67). To summarise: First we calculated overall mean scores for all three subscales and stratified these for academic ranks and disciplinary fields. Second, we assessed whether there were differences between particular academic ranks or disciplinary fields using Bonferroni corrected *F*-tests and Mean Differences (*MD*) with 95% Confidence Intervals (*CI*). Third, we build multivariable regression models to test whether academic rank and disciplinary field were associated with PPQr subscale mean scores. In these regression models, we also looked for evidence of confounding and interaction. Estimates corrected for confounding are provided and instances of interaction were reported. All analyses were conducted using SPSS Statistics. # Results ## Response rate and inclusion From the 7548 researchers that were invited to participate, 30% (*n* = 2274) followed the link to the online survey. 1073 of invitees filled in the PPQr (response rate = 14%), demographic information is listed in. About 2% of the invitees filled in the non-response questionnaire. See. Overall, we find academic researchers in our sample to score highest on Attitude (*M* = 3.59). This indicates that the negative attitude towards the publication climate is substantial. There is on average a somewhat lesser degree of Publication Stress (*M* = 3.22) and a relatively small lack of Publication Resources (*M* = 2.21). Stratified and total sample mean scores can be found in. ## Publication pressure by academic rank Pairwise *Bonferroni* and confounding-corrected (disciplinary field and gender) mean differences between academic ranks indicate that postdocs and assistant professors perceive significantly more publication stress than both PhD students and associate and full professors. Besides, both PhD students as well as postdocs and assistant professors have a more negative attitude towards the publication culture compared to full professors. Furthermore, PhD students perceive a significantly greater lack of resources than both postdocs and assistant professors as well as associate and full professors. Finally, postdocs and assistant professors perceive less resources than associate and full professors. See. Crude and *Bonferroni* corrected mean differences between pairs of groups can be found in. For crude and corrected association models between academic rank and the PPQr subscales, see. ## Publication pressure by disciplinary field Pairwise *Bonferroni* and confounding-corrected (academic rank and gender) mean differences indicate that researchers in the humanities perceive more publication stress than both biomedicine and the natural sciences. Yet the researchers from the social sciences perceive more publication stress than their biomedical colleagues. There were no statistically significant differences between disciplinary fields on attitude scores. Finally, researchers in biomedicine as well as social sciences perceive a significantly greater lack of publication resources than researchers in the natural sciences. See. Crude and Bonferroni corrected mean differences between pairs of groups can be found in. For crude and corrected disciplinary field association models, see. ## Effect modification We only found effect modification by disciplinary field of the differences between academic ranks’ Publication Resources scores. Differences between PhD students and senior academic researchers in perceived Publication Resourced are greater in natural sciences compared to other disciplinary fields. Stratified results are displayed in. ## Effect sizes We found 12 significant differences between pairs of groups and since we performed many statistical tests, it is likely that some of the significant differences are in fact due to chance. To provide the reader with some guidance on which effects are relevant, we calculated effect sizes of each difference. This analysis was not preregistered and thus should be considered exploratory. The effect sizes range from small to very large using Cohen’s effect size criteria, see. To prevent overinterpreting small differences, we will focus further discussion on differences with an effect size of medium or above. # Discussion We assessed the level of perceived publication pressure in the four academic institutions in Amsterdam and whether the pressure to publish differed between academic ranks and disciplinary fields. Overall, there is a negative attitude towards the publication climate. Hence the ‘publish or perish’ mantra from the late 20^th century may turn into ‘publish *and* perish’, since even when a researcher publishes reasonably, chances for tenure in academia may still be low. Below we elaborate on the differences of effect sizes that were medium or above or on those where we found interaction effects. ## Academic rank differences Postdocs and assistant professors perceive most publication stress and have the most negative attitude towards the current publication climate, which is in line with previous studies assessing perceived publication pressure in biomedicine and organisation science. This finding seems intuitive as this particular group aims for a (tenured) position and promotion criteria are to a large extent based on quantitative publication indicators. Associate and full professors have already an established position, and consequently may perceive less publication pressure. PhD candidates’ likelihood of successfully defending their thesis is usually not dependent on the number of publications. This may explain why their publication pressure level is somewhat lower. Besides, some PhD students may not aspire an academic career and will therefore presumably perceive less publication pressure then their colleagues who wish to pursue an academic career. However, PhD candidates perceive the greatest lack of resources. This is both alarming and understandable. Arguably, PhD students are inexperienced in handling difficulties that may arise when working on a publication. The same holds for starting postdocs. Consequently, junior researchers could benefit most from supportive colleagues and supervisors. Unfortunately mentoring may be suboptimal. ## Disciplinary field differences Differences between disciplinary fields were significant but small. Hence, we focus here on the interaction between disciplinary field and academic rank in perceived Resources. Researchers from the natural sciences perceived most publication resources, which may be due to their typical organisation (large) research teams where collaboration is vital for discovery. However, PhD students in the natural sciences perceive a lack of resources that is similar to PhD students from the other disciplinary fields. It may be that insufficient mentoring in the publication process makes them feel incompetent and insecure. ## Strengths This is the first study that comprehensively measured publication pressure with a validated measurement instrument. The three dimensions, stress, attitude and resources, respectively, are meaningful components when conceptualising publication pressure. Also, these three dimensions are sufficiently distinctive in the data reported here. Second, this is the first study to investigate publication pressure across academic ranks and disciplinary fields. It can serve as a benchmark for future studies. We managed to include a substantial number of participants in our study that increases the reliability of the differences found. ## Limitations Our study also has some limitations we would like to address. First, we have a relatively low completion rate (14%) which may be an indication of response bias, although our completion rate is similar to other web-based surveys. Only 2% of our invitees filled in the non-response questionnaire, which we consider to be too little to assess whether non-responders differed from responders. Perhaps invitees chose not to respond because they were to focused on their publications, leading to an underestimation. Related, simply mentioning that our study investigated the publication culture could have prompted negative connotations with the publication culture, as it has not gone unnoticed in the public debate in The Netherlands. To assess the representativeness of our sample, we first looked into the population characteristics. In our sample, 56% of completers indicated working in the biomedical field, whereas 53% of our invitees was employed at one of the Amsterdam University Medical Centers, indicating a small overrepresentation from biomedicine. Statistics on PhD students employed at both universities in Amsterdam indicated that PhD students make up 30% of the academic workforce, whereas PhD students formed 41% of our sample. Likewise, 44% of academic researchers in Amsterdam is female, yet women made up 57% of our sample, indicating overrepresentation of both PhD students and women. However, we corrected for the potential gender bias by adjusting our estimates for confounding variables. Besides, we found no effect modification from gender. To conclude, it is unlikely that the selectivity of our sample biased our results. To assess possible response bias, we conducted a wave analysis. We used late responders–those who responded after the last reminder–as a proxy for nonresponders and compared these to early responders–those who responded after the initial invitation–as described by Phillips’ and colleagues. Differences were.13.07 and.02 for Stress, Attitude and Resources, respectively. These differences were then multiplied by the proportion of non-responders, in our case 86%. Consequently, the non-response bias was.11.07 and.02 for Stress, Attitude and Resources, respectively. This was found to be small compared to the difference that we observed between the subgroups which ranged from.18 to.65. It is therefore unlikely that non-response affected our conclusions. Besides, the PPQr focuses exclusively on publication pressure. However, research is not conducted in a vacuum and if teaching or other professional duties put excessive demands on a researcher, then naturally there is less time left for publishing, which could lead to elevated levels of publication pressure. This can be labelled as role-conflict: you are expected to meet different obligations, i.e. teaching, research, and professional duties, in a naturally limited amount of time. How much stress is due to *just* publication pressure is unclear (see also). Relatedly, universities have been subject to neoliberal and Taylorist reforms that were—in a nutshell—intended to make universities more competitive and were accompanied with an excessive focus on researchers’ performance management, perhaps at the expense of traditional hallmarks of the academia such as teaching and collegiality. A full review of Neoliberal and Taylorist reforms in academia is beyond the scope of this paper (the reader is referred to Lorenz’ excellent paper that includes specific examples of reforms in Dutch academia) but it seems feasible to reason that publication pressure is one of its consequences, although the exact relation has, to our knowledge, not been studied systematically. Finally, since this is the first study conducted with the PPQr, it’s rather difficult to interpret the absolute levels and differences in publication pressure we found. ## Future research Future work should aim to explore if the differences we found generalise internationally. Publication climates in the USA and Asian countries may be different as their funding systems greatly differ. However, the same could apply to closer examples such as Germany and Belgium, as those funding systems are also somewhat different from those in the Netherlands. Interestingly, a study with an previous version of the PPQ found Flemish biomedical researchers to experience more publication pressure than their Dutch colleagues. Besides, it will be informative to study publication pressure longitudinally to see if it is associated with burn-out and research misbehaviour. Finally, it would be intriguing to investigate qualitatively what it means for researchers to experience high publication pressure and how it impacts their academic work. ## Conclusions Taken together, publication pressure concerns researchers from all disciplinary fields and seems to be a particularly detrimental stressor for postdocs and assistant professors. In addition, PhD students perceive a significant lack of resources that may hamper their development into responsible researchers. The amount of resources is perceived to be better among researchers from the natural sciences, but PhD students in this disciplinary field nevertheless would also benefit from more support from their senior colleagues. Our findings emphasize the need to move the debate forward towards a healthy publication climate, where researchers are incentivised to focus on the quality and the integrity of their publications and feel supported to conduct responsible research. # Supporting information We would like to acknowledge all members of the steering committee of the Academic Research Climate Amsterdam project (Frans Oort, Gerben ter Riet, Hanneke de Haes, Guy Widdershoven and René van Woudenberg) for their continuous critical input. Frans Oort has been crucial in the creation of the PPQr. Peter van de Ven has been indispensable in discussing statistical tests and models. We also wish to acknowledge Gowri Gopalakrishna and Venetia Qendri for the independent re-analysis of the survey data. [^1]: The authors have declared that no competing interests exist.

# Introduction Schizophrenia is a severe and complicated mental disorder that seriously impairs human independence and imposes a significant burden on society. Both hereditary and environmental factors contribute to this disease. Much has been done in the past decades to unravel the pathogenesis of schizophrenia, leading to various hypotheses. The glutamate hypothesis focusing on N-methyl-D-aspartate (NMDA) glutamate receptor hypofunction has shown a number of promising leads. NMDA receptor mediates glutamate-related cell signaling among neural cells. When the receptor is activated, transcription factors such as CREB (cAMP response element-binding) is mobilized to modulate long-term potentiation, long-term memory, synaptic plasticity and cell survival status. Based on this hypothesis, animal models treated by noncompetitive NMDA receptor antagonists, such as dizocilpine (MK-801) and phencyclidine, are widely used in schizophrenia research. It has been shown that MK-801 treated rats demonstrate both positive and negative symptoms of schizophrenia. Our previous proteomic study scrutinized cortical synaptosome proteins in subchronic MK-801 treated rats and revealed dysfunctions in energy metabolism in these rats. Although alterations in brain energy metabolism have been found in human proteomic studies for schizophrenia , the exact metabolism pathways involved in the dysfunction have not been identified yet. This prompted us to further investigate metabolite levels in the same rat model to delineate the involved pathways which would provide insights to the pathology of schizophrenia. In the past, several studies concerning with certain metabolites have been conducted with the brain tissue extract of the MK-801 treated rats, finding that neurotransmitter metabolism in glial–neuronal interactions was impaired –. Metabolomics, as a modern systems biology approach, is different from the studies focusing on individual metabolites. It monitors entire pattern of low molecular weight compounds and models the global metabolic status of the samples. In the present study, we used proton magic angle spinning nuclear magnetic resonance (1H MAS NMR) spectroscopy to scan the overall metabolite signals in cortex and hippocampus of MK-801 treated rats. 1H MAS NMR spectroscopy has the advantage of a nondestructive procedure that can detect metabolites directly in the intact tissues. Cortex and hippocampus are two brain tissues that are rich of NMDA receptors and thus are responsive to MK-801, which helps us to identify the typical metabolism dysfunctions induced by MK-801. Multivariate statistics and ingenuity pathways analyses (IPA) were employed in data processing. The result was further combined with our previous proteomic data in IPA for a more systematic view on metabolomic observations. # Materials and Methods ## Animal Model and Ethics Statement All animal handling and procedures were performed in accordance with the Guide for the Care and Use of Laboratory Animals once the study received approval by the Institutional Animal Care and Use Committee at Shanghai Jiao Tong University Bio-X institutes, Shanghai, China. All surgery was performed aseptically and every attempt was made to minimize pain and discomfort. 23 male Sprague-Dawley rats (220–250 g) were randomly divided in two groups. Rats in the control group (n = 11) were injected subcutaneously with physiological saline 3.5 ml/kg (0.9% wt/vol NaCl \[aqueous\]) and those in the treatment group (n = 12) with 0.7 mg/kg MK-801 (Research Biochemicals, Natick, Massachusetts) (saline as vehicle) for 10 days. We chose the dose of 0.7 mg/kg as it produced the proper animal model and it was the same dose as our previous work. The volumes of MK-801 or saline were adjusted according to the body weight of each individual animal. The rats were kept in a 12∶12-hour light/dark cycle with food and water available ad libitum. On day 11, approximately 24 hours after the final injection, the rats were killed by cervical dislocation (<http://www.ccac.ca/en_/standards/guidelines>). The brains were quickly removed and the frontal and parietal lobe of cortex and hippocampus were excised from the brain and immediately snap-frozen in liquid nitrogen and stored at −80°C pending analysis. These operations were typically processed within 5–10 min to limit post-mortem changes in the metabolite content of the samples. ## 1H MAS NMR Spectroscopic Analysis Each frozen 15–20 mg intact sample was rinsed with D₂O solution (1 mg/ml) and then rapidly inserted into a zirconia 4 mm outer diameter rotor (Bruker Analytische GmbH, Rheinstetten, Germany). D₂O provided a field-frequency lock. 1H MAS NMR data were recorded on a Bruker AVANCE spectrometer with a field strength of 500.13 MHz. Samples were spun at 5 KHz and maintained at 298 K throughout the experiment to minimize temperature-dependent metabolic changes. In order to suppress broad signals from macromolecules, such as proteins, and hence to focus the subsequent analysis on the relatively small molecules, Carr–Purcell–Meiboom–Gill (CPMG) spin-echo pulse sequence \[D–90°–(τ–180°–τ)n–FID, where FID is free induction decay\] with a fixed spin–spin relaxation delay, 2 nτ of 64 ms (n = 128, τ = 400 µs), was applied to acquire 1H MAS NMR spectra of all samples. Typically, 256 transients were collected into 64 K data points with a spectral width of 30 ppm and an acquisition time of 2.18 s per scan. Prior to Fourier transformation, the FIDs were multiplied by an exponential weighting function corresponding to a line broadening factor of 1 Hz. Manual phase and baseline correction was performed using TOPSPIN software (Bruker Biospin GmbH, version 2.1). The spectra were referenced to lactate (CH3 δ = 1.325). Metabolites were identified with reference to the literatures – and the standard spectra database in HMDB (<http://www.hmdb.ca>). ## Data Reduction and Statistical Analyses A bucket size of 0.01 ppm was chosen to reduce the spectra data using AMIX software (Bruker Biospin GmbH, version 3.8.6). The regions 0–0.6 ppm (no signal peaks), 1.1–1.23 ppm (ethanol), 3.62–3.7 ppm (ethanol), 4.54–5.0 ppm (water) and 8.3–20 ppm (no signal peaks), were excluded. The reduced spectral data were then normalized to a constant sum for each spectrum. The univariate Student’s *t*-test was applied to each bin to evaluate its variation between groups. To account for multiple comparisons, the p-value from each *t*-test was mapped to a Storey-Tibshirani’s q-value using the “qvalue” package in R platform (<http://www.r-project.org>) to estimate the false discovery rate of the test when it’s called significant. The bins with q-values lower than 0.2 were regarded as significantly changed bins. For multivariate statistical analysis, bucket tables were imported to SIMCA-P software (version 11.5; Umetrics, Umea, Sweden). In the software, the univariance scaling method was employed to avoid over-weighting of peaks from metabolites of high concentrations. Orthogonal Partial Least Squares-Discriminant Analysis (OPLS-DA) was conducted to separate MK-801 group from the control group, optimizing the discovery of treatment-related metabolites. For each PLS model, the explained variation (R²) and goodness of prediction (Q²) were given by the software for model evaluation. A cross validation was additionally conducted to test the predictability of the models. Firstly, a test set was constructed using three observations from each class. The left observations constituted the training set. A model was built with the training set and was used to predict the test set’s class membership with a cutoff of 1.5 (1 for the treated & 2 for the control). This was repeated for four times and the average percentage of correct classification was calculated. To better interpret the results from OPLS-DA, back-scaled coefficient plots were drawn using R software (<http://www.r-project.org/>): firstly, the coefficients of the first OPLS component were back-transformed by multiplying all values by the respective variable standard deviation; secondly, the back-transformed coefficients were plotted and colored according to respective VIP (Variable Importance for the Projection) values generated by SIMCA-P software for the model. The scale range of the colors was set using the maximum and the minimum of the VIP values. Variables (bins) that with a Student’s *t*-test q-value lower than 0.2 or an OPLS-DA VIP value higher than 1.5 were selected as the MK-801 treatment related bins. Those bins were assigned with corresponding metabolites. An average fold change of the bins from the same metabolite was calculated as the ratio of the average bin value in treated group to that in control group, indicating an overall change direction of the treatment related metabolite. This result was then imported in Ingenuity Pathways Analysis (IPA) software for molecular pathway and network analysis. ## Molecular Pathway and Network Analysis in IPA Ingenuity Pathways Analysis (IPA; <http://www.ingenuity.com>) is a web-based software application that identifies biological pathways and functions relevant to bio-molecules of interest. To scrutinize the systematic influence of the treatment related metabolites, we uploaded the metabolite lists (with KEGG IDs) and the change directions of these metabolites onto an IPA server. Canonical pathways and molecular interaction networks were generated based on the knowledge sorted in the Ingenuity Pathway Knowledge Base. A ratio of the number of metabolites that map to the canonical pathway divided by the total number of molecules that map to the pathway was displayed. Fisher’s exact test was used to calculate a p-value determining the probability that the association between the metabolites and the canonical pathway was explained by chance alone. The network score was based on the hypergeiometric distribution and was calculated with the right-tailed Fisher’s Exact Test. The higher a score was, the more relevant the eligible submitted molecules were to the network. Integrated analysis of results from the present study and our previous proteomic study was also conducted in IPA by uploading a combined list of the treatment related metabolites and proteins onto the IPA server. # Results The 1H MAS NMR spectra from cortex were similar to that from hippocampus. After data reduction, 325 bins (variables) were obtained from the spectra. Bins from the MK-801 treated group and the control group were compared by Student’s *t*-test. 48 bins had p-values lower than 0.05 in the cortex, of which 44 had q-values lower than 0.2. 34 bins had p-values lower than 0.05 in the hippocampus, of which 11 had q-values lower than 0.2 ( &). Compared with bins in the hippocampus, more bins in the cortex showed statistically significant changes. Multivariate OPLS-DA analysis was implemented to directly search for treatment related metabolites and the results were displayed in the forms of score plots and back-scaled loadings plots. The score plots showed a clear separation between the MK-801 treated group and the control group in both cortex and hippocampus (with R²X = 0.441, Q²Y = 0.413 and R²X = 0.698, Q²Y = 0.677, respectively). Further validation showed that cortex models could predict class membership well with an accuracy of 83.3% and hippocampus models with an accuracy of 82.6%. In the cortex of the MK-801 treated rats, the back-scaled loading plot shows increased levels of lactate, acetate, L-alanine, L-aspartate, GABA, NAA, scyllitol, L-serine and succinate, and decreased levels of citrate, glutamine, glutamate, myoinositol, choline, phosphorylcholine, creatine and taurine. Similar result was seen in the hippocampus but with some differences, such as acetate and L-aspartate levels which were elevated in the cortex but decreased in the hippocampus. The VIP values of the bins can be roughly judged from the colors indicated in the back-scaled loading plots. Warm colored bins (e.g. bins of GABA in red) with high VIP value contributed more than the cold colored ones (e.g. bins of myoinositol in blue) in the inter-group discrimination. We listed all the treatment related variables (bins) with either VIP value \>1.5 or q-value \<0.2 ( &). Those variables (bins) with higher VIP values in OPLS analysis tended to have lower p-values and q-values in the Student’s *t*-tests. Most of the bins could be assigned to corresponding metabolites. is an extract of & that lists all these treatment related metabolites found in cortex and hippocampus of MK-801 treated rats. The change direction of a metabolite was indicated along with an average fold change value of the bins of the same metabolite. Among these metabolites, GABA, succinate and NAA were up-regulated in both the cortex and hippocampus; levels of myoinositol and glutamine were consistently decreased in the two brain tissues; concentrations of L-aspartate, phosphocholine (PC) and L-serine changed differently in the cortex and hippocampus, suggesting a variation in response to MK-801 in different brain areas. IPA analysis was applied with treatment related metabolites to explore systematic influences of subchronic MK-801 treatment. The top ten altered pathways were generated and are listed in. The common pathways shared by the two brain regions were alanine and aspartate metabolism, glutamate metabolism, GABA receptor signaling, nitrogen metabolism and glycine, serine and threonine metabolism. In the network function analysis, treatment related metabolites in cortex and hippocampus tended to gather into one single network, respectively. The two networks were similar and shared the same functions, i.e., amino acid metabolism, molecular transport and small molecule biochemistry. Our previous proteome study revealed 49 proteins altered in the cortical synaptosomes of subchronic MK-801 treated rats. We combined those differentially expressed proteins with treatment related metabolites in the cortex of this study and carried out an additional IPA analysis. The top network function was still the amino acid metabolism, molecular transport and small molecule biochemistry, while the top canonical pathway switched to the Krebs cycle. # Discussion In this study, we employed modern metabolomic method on the platform of 1H MAS NMR to scrutinize metabolite traits in cortex and hippocampus of subchronic MK-801 treated rats, a NMDA receptor hypofunction animal model for schizophrenia. We found that metabolites, not only neurotransmitters but also those involved in energy metabolism, were altered in this schizophrenia animal model. ## NMDA Receptor Hypofunction Causes Disturbance to Glutamate Homeostasis Glutamate was reduced in the hippocampus and had a trend of decrease in the cortex in our study ( &). The change of glutamate in the brain of schizophrenia patients has been the subject of discussion since 1980 but no consensus has so far been achieved. Animal models offer valuable evidence in this field. Acute injection of MK-801 in rats has been shown to cause an elevation of glutamate in certain brain regions. However, in line with our result, a mouse model subject to 7-day subchronic MK-801 injection showed decreased extracellular glutamate level in the prefrontal cortex. Similar results were obtained in most of brain subareas of MK-801 treated Sprague-Dawley rat model. This implies a potential dynamic regulation of glutamate level in response to the length of MK-801 treatment, which is suggestive for human studies since a comprehensive down- regulation of glutamate synthesis was found in chronic schizophrenia patients. ## Glutamate Related Metabolic Pathway Involving Energy Metabolism was the Top Altered Pathways IPA analysis revealed 5 commonly altered pathways in the rat cortex and hippocampus following MK-801 treatment. Based on their biochemical relationships, we integrated the pathways into a brief plot that included most of the altered metabolites. This plot shows that glutamate and glutamine were concurrently down-regulated, but GABA was up-regulated in both brain areas. GABA is a typical inhibitory neurotransmitter and GABAergic neurons are sensitive to glutamate elevation. Thus, an elevation of glutamate can stimulate GABAnergic neurons to release GABA which inversely inhibits glutamate synthesis. This kind of negative feedback might be involved in the dynamic regulation of glutamate in respond to MK-801 treatment as previously mentioned. Moreover, like the glutamate, GABA’s direction of disturbance in schizophrenia or related animal models is still unclear. Discrepancies were found among studies with different samples. For instance, similar to this study, elevated GABA level has been found in chronic schizophrenia patients but no differences in the density of parvalbumin-immunoreactive(PV-ir) GABAergic neurons in cortex was seen in a postmortem study of schizophrenia. Moreover, decreased GABA level was found in rat’s prefrontal cortex after 5-day repeated treatment of phencyclidine which is another NMDA receptor antagonist. Repeated phencyclidine treatment also reduced density of PV-ir GABAergic neurons in rat’s hippocampus 6 weeks after the dosing. Another postmortem study of schizophrenia has identified deficit of GABAergic neurons in frontal cortex. To ravel these conflicts, more systematically designed studies are required to delineate the dysfunction of GABAergic neurons in schizophrenia. Besides neurotransmitter dysregulation, we also observed alterations in the Krebs cycle: the succinate was elevated, while citrate was declined. Succinate is an important intermediate in the Krebs cycle, and can be formed from GABA through GABA shunt. GABA shunt is a characteristic pathway in GABAergic neurons. It allows GABA carbon skeleton to enter the Krebs cycle via succinate. The elevation of succinate is possibly associated with the excess of GABA. A number of enzymes in Krebs cycle have been found abnormal due to MK-801 treatment, such as citrate synthase, malate dehydrogenase and aconitase. Our previous proteome study also found alterations in the Krebs cycle in the cortical synaptosome of subchronic MK-801 treated rats, pointing to a dysfunction of brain mitochondrial energy metabolism caused by MK-801. Lactate was increased in our study, which may result from a deficiency of energy supply. Aspartate and alanine metabolism was also altered in the treated rats. L-aspartate was increased in the rat cortex but decreased in the hippocampus, while L-alanine was enriched in both brain areas. It has been reported that aspartate and alanine metabolism was disturbed in the dorsal prefrontal cortex of schizophrenia patients. The elevation of L-alanine may play different roles in different types of neuron. Through transamination reactions, L-alanine can be converted into pyruvate and utilized as a metabolic fuel in GABAergic neurons. Considering the enhanced GABA shunt and impaired Krebs cycle identified in our study, the elevation of alanine level might be required to meet the energy demand in activated GABAergic neurons. L-alanine is also regarded as a carrier of ammonia nitrogen from glutamatergic neurons to astrocytes using a flux of lactate in the opposite direction to account for the balance of C3 carbon skeleton. From this aspect, the increase of L-alanine along with lactate indicates a stirring transamination activity in glutamatergic neurons and glials, which has already been suggested above as a glutamate centered amino acid metabolism disturbance. We found that levels of acetate and L-serine increased in cortex but decreased in hippocampus, suggesting that the response to MK-801 varied in different brain tissues. L-Serine is required for the synthesis of glycine and D-serine, both of which are NMDA receptor co-agonists. Potential roles of L-serine have been suggested in schizophrenia. Acetate is a common anion in biology and is mainly utilized in the form of acetyl coenzyme A for energy metabolism or acetylizations. Acetylization of L-aspartate turns out N-Acetylaspartate (NAA), an abundant metabolite in brain neurons. NAA was up-regulated in our study both in the cortex and hippocampus. Increased NAA was also found in the hippocampus of schizophrenia patients in a MRS study. NAA was recently reported to be a major storage and transport form of acetyl coenzyme A in the nervous system, linking the accumulation of NAA to insufficient downstream utilization, i.e., the impaired Krebs cycle indentified in this study. ## Result in the Network Function Analysis of IPA In the literature-based network analysis, i.e. IPA analysis, the main functions of our two networks constructed from the cortex and hippocampus respectively are of the same function, i.e., amino acid metabolism, molecular transport and small molecule biochemistry. The networks encompass many biomolecules important in schizophrenia research, such as kynurenic acid which is an endogenous glutamate antagonist. HTT (huntingtin) is one of the core molecules appearing in both networks. It has also been featured in our previous proteomic study. HTT is a primary protein involved in Huntington’s disease. Wild-type HTT protects neurons from NMDA receptor-mediated excitotoxicity, while polyglutamine-expanded HTT sensitizes NMDA receptors toward excitotoxicity and hampers glial glutamate transport capacity. Given the close interaction of HTT with the NMDA receptor mediated glutamate signaling system, its relationship to schizophrenia, though currently unclear, deserves future investigation. ## Metabolite-protein Integrated IPA Analysis We did metabolite-protein integrated IPA analysis in order to expand our vision from the metabolic “dimension” to the protein-metabolic “space” and to reveal otherwise hidden implications using solely metabolomic data. The final top network function was the same as the metabolomic result while the top canonical pathway turned out to be the Krebs cycle. This result not only agrees well with our metabolomic conclusion that glutamate related and energy metabolism were the top altered pathways responding to subchronic MK-801 injection but also stresses the involvement of energy metabolism such as the Krebs cycle in the MK-801 induced dysfunctions. Xiao et.al came up with a similar result through a LC-MS based metabolomic study of the prefrontal cortex of subchronic phencyclidine treated rats. Phencyclidine is another widely used NMDA receptor antagonist. Compared with our study, though different technical platforms for metabolite detecting and different algorithms for data mining were employed in Xiao’s study, they also found disturbances of metabolites such as GABA, glutamate and glutamine in the treated rats and concluded that subchronic phencyclidine treatment would induce compromised glutamatergic neurotransmission as well as disruption of metabolic pathways linked to glutamate in the rats model. ## Conclusion Our study revealed a series of treatment related metabolites in the cortex and hippocampus of subchronic MK-801treated schizophrenia rat model. The disturbed pathway was a highly interconnected glutamate and energy metabolic pathway characterized by down-regulated glutamate synthesis and disturbed Krebs cycle. The disturbances on glutamate neurotransmitter system and energy metabolism are both well-recognized hypotheses of schizophrenia. This study reveals innate biochemical connections between those two theories. Compared to the glutamate hypothesis, energy metabolism dysfunction hypothesis is less discussed and deserves more attention in schizophrenia researches. It also suggested that future studies focusing on either hypothesis take the other one into consideration for a better understanding of the biochemical basis of schizophrenia etiology. In addition, this study proved that metabolomics, as a modern systems biology method, is an efficient and robust knowledge discovery approach for disease studies. Seeing that only cortex and hippocampus which are MK-801 high-binding regions were used here, studies concerning with other MK-801 low-binding brain areas, such as cerebellum and brainstem, are recommended for further validations and explorations. # Supporting Information The authors are grateful to Dr. Huiru Tang and Dr. Genjin Yang for instructions and assistance on data analysis. Thanks to Ran Huo for comments on manuscript writing. [^1]: The authors have declared that no competing interests exist. [^2]: Revised the manuscript: CW ZW KW. Conceived and designed the experiments: CW LH. Performed the experiments: JL HZ LS BJ. Analyzed the data: LS ZZ KZ JY YL. Contributed reagents/materials/analysis tools: MZ LY GH LG XH WL LT YY. Wrote the paper: LS CW.

# Introduction Occupational happiness reflects the subjective well-being of individuals at the workplace, and refers to the positive and negative emotional feelings of employees towards their jobs as well as their cognitive evaluations of their jobs. On the one hand, researchers use indicators, such as job satisfaction, work engagement, positive emotional experience at the workplace, as indirect measures of employee well-being. On the other hand, researchers are paying increasing attention to other indicators, such as turnover intentions, job strain, job burnout, as negative predictors of employee well- being at the workplace. Job burnout is a state of mental and physical exhaustion that results from prolonged heavy workload and stress. In 1974, American psychiatrist Freudenberger first proposed the concept of “job burnout” to describe unhealthy physical, psychological, and emotional states, such as fatigue, decreased work engagement, reduced sense of accomplishment, which were experienced by people working in human service professions due to long working hours, heavy workload, and excessive work intensity. Maslach and Jackson further defined “job burnout” as long-term stress response of an individual to prolonged exposure to emotional and interpersonal stressors at work, which encompass emotional exhaustion, depersonalization, and reduced personal accomplishment. Emotional exhaustion refers to feelings of extreme emotional fatigue and lack of enthusiasm and vigor towards work. Depersonalization refers to the deliberate attempt to maintain distance between the self and work as well as the exhibition of passive, indifferent, and cynical attitudes and emotions towards others at work. Reduced personal accomplishment is manifested in low sense of self-respect and in even more negative evaluation of work, inability to experience pleasure, satisfaction, and a sense of accomplishment associated with performing the job. From a compositional perspective, job burnout comprises emotional experiences and cognitive components. If an individual is not passionate for his or her job, has poor interpersonal relationships at work, and cannot find a sense of self- worth from performing his or her job, then the person will not experience happiness in his or her work. Therefore, we think that job burnout can serve as a strong negative measure of employee occupational happiness, and enhancing employee occupational happiness requires the reduction of job burnout. How can we reduce job burnout? This study focuses on both individual and organizational dimensions, namely, psychological capital and organizational commitment, in reducing job burnout. In 2004, Luthans et al. proposed the concept of psychological capital within the framework of positive psychology and positive organizational behavior. They defined psychological capital as “a positive state of mind exhibited during the growth and development of an individual” that includes four core components of self-efficacy, optimism, resiliency, and hope. Self-efficacy refers to the confidence in being able to execute a task, ability to face challenges, and the will to succeed. An optimistic individual makes attributions for positive events and maintains a positive attitude towards the present and future. Resiliency refers to the ability to recover quickly, or even change and grow, from adversities, setbacks, and failures. Hope is the positive motivational state that helps in achieving the intended goals through various means. Several studies in management have shown that psychological capital and its various dimensions can have positive effects on work performance and attitudes of both leaders and employees. For instance, Avey, Patera, and West showed that each of the four dimensions of psychological capital, i.e., self-efficacy, optimism, resiliency, and hope, is negatively correlated with employee absenteeism and turnover intentions. Luthans et al. conducted an empirical study of the relationship between psychological capital and job performance in 422 Chinese employees and found a positive correlation between the three positive mental states of hope, optimism, and resiliency, and the evaluation of work performance of these Chinese employees by their managers. Moreover, the positive correlation between psychological capital and work performance was even stronger, and psychological capital and employee performance-based pay were also positively correlated. Currently, few studies have explored the relationship between psychological capital and job burnout. Luthans et al. believed that psychological capital plays a pivotal role in the development of job burnout and that it may effectively reduce the extent of burnout. Using Chinese nurses as participants, Luo and Hao found preliminary evidence for the preventative effect of psychological capital on job burnout. In general, however, the body of research concerning the relationship between psychological capital and job burnout remains relatively small. Organizational commitment refers to the attitude towards the organization. This attitude is a psychological bond in the relationship between an employee and the organization that affects the degree to which the individual identifies with the goals and values of the organization, exerts effort to achieve organizational goals, and desires to remain in the organization. The three-factor model of organizational commitment proposed by Meyer and Allen, which includes affective, continuance, and normative, has been well-received and supported by international studies. Affective commitment refers to the emotional attachment that employees feel towards their organization. Continuance commitment refers to the awareness of employees of the cost of losing their membership in the organization. Normative commitment refers to the level of consistency between the values of the individual and the organization or the sense of responsibility for the organization, which is shaped by the long-term influence of society on the sense of social responsibility of an individual. A large body of research supports the relationship between organizational commitment and job burnout. King and Sethi found that organizational commitment has a moderating effect on the relationship between stress and job burnout. Tan and Akhtar reported that when age, tenure, organizational level, and work perceptions of Chinese employees were controlled, organizational commitment had a significant effect on experienced burnout. Wright and Hobfoll further showed that organizational commitment has an effect on every dimension of job burnout. A significant correlation also exists between psychological capital and organizational commitment. Luthans and Jensen demonstrated a very strong positive correlation between the psychological capital and the assessment of the commitment of nurses to the mission, values, and goals of the hospital. In a study of 74 employees, Larson and Luthans reported a significant positive correlation between psychological capital and job satisfaction and organizational commitment. Zhong studied the influence of psychological capital on work performance and organizational commitment in the Chinese cultural context and found, after controlling for gender and age, all three positive mental states of hope, optimism, and resiliency, had a positive influence on employee work performance, organizational commitment, and organizational citizenship; hence, psychological capital, which is composed of hope, optimism, and resiliency of an employee, had a positive impact on work performance, organizational commitment, and organizational citizenship. Psychological capital can be seen as an important human resource that has significant effects on organizational commitment and job burnout. Organizational commitment can affect job commitment because an employee who has a sense of belonging and commitment to the organization is unlikely to tire from the job, and this effect can be even more significant in collectivist cultures. While the relationship between any two of the three variables is relatively clear, the relationship of these three variables has not been explored. Nursing is a high risk, high pressure, and labor-intensive profession, and thus, a high incidence of job burnout exists among nurses. A survey conducted in five countries, including the United States, revealed that job burnout is a very serious phenomenon within the nursing profession. Among the five countries surveyed, 40% of the nurses in four countries experienced job burnout. Nursing was also first in the job burnout profession ranking released in China in 2006. In the present study, we studied the relationship among psychological capital, organizational commitment, and job burnout in Chinese nurses. # Methods ## 2.1: Participants and Procedure The participants comprised 473 female nurses from four large general hospitals in Xi’an, China. Their ages ranged from 20 years to 39 years, with a mean of 26.23 years (SD = 3.60). At the time of the gathering of data, the nurses had worked in hospitals from 6 to 220 months. The participants completed the questionnaire in a classroom environment. All participants signed informed consent forms before completing the measures. The research described in this paper meets the ethical guidelines of the Fourth Military Medical University and was approved by the ethics committee of Xijing Hospital. Participants were told they were engaging in a psychological investigation in which there were no correct or incorrect answers. We distributed 473 questionnaires, which were all collected and validated. Participants received ¥40 in compensation. ## 2.2: Instruments ### 2.2.1: Psychological Capital Questionnaire The Psychological Capital Questionnaire (PCQ), developed by Luthans et al., is a 24-item self-report scale that includes four dimensions, namely, self-efficacy, optimism, resiliency, and hope. The items are rated from 1 (strongly disagree) to 6 (strongly agree). Some of the items are “I usually take stressful things at work in stride” and “I always look on the bright side of things regarding my job.” Scale scores are the sum of items with reverse coding of relevant items. In this study, the Cronbach’s alpha coefficient for the PCQ was 0.846. ### 2.2.2: Maslach Burnout Inventory-General Survey The Maslach Burnout Inventory-General Survey (MBI-GS), developed by Maslach, is a 15-item self-report measure of job burnout that includes three dimensions, namely, emotional exhaustion, depersonalization, and reduced personal accomplishment. The items are rated from 1 (never) to 7 (every day). Some items are “I have become less enthusiastic about my work,” and “I have become more cynical about whether my work contributes anything.” In this study, the Cronbach’s alpha coefficient for the MBI-GS was 0.884. ### 2.2.3: Organizational Commitment Scale The Organizational Commitment Scale (OCS), developed by Allen and Meyer, comprises 18 items and three dimensions, namely, affective, normative, and continuance. The items are rated from 1 (strongly disagree) to 6 (strongly agree). Some items are “I am very happy being a member of this organization,” “I worry about the loss of investments I have made in this organization,” and “I feel that I owe this organization quite a bit because of what it has done for me.” Scale scores are the sum of items with reverse coding of relevant items. In this study, the Cronbach’s alpha coefficient for the OCS was 0.891. ## 2.3: Data Analysis The mediation test is emphasized because we aimed to explore the trilateral relations among psychological capital, organizational commitment, and job burnout. Let X, M, and Y be the independent, mediating, and dependent variables, respectively, and let Y=cX+ e1, M=aX+e2, Y=c’X+bM+e3. The mediation effect is an indirect effect, in which the effect of independent variable on dependent variable goes through a mediator, which can be operationalized as a×b not equal to zero. The commonly employed method for examining the statistical significance of a mediation effect is the Sobel test, which involves computing the ratio of a×b to its estimated standard error (Z= ab/$\sqrt{a^{2}\text{s}_{\text{b}}^{\text{2}}\text{+}b^{2}s_{a}^{2}}$). However, a Sobel test requires that a×b follows a normal distribution; otherwise, statistical efficacy would be reduced. The bootstrap test implemented by Preacher and Hayes tested the null hypothesis a×b=0 in another way. This test takes sample size N and draws substitute N values of (X, M, Y) to create a new sample. If the option is repeated, for example, 1000 times, then 1000 estimations of a×b can be calculated. The bootstrap test relies on 95% confidence intervals from the empirical distribution of a×b estimates. In the current study, psychological capital, organizational commitment, and job burnout were regarded as latent variables, and thus, a two-step procedure introduced by Anderson and Gerbing was adapted to analyze the mediation effect. First, the measurement model was tested to assess the extent to which each of the three latent variables was represented by its indicators. If the confirmatory measurement model were acceptable, then the maximum likelihood estimation would be used to test the structural model in the AMOS 17.0 program. The following four indices were used to evaluate the quality of the fit of the model: (a) Chi square statistic (χ²), (b) the Standardized Root Mean Square Residual (SRMR), (c) the Root Mean Square Error of Approximation (RMSEA), and (d) the Comparative Fit Index (CFI). A model was considered to have a good fit if all the path coefficients were significant at the level of 0.05, the SRMR was below 0.08, the RMSEA was below 0.08, and the CFI was 0.95 or higher. Based on above, the bootstrap test and structural equation modelling were both used to test the mediation effect. # Results ## 3.1: Measurement Model Confirmatory factor analysis was used to examine whether the measurement model adequately fit the sample data. The measurement model included 3 latent constructs and 10 observed variables. The results indicated that the measurement model fit the data very well. For psychological capital: χ² (239, N=473) =592.07, p\<0.001; RMSEA=0.056; SRMR=0.076; and CFI=0.968. For organizational commitment: χ² (128, N=473) =324.85, p\<0.001; RMSEA=0.057; SRMR=0.048; and CFI = 0.951. For job burnout: χ² (85, N=473) =258.72, p\<0.001; RMSEA=0.056; SRMR=0.066; and CFI=0.967. All the factor loadings for the indicators on the latent variables were significant (P\<0.001), indicating that the measurement model is acceptable. ## 3. 2: Correlation Analyses shows all the latent variables, namely, psychological capital, organizational commitment, and job burnout, significantly correlated each other. ## 3. 3: Structural Model and Bootstrap Test In the first step, the direct effect of the predictor (psychological capital) on the dependent variable (job burnout) in the absence of mediator (organizational commitment) was tested. The directly standardized path coefficient was significant, β=-0.60, p\<0.001. After that, a partially mediated model (Model 1) that contained mediators (organizational commitment) and a direct path from psychological capital to job burnout was tested. The results showed that the model did not fit the data very well, χ² (23, N=473) =190.02, p\<0.001; RMSEA=0.084; SRMR=0.102; and CFI=0.921. However, an examination of parameter estimates indicated that the standardized path coefficient from psychological capital to job burnout and organizational commitment and from organizational commitment to job burnout were all significant. Thus, according to the modification indices in the Model 1, Model 2 was created by adding the correlations of residual terms between self-efficacy and optimism, between emotional exhaustion and reduced personal accomplishment, and between continuance commitment and normative commitment. After adding the correlations of the residual terms, the final meditational model, shown in, was analyzed. The final meditational model showed a good fit to the data: χ² (26, N=473) =94.68, p\<0.001; RMSEA=0.054; SRMR=0.073; and CFI=0.966. Theoretically, the correlations of self-efficacy-optimism, emotional exhaustion-reduced personal accomplishment, and continuance commitment-normative commitment were easy to explain, and we deemed that the final model was acceptable. Lastly, the mediating effect of organizational commitment between psychological capital and job burnout was tested by adopting the bootstrap estimation procedure with AMOS (a bootstrap sample of 1,500 was specified). shows psychological capital had significant direct effect on job burnout (β=-0.262, p\<0.001). In addition, the indirect effect through organizational commitment was also significant (β=-0.321, p\<0.001). The indirect effect made up 55.1% of the total. # Discussion The results of this study, with Chinese nurses as participants, indicated that through the partial mediation of organizational commitment, psychological capital could affect job burnout. The finding that psychological capital can negatively affect job burnout is consistent with previous studies. According to the Conservation of Resource Theory, job burnout results from either the lack of resources and inability to meet the job requirements or the imbalance between individual effort and payout. Understandably, psychological capital, which is an important human resource, can prevent job burnout. As a composite concept, each dimension of psychological capital also correlates significantly with job burnout. For instance, Luthans et al. investigated Chinese employees and reported a positive correlation between the three positive mental states of hope, optimism, and resiliency, and work performance. Evers, André Brouwers, and Tomic explored the relationship between self-efficacy and job burnout among teachers, and their results indicated that teacher self-efficacy was related to their burnout level, in which teachers with strong self-efficacy seemed to be more prepared to experiment and to implement new educational practices. Optimistic individuals are less likely to experience burnout, as they make more attributions for positive events, are able to posit positive explanations for work events, possess positive attitudes, and are able to cope more easily in the face of different types of occupational stress. Resilience is the ability of an individual to adjust positively to adversity, such that individuals with higher levels of resiliency recover more easily from frustration and failure, and this adaptive capability can significantly help one withstand the fatigue and emotional exhaustion caused by work stress. Hope is the process of thinking about goals, the ways to achieve those goals, and the motivation to accomplish those goals. Research suggests that individuals with more hope typically have clear work objectives, formulate practical action plans to meet their objectives, and work hard to meet these objectives; hence, they are less likely to experience the negative effects of burnout. Since psychological capital is the generic concept of self-efficacy, optimism, resiliency, and hope, nurses with greater psychological capital would naturally experience lower levels of burnout. This finding suggests the importance of nurturing self- confidence and positive coping styles among nurses at work and training them to handle stressful events. When nurses are hopeful about their jobs, have optimistic attitudes, and possess high levels of endurance and adaptability, their physical, emotional, and psychological depletion will be minimal, and they will not easily experience job burnout. The significant correlation between organizational commitment, on the one hand, and psychological capital and job burnout on the other, is consistent with the results of previous studies. In the current study, we focused on the moderating function of organizational commitment between psychological capital and job burnout. As attitude towards the organization, organizational commitment refers to the ground for the willingness of an individual to remain in the organization. Employees with greater levels of psychological capital are more likely to be dedicated to their assignments, to have a strong sense of duty, and to respond resolutely to adversity. They also identify more strongly with the team, enjoy more harmonious interpersonal relationships, and are more willing to contribute to the organization. Thus, such employees possess greater organizational commitment. Wegge et al. found that a strong team identity is correlated with lower levels of emotional exhaustion and cynicism and higher levels of individual accomplishment. We believe that individuals with greater psychological capital are more willing to contribute to the organization and to identify more strongly with the team. Therefore, they are able to receive all kinds of support from the organization and have higher levels of occupational happiness. The present study provides evidence in confirmation of this pathway of influence. Based on the results of the present study, we can conclude that enhancing psychological capital is an effective strategy for increasing organizational identity and reducing job burnout. As the psychological resource of an individual, psychological capital can indeed be developed and enhanced. Luthans et al. proposed a number of possible methods such development. For example, an effective strategy for developing self-efficacy and self-confidence among employees is to allow employees to experience success. For example, management can help employees achieve goals or allow them to observe the successful results of continued efforts of others with similar backgrounds and circumstances. Another strategy is to train employees to use the stepping method to decompose personal goals into more manageable sub-goals, to enjoy the process of realizing the goal rather than to be solely preoccupied with the final result, and to be ready and willing to persevere when faced with obstacles and difficulties. Employees can also be trained to tolerate past events and accept their past mistakes, failures, and setbacks. Employees should be encouraged to appreciate the present and to be grateful for the positive aspects of their present lives. They should seek opportunities for future improvement and development, view future uncertainties as opportunities for development and progress, and face tomorrow with a positive, welcoming, and confident attitude. The present study has some limitations. We employed a cross-sectional rather than longitudinal design, and we do not have evidence for the dynamic influence of psychological capital on organizational commitment and job burnout. As China is a collectivist nation, organizational factors have a strong impact on employee occupational happiness. Further, whether the mediating function of organizational commitment between psychological capital and job burnout holds cross-culturally is a question well worth exploring in future studies. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JP YZ DM. Performed the experiments: JP XJ JZ RX YS XF. Analyzed the data: JP JZ. Contributed reagents/materials/analysis tools: JP XJ DM. Wrote the manuscript: JP YZ DM. Designed the software used in analysis: JZ.

# Introduction Menarcheal age is the most widely used indicator of sexual maturation and can be used as an indicator of female health, growth and development, and the capacity to reproduce. Onset of maturation and age at menarche, are influenced by several factors, e.g. genetics, ethnicity, height, weight, body mass index (BMI), and socioeconomic circumstances. ; ; – Several studies from various developed countries worldwide have shown a systematic decrease in median age at menarche in the past 160 years ; ;. In Europe the median age at menarche decreased with 2 to 3 months per decade from 16.5 years in 1840 to about 13.0 years in the 1960s, with 0.5 years variation between countries. ; ; ; In Scandinavian countries the secular trend in age at menarche and stature has stopped, and possibly even reversed in the past three decades ; ;. Based on three previous Dutch Growth Studies ;, Mul et al. concluded that between 1980 and 1997 the secular change towards earlier puberty and age at menarche had stabilized in the Netherlands as well. Several studies suggest a relationship between increasing obesity and decline of median age at menarche ;. Others suggest that menarche is accompanied or quickly followed by a rapid increase in body weight. In the Fourth Dutch Growth Study (1997) a correlation between BMI and menarche was only present in normal weight or underweight girls. Beyond a BMI of +1 Standard Deviation Score (SDS) there was no decrease in age at menarche with increasing BMI. The aim of this study is: 1) to assess median age at menarche in the Netherlands, in three ethnic groups (i.e. girls of Dutch, Moroccan and Turkish descent); 2) to assess whether there is a secular trend by comparing data on age at menarche from previous Dutch Growth Studies; and 3) to investigate the differences in BMI-SDS in premenarcheal and postmenarcheal Dutch girls in the 1980, 1997 and 2009 studies. # Materials and Methods This study is part of the Fifth Dutch Growth Study, a cross-sectional study performed in 2008/9 in the Netherlands. The study protocol was approved by the Medical Ethical Review Board of Leiden University Medical Centre. Written informed consent was not needed. Measurement of growth in children is part of routine Youth Health Care in the Netherlands. ; Data on growth and development were collected in children 0–21 years of age in the Netherlands by trained youth health care professionals. The study design is identical to the Fourth Dutch Growth Study, which in turn was based on the Dutch Growth Studies in 1955, 1965, and 1980. The design and methods of ‘the Fifth Dutch Growth Study’ have been described elsewhere. Pubertal staging was not registered in the Fifth Dutch Growth Study, as there were no significant differences in onset or pace of secondary sexual characteristics between 1980 and 1997. All Dutch Growth Studies consist of a representative sample of children of Dutch origin. Because of differences in growth between Dutch children and children of Turkish and Moroccan descent, the two largest minority groups in the Netherlands, cohorts were included in the 1997 and 2009 study. ; Turkish and Moroccan children were predominantly from the four largest cities in the Netherlands. If the girls, both (biological) parents, or the four (biological) grandparents were born either in Turkey or Morocco, they were defined as first, second or third generation, respectively. To determine age at menarche, each girl aged 9 years and older was asked if she had had her first menstruation (status quo method). BMI was defined as weight in kilograms divided by the square of height in meters, and classified as normal weight, overweight or obesity according to the criteria of the International Obesity Task Force (IOTF). The educational level of the parents was defined as the educational level of the highest educated parent and categorized into low, middle and high level. ## Statistical Analyses Reference curves for age at menarche were obtained from the model with the best fit based on the analysis of variance for each ethnicity and year of study separately. We tested generalized linear and additive logistic models with a binomial family and probit and logit link functions and a smoothing spline fit s on age with equivalent degrees of freedom ranging between 2 and 4 (default). These models describe the probability of having menarche or not, as a function of age. Differences in these probabilities between year of study and ethnicities were tested with analysis of variance with a model with and a model without an additional independent variable year of study or ethnicity. Previous models were also tested adjusted for BMI-SDS to find out if differences can be explained by differences in BMI-SDS between year of study or ethnic groups. BMI-SDS was calculated using the previously published BMI-for-age references. To test differences in mean BMI-SDS between postmenarcheal and premenarcheal girls within age and year of study, linear regression analyses were performed with BMI-SDS as dependent variable and menarche (two categories), age (linear), year of study (three categories) and the interactions BMI-SDS with age and year of study as independent variables. P-values smaller than 0.05 (two-sided) were considered statistically significant. The statistical analyses were performed in R Version 2.12.0 (The R Foundation for Statistical Computing). # Results 12,005 children of Dutch origin (6,270 female), 2,582 of Turkish origin (1,267 female), 2,616 of Moroccan origin (1,328 female) were included in the umbrella study. Age at menarche in girls 9 years and older was obtained from 2138 girls of Dutch, 282 girls of Turkish and 295 girls of Moroccan origin. depicts the 10^th, 50^th and 90^th centiles and the 95% confidence intervals for age at menarche, specified by ethnicity and data from the previous Dutch Growth Studies (1955, 1965, 1980 and 1997) are added for comparison. The proportion of Dutch girls that reached menarche at different ages is shown in for Dutch girls and in for Turkish and Moroccan girls. In 2009, median age at menarche in Dutch, Turkish and Moroccan girls was 13.05, 12.50 and 12.60 years, respectively. Median age at menarche occurred at a significantly earlier age than in 1997 for all three ethnicities. In Dutch girls menarche occurred 1.2 months (0.1 years) earlier (p = 0.002; ANOVA); this also holds when adjusted for differences in BMI-SDS (p = 0.004; ANOVA) and in both Turkish and Moroccan girls 3.6 months (0.3 years) earlier (p = 0.02 and p = 0.04, respectively; ANOVA). These differences remain significant after adjustment for BMI-SDS (both p = 0.01; ANOVA). There were significant differences in median age at menarche between Dutch girls and girls of Turkish (p\<0.001) and Moroccan (p = 0.007) descent. Similar results were found when adjusted for BMI-SDS (p\<0.001 and p = 0.009 respectively; ANOVA). In 1997, only the difference in median age at menarche between Dutch (13.15 years) and Turkish girls (12.80 years) was significant (p = 0.009; ANOVA). However, after adjustment for BMI-SDS, this difference was not significant any more (p = 0.85; ANOVA). There was no significant difference in median age at menarche between girls with high or low educated parents (mothers: 12.9 vs 13.0 years, and fathers 13.2 vs 13.1 years; p = 0.4 and p = 0.7, respectively). Over 90% of parents of the Turkish and Moroccan girls have a low educational level. shows that the mean BMI-SDS in premenarcheal Dutch girls is lower than in postmenarcheal Dutch girls irrespective of age category (data from the Third, Fourth and Fifth Dutch Growth Study combined). There is a significant difference (Δ) of 0.58 in mean BMI-SDS between postmenarcheal and premenarcheal girls, independent of age and year of study (p\<0.001). In is shown that across different ages and studies the differences are similar (Δ 0.66–1.08). The differences (Δ) in the 1980, 1997, and 2009 studies for the age category 13.0–13.9 years are identical (0.84). The proportion of girls of Dutch, Turkish and Moroccan origin that have reached menarche is given by age in. Eight percent of the girls of Dutch, 11% of the girls of Moroccan, and 33% of the girls of Turkish descent have reached menarche before the end of primary school (12 years of age). # Discussion Between 1955 and 1965 age at menarche in Dutch girls decreased with 2.5 months per decade. Since 1965, median age at menarche still decreased significantly but with a lower pace of approximately about one month per decade to 13.05 years in 2009. Median menarcheal age in girls of Turkish (12.50 years) and Moroccan (12.60 years) descent was 3.6 months earlier in 2009 than in 1997, a decrease of approximately three months in a decade suggesting a secular trend. In the same period the difference in median age at menarche between Dutch girls and girls from Turkish and Moroccan descent increased with 2.4 months to 6.6 months and 5.4 months, respectively. ## Age at Menarche and Secular Trend in Different Countries Around the World Is the secular trend in earlier menarche proceeding in other countries as well? In the United States and Denmark onset of sexual maturation in girls decreased the past decades ; ; ; ;. This trend is seen in girls of all different ethnicities (African-American/Caribbean, Hispanics, Indo-Pakistani and Caucasian) ;. In contrast, the secular trend in age at menarche stabilized in German girls. In Swedish girls it even reversed. However this may be explained by different study designs making comparison of data difficult. Menarcheal age in girls in Turkey varies in different studies and study populations: 12.36 years in 1975, 13.28 years in 1996, 12.2 and 12.41 years in 2008 and 12.74 years in 2011. Bundak et al. as well as Atay et al. found no evidence for a secular trend in age at menarche in girls in Turkey in the last three decades. Age at menarche in Turkish girls living in Germany in 1985 was 12.90 years, this is comparable to median age at menarche (12.80 years) in girls of Turkish descent in our study 12 years later (1997). Interestingly, in girls from Turkish and Moroccan descent living in the Netherlands age at menarche decreased at a higher pace than it did in Dutch girls. A higher BMI and improved socioeconomic circumstances compared to their country of origin may explain this phenomenon ; ; ; ;. To the best of our knowledge, there are no recent data available on menarcheal age in girls in Morocco. How can we explain the ongoing secular decline in age at menarche? ## Ethnicity and Menarche Hughes et al. stated that genes remain important in determining the pace of maturation and age of menarche as they account for about 75% of the variance in identical twins, mother/daughter pairs, and girls of the same ethnicity. In industrialized countries girls of African descent have the lowest median age at menarche, followed by Hispanic-Latino and Hindo-Pakistani girls, while Caucasian girls have the highest median age at menarche. ; ; ; ; ; ; Compared to Dutch girls we found significantly lower median age at menarche even after adjustment for BMI-SDS in girls from Turkish and Moroccan origin. ## Socioeconomic Circumstances Improved health, hygiene, nutrition, housing and employment are assumed to be responsible for most if not all the decline in age at menarche. In classical times until the mid 1950s age at menarche in industrialized countries was lower in girls with high socioeconomic status. Since then the situation reversed and age at menarche in Caucasian girls with low socioeconomic status (SES) compared to girls with high SES became lowest. Studies indicate that girls of African descent in industrialized countries and girls in Turkey with high SES (still) have an earlier age at menarche than girls with low SES ; ;. Deprivation or malnutrition may be (part of) the explanation. In this study we found no significant differences in median age at menarche between girls from parents with low and high education. In contrast to girls in Turkey the mean BMI-SDS of Dutch children whose parents were low educated was higher in 2009 than those of higher educated parents ; ;. ## Body Mass Index, Weight, Height and Menarche Several studies suggest that obese girls mature earlier. However, this discussion is controversial – The decline in age at menarche is suggested to be related to the obesity epidemic, but obesity can not solely held responsible, because the secular change was also found in normal weight girls. The link between obesity and age at menarche may be due to a mechanism ensuring that pregnancy will not occur unless there are adequate fat stores to sustain both mother and foetus. Although weight is involved in age at menarche we note that after adjustment for BMI-SDS the decline in median age at menarche remained significant. Because of the cross sectional study design we can not differentiate between co-occurrence and causality. Differences in BMI-SDS in premenarcheal and postmenarcheal girls are previously observed by O’Dea et al., Kaplowitz and others ;. Our data of mean BMI-SDS in premenarcheal and postmenarcheal Dutch girls in different age categories in the 1980, 1997 and 2009 studies show a positive shift in BMI during the maturation process. The differences in mean BMI-SDS between pre- and postmenarcheal girls in 1980, before the obesity epidemic in the Netherlands, are almost identical in 1997 and 2009. Age at menarche is also thought to be dependent on height. The secular trend in stature stopped decades ago in the United States and Scandinavia; ;, while in the Netherlands it stopped only recently. The assumption was that the secular trend in age at menarche would stop too, but it did not. Thus, increase of height is no longer (part of) the explanation of the decline in age at menarche in 2009. ## Consequences of Early Menarche for Public Health Early menarche is associated with psychosocial and health problems ;. Already one third (33%) of the girls of Turkish descent in primary schools (up to 12 years of age) is menstruating. The continuing younger age at menarche requires changes in sexual education in primary schools, In addition primary schools have to provide facilities for menstruating girls. Furthermore, adult women with early menarche have a higher risk of breast cancer, metabolic syndrome, depression, and cardiovascular disease ; ;. This has profound Public Health implications. ## Strengths and Limitations of the Study A major strength of the present study is the large study population. The Netherlands have an outstanding tradition of nationwide Growth Studies in 0–21 year old children. This series of growth studies is unique in the world because of the identical design and the national representative samples. By repeating the Growth Study every ten to fifteen years secular trends can be made clear. A limitation of our study is not having the actual menarche dates of the girls. However the status quo method in a large study population like we used in our study is considered to be even more reliable than the recall method for obtaining menarche dates. A further limitation of the study is the cross sectional character prohibiting conclusions of a causal relationship between age at menarche and BMI. In the future, monitoring secular trends in children of Dutch, Turkish and Moroccan descent in the Netherlands remains important to assess the differences between ethnicities. Longitudinal studies are necessary to further elucidate the determinants of age at menarche. We thank all participating well-baby clinics and Youth Health Care departments of Municipal Health Services. The help of Joana Kist-van Holthe (EMGO+/VUmc) in editing the manuscript is very much appreciated. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: HT YS PVD BB SVB RAH. Performed the experiments: HT YS PVD SVB. Analyzed the data: PVD SVB YS. Contributed reagents/materials/analysis tools: BB RAH. Wrote the paper: HT YS PVD BB SVB RAH.

# Introduction Several species of mosquitoes are responsible for the transmission of various pathogens to humans and animals. For example, *Aedes aegypti* (the yellow fever mosquito) is a vector of the yellow fever virus, dengue virus, chikungunya virus, and zika virus, whereas *Aedes albopictus* (the Asian tiger mosquito) transmits chikungunya virus, dengue virus, and dirofilaria parasites and is considered a potential vector of the zika virus. *Aedes vexans* is an important vector that transmits West Nile virus (WNV) from birds to mammals. *Culex quinquefasciatus* (the southern house mosquito) is a vector of the filarial nematode, *Wuchereria bancrofti*, Rift Valley fever virus, and WNV. *Culex vishnui* is the most common vector of Japanese encephalitis virus (JEV) in South-East Asia. *Anopheles tessellatus* has been found naturally infected with human malaria parasites in Indonesia, Sri Lanka, along with JEV in Taiwan. Both entomological surveillance and the monitoring of mosquitoes are very important for the prevention and control of mosquito-borne diseases. The timely detection of vector species can save many lives and reduce economic losses. The classification accuracy of traditional models, as represented by the support vector machine (SVM), and those of a few deep learning models, including AlexNet, VGGNet, and ResNet, have been carried out. As for SVM, twelve types of features had to be extracted beforehand. The three deep neural networks, however, did not need feature extraction. The maximum accuracy of SVM was 82.4%, and that of the deep learning models was found to be up to 13.1% higher. The effectiveness of the three convolutional neural networks (CNN): AlexNet, LeNet, and GoogLeNet was evaluated. Results demonstrated that GoogLeNet yielded the best result, having an accuracy of 76.2%. AlexNet and LeNet attained an accuracy of 51.2% and 52.4%, respectively. This work suggests that networks with more complexity are likely to be more accurate. Regarding the mosquito classification, conventional CNN models such as VGG16, ResNet, and SqueezeNet with data augmentation and transfer learning have achieved state-of-the-art results. The oldest VGG16 yielded the best results. Furthermore, this work revealed that morphological keys used by human experts have been found to be similar to the heatmap of the CNN viz. patterns of the body, legs, and proboscis. This work published an image set, including five classes of disease vectors and one class of non-vectors. Thereafter, we would refer to it as *the reference dataset.* Further, two deep convolutional neural networks (DCNN), YOLOv3 and Tiny YOLOv3, were deployed to classify five vector classes and a non-vector class. This work reused images from *the reference dataset* along with additional unpublished non-vector images. It is seen that YOLOv3 and Tiny YOLOv3 achieved 97.7% and 96.9% accuracy, respectively. Subsequently, the authors amplified the performance of the YOLOv3 model using four augmentation conditions. Results showed that when *default* + blur + noise augmentation were used, such conditions increased the accuracy of classification up to 99.1%. Recently, two DCNN models including VGG16 and ResNet50 with initial weights from ImageNet, were investigated in. The models reused the dataset published. Two of the networks i.e. VGG16 and ResNet50 were employed to separate *Aedes albopictus* species from non-*Aedes albopictus*. Thus, it was found that VGG16 achieved more than 94% accuracy on the test set. Furthermore, the authors identified that the white band stripes in the legs, the abdominal patches, the head, and the thorax were features used by *the classifier*. A heatmap visualization of discriminative regions was also provided. This recent work found that VGG16 again yielded the best result. It is evident that the VGG16 model appears to be the most promising, as it provided the best results. Moreover, in situations where the quality of images cannot be guaranteed, model robustness is also crucial. In the method section, *the classifier* is composed of three multi-view models (A), (B), and (C), as well as an ensemble model (D) to predict mosquito species from three images rather than one. Note that photographing a mosquito three times does not require much more work than photographing it once. We assume that by applying three images, *the classifier* can produce more accurate and more robust results. # Materials and methods ## Ethics statements This study was approved by the animal research ethics committee of Chulalongkorn University Animal Care and Use Protocol (CU-ACUP), Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand (COA No. 029/2564). ## Sources of mosquitoes and species identification *Ae. aegypti, Ae. albopictus*, and *Cx. quinquefasciatus* mosquitoes have been maintained continuously for many consecutive generations in an insectary at the Department of Parasitology, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand. Some of them were used in this study. Live engorged female mosquitoes were collected from six provinces of Thailand, including Chiang Mai (18.891369, 99.112922; Northern), Bangkok (13.7253188, 1007529578; Central), Chai Nat (15.2233373, 100.1907276; Central), Pratum Thani (14.0619100, 1005437231; Central), Samut Prakan (13.5770855, 1008557932; Central), and Nakhon Si Thammarat (8.2169321, 99.8872206; Southern), from October 2021 to May 2022. Each of them oviposited eggs in an isolated ovipot to establish an iso-female line in the insectary using the methods described. To identify species, a representative of the F1 progeny of each isoline was subjected to morphological and molecular identification using keys published and a PCR method using the barcoding primers of the mitochondrial cytochrome c oxidase subunit I (*COI*) gene, LCO1490 (5′–GGT CAA CAA ATC ATA AAG ATA TTG G–3′) and HCO2198 (5′–TAA ACT TCA GGG TGA CCA AAA AAT CA–3′). PCR products were purified using the PureLink™ PCR Purification Kit (Thermo Fisher Scientific, Vilnius, Lithuania) and sequenced using a 23 ABI 3730XLs sequencer (Macrogen Inc., Seoul, South Korea). To generate the image dataset, 231 *Ae. aegypti*, 223 *Ae. albopictus*, 197 *Ae. vexans*, 160 *Cx. quinquefasciatus*, 269 *Cx. vishnui*, 178 *An. tessellatus*, and 236 *Cx. aculaetus* females, or about 200 specimens each from the F1 progeny, were sampled. Additionally, 100 females of each laboratory strain, including *Ae. aegypti, Ae. albopictus*, and *Cx. quinquefasciatus*, were used. Each and all mosquitoes were photographed a few times. All photos would then be preprocessed, and the resulting images would comprise an image collection that would henceforth be referred to as *the new dataset*. ## Image datasets Two image datasets were utilized: (1) *the reference dataset* of mosquitoes in South Korea and (2) *the new dataset* of mosquitoes in Thailand (generated in this study). Using the sources described previously, the dataset containing images of mosquitoes in Thailand was generated in two distinct settings. In the first setting, 4,590 high-quality photographs were taken using a flagship Vivo V21 (Guangzhou, China) smartphone equipped with a phone holder, an LED ring light, and a clip-on macro lens, as depicted in. The 64MP smartphone was set to produce photos at its highest resolution of 9,248 × 6,936. The clip-on lens was used to shorten the focus distance, magnifying the mosquito in the photograph. This collection of images was used to evaluate the accuracy of our proposed model. In the second setting, ten or more mosquitoes per species were resampled. They were captured by mid-range smartphones: Samsung A52s (Seoul, South Korea) and a low-end Vivo Y21, which produced 459 and 564 images, respectively. The photographer held one of the mobile devices with the same clip-on lens while photographing mosquitoes under variable indoor lighting conditions from different angles. The 64MP mid-range and the 13MP low-end phones produced images with resolutions of 9,248 × 6,936 and 4,128 × 3,096, respectively. Note that the quality of the lens and image sensors on the phones’ cameras has a greater impact on the resulting photos than the resolutions of the cameras. In this simple setting, the photographs generated may be blurry, feature shading, or have a high level of shot noise. This second collection of photos was utilized to assess the robustness of the proposed models and to evaluate the results of retraining. The difference in the quality of images between the first and second settings can be clearly seen (left and middle images) in. lists the number of females and photos captured by three smartphones from each biological species, along with their collection sites. Despite our efforts to minimize the distance between the camera and the subject, the mosquitoes in the photographs are quite small. A YOLOv5 that had been trained to estimate the boundaries of mosquitoes (if any) was deployed to crop and resize each image to 512 × 512 pixels. All of the photos captured in the first setting were successfully cropped and resized. Of about one thousand photos taken in the latter setting, only 31 photos (3%) were found to be of no use due to their very low quality (the YOLOv5 could not find a mosquito in it), and thus were discarded. ## Techniques to improve accuracy ### Transfer learning Transfer learning is a machine learning technique in which the parameters of a well-trained model are used as the initial parameters of another model before it is trained, provided that both models have identical or comparable structures. *ImageNet* offers multiple sets of parameters for well-known models, including the VGG16. In this study, ImageNet’s parameters were used to begin training all of our modified VGG16-based models when high-quality images from *the reference dataset* and from *the new dataset* were applied. Later, when the low-quality images from *the new dataset* were applied, the parameters obtained from our trainings (using the high-quality images) would be transferred to retrain the models. ### Data augmentation It is well-known that, as the amount of data increases, the performance of the deep learning model also improves. In general, data augmentation is a technique of altering existing data to create a lot more data for model training purposes. As the number of mosquito specimens in our lab are limited, only 600 *original* images generated from photos taken in our well-lit lab were available. We employed data augmentation or *image augmentation* to expand the original dataset *artificially* to train our deep neural networks. Herein, the augmented images are representative of the quality of images that we get from users who take photos in various conditions. - **Random zoom**. Due to the fact that YOLOv5 was already deployed to resize images, so that they all have the same dimensions of 512 by 512 pixels, the mosquitoes on the images are not proportionally scaled. In addition, lab- reared mosquitoes are typically larger than those found in their natural habitat due to superior nutrition. Therefore, size cannot ever be a defining characteristic. We want each model to observe mosquitoes of varying sizes. All images of mosquitoes were randomly resized to within ±20% of their original dimensions. This method ensures that models are insensitive to the size of mosquitoes depicted in images. In other words, the models can learn the appearance of a species regardless of the size of the mosquitoes. - **Image rotation**. Similar to variation in size, we cannot ask our prospective users to take photos of mosquitoes at a specific angle. Therefore, the image should be randomly rotated within a range of ±180° so that the models are invariant to the rotation of the mosquito images. - **Random crop**. It is noted that YOLOv5 helps us to zoom in and resize the images of the mosquitoes. Most of the time, the bounding box is slightly off center. So after resizing, the mosquito appears near the center of the image (but not exactly there). In addition, 512 × 512-pixel images are kept in a database for future use; in this work, the images were resized further down to 256 × 256 pixels before use. Then, the images are randomly cropped, providing 224 × 224 pixels to let the models learn the mosquitoes while they move around the center of the images. - **Random brightness**. Modern mobile phones usually handle the brightness of photos very well, especially in not-so-bad lighting conditions. Therefore, we randomly change the brightness of images only in a range of ±10%. - **Random hue**. Users are advised to take photos in good lighting conditions. In terms of brightness, most users know what is “good”. However, it is not the case for hue or color temperature; the background of images may appear reddish, yellowish, or blueish. Thus, it is sensible to randomly rotate the hue: fully 360° to let the models see the same mosquito images in various background colors. - **Gaussian noise injection**. When taking photos, shot noise or image noise is always present, especially on mobile phones with low-quality cameras. In addition, it is well-known that training a deep model with noise can reduce the likelihood of overfitting, particularly when the dataset is relatively small in comparison to the number of parameters. Before employing any image, we intentionally added random Gaussian noise with a standard deviation of 0.1. ### Spatial dropout Training deep learning models on relatively small datasets can lead to overfitting (to the training data, but not to the general data); training them for too many epochs guarantees that they will become overfit as well. In, the architecture of VGG16 consisting of five convolutional blocks, is depicted. Each block contains two or three consecutive convolutional layers followed by a max pooling layer. The total number of parameters exceed 138 million. Even though the number of images have been increased from 4,600 to 20,000 via image augmentation described previously, the dataset is still relatively small. So there is still an opportunity for the model to overfit. To reduce the chance, a dropout layer may be inserted between two layers. In the training phase, it randomly omits values from nodes in the previous layer to nodes in the next layer, thereby ensuring that the output(s) of the models will never be an exact match for the dataset. Despite higher loss during training, greater accuracy is obtained during testing. It is noted that no values are dropped during testing or inference phase; rather, they are averaged out. *Spatial dropout* evolves from *dropout*. However, for images, dropout may not be effective enough due to the high mutual correlations between the adjacent 2-D pixels. In the case of classification, dropping any of them may produce the same outcome. Instead of blocking random nodes within and between channels, spatial dropout blocks all nodes of some random channels because correlation between channels is significantly lower than correlation within channels. The distinction between standard dropout and spatial dropout is shown, as in. In this paper, we propose to insert five spatial dropout layers, each of which is after the max pooling layer of each convolution block. The architecture of our modified VGG16, designated as “SDVGG16”, is depicted in. ### Multiple inputs To increase the network’s robustness, we propose employing models that receive two images rather than one, as is the case with the SDVGG16 network. If one image lacks the part(s) that specifies the species, or if one image is of poor quality, the models having two inputs would have a greater chance of identifying the correct species. As shown in, the dropout concept of the spatial dropout is expanded upon so that two layers from two input images can be merged into a typical VGG16 layer. It is seen that the odd channels of the first image are blocked, while the even channels of the second image are dropped. The unblocked channels are interleaved to form a layer with the same shape as the conventional single input model. This study focuses on the following three dual-view models: - **Early-combined model**. Two branches from two inputs are combined just after the *first* convolution block. Later, it may be referred to as Model (A). - **Middle-combined model**. Two branches from two inputs are combined after the *third* convolution block. Later, it can be called Model (B). - **Late-combined model**. Two branches from two inputs are combined after the *fifth* convolution block. It may henceforth be referred to as Model (C). The architectures of the early-, the middle-, and the late-combined models are shown in Figs, respectively. ## Classifier One of the principal outcome of this study is our *classifier* as illustrated in. It is composed of the three dual-view models: Models (A), (B), and, (C) and an ensemble model described below. Our classifier emphasizes its robustness so it requires three images instead of one. All three images are sent to each and all three models. As such, this produces three pairs from three inputs, $\binom{3}{2}$. Hence, each model produces three prediction results, which are three sets of confidence scores. Totally, the three models give out nine sets of confidence scores. ### Ensemble model After the three dual-view models have predicted the species from three images, the accuracy of *the classifier* can be enhanced by fusing nine sets of confidence scores into one using an ensemble model, which may be henceforth called Model (D). Each dual-view model contributes three sets of confidence scores to the ensemble model. The species that receives the highest fused score is the optimal answer of both Model (D) and *the classifier*. In this study, we propose to make use of a simple but effective artificial neural network (ANN) with a 6-node hidden layer as the optimal result combiner. # Results and discussion Multiple experiments were designed to assess the performance of our proposed models. All the models were written in Python with the Keras API, which is built on top of the TensorFlow Framework. KerasCV provides some image processing functions. In all experiments, the same codes were executed ten times. Performance of the models was investigated in accordance with the scores gained via their means and standard deviations of their classification accuracy. Herein, results are provided in the following tables (Tables –)). The best, second-best, and worst of means and standard deviations are indicated by the figures in bold, italic, and underline, respectively. Our Python codes and simulation outputs in the form of the well-known Jupyter notebook are available on GitHub, whose link is given in. The link to download the dataset produced in this study can also be found on. ## Experiments on the reference dataset A couple of experiments utilizing *the reference dataset* have been set up. The first one attempts to determine the optimal spatial dropout rate. The second experiment aims to evaluate the precision and the robustness of the proposed models. In the same article that published *the reference dataset*, a Python code for its VGG16 version, henceforth referred to as *the reference model*, was also given. All the Python codes for all models were carried out ten times. At the beginning of each run, the dataset was randomly divided into 8:2 proportions of training and test datasets; the models were initialized with parameters from ImageNet. It is noted that when either an augmented or original image was selected for the training dataset, all of the augmented/original versions were chosen too, due to their high degree of correlation. ### SDVGG16 In this experiment, the dropout rate of the SDVGG16 model varied from 0.1 to 0.6 with a spacing of 0.1. Results were then compared with those of *the reference model* that had no spatial dropout layer. Their classification accuracy gained from ten executions are shown in. As expressed in, the accuracy of *the reference* model are significantly lower than those stated in its article. It is claimed that the *fine-tuned* accuracy is 97.19%, but based on our ten tests *without fine tuning*, the average accuracy is only 83.26%. Moreover, its standard deviation is higher than any of the SDVGG16. The outcome of the SDVGG16 model is extremely promising. With our data augmentation and spatial dropout techniques, the accuracy is increased by greater than 15% compared to that of the reference model using the same dataset. The SDVGG16 with dropout rates of 0.1 and 0.2 produces comparable results, as the mean of the one with a dropout rate of 0.1 is negligibly higher, but the one with a dropout rate of 0.2 is more consistent, as indicated by a little bit less SD. On the other hand, outcomes become progressively worse as its dropout rates increase from 0.3 to 0.6. Thereafter, if the dropout rate is not mentioned, it is 0.2. ### Multi-view models In this experiment, the train and test datasets are further divided into groups of two images for dual-view inputs in order to train and test the early-, medium-, and late-combined models (Models (A), (B), and (C)). It is noted that the dropout rate is fixed at 0.2. To train the ensemble model, the train/test dataset is re-split into groups of three images that are consumed by all three models. The three models confidence scores serve as the ensemble’s inputs. After being trained, *the classifier*, which is composed of the three dual-view models and the ensemble model, predicts the species using the unseen test dataset as its inputs. In, the accuracy results of *the reference model*, Models (A), (B), (C), and *the classifier* are shown. demonstrates that the accuracy of the dual-view models using the same dataset proved to be significantly higher than that of *the reference model*. In particular, the results of *the classifier*, which are the outputs of the three models plus that of the ensemble, are nearly perfect and has a standard deviation of only 0.124 percent. This demonstrates that our proposed methods increased accuracy and decreased uncertainty caused by various inputs (lower SD). It should be noted that all of our proposed models are not fine-tuned, so they can be retrained without the assistance of AI specialists when new data becomes available. ## Experiments on our dataset To show that our models do not depend on a specific dataset, all of our proposed models again attempted to guess the species using the mosquito images from the dataset produced in this study. *The reference dataset* contained five vector species and a non-vector counterpart, but our dataset contained six vector species and a non-vector counterpart. ### Accuracy tests on high-quality images To recheck the accuracy on another high-quality dataset, all dual-input models were initialized with ImageNet parameters, and only high-quality images from the first setting were used as input. Once more, samples were chosen at random for train/test datasets with a ratio of 8:2. In order to train and evaluate the early-, medium-, and late-combined models, the samples were further divided into groups of two images for dual-view inputs. The dropout rate was 0.2. In, the classification accuracy results for our four models are displayed. The repeated experiments on our dataset yielded similar results. The average accuracy of the three multi-view models decreased slightly, but all were found to be over 95% with a standard deviation of less than 2%. The lower accuracy may be due to the increased number of species (from 6 to 7) that the models must differentiate. However, the ensemble/classifier results are still nearly perfect and extremely consistent (SD = 0.324%). This demonstrates that our proposed models do not depend on a specific dataset. ### Accuracy tests on low-quality images In this experiment, one of the ten trained models from the previous experiment was randomly selected and retrained using all of the high-quality images. Only 20% of the input images were unseen, so the model was trained in a few epochs. To evaluate accuracy, ninety percent of the low-quality images (from the second setting) were randomly selected and used as the testing input for the models, which were obtained by retraining. In, the classification accuracy of the ten tests conducted is shown. It is not surprising that overall precision would decline. However, the results demonstrate the robustness of the proposed models when tested on datasets that are comparable but inferior. The models can still perform, albeit, with diminished precision. The ensemble’s mean accuracy was found to be 86.14 percent, which is still quite high. This experiment reveals the classifier’s true potential. In previous experiments, the difference in accuracy was only a few percentage points. However, in this experiment, the ensemble’s accuracy is seen to be significantly higher than that of the dual-view models. The SDVGG16 model proved to be more accurate than the dual-view models by 3–10% but less accurate than *the classifier* by 6%. ### Accuracy tests after retraining The aim of this experiment is to check if the proposed models can be retrained to learn more mosquito images, captured from various conditions. In this experiment, models trained with 80% of high-quality images would be retrained using ten random selections of 80% of both high- and low-quality images. All of the unselected images–both the high- and low-quality ones–would be used to evaluate the retrained models. displays the accuracy together with means and standard deviations after the retraining. The results presented in Tables and suggest that the classification accuracy after deployment may not be high immediately because the models were only trained on lab-quality images and the image quality of the photographs submitted by prospective users cannot be guaranteed. However, after tedious labeling and retraining, the accuracy of the models was restored. We hope that after a few retrainings, their inference accuracy will decrease less as they are exposed to more *generic* images. ### Robustness tests In this experiment, all models were trained on high-quality images before being tested on poor-quality photographs. Similar to a previous experiment, but one or two input images were purposefully blurred to diminish their quality. Thus, the one or two input images were degraded using a 2-D Gaussian filter with a kernel size of 11 × 11 (note that the size of images in the dataset is 512 × 512) and a sigma of 1.5 (the higher the sigma, the blurrier the image). In, the difference in image quality before and after blurring is quite evident, as seen in the middle and right images. It is noted that the SDVGG16 model only accepts a single image. Dual-view models require two images, and *the classifier* (a combination of the dual-view and ensemble models;) requires three images. Therefore, only dual-view models and *the classifier* would be tested with two bad inputs. When the single-view SDVGG16’s only input is of poor quality, the accuracy of the model drops considerably from 80.35% to 54.43% on average. In contrast, the accuracy of the multi-view models drops much less, as evidenced by a comparison of the results in Tables and. In particular, only 0.4 percent is lost for the late-combined model (Model (C)). *The classifier* is still found to be the most reliable because two out of its three inputs are good. The results in demonstrate that the dual-view models perform poorly when both of their inputs are of bad quality. The classifier’s accuracy still holds up at about 72% despite the fact that only one of its three inputs is good. # Conclusion Herein, using multi-view and spatial dropout techniques, the performance of the well- known VGG16 was enhanced. Each max pooling layer of the VGG16 was followed by a spatial dropout layer; its architecture was modified to allow the model to simultaneously accept two images. Furthermore, an ensemble model was proposed to receive the results from the three models and yield the most accurate answer. A reference dataset containing mosquito vectors in South Korea and our own dataset containing vectors in Thailand were utilized. Using the reference dataset, the ten-run accuracy increased from 83.26% with a standard deviation of 2.602% to 99.77% with a standard deviation of 0.124%. Our dataset’s classification accuracy exceeded 99% with a small standard deviation. When low-quality images were applied, our models proved to be retrainable and robust. In this study, the classifier system is only in its prototype stage. In the real-world operation, classification accuracy will not be this high as the quality of images from users nationwide can vary hugely. Hence, producing models that are more robust to generic image quality will be our aim in further work. The ensemble model in this study combines only the results from the dual-view models, but the results from the SDVGG16 can be utilized too. Combining results from various models to increase robustness even further will also be our objective. # Supporting information 10.1371/journal.pone.0284330.r001 Decision Letter 0 P.P. Abdul Majeed Anwar Academic Editor 2023 Anwar P.P. Abdul Majeed This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 6 Feb 2023 PONE-D-22-32464Enhancement of VGG16 model with multi-view and spatial dropout for classification of mosquito vectorsPLOS ONE Dear Dr. Pora, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Mar 23 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Anwar P.P. Abdul Majeed Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at  <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and  <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Please update your submission to use the PLOS LaTeX template. The template and more information on our requirements for LaTeX submissions can be found at <http://journals.plos.org/plosone/s/latex>. 3\. Thank you for stating the following in the Acknowledgments Section of your manuscript:  "This study is funded by Thailand Science Research and Innovation Fund Chulalongkorn University (CU FRB65 hea(45) 052 30 33)." We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form.  Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows:  "This study is funded by Thailand Science, Research and Innovation Fund Chulalongkorn University (CU_FRB65_hea(45)\_052_30_33). This fund is managed by Thailand science, research and innovation office (<https://www.tsri.or.th>).  Its principal investigator and investigator are Narissara Jariyapan and Wanchalerm Pora, respectively.  The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." Please include your amended statements within your cover letter; we will change the online submission form on your behalf. 4\. In your Data Availability statement, you have not specified where the minimal data set underlying the results described in your manuscript can be found. PLOS defines a study's minimal data set as the underlying data used to reach the conclusions drawn in the manuscript and any additional data required to replicate the reported study findings in their entirety. All PLOS journals require that the minimal data set be made fully available. For more information about our data policy, please see <http://journals.plos.org/plosone/s/data- availability>. Upon re-submitting your revised manuscript, please upload your study’s minimal underlying data set as either Supporting Information files or to a stable, public repository and include the relevant URLs, DOIs, or accession numbers within your revised cover letter. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data-availability#loc-recommended- repositories>. Any potentially identifying patient information must be fully anonymized. Important: If there are ethical or legal restrictions to sharing your data publicly, please explain these restrictions in detail. Please see our guidelines for more information on what we consider unacceptable restrictions to publicly sharing data: <http://journals.plos.org/plosone/s/data-availability#loc- unacceptable-data-access-restrictions>. Note that it is not acceptable for the authors to be the sole named individuals responsible for ensuring data access. We will update your Data Availability statement to reflect the information you provide in your cover letter. 5\. We note that you have referenced (ie. Jomtarak R, Kittichai V et al. \[17\]) which has currently not yet been accepted for publication. Please remove this from your References and amend this to state in the body of your manuscript: (ie "Jomtarak R, Kittichai V  et al. \[Unpublished\]") as detailed online in our guide for authors: <http://journals.plos.org/plosone/s/submission-guidelines#loc-reference-style> 6\. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Thank you for the interesting study. The manuscript has been presented well. However, there are improvements that can be made. Would you please mention the resolution of the high-quality photographs mentioned in the subchapter “Image datasets”? What is the resolution of the photographs captured by the mid-range and low-end smartphones mentioned in the subchapter “Image datasets”? This may or may not affect the input image details as it will be resized accordingly, but it is worth to be mentioned. Thank you. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No \*\*\*\*\*\*\*\*\*\* \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0284330.r002 Author response to Decision Letter 0 28 Feb 2023 There is only one reviewer and (s)he gave only two comments: Comment 1. Would you please mention the resolution of the high-quality photographs mentioned in the subchapter "Image datasets"? Response: Thank you so much for bringing this to our attention. We added a sentence: The 64MP smartphone was set to produce photos at its highest resolution of 9,248 x 6,936. Hopefully, it will provide additional information to the audience. Comment 2. What is the resolution of the photographs captured by the mid-range and low-end smartphones mentioned in the subchapter "Image datasets"? Response: Again, thank you for pointing this out. So we added a sentence: "The 64MP mid-range and the 13MP low-end phones produced images with resolutions of 9,248 x 6,936 and 4,128 x 3,096, respectively." To remind the audience that the quality of the image does not depend solely on the image resolution, we added further a sentence: "Note that the quality of the lens and image sensors on the phones' cameras has a greater impact on the resulting photos than the resolutions of the cameras." Despite having the same image resolution, the Vivo V21 (the high-end) and Samsung A52s (the mid-range) received DXOMark (still camera) scores of 105 and 88, respectively. This shows that the cameras on the Vivo V21 are generally better than those on the Samsung A52s. Note that DXOMark.com did not give the Vivo Y21 (the low-end) a camera test. 10.1371/journal.pone.0284330.r003 Decision Letter 1 P.P. Abdul Majeed Anwar Academic Editor 2023 Anwar P.P. Abdul Majeed This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 28 Mar 2023 Enhancement of VGG16 model with multi-view and spatial dropout for classification of mosquito vectors PONE-D-22-32464R1 Dear Dr. Pora, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Anwar P.P. Abdul Majeed Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Thank you for the responses. All comments has been addressed. Please prepare the manuscript for the final proof by the publisher. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No \*\*\*\*\*\*\*\*\*\* 10.1371/journal.pone.0284330.r004 Acceptance letter P.P. Abdul Majeed Anwar Academic Editor 2023 Anwar P.P. Abdul Majeed This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 4 Apr 2023 PONE-D-22-32464R1 Enhancement of VGG16 model with multi-view and spatial dropout for classification of mosquito vectors  Dear Dr. Pora: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Anwar P.P. Abdul Majeed Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist. [^2]: ‡ NK, KT, RA and PS also contributed equally to this work.

# Introduction Following the epidemic of obesity and metabolic syndrome recorded in children and adolescents in the last couple of decades, nonalcoholic fatty liver disease (NAFLD) has become the main cause of chronic liver disease in these groups. In Western countries, the prevalence of NAFLD is 20–30% in the pediatric population and 70–80% in obese children. NAFLD is considered a “multi-hit” disorder, in which genetic, epigenetic and environmental factors interact causing the onset and progression of liver damage. Recent studies have demonstrated that approximately 25% of children with NAFLD have NASH and interesting data derived from longitudinal studies have indicated that hepatic fibrosis is the most important prognostic marker of progression of liver disease. The rate of progression of liver disease in NAFLD is slow with an estimated average of 7 years elapsing between the development of NASH with fibrosis in patients who had prior nonalcoholic fatty liver. Because the presence of liver fibrosis predicts liver-related outcomes and mortality, blocking mechanisms of fibrogenesis is a key therapeutic goal in the treatment for NASH. Fibrosis is characterized by an excessive deposition of extracellular matrix (ECM), with increases in total collagen content and in fibril-forming collagens (Type I, III and IV). These changes induce dysfunction and activation of the hepatic stellate cells (HSCs) with development and progression of fibrogenesis. Lifestyle interventions, consisting of a weight decreasing diet and increases in physical exercise, remain the cornerstone of treatment of pediatric NAFLD, even if several studies indicate improvement only in metabolic parameters and liver steatosis. Consequently, in the last decade, several pharmacological approaches have been tested that are focused on ameliorating mechanisms of liver damage. Unfortunately, none of the tested drugs to date has produced unequivocal results with the most effective treatments showing limited efficacy and worrying side effects in studies in adults. Omega-3 fatty acid treatment is potentially safe in adults and children, and docosahexaenoic acid (DHA) treatment in children and omega-3 fatty acid treatment producing \> 2% DHA tissue enrichment in adults has shown promising results to decrease liver fat in patients with NAFLD. Recently, vitamin D deficiency (VDD) has been associated with obesity, metabolic syndrome and cardiovascular risk in adults and children. VDD occurs frequently among healthy children, with a rate of 55% in the American pediatric population. Moreover, several studies have reported that VDD is common in patients with NAFLD and importantly, VDD is associated with increased risk of steatosis, necroinflammation and fibrosis in both adults and children with biopsy-proven NAFLD. Several studies in humans and in animal models indicate that VDD contributes to increased oxidative stress and systemic inflammation. Furthermore, emerging evidence suggests a role for VDD in fibrogenesis, with the potential therefore for an anti-fibrotic effect of vitamin D treatment. The available data to date suggests that vitamin D may reduce fibrotic processes inhibiting the expression of transforming growth factor beta (TGF-β) and suppressing the deposition of collagen Iα1 and the activation of alpha-smooth muscle actine (α-SMA) positive HSCs. Therefore, given the potential benefits of both DHA treatment and vitamin D treatment to ameliorate the features of NAFLD/NASH, the aim of the present study was to undertake a proof of concept, randomized double blind placebo-controlled trial (RCT) to test the potential efficacy and tolerability of a mixture of DHA and Vitamin D in children and adolescents with vitamin D deficiency and biopsy- proven NAFLD # Materials and Methods ## Study population and design An RCT was undertaken to examine the efficacy and safety of a mixture of vitamin D (800 IU) and DHA (500 mg) orally once daily, versus identical placebo for 24 weeks on hepatic histology and metabolic parameters in children and adolescents with biopsy-proven NAFLD. Sixty-six white European patients (4–16 years) with liver biopsy-proven NAFLD, referred to the Hepato-Metabolic Department of “Bambino Gesù” Children’s Hospital (Rome, Italy) between March 2014 and April 2015, were evaluated for the present study. Patients were recruited and studied between March 2014 and April 2015. Children were eligible for the study if they were between: 4 and 16 years of age, had a liver biopsy result consistent with a diagnosis of NAFLD/NASH, and also had decreased serum vitamin D levels (\< 20 ng/ml), aminotransferases (ALT) levels \<10 upper limit of normal (ULN), and no laboratory and/or clinical signs of liver decompensation. Moreover, in all children other causes of liver disease, such as viral liver disease, autoimmune hepatitis, Wilson's disease, α-1-antitrypsin deficiency, celiac disease, alcohol consumption (any quantity), use of drugs known to induce fatty liver, were also excluded. Patients were randomized to receive capsules combining 500 mg of docosahexaenoic acid and 800 IU of Vitamin D (Treatment arm) or identical capsules as placebo (Placebo arm). The dosage of the DHA and vitamin D intervention were determined based on the available evidence in obese patients with NAFLD. As previously reported by our group, DHA supplementation improves liver steatosis and insulin sensitivity in children with NAFLD with similar effects for doses of 250 and 500 mg/day. As for vitamin D, several expert groups, including the American Academy of Pediatrics, have recently revised the recommended supplementation dosages. In this position paper, 600–1.000 UI/day of vitamin D have been recommended in adolescents with risk factors for vitamin D deficiency, such as obese adolescents group. Based on these finding, we treated our patients with 800 UI/day in the treatment arm. A computer-generated randomization sequence assigned participants in a 1:1 ratio to treatment with Vitamin D plus DHA (Treatment arm) or placebo (Placebo arm). A statistician, who was blinded to participants' clinical data and did not participate in patients' clinical care, generated the allocation sequence and assigned participants to their group. Only the statistician had access to the treatment codes. The capsules were taken every day for 24 weeks. Additionally, all patients were included in a lifestyle intervention program consisting of a hypocaloric diet (25–30 Kcal/kg/day) and regular physical exercise (twice weekly 1-hour physical activity). Participants and investigators were blinded to the treatment for the duration of the study. Capsules were dispensed at the baseline visit, and after three months. The compliance with treatment was monitored at each visit by counting the returned capsules. Moreover, at each visit, adverse effects were recorded by the Principal Investigator. Anthropometric measurements and laboratory data were collected at each visit (at baseline, 6 and 12 months). Liver biopsy was performed at baseline and after 12 months, only in the treatment arm. For ethical reasons and according to the Position Paper of the Hepatology Committee of ESPGHAN (European Society of Pediatric Gastroenterology, Hepatology and Nutrition) at the end of study, it was decided to repeat liver biopsy only in treated patients and patients in the placebo group did not undergo an end of study biopsy. We defined changes in NAS as the primary outcome of the present proof of concept trial because several studies have showed liver histology as the most appropriate endpoint to define efficacy in clinical trials in NAFLD. The secondary outcomes were the improvement of metabolic parameters, such as gluco- insulinemic profile and serum lipid concentrations. ## Anthropometrical and biochemical measurements Anthropometric measurements and laboratory tests, including liver enzymes, gluco-insulinemic profile and lipids were performed at baseline and repeated at 6 and 12 months. The body weight and height were measured with the patients wearing underwear. Body mass index (BMI = kg/m²) and standard deviation score (Z score) were calculated. Serum glucose, lipid profile \[triglycerides, cholesterol-total, high-density lipoprotein (HDL) cholesterol and low-density lipoprotein (LDL)\], liver function tests (aspartate- (AST) and alanine- (ALT) aminotransferases, gamma- glutamyl-transpeptidase (GGT), albumin and International Normalized Ratio (INR)), fasting plasma glucose and insulin were measured in all patients after an overnight 12-h fasting. In all patients, Oral Glucose Tolerance Tests (OGTT) were performed. Insulin-resistance (IR) was assessed by the homeostatic model assessment (HOMA) \[HOMA-IR = (insulin0 (μIU/ml) x glucose0 (mmol/l))/22.5)\]. A cut-off value of \> 2.5 was considered as an index of insulin resistance. In all patients, serum 25-hydroxyvitamin D \[25(OH)D, vitamin D\] concentration was measured by radioimmunoassay (IDS Immunodiagnostics, IDS Limited, Tyne and Wear, UK). Subjects were categorized as having either low vitamin D levels (\<20 ng/mL), or normal vitamin D levels (≥20 ng/mL). ## Determination of total monthly hours of sunlight The mean hours of sunshine was determined using the “Italian atlas of solar radiation” from the ENEA center ([http://www.solaritaly.enea.it](http://www.solaritaly.enea.it/)). The formula for estimating mean hours of sunlight was: % Sunshine x \[(Clear days x 0.85) + (Partly Cloudy days x 0.45) + (Cloudy day x 0.10) x 24\] 1. Sunshine % = the percentage of the daylight hours for Rome during that month; 2. Clear days = defined as 70%-100% of sunshine; was used for the mean value of85% or 0.85 in the formula; 3. Cloudy days = defined as 30%-60% of sunshine; was used for the mean value of 45% or 0.45 in the formula; 4. Cloudy Days = defined as 0–20% of sunshine; was used for the mean value of 10% or 0.10 in the formula. ## Liver biopsy Echo-guided liver biopsy was performed using an automatic core biopsy device (Biopince, Amedic, Sweden) with an 18-G needle, under general anesthesia. A single experienced pathologist evaluated liver specimens. The histological features of steatosis (0–3), lobular inflammation (0–3), and hepatocyte ballooning (0–2) were combined in the NAFLD activity score (NAS), ranging from 0 to 8 using the criteria of NAFLD Clinical Research Network. ## Assessment of fibrillar collagen deposition in liver biopsies The assessment of fibrillar collagen deposition within the liver biopsy was evaluated in Sirius Red (SR) stains, as previously. Briefly, SR stained slides were scanned by a digital scanner (Aperio Scanscope CS System, Aperio Technologies, Inc, Oxford, UK) and processed by ImageScope. An image analysis algorithm has been used to quantify the proportion of SR-stained area. The algorithm was applied on the entire section (Part A and B of). The extent of collagen deposition was expressed as the proportion (%) of SR-stained area with respect to the total biopsy area, providing a quantitative value on a continuous scale. Only biopsies containing at least 5 portal tracts were considered. In order to establish reference values for fibrillar collagen in normal liver samples, specimens from 6 lean, non-diabetic children (boys, 4; girls, 2; median age: 13 years, range, 12–16 years) without liver disease were used as controls, as previously. These fragments were obtained from patients who underwent laparotomy or laparoscopic procedures (for cholecystectomy), from liver donors (orthotopic liver transplantation) or incidental “normal” liver biopsies (children exhibiting persistent/intermittent elevations of liver enzymes for \>6 months). Informed consent in writing was obtained from next of kin, caretakers, or guardians on behalf of the children enrolled in this study. ## Immunohistochemistry for α smooth muscle actin and evaluation of hepatic stellate cell/myofibroblast pool Sections were incubated overnight at 4°C with primary antibodies against α smooth muscle actin (αSMA: Dako, mouse monoclonal, code: M0851, dilution: 1:50). Samples were then incubated for 20 minutes at room temperature with secondary biotinylated antibody and, successively, with streptavidin-Horse radish peroxidase (LSAB+, Dako, code K0690). Diaminobenzidine (Dako, code K3468) was used as the substrate and the sections were counterstained with hematoxylin. For all immunoreactions, negative controls (the primary antibody was replaced with pre-immune serum) were also included. Sections were examined with a Leica Microsystems DM 4500 B Microscopy (Weltzlar, Germany) equipped with a Jenoptik Prog Res C10 Plus Videocam (Jena, Germany). Observations were processed with an Image Analysis System (IAS, Delta Sistemi, Rome, Italy) and were independently performed by 2 researches in a blinded fashion. Only biopsies containing at least 5 portal tracts were considered. The activation of Hepatic Stellate Cell (HSC)/Myofibroblast (MF) pool was evaluated by counting the number of αSMA-positive cells per high power field (HPF: at 40x). Perisinusoidal HSCs and portal/septal MFs were separately evaluated; αSMA-positive HSCs were recognized in accordance with their stellate/spindle shape and their perisinusoidal location within the parenchymal lobule; besides, portal/septal MFs were considered as stellate- or spindle- shaped (αSMA-positive cells) located at the interface between parenchyma and portal tract or between parenchyma and septa, and those residing in the portal tracts and the fibrotic septa. The number of αSMA-positive HSCs and MFs was counted and expressed as number of positive cells per HPF. Only the cells which displayed nuclei on the section were considered. For each slide, at least 15 non-overlapping microscopic HPFs were randomly chosen. ## Ethical Approval The trial was fully approved by the Ethics Committee of the Bambino Gesù Children's Hospital in January 2014; protocol number: 791.13/0PBG, see), according to the Declaration of Helsinki (as revised in Seoul, Korea, October 2008) and CONSORT guidelines. A written informed consent to the study protocol and to publication of results was obtained from the parents or legal guardians of the children. This study was registered on March 24, 2014 in ClinicalTrials.gov (Registration Number: NCT02098317 –see). The authors confirm that all ongoing and related trials for this drug/intervention are registered. ## Statistical analysis The data were analyzed using a STATISTICA (version 2010, Chicago, IL, USA). Continuous variables were expressed as mean ± standard deviation (SD). Data distribution was checked for normality by the Kolmogorov-Smirnov test. Data were analyzed using the intention-to-treat principle and the values recorded at baseline were compared to values recorded at 6 and 12 months in all patients, regardless of treatment duration. Baseline and follow up characteristics were tested for differences by Student’s *t*-test (p\<0.05). The change of anthropometrical and laboratory values, between placebo and treatment groups, was evaluated using analysis of variance (ANOVA) with repeated measures. Difference between proportions were tested using the Chi-square test. Univariate correlations were investigated with Pearson’s correlation. Multivariable logistic regression analysis was used to test the independence of associations between end of study vitamin D concentrations as the key exposure and histological characteristics after adjusting for BMI, change in BMI between baseline and follow up and basal values of Vitamin D. # Results ## Baseline characteristics In our study, between March 2014 and April 2015, 66 patients were screened and 43 of these with biopsy-proven NAFLD were enrolled. The patients were enrolled with similar proportions recruited during the winter (20/43, 47%) and spring (23/43, 53%) months. Twenty patients received an oral dose of 500 mg of DHA and 800 IU/day of Vitamin D (treatment arm) and 23 children received the capsules of placebo for 6 months (placebo arm). Forty-one patients completed the study, with two patients from the treatment arm being lost to follow up. There were no significant adverse events. The dropouts from the treatment arm were not associated with any study adverse events, but were due to the refusal by parents to consent to the second liver biopsy at 12 months. The two groups had similar baseline characteristics, as shown in. ## Effects on anthropometric, clinical and laboratory parameters shows the anthropometric and laboratory characteristics for each arm of the study. At 12 months, the placebo group showed no significant improvements for any anthropometric and laboratory parameters. In the treatment arm, at 12 months, there was a decrease in BMI (28.42 to 24.58 kg/m², p = 0.04), serum triglyceride concentration (174.5 to 102.15 mg/dl; p = 0.001) and in the measure of insulin-resistance were observed (HOMA-IR 4.59 to 3.42; p = 0.03). Repeated measures ANOVA showed both treatment and placebo decreased BMI (F_(5,18) = 6.55; p = 0.0001), ALT (F_(5,18) = 4.34; p = 0.0003), Triglycerides (F_(5,19) = 10.1; p = 0.0001), insulin-‘120 (F_(5,19) = 2.97; p = 0.015) and vitamin D (F_(5,18) = 16; p\<0.0001). ## Vitamin D supplementation At baseline, all patients showed vitamin D deficiency (VDD), with median values of vitamin D of 16.01±3.98 ng/dL. The values of vitamin D were normalized to the hours of sunlight. In the placebo arm, the values of vitamin D did not change during the study, with persistent VDD. In contrast, in the treatment group, a persistent and significant increase of Vitamin D concentration was observed (baseline = 15.98; 6-months = 29.7 and 12-months = 25.42 ng/dL; p = 0.02). None of treated patients developed hypercalcemia and/or nephrotoxicity. ## Effects on liver histology Improvement in liver histology was the primary outcome of the trial. showed all histological features and NAS scores assessed at baseline in both groups and after 12 months in the treatment group. The data were similar between the two groups at baseline for steatosis, ballooning, portal and lobular inflammation and fibrosis. Lower levels of 25 (OH) D3 were associated with greater fibrosis and steatosis. Before randomization, the biopsies were classified, in accordance with the NASH CRN-criteria, into Not-NASH (N = 6) and definite NASH (N = 14). After treatment with DHA and Vitamin D, the classification of biopsies indicated a decrease of definite NASH (N = 3) and an increase of not-NASH diagnosis (N = 17). Moreover, NAS improved (from 5.40 to 1.92; p\<0.001), and steatosis (from 2.25 to 1.0; p = 0.002), ballooning (from 1.6 to 0.46; p = 0.001), lobular inflammation (from 1.5 to 0.88; p = 0.04) and portal inflammation (from 1.6 to 1.0; p = 0.05)\], whilst there was a trend toward a decrease in fibrosis (from 2.0 to 1.5; p = 0.06). Fibrosis severity at baseline was: stage 1c in 13 samples, stage 2 in 6, stage 3 in 1, and there were no biopsies that were classified as fibrosis stage 4. After treatment, no statistical significant changes were present in fibrosis stage \[stage 1c: 12 patients; stage 2: 6 patients; (stages 3–4: no patients)\]. Binary logistic regression analysis showed that change in Vitamin D level with treatment was independently associated with features of NAFLD as dichotomous outcomes \[fibrosis (OR = 2.96, 95% CI = 1.9–4.69, p-value = 0.003), steatosis (OR = 3.53, 95% CI = 1.33–3.4, p-value = 0.001) and NAS (OR = 2.75, 95% CI = 1.2–3.32, p-value = 0.005)\]. ## Fibrosis and collagen deposition assessment in liver biopsies Since the available evidence suggests that vitamin D may have a beneficial effect on fibrogenesis and as we observed a trend toward an improvement in fibrosis score with treatment, further exploratory analyses were undertaken to examine the effects of treatment on factors involved in the fibrogenetic process. The fibrillar collagen content was assessed in SR stained biopsies at baseline; overall, NAFLD biopsies at baseline showed increased but not statistically significant values of fibrillar collagen content (2.60 ± 1.76), compared with normal controls (1.44 ± 0.41; p = 0.088). Only eleven out of twenty NAFLD biopsies at baseline showed increased content of collagen fibers (3.51 ± 1.66) in comparison with normal samples (p\<0.01). Moreover, biopsies with a fibrosis score = 2/3 (N = 7) had higher fibrillar collagen content (4.17 ± 2.10) in comparison with those obtained from patients with fibrosis score = 1 (1.90 ± 1.11; p\< 0.05). At the baseline, the fibrillar collagen content calculated in SR stained slides was significantly correlated with fibrosis stage (r = 0.647; p\<0.02) and NAS score (r = 0.736; p\<0.01). Patients with an increased fibrosis content at the baseline (N = 11) showed a significant decrease in fibrillar collagen content at the end of the treatment (1.59 ± 1.37 v. x; paired t-test: t = 3.86 p = 0.003;). ### Activation of HSC/MF pool The activation of HSC/MF pool was evaluated at the baseline and at the end of the treatment by immunohistochemistry for αSMA. At the end of the treatment, the number of αSMA+ HSCs/MFs was significantly reduced (pericentral HSC = 2.72±2.51 and periportal MFs = 2.05±1.49) compared with biopsies at baseline (paired t-test: t = 4.60 p\< 0.01 and t = 3.53, p\<0.05). # Discussion To the best of our knowledge, this is the first RCT evaluating the efficacy of treatment with DHA plus Vitamin D in NAFLD/NASH patients with vitamin D deficiency, using changes in liver histology as the primary end-point. In accord with a previous study testing the effect of DHA treatment in pediatric NAFLD, the results of our study show that the administration of a mixture of DHA and vitamin D was associated with an improvement in insulin resistance with a concomitant reduction of serum triglyceride concentration and an improvement in ALT concentration. In we have compared the effect of treatment in the presented trial with that of our previous DHA trial in pediatric NAFLD, in order to test whether there were more marked effects in the DHA plus vitamin D intervention. These data show there were no significant differences between the trials for differences in triglyceride concentrations, HOMA-IR or ALT levels; thus, these comparative data suggest that treatment with DHA plus vitamin D is not better than DHA treatment alone in producing an improvement in these parameters (that are often abnormal in patients with NAFLD). Therefore, the data suggest that the amelioration of the metabolic profile observed in our patients in the current trial is probably related to the DHA treatment alone, rather than to the vitamin D treatment. The presented data are also in accord with other studies in which supplementation of vitamin D in obese children did not affect the lipid profile and markers of insulin resistance and inflammation. In contrast to our previous trial, in the presented study we observed a reduction of BMI in the treatment arm at the end of the trial (12 months). The greater weight decrease in the treatment arm may be due better adherence in this group of children to the therapeutic lifestyle advice that was given to all participants. It is well accepted that weight loss can improve the early features of NAFLD, but it is important to note that the benefit of the intervention was independent of weight loss in the treatment arm of the study. Regarding our primary outcome, NAS improved in all treated patients, with a significant reduction of steatosis, ballooning, portal and lobular inflammation. In fact, 14/20 patients with NASH at baseline improved with treatment. This improvement in NAS was similar to that observed in our previous trial testing the effects of DHA treatment alone in pediatric NASH (p\<0.05). Moreover, bearing in mind the potential for benefit of vitamin D treatment on fibrosis in NAFLD and the known prognostic implication of liver fibrosis for serious chronic liver disease-related outcomes, we evaluated the effects of treatment on changes in fibrosis score in the treatment arm. In recent years, the role of vitamin D in metabolic syndrome and cardiovascular risk has attracted considerable attention and several reports suggest a crucial role for vitamin D in NAFLD development and progression. A recent systematic review demonstrated that patients with NAFLD were 1.26-times more likely to be vitamin D deficient compared with controls. Moreover, both in adults and in children, studies show that low levels of vitamin D are associated with NAFLD, independently of known metabolic risk factors. Deficiency of vitamin D may play a role in the development of fibrosis in NAFLD. For example, Zhu et al reported that long-term vitamin D deficiency can provoke chronic liver inflammation, inducing apoptosis and activation of hepatic stellate cells (HSC) to initiate liver fibrosis. There is also evidence indicating that vitamin D is able to modulate HSC activation in vitro and to reduce liver fibrosis in experimental models of liver injuries. Despite a clear and marked improvement in the NAS, there was only a non significant trend toward an improvement in fibrosis score. In keeping with this trend toward an improvement in fibrosis score, our results indicate that vitamin D administration reduces the activation of HSC/MF pool and, in patients with increased fibrillar collagen content, we observed signs of total collagen content reduction at the end of treatment. Results from experimental cirrhosis in rats indicate that vitamin D treatment is able to prevent liver fibrosis but does not ameliorate established cirrhosis. In keeping with these data, no patients included in our trial presented with bridging fibrosis or established cirrhosis at baseline. Consequently, our results are relevant only to the early phases of fibrogenesis and the data suggest there is a benefit of reduced activation of fibrogenetic cells (HSC/MF pool) after the treatment with vitamin D. The presented study has some limitations. The first is the lack of an end of study liver biopsy in the placebo group. For ethical reasons, bearing in mind that our patients are children, the liver biopsy was not repeated at 12 months in this group. Additionally, it was not possible to test separate effects of DHA and vitamin D in this trial by using a 2x2 factorial study design and as this was a proof of concept study that lacked an end of study biopsy in the placebo group, we did not attempt a sample size calculation. A second limitation is that none of our patients showed bridging fibrosis or cirrhosis at baseline and, thus, the observed results are limited to the early stages of fibrogenesis. Therefore, it is remains uncertain whether this treatment is effective in modifying the fibrogenetic pattern in more advanced stages of liver fibrosis (F3-F4). Another limitation could be the dosage of vitamin D used in our study (800 IU daily), which although twice the average daily requirement of vitamin D, is lower than the dosage prescribed in previous clinical trials in adults with NASH. Actually, data regarding the safety of vitamin D supplementation in pediatric NAFLD is lacking. Consequently, we considered it necessary to use this dosage of vitamin D for only six months, in order to avoid possible adverse effects. For ethical reasons, we repeated the liver biopsy in our patients after one year (and not at 6 months) after randomization. In conclusion, the results of our proof of concept study have shown beneficial effects of DHA plus vitamin D treatment on insulin-resistance, ALT triglyceride concentration and NAS score in VDD patients with biopsy-proven NAFLD. The combination of 500 mg o.d. of DHA and 800 IU o.d. vitamin D was safe over 6 months of intervention. The supplementation of a mixture of DHA and vitamin D in VDD obese children and adolescents with NAFLD may induce a remodeling of the fibrogenetic pattern with a reduction of the activation of the HSC/MF pool and of collagen content. We suggest that further longer-term studies are now warranted in both adults and children, including a greater number of patients with more advanced stage of fibrosis, in order to confirm our preliminary results. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** VN, CDC, CDB, GC, EG. **Formal analysis:** AM, CDB, AL, MR. **Investigation:** AL, RDV, GC, DO, EG, VN. **Methodology:** AL, AM, CDS, LS, SC. **Resources:** CDC, VN. **Supervision:** CDB, VN, CDC; GC. **Validation:** AL, RDV, DO, SC. **Writing – original draft:** CDC, GC, VN, CDB. **Writing – review & editing:** EG, SC, LS.

# Introduction Simvastatin, a HMG-CoA reductase inhibitor (statin), is prescribed worldwide to patients with hypercholesterolemia to prevent cardiovascular disease and death. Although simvastatin is well-tolerated, side effects like myotoxicity have been reported, ranging from fatigue to life-threatening rhabdomyolysis. Several hypotheses explaining the statin-induced myotoxicity have been put forward, but the underlying mechanism is still poorly understood. The mechanisms for statin- induced myotoxicity are probably multifactorial and at least partly due to a combination of impaired isoprenylation of trafficking proteins, altered Ca²⁺ homeostasis and impaired mitochondrial respiratory function. The mitochondrial respiratory dysfunction observed in statin-treated patients caused glucose intolerance. Also, glucose intolerance was observed in tumor cells after exposure to statin, resulting in decreased glucose uptake and a higher glucose concentration in the conditioned cell medium. Regarding impaired mitochondrial function, many biochemical processes are affected by impaired glucose oxidation and low ATP production, like the activity of the lysosomal H⁺-ATPase, pH of the lysosomes and processing of lysosomal proteases. Legumain (asparaginyl endopeptidase) is a cysteine protease mainly localized to the lysosomes and was first characterized in mammals in 1997. Legumain is ubiquitously expressed in mammalian tissue, and over-expression is associated with atherosclerotic plaque instability and cancer malignancy. In cancer malignancy legumain has been reported to translocate from the lysosomes to the cell nucleus. Legumain has also been described to participate in apoptosis of *Blastocystis* and neural cells of mice. Furthermore, legumain contributes to the maturation process of cathepsin B and L, two other cysteine proteases. Recently, down-regulation of legumain mRNA in macrophages caused by atorvastatin has been reported as well as decreased cathepsin L activity in statin-treated patients with aortic aneurysms. Legumain cleaves peptide bonds carboxyterminally to asparagine, as well as at aspartate residues at pH below 5 and thus acquiring caspase-like properties. The protease is expressed as a 56 kDa proform, which is autoactivated at acidic pH to 47/46 kDa intermediate forms. The intermediate legumain forms are further enzymatically processed to the mature active 36 kDa form. Also, prolegumain has been reported to be secreted as well as being associated with integrins, and can be internalized and subsequently autoactivated. Although the biological and pathological roles of legumain are starting to be elucidated, much is still unknown. Interestingly, increasing the concentration of glucose to the media of human monocytes and murine macrophage- like J774A.1 cells are reported to down-regulate the activity of cathepsin B, D, L and S, thus indicating an interesting regulatory role of glucose on lysosomal enzymes. The overall aim of this study was to investigate effects of simvastatin on glucose metabolism in human myotubes. Skeletal muscles are the major organ for glucose metabolism, and any alteration caused by simvastatin on glucose metabolism in human myotubes could shed light on mechanisms involved in adverse effects and toxicity of statins. Also, the effects of simvastatin on regulation of the cysteine protease legumain were studied in this context. # Materials and Methods ## Materials Dulbecco's modified Eagle's medium (DMEM-Glutamax™, 5.5 mM glucose), foetal bovine serum, Ultroser G, penicillin–streptomycin (P/S), amphotericin B, DAPI, XCell SureLock® Mini, NOVEX Tris-Glycine Native Sample Buffer (2X), NOVEX Tris- Glycine Native Running buffer (10X), NuPAGE Bis-Tris 4–12% gels, NuPAGE MOPS SDS running buffer (20X), NuPAGE LDS sample buffer (4X), Alexa®568 donkey anti-mouse (cat. no. A10037), Alexa®488 donkey anti-goat (cat. no. A11055) and ProLong® Gold antifade reagent with DAPI (cat. no. P36935) were obtained from Life Technologies (Paisley, UK). \[¹⁴C-(U)\]glucose (107.3 GBq/mmol), \[³H\]deoxyglucose (37 MBq/ml) and \[1-¹⁴C\]oleic acid (2 GBq/mmol) were purchased from PerkinElmer NEN® (Boston, MA, USA). Simvastatin was obtained from Toronto Research Chemicals (Ontario, Canada). Insulin Actrapid was from Novo Nordisk (Bagsvaerd, Denmark). Culture plates (6-, 12- and 96-wells) and 25 cm² flasks were obtained from Corning Life-Sciences (Schiphol-Rijk, The Netherlands). OptiPhase Supermix and UniFilter®-96 GF/B were delivered by PerkinElmer (Shelton, CT, USA). CHAPS, DL-dithiotreitol (DTT), trypan blue, triton X-100, mevalonolactone, geranylgeranyl pyrophosphate ammonium salt, farnesyl pyrophosphate ammonium salt, rotenone, oligomycin A, carbonyl cyanide 4-(trifluoromethoxy)-phenylhydrazone (FCCP) and antimycin A were purchased from Sigma-Aldrich (St. Louis, MO, USA). Sodium pyruvate solution, foetal calf serum (FCS) and trypsin-EDTA were purchased from PAA Laboratories GmbH (Pasching, Austria). Nitrocellulose membranes were from Hybond ECL (Amersham Biosciences, Boston, MA, US). Cytoslides (cat. no. 154534), SuperSignal West Dura Extended Duration Substrate and Restore Western Blot Stripping Buffer were purchased from Thermo Fisher Scientific (Rockford, IL, USA). Protein assay reagent, Tween 20, SDS, Precision plus protein standards, goat anti-rabbit IgG HRP-conjugate (cat. no. 170-6515) and goat anti-mouse IgG HRP-conjugate (cat. no. 170-6516) were purchased from BioRad (Copenhagen, Denmark). Goat anti-human legumain (cat. no. AF2199), goat anti-human cathepsin L (cat. no. AF952), goat anti-human cathepsin B (cat. no. AF953) and mouse anti- human arylsulfatase B (cat. no. MAB4415) were purchased from R&D Systems (Abingdon, UK). MitoProfile® Total OXPHOS Human WB Antibody Cocktail (cat. no. ab110411) and rabbit anti-human GLUT1 (cat. no. ab15309) were from Abcam (Cambridge, UK). Rabbit anti-goat IgG HRP-conjugate (cat. no. P0160) was purchased from DAKO (Glostrup, Denmark), whereas mouse anti-human α-tubulin (cat. no. CP06) was obtained from Calbiochem (San Diego, CA, USA). Mouse anti- human GAPDH (cat. no. sc-47724) and mouse anti-human LAMP-2 (sc-18822) were from Santa Cruz (Heidelberg, Germany). Z-Arg-Arg-AMC and Z-Ala-Ala-Asn-AMC were purchased from Bachem (Bubendorf, Switzerland). Non-fat dry milk was from Normilk (Levanger, Norway). Qproteome Cell Compartment Kit was purchased from Qiagen (Hilden, Germany). CellTiter 96® aqueous one solution cell proliferation assay (MTS assay) was obtained from Promega (Madison, Wisconsin, USA). All other chemicals used were standard commercial high-purity quality. ## Ethics Statement The biopsies were obtained with informed written consent and approval by the Regional Committee for Medical and Health Research Ethics (Oslo, Norway). The research performed in this study was approved, as a part of a larger project, by the Regional Committee for Medical and Health Research Ethics (Oslo, Norway). ## Cell Culturing Satellite cells were isolated from the *Musculus obliquus internus abdominis* of healthy human donors with no history of statin treatment. The cells were isolated, cultured, proliferated and differentiated as described elsewhere. Briefly, cells were cultured in wells or flasks at a density of approximately 5000–30000 cells/cm² in medium containing DMEM-Glutamax (5.5 mM glucose), 10% FCS, 50 units/ml penicillin/streptomycin (P/S) and 1.25 µg/ml amphotericin B. This medium was substituted after one day with medium containing DMEM-Glutamax, 2% FCS, 2% Ultroser G, 50 units/ml P/S and 1.25 µg/ml amphotericin B and changed every 2–3 days until 80–90% confluence. Myoblast differentiation to myotubes was then induced by changing medium to DMEM-Glutamax with 2% FCS, 34 pM insulin, 50 units/ml P/S and 1.25 µg/ml amphotericin B. The cells were cultured, proliferated and differentiated in humidified 5% CO₂ atmosphere at 37°C. Incubation with 0–40 µM simvastatin with or without 0.05 or 1 mM mevalonolactone (ML), 3 µM farnesyl pyrophosphate (FPP) or 3 µM geranylgeranyl pyrophosphate (GGPP) in the differentiation medium were introduced on day 5 of differentiation. After 48 h of incubation, cells were either harvested in lysis buffer containing 100 mM sodium citrate, 1 mM disodium-EDTA, 1% n-octyl-β-D-glucopyranoside, pH 5.8 or used for further experiments described later. Cell lysates were freeze-thawed 3 times before analysis by immunoblotting, enzyme activity and total protein measurements. Total protein concentrations were determined by a procedure described elsewhere and standard curves were established using albumin. ## Cysteine Protease Activity Measurements Legumain activity was measured by recording the cleavage of the peptide substrate Z-Ala-Ala-Asn-AMC. Briefly, 20 µl of cell lysate was added to black 96-well microtiter plates. A kinetic measurement based on increase in fluorescence over 10–60 min was performed after addition of 100 µl legumain assay buffer and 50 µl peptide substrate solution (10 µM Z-Ala-Ala-Asn-AMC) described elsewhere. Cathepsin B activity was measured in a similar way except of using the peptide substrate Z-Arg-Arg-AMC. Briefly, 20 µl of cell lysate was added to black 96-well microtiter plates. Cathepsin B assay buffer, and peptide substrate solution (20 µM Z-Arg-Arg-AMC) was added and fluorescence measured. Temperature was kept at 30°C and all measurements were done in triplicate. Linarites of the assays were established by measuring the initial substrate cleavage rates and limiting the substrate consumption to less than 2% during the measurements. Enzyme activity is presented as unit/mg total proteins (µmol/(min·mg)). ## Immunoblotting Samples of cell lysate were prepared for NuPAGE electrophoresis according to the manufacturer’s recommendations (Life Technologies). Briefly, samples were mixed with 0.5 M DTT and NuPAGE LDS sample buffer and run along with 5 µl Precision plus protein standard on NuPAGE 4–12% gels in a container with NuPAGE MOPS SDS running buffer. Blotting was performed using 20% methanol, 25 mM Tris, and 0.2 M glycine, pH 8.3. Nitrocellulose membranes were blocked with 5% non-fat milk in Tris-buffered saline containing 0.05% Tween 20 (TBS-T) for 1–2 h at room temperature, and then incubated overnight at 4°C with goat anti-human legumain (1∶1000), goat anti-human cathepsin B (1∶10000), goat anti-human cathepsin L (1∶5000), mouse anti-human GAPDH (1∶10000), mouse anti-human LAMP-2 (1∶500), mouse anti-human α-tubulin (1∶5000), rabbit anti-human GLUT1 (1∶1000), mouse anti-human total OXPHOS cocktail (1∶500) or mouse anti-human arylsulfatase B (ARSB; 1∶500). Further incubation for one hour was performed with appropriate HRP-conjugate of secondary antibodies. After four ten-minute washes in TBS-T, immunoreactive bands on the membranes were detected by SuperSignal West Dura Extended Duration Substrate. Membranes were reprobed after stripping in Restore Western Blot Stripping Buffer as described by the manufacturer (Thermo Fisher Scientific). Immunoband intensities were analyzed by Image 4.0 (BioRad). ## Confocal Imaging Cells (5×10⁴) were seeded on cytoslides, cultured, differentiated and treated as described above. On day 7 after start of differentiation, cells were fixed with 4% paraformaldehyde in PBS for 10 min on ice and washed twice in PBS before being permeabilized for 5 min with 0.2% Triton X-100. Cells were washed three times with PBS, 0.1% BSA, 0.2% Triton X-100 and 0.05% Tween 20, and then blocked with 10% horse serum for 1 h. Then, immunocytochemical staining was performed using goat anti-human legumain (1∶50) or goat anti-human cathepsin B (1∶100) and mouse anti-human ARSB (1∶50) primary antibodies for 1 h. After three washes, corresponding secondary antibodies were applied (donkey anti-goat; Alexa 488; 1∶250 or donkey anti-mouse; Alexa 568; 1∶500, respectively). The coverslips were mounted in ProLong Gold antifade reagent with DAPI. The cells were observed using a laser-scanning confocal imaging system LSM710 (Carl Zeiss) with equal settings in all experiments. ## ELISA Established ELISA procedure given by the manufacturer was performed to measure the concentrations of legumain (R&D Systems; MAB21992) in conditioned media from myotubes. ## Cell Viability (MTS) Cell viability assays were carried out using the manufacturer’s protocol. Briefly, 5×10⁴ cells were cultured, proliferated, differentiated and treated with simvastatin in quadruplicates in 96-wells culture plates. After 24 h, 20 µl of 3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxy- methoxyphenyl)-2-(4-sulfophenyl)-2H-tetrazolium (MTS reagent) were added to each well and incubated for 2 h before absorbance was measured at 490 nm in a microplate reader, Wallac Victor³ (PerkinElmer). ## Subcellular Fractionation Samples of different subcellular fractions were prepared using the Qproteome Cell Compartment Kit according to the manufacturer’s protocol. Purity of fractions was checked with α-tubulin, LAMP-2, ARSB and Lamin B as cytosolic, membrane, soluble membrane and nuclear markers, respectively. ## Uptake and Oxidation of Glucose Cells were cultured, proliferated and differentiated on 96-well CellBIND® microplates. First, 48 h incubation with simvastatin was started on day 5 of differentiation before analysis with cell-based multiwell assay was performed as described elsewhere. Briefly, medium was removed before addition of \[¹⁴C-(U)\]glucose (37 kBq/ml, 0.2 mM) in Dulbeccòs PBS (DPBS) with 10 mM HEPES, containing either 0.1% DMSO, 0.1 µM rotenone, 1 µg/ml oligomycin A, 0.1 µM antimycin A, or 1 µM FCCP. A 96-well UNIFILTER® microplate presoaked with 20 µl 1 M NaOH was mounted on top of the CellBIND® plate, and the cells were incubated at 37°C and 5% CO₂ for 4 h. The CO₂ trapped in the filter was then counted by liquid scintillation in a MicroBeta™ Trilux scintillation counter (PerkinElmer). The remaining cell-associated radioactivity was also assessed by liquid scintillation, and the formation of CO₂ and cell-associated radioactivity was considered as total glucose uptake while the formation of CO₂ was considered as glucose oxidation. ## Uptake of Deoxyglucose The cells were cultured, proliferated and differentiated on 12-well plates. First, 5 days after onset of differentiation, treatment with 5 µM simvastatin with or without 1 mM ML were added. After 48 h of incubation, the cells were washed and incubated for 1 h with 140 mM NaCl, 20 mM HEPES, 5 mM KCl, 2.5 mM MgSO₄ and 1 mM CaCl₂, pH 7.4, before \[³H\]deoxyglucose (37 kBq/ml, 10 µM) was added and incubated for 15 min. Then cells were washed 3 times with PBS and harvested in 250 µl 0.1 M NaOH. The lysates were counted by liquid scintillation. ## Statistics The data are represented as mean ± SEM. Student t-test or student paired t-test were performed when appropriate, and statistical significance was considered at p\<0.05. All experiments were performed on cells from at least three donors and at least triplicate measurements. # Results ## Simvastatin Reduced Glucose Uptake and Oxidation in Human Myotubes Previous reports have shown that simvastatin reduces glucose metabolism both in cancer cells and adipocytes. In our study using differentiated human myotubes, simvastatin significantly reduced uptake of \[¹⁴C\]glucose in a dose- dependent manner with an IC₅₀ value of approximately 8 µM. The myotubes seemed to be more sensitive to reduced glucose uptake by simvastatin than the embryonic kidney cell line HEK293. Whereas 5 µM simvastatin significantly decreased glucose uptake in the myotubes, no effect was observed in the HEK293 cells. Reduced glucose uptake caused by 5 µM simvastatin was confirmed using \[³H\]deoxyglucose giving approximately 45% less uptake. Treatment with a combination of simvastatin and mevalonolactone (ML; 1 mM) totally prevented the decreased deoxyglucose uptake. Since the experiments were done in absence of insulin and GLUT1 is the predominant glucose transporter in human myotubes, the GLUT1 expression after exposure to simvastatin was studied. The observed effects of simvastatin were not due to differences in GLUT1 expression. To verify that the observed effects of simvastatin were not due to cell death of myotubes, cell viability, total proteins concentrations and caspase-3 expression were studied. There were no differences in cell viability, total protein concentrations or immunoband of active caspase-3 after simvastatin treatment. Knowing that the myotubes were fully viable at the simvastatin concentrations used, the observed reduction in glucose uptake could be an indirect effect of impaired oxidation of glucose. Therefore, an uncoupler of oxidative phosphorylation (FCCP) was introduced and shown to increase oxidation of glucose by approximately 3-fold. This reflected a high reserve capacity for glucose oxidation in myotubes, calculated as the difference between oxidation in the presence or absence of FCCP. The reserve capacity for glucose oxidation after simvastatin (5 µM) treatment was reduced by approximately 40%. To further study the mechanism of simvastatin (5 µM) on oxidative phosphorylation, different inhibitors of the respiratory chain and ATP formation were introduced. Rotenone, oligomycin A, and antimycin A were used to inhibit complex I, ATP synthase and complex III, respectively. No significant effects on glucose oxidation were observed in myotubes treated with any of these agents with or without simvastatin, whereas treatment with FCCP reflected the observation already described above. Finally, the effects of simvastatin on the expression of complex I, complex II subunit 30 kDa, complex III core 2, complex IV and ATPase- α-subunit were studied. No significant differences were observed, but there was a tendency of reduced expression of complex I-IV of the respiratory chain by simvastatin. ## Reduced Legumain Activity and Expression in Simvastatin-treated Myotubes The glucose concentration in the cell culture media has been reported to regulate the activity of cathepsin B, D, L and S in human monocytes and murine macrophage-like J774A.1 cells. Also, atorvastatin has been shown to decrease legumain mRNA in monocytes. Since glucose metabolism was decreased in human myotubes after exposure to simvastatin, this interesting regulatory role of glucose on lysosomal enzymes led us to investigate whether legumain was affected by simvastatin in human myotubes. Initially, expression and activity of legumain were studied during myotube differentiation and compared to cathepsin B, which is reported to participate in myotube differentiation. Undifferentiated myoblasts (differentiation day 0) showed both legumain and cathepsin B activity at a level of 0.9 (±0.2) and 5.8 (±1.1) µUnit/mg, respectively. Legumain activity was significantly increased both at differentiation day 2 and 5 compared to day 0. Increased activity was due to increased expression of the mature active form (36 kDa) as reflected by immunoblotting. Cathepsin B activity and expression of the active two-chain form (23 kDa) also showed increasing tendencies throughout myotube differentiation (B and D). Treatment of myotubes on day 5 of differentiation with increasing concentrations of simvastatin for 48 h showed reduced legumain activity in a dose-dependent manner with an IC₅₀ value of about 25 µM. There was also a tendency of reduced cathepsin B activity caused by increasing simvastatin concentration (data not shown). To study whether inhibition of the HMG-CoA reductase was involved in the reduced legumain activity observed by simvastatin, intermediates of the mevalonate pathway including mevalonolactone (ML), geranylgeranyl pyrophosphate (GGPP) and farnesyl pyrophosphate (FPP) were introduced. The effect of simvastatin (30 µM) on legumain activity was partly prevented by ML (1 mM) or FPP (3 µM;). Also, the legumain activity measurement was reflected by the expression of the 36 kDa immunoband, representing the mature active form, which was significantly reduced after treatment with simvastatin alone ( alone). In addition, a concomitant accumulation of the 56 kDa immunoband was seen, reflecting reduced prolegumain processing. No significant changes were observed by addition of ML, GGPP or FPP. Furthermore, fully differentiated myotubes secreted legumain to the conditioned media at a rate of 0.02 (±0.005) pg/cell/day, and the secretion was not affected by treatment with simvastatin (30 µM). To study whether simvastatin could alter the intracellular distribution of lysosomal cysteine proteases, subcellular compartments of myotubes were isolated and analyzed for legumain, cathepsin B and L. Immunoblots showed that both legumain, cathepsin B and L in untreated myotubes were located only in the membrane compartment, comprising cytoplasmic organelles. After simvastatin treatment, there was no altered subcellular localization of either legumain, cathepsin B or L, as no immunbands were detected in the cytosolic or nuclear compartments (data not shown). Subcellular compartment purities were verified by LAMP-2 (membranes), α-tubulin (cytosol) and lamin B (nucleus; not shown), respectively. Simvastatin (30 µM) reduced the expression level of the mature form of legumain (36 kDa) as well as increased the level of prolegumain (56 kDa) in the membrane fraction, confirming the observation in whole myotube lysates. Also, subcellular presence of legumain and cathepsin B was confirmed by confocal imaging. Legumain seemed to be co-localized in lysosomes with arylsulfatase B (ARSB, a soluble lysosomal enzyme) in untreated myotubes (yellow;) and treatment with simvastatin did not seem to alter legumain distribution. Cathepsin B also seemed to be vesicular although not distinctly co-localized with ARSB, and simvastatin caused a more diffuse staining. # Discussion In this study we observed decreased uptake and oxidation of glucose in human myotubes caused by treatment with simvastatin. Also, legumain (a cysteine protease) was for the first time characterized in human myotubes and decreased legumain activity and expression was observed by simvastatin. The reduction in legumain activity was caused by decreased processing of the 56 kDa prolegumain due to inhibition of the HMG-CoA reductase by simvastatin. These effects of simvastatin on differentiated human myotubes may contribute to the understanding of the pharmacology and toxicology of statins. The glucose metabolism in human myotubes decreased upon simvastatin treatment, and it is tempting to speculate if this could contribute to hyperglycemia, since some patients taking statins develop decreased insulin sensitivity, insulin resistance and glucose intolerance. Furthermore, the observed decrease in glucose uptake was prevented by concomitant addition of mevalonolactone, suggesting that the mechanism was due to direct inhibition of the HMG-CoA reductase. Although no difference in expression of GLUT1 was observed after simvastatin treatment in this study, regulation of activity or translocation of the GLUT1 transporter or other GLUT transporters could nevertheless account for the effects observed but was not investigated further. Previously, simvastatin has been reported to impair complex I and II in the respiratory chain resulting in ROS accumulation in primary human myotubes established from satellite cells isolated from another source *(Musculus vastus lateralis)*. In our study we detected no differences in CO₂-production by simvastatin in presence of rotenone (complex I inhibitor), oligomycin A (ATP-synthase inhibitor) or antimycin A (complex III inhibitor). Deviant observations could be due to differences in both origin of myotubes and methods used. Surprisingly, simvastatin reduced FCCP-induced glucose oxidation and oxidative reserve capacity, indicating some effects of simvastatin on cell respiration and mitochondrial function. Therefore, the expressions of proteins in the respiratory chain (complex I-IV and ATP synthase) were studied after treatment with simvastatin. Although no statistically significant changes in the expressions of the analyzed complexes were observed, the expressions tended to decrease and could account for some of the effects. There may be a link between the two effects observed of simvastatin, impaired glucose metabolism and prolegumain processing, as reduced supply of glucose could cause ATP-depletion in the cell. Decreased ATP levels due to mitochondrial dysfunction have previously been reported in skeletal muscle cells from patients with type 2 diabetes and in rat L6 GLUT4myc myotubes acquiring impaired glucose metabolism. We have recently shown that bafilomycin A1 (a strong inhibitor of the vacuolar type H⁺-ATPase) also reduced the activity of legumain. Since the lysosomal H⁺-ATPase needs ATP to accomplish acidic lysosomal pH, our results may indicate that simvastatin could reduce the H⁺-ATPase activity due to possible ATP-depletion resulting in increased lysosomal pH and thus reduced prolegumain processing. Extracellular glucose is reported to regulate the activity of the cysteine proteases cathepsin B, D, L and S in human monocytes and murine macrophage-like J774A.1 cells. This, together with our observations that simvastatin decreased glucose metabolism in myotubes and the reported down-regulation of legumain mRNA by atorvastatin observed in monocytes, made us study the regulatory effects of simvastatin on legumain. We observed a reduced activity of legumain caused by simvastatin and speculated if this could be due to a translocation from the lysosomes to the cytosol since such translocation of legumain has been reported in apoptosis. As a HMG-CoA reductase inhibitor, simvastatin blocks the rate limiting step in the cholesterol synthesis. Cholesterol is a major component of lipid bilayers and cholesterol removal from lysosomal membranes increases the permeability of especially ions and protons, resulting in osmotic imbalance, destabilization and potential leakage of proteins. No lysosomal translocation of legumain was observed using 30 µM simvastatin and could thus not explain the reduced legumain activity observed since the activity is expected to be abolished at neutral cytosolic pH. Cholesterol depletion has also been reported to lead to improper subcellular compartmentalization of phosphoinositides, which are essential for trafficking of intracellular vesicles. The delocalization of phosphatinositoles is partly due to regulation by cholesterol of the activity of various phosphatinositol kinases at distinct intracellular compartments e.g. type II phosphatidylinositol 4-kinase IIα at the Golgi membrane (PI4KIIα). Prolegumain needs acidic pH to autoactivate. Therefore, it is possible that some of the effect of simvastatin on inhibition of legumain activity could be due to dysfunctional transport of prolegumain from the Golgi to vesicles with acidic pH, like the late endosomes/lysosomes. Also, depletion of isoprenoids, mainly the geranylgeranyl pyrophosphate (GGPP) and farnesyl pyrophosphate (FPP), have been suggested to be critical for statin-induced myopathy due to their role in prenylation of small GTPases. Inhibition of prenylation will lead to improper intracellular trafficking since functionally small GTPases such as Ras and Rab are essential for targeting, tethering, uncoating and formation of intracellular vesicles. GGPP has been shown to prevent the decreased prenylation of Rab1 GTPase by fluvastatin. Here, down-regulation of legumain activity by simvastatin was significantly prevented by FPP or mevalonolactone (ML), but not by GGPP, indicating that inhibition of prenylation of Rab1 GTPase may not account for the effects observed. The activity of cysteine proteases are strictly controlled within a cell, and uncontrolled legumain activity is associated with serious diseases like atherosclerotic plaque instability and malignant cancer,. A pro-survival role of legumain has been reported in the parasite *Blastocystis* of which inhibition of legumain activity is associated with increased programmed cell death. In this study legumain activity and expression were decreased by simvastatin, but not cell viability, cell total protein content or caspase-3 expression indicating that proteases other than legumain are involved in statin-induced cell death. Down-regulation of legumain could still contribute to a general distortion of protease/kinase activity possibly leading to toxicity, but this needs to be further investigated. Simvastatin was introduced 5 days into myotube differentiation to mimic an *in vivo* condition since reports have demonstrated that toxic effects of statins mainly affect differentiated myotubes and not myoblasts. Simvastatin was used in our study since this statin has the highest frequency of reported myotoxic effects, and the lactone forms of statins are more potent in causing myotoxicity. The concentration of simvastatin used in present study is in the micromolar range, but the achieved plasma concentration after administration of a clinical simvastatin dose is in the nanomolar range. Still, the statin concentrations used in this study are in range with other published *in vitro* experiments. Since the exposure of simvastatin *in vitro* must be for shorter periods of time (days) compared to long term therapeutic *in vivo* use, and the primary cell culture has limited life span, it can be argued that higher statin concentrations are needed to detect cellular responses. Nevertheless, in our study and as mentioned above, the simvastatin concentrations used in this study (5–40 µM) caused no significantly effects on cell viability, cell total protein contents or expression of active caspase-3. In contrast, it has previously been reported reduced myotube viability by 5 µM simvastatin. In conclusion, this study shows that simvastatin reduced both glucose metabolism and legumain activity in myotubes. Both phenomenon are of importance to fully understand the pharmacology and toxicology of statin treatment and needs to be further investigated. # Supporting Information The authors would like to thank Hilde Nilsen for excellent technical assistance. We will also thank Ellen Skarpen for valuable advices during the confocal imaging. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: R. Smith R. Solberg ACR GHT HTJ. Performed the experiments: R. Smith LLJ ALV. Analyzed the data: R. Smith LLJ ALV. Contributed reagents/materials/analysis tools: R. Smith R. Solberg ACR GHT HTJ. Wrote the paper: R. Smith. Read and approved the paper: R. Smith R. Solberg LLJ ALV ACR GHT HTJ. [^3]: Current address: Department of Pharmaceutical Biosciences, School of Pharmacy, University of Oslo, Oslo, Norway

# Introduction Research in the last three decades has established that the brain actively deploys lexical and contextual information to facilitate word processing during language comprehension. For example, previous work has shown that neural responses to a word are decreased when the word is presented a second time or preceded by a related item in a word list. This phenomenon is known as lexical priming and has been attributed to eased access to a word in long-term semantic memory after its level of activation has been boosted by the first presentation of that word or a related word. Lexical priming has been observed across different tasks, and it occurs even when the prime is masked from consciousness or when it is embedded in a sentence or discourse. As such, the mechanism underlying lexical priming is thought to be bottom-up, highly automatized and reflective of the organization of long-term semantic memory. On the other hand, neural responses to words within a sentence context are also strongly modulated by the fit between these words and their context, such that words that are more predictable in context are processed more easily. The effect of predictability, as in the case of lexical priming, also results in facilitated access to words in long-term semantic memory. However, unlike lexical priming, this facilitation is thought to result from the semantic interpretation of the preceding sentence or discourse. Previous research has demonstrated that comprehenders incrementally compute a semantic interpretation of the sentence context and integrate it with their world knowledge to anticipate and pre-activate likely upcoming words on the fly. Further, previous research has also shown that the effect of prediction is not reducible to lexical semantic priming. As such, sentence context is thought to facilitate lexical semantic processing through a top-down mechanism that, in contrast with lexical priming, relies on the combinatorial computation of contextual information and is sensitive to task demands and subject to strategic control. However, less is known about how the mechanisms that compute lexical and sentence-level information work in tandem to facilitate lexical semantic processing in comprehension. Studying their combined effect is important because discourses in natural language are structured around semantically coherent structures or topics, and there are systematic topic-to-word and word-to-word relationships. For example, if a text or passage has “finances” as its topic, the probability of occurrence of a word like “bank” will be high, and the appearance of this word might in turn make it more likely that words like “federal” and “reserve” will also appear in the discourse. Therefore, examining how lexical and contextual information combine to facilitate the processing of words in sentences should lead to a better understanding of how these sources of information are used in everyday language situations, where both are expected to jointly aid in the comprehension of discourse. In the current study, we address this issue by examining the interaction of word repetition and word predictability. In particular, we ask: how does recent exposure to a word interact with contextual information during comprehension? We consider three possible ways in which previous exposure to a word and contextual information can work together during reading comprehension. One possibility is that word processing in sentences is primarily modulated by the information provided by the sentence context, such that repetition only exerts an influence when contextual information is limited and does not strongly predict a word. For example, it has been previously observed that word-level variables such as concreteness and lexical frequency have a weaker impact on neural responses when words are embedded in meaningful sentence contexts (see for discussion). Relatedly, semantic priming effects are larger for words that appear in word lists or in less constraining contexts than in very constraining sentence contexts (see for review). Alternatively, repetition might affect word processing by strengthening predictions that are already licensed by contextual information. Although we are not aware of a proposal of this nature in the domain of word recognition, analogous proposals have been made in the domains of speech perception and word learning. For example, in speech perception it has been shown that acoustic information is used more effectively if listeners are given contextual information about the type of stimuli they are tested on. Similarly, infants’ success in using phonetic information to discriminate novel words (e.g. “bin” vs. “din”) in a word recognition task greatly improves when the words’ referential status is provided (e.g., by being paired with an object), as compared with when the words are presented in isolation. Both these examples suggest that low-level perceptual information might be more useful when deployed together with rich contextual information. Analogously, repetition and sentence predictability may facilitate lexical semantic activation in a supra-additive fashion, with the effect of repetition being larger for words that are more predictable in a given sentence context. We contrast these alternatives with the possibility that word repetition and context predictability have independent and additive effects on lexical semantic activation. According to this view, recent exposure to a word always results in facilitated processing of that word in a sentence context, regardless of its predictability. This view predicts that word repetition and context predictability should have an additive influence on lexical semantic activation even if the word previously appeared outside of the sentence context. One class of computational models that could predict such a pattern are cache-based natural language models, in which information about which words have appeared recently in the discourse is maintained in a running cache and combined with information about the current sentence context to yield an estimate of word probability. A recent example of independent contributions of lexical and sentential context in comprehension comes from work suggesting that the frequencies of the multiple entries of category-ambiguous words are computed independently of the context they appear in. We examine these alternatives using the N400 component as a measure of lexical semantic activation. The N400 is a broad negative deflection of the event- related potential (ERP) that starts 200–300 ms after a word has been presented and peaks after approximately 400 ms. Although a precise interpretation of the processes indexed by this component is still under debate (see for review) we adopt here the proposal that the N400 reflects activation of the semantic features of the long-term memory representations that are associated with a lexical item. According to this view, the N400 response to a word indexes how easy or hard it is to retrieve this word from long-term semantic memory. Correspondingly, words that have been encountered recently or that are more expected are associated with a reduction in the N400 amplitude, with localization evidence suggesting that this differential activity is generated in regions of temporal cortex involved in representing lexical and conceptual information. Therefore, the N400 provides a good implicit measure for examining the mechanism by which lexical repetition and contextual information combine to impact word activation during comprehension. ## Previous Work Many studies have reported N400 reductions due to word repetition (e.g.,) and predictability. A few studies have examined the joint influence of these factors, and they have mainly concluded that repetition effects are context- dependent. One group of studies has focused on cases where repetition violates sentence-level constraints, in examining how repetition of proper names is modulated by co-referential constraints. These studies have shown that a repeated name like “Daniel” in the sentence “*At the office Daniel moved the cabinet because <u>Daniel</u>* …” elicits a larger N400 than in the control sentence “*At the office Daniel and Amanda moved the cabinet because <u>Daniel</u>* …”. The larger N400 for the repeated name in first sentence has been attributed to a violation of coreference constraints, under which when a referent is prominent in the discourse (as in the first sentence) it should be referred to using a pronoun instead of a proper name. The N400 effect elicited by repeated names that infelicitously refer to a prominent antecedent has been called the *repeated name penalty*. This finding has been taken to argue that repetition effects are context-dependent, and that they can be overriden by higher level processes. Similarly, a previous study using a memory paradigm found that when both a target word and its context are repeated, word repetition reduces the N400 to incongruous words only, consistent with repetition effects being dependent on context. In this study, participants read sentences in which the final word was either highly predictable in a congruous sentence or highly unpredictable in an incongruous sentence. Predictability was assessed offline using the cloze procedure, in which a separate group of participants were asked to provide completions to sentence fragments ; in this study the sentence-final words had a cloze probability greater than 0.75 in the predictable conditions and a cloze probability of 0 in the unpredictable conditions. Participants were asked to memorize the sentence-final words (first presentation) and were then given the sentence frames and asked to recall the missing final words (recall test). Afterwards, they read the same set of sentences for a second time (second presentation). Comparing ERPs across the first and second presentations, the authors found that repetition led to a reduced N400 response for incongruous words but not for congruous ones, yielding a statistical interaction between predictability and repetition. They suggested that lexical repetition does not affect normal sentence processing, but it can facilitate the processing of incongruous words. However, one potential concern about these previous studies is that they manipulated word repetition in such a way that the second occurrence of the word was infelicitous, either because it violated a pragmatic constraint, or because the entire sentence context was repeated in a rather artificial way. In contrast, word repetition in typical language comprehension presumably takes place in congruous sentences, and while repetition of words is fairly common, repetition of whole sentences is not. More importantly, these studies manipulated lexical repetition and contextual variables in a non-orthogonal manner, such that the effects of the manipulated variables could not be clearly dissociated. Specifically, in the repeated name penalty studies, repetition rendered the sentences infelicitous, and therefore the observed N400 patterns may have reflected both the effect of repetition and of the incongruity of the target words in context. In the study by Besson and colleagues, since the target words appeared in identical sentences across both presentations, the repetition of the contexts likely changed the predictability of the target words on second presentation. Thus, unpredictable words might have become much more predictable when repeated in the same context. For example, even though the word ‘socks’ is unexpected in a sentence like ‘*I like my coffee with cream and…*’, it is likely to be more expected when the same sentence appears for the second time. Therefore, it is unclear whether the observed N400 reduction was due to lexical repetition of the target, to its increased predictability in the sentence context, or both. In order to dissociate between the effects of word and sentence repetition, in a later study Besson and Kutas examined the effects of repetition of low cloze probability words (cloze probability \<28%) and varied whether they appeared in the same or in different sentence contexts across presentations. They showed that when the sentence context was changed, word repetition did not modulate the N400. As a result, they concluded that word repetition led to a reduced N400 response only when contexts are also repeated. However, since the predictability of the target words was not manipulated in the study, these results cannot address our question about the combined effects of word repetition and predictability on lexical semantic activation. Furthermore, the finding that lexical repetition leads to an N400 reduction only when the word appears in identical contexts across presentations seems at odds with previous findings in the single-word literature, where words repeated in word lists are consistently associated with reduced N400 responses,. Although target words are preceded by different words across the first and second presentations, repetition consistently leads to an N400 reduction, suggesting that repetition can facilitate lexical semantic activation even when it occurs across different contexts. In summary, although previous work has established that lexical repetition and contextual information can impact lexical semantic activation during comprehension, questions about how they work in tandem remain unanswered as there are discrepancies across studies and paradigms. Below we present a new paradigm that addresses some of these concerns in order to examine the interaction between lexical and contextual information during sentence processing. ## The Present Study The current study investigated whether recent exposure to a word interacts with contextual information during comprehension by examining the joint effect of word repetition and predictability on the N400 amplitude. We devised a novel paradigm in which word repetition always occurred in non-repeated contexts. Since word repetition and predictability can be manipulated orthogonally in this paradigm, we can avoid some of the ambiguities in interpretation encountered by previous studies. As illustrated in, the paradigm consisted of a familiarization phase, a reading comprehension phase and a recognition test phase. In the familiarization phase participants were asked to study a set of words for a later recognition memory test. Since differences at initial memory encoding have been shown to affect the amplitude of the N400 during later recognition, we presented the words in isolation rather than in sentence contexts to minimize systematic encoding differences. In between the familiarization and the recognition memory test phase, participants read a list of sentences for comprehension while their electroencephalogram (EEG) was recorded. We manipulated whether the target words had been studied in the familiarization phase or not (old vs. new) as well as their predictability in the sentence context (expected vs. unexpected), which was operationalized as their cloze probability. In order to avoid floor effects on the N400 amplitude (cfr.), we used expected target words of intermediate cloze probability (7.9–39.5%). Representative sample items are shown in (1) to (3) with the target word underlined with expected targets presented to the left of unexpected targets: 1. Vivian wanted to leave the party because she couldn’t stand the <u>noise/drinks</u> and the rowdy crowd. 2. Brian looked all over the house for his missing <u>keys/watch</u> before leaving for work. 3. The doctor realized that the patient would need a <u>transplant/miracle</u> in order to survive. We divided the experimental session into 16 short blocks, each containing 8 words for familiarization, 8 sentences for comprehension, and 8 words for the recognition memory test. This was motivated by two main considerations: we reasoned that while a long delay might substantially weaken the effects of repetition, a delay that is too short (i.e., one with too few trials per block) might make it apparent to the participants that half of the studied words would reappear in a sentence within the same block. This might have encouraged participants to predict that the studied words would reappear during sentence comprehension, rendering our experimental manipulations non-orthogonal, as repeated words would also have been more predictable. For these reasons, we piloted the experiment with different block sizes. We decided to pursue 8-trial blocks since they were short enough for participants to perform the memory recognition task well above chance-level without making apparent whether and/or when a studied word would reappear in the sentences within the same block. Further, since multiple-trial blocks afford variable time lags between the first and second presentations of a word, they allow us to better model the fact that word repetition may occur over any number of sentences in written texts or speech in real-life settings. Different patterns of results are predicted by each of the three hypotheses discussed in the Introduction about how lexical repetition and contextual information combine to impact lexical semantic activation. If recent exposure to a word and predictability affect lexical semantic activation independently, they should have an additive effect on the N400 amplitude. Alternatively, if lexical repetition facilitates lexical semantic activation only when a word is not predictable in a given context, then the effect of repetition on the N400 should be larger for unexpected words than expected words. This would be consistent with observations in previous studies on semantic priming. Lastly, if word repetition can strengthen or reinforce the predictions that arise from the sentence context, then the effect of repetition on the N400 should be larger for expected words than unexpected words. # Methods ## Participants Twenty-four students (12 female, mean age = 21.1 years, range = 18–28 years) from the University of Maryland, College Park participated in the current study. Informed consent was obtained in all cases. Participants were right-handed, native English speakers with normal or corrected-to-normal vision. ## Ethics Statement This study was conducted with the approval of the University of Maryland, College Park Institutional Review Board (UMCP IRB). All participants gave written consent and were paid 20 USD for their participation in accordance with the policies of UMCP IRB. ## Materials Stimuli for reading comprehension consisted of 128 sentence item sets. We orthogonally manipulated the repetition status (old vs. new) and the predictability (expected vs. unexpected) of the target word in the sentences. A target word was considered ‘old’ if it was presented in the preceding familiarization phase and ‘new’ otherwise. Cloze probabilities for the experimental sentences were obtained in a norming study. Participants were 114 student volunteers at the University of Maryland, College Park. They were asked to provide the best continuations for 220 sentence frames. A total of 180 sentence fragments for which the maximum cloze probability was below 40% were selected to form the sentences for the EEG study. Expected target words had an average of 22.5% cloze probability (range = 7.9–39.5%); unexpected plausible target words were selected from words that were provided exactly once (0.9% cloze probability) in the norming study. Thus, all experimental sentences were semantically congruous. The sentences were extended beyond the target word in order to avoid wrap-up effects. The stimuli for the familiarization phase consisted of 128 words. Half of these words were presented as target words in sentences for reading comprehension and the other half were words selected to match the lexical frequency of the target words (average log frequency of targets: 3.20; fillers: 3.16;). For the recognition task the stimuli also consisted of 128 words, half of which were presented in the familiarization phase, while the other half were new words matched in average frequency. Thus, only half of the 64 studied words appeared as targets in sentences for reading comprehension. The sentence items and the words for the familiarization and recognition tasks were distributed in four presentation lists using a Latin square design. Each list contained 128 words for familiarization and recognition respectively, along with 128 sentences (32 per condition), each paired with a corresponding Yes/No comprehension question. The overall ratio of Yes/No target response was 1∶1 in each presentation list. Each list was presented to six participants. The materials were presented in 16 blocks, each containing 8 words for familiarization, 8 sentences for reading comprehension and 8 words for recognition. The order of blocks and the materials within each phase in each block were pseudorandomized across participants. ## Procedure As illustrated in, participants were instructed to memorize the words presented during the familiarization phase for a later recognition task. In between each familiarization and recognition phase they were asked to read sentences attentively and to answer comprehension questions about those sentences. In the familiarization phase words were presented in two sequences of four, and each sequence was followed by a screen on which all four words were presented together. In each sequence words were presented individually at the center of the screen for 600 ms, followed by 400 ms of blank screen. At the end of each sequence the four words reappeared on the screen together. Participants were told to press a button to proceed to the second sequence (or to the next phase if they had seen both sequences) when they had memorized the words. In the sentence comprehension phase, sentences were presented one word at a time at the center of the screen. Each sentence was preceded by a fixation cross that appeared for 500 ms. Each word appeared on the screen for 300 ms, followed by 230 ms of blank screen. The last word of each sentence was marked with a period, followed by a comprehension question 1000 ms later. Participants were instructed to avoid eye blinks and movements during the presentation of the sentences and to answer the comprehension questions by pressing one of two buttons. In the recognition phase, each trial consisted of a word presented at the center of the screen. Participants indicated whether the word had been presented in the familiarization phase of that block by pressing one of two buttons. The next trial began automatically after they had responded. Prior to the experimental session, participants completed a practice block with 8 words for familiarization, 4 sentences for reading comprehension and 8 words for recognition. The experimental session was divided into 16 blocks, with short pauses in between. Including set-up time, an experimental session lasted between 1.5 and 2 hours. ## EEG Recording EEG was recorded continuously from 29 AgCl electrodes mounted in an electrode cap (Electrocap International): midline: Fz, FCz, Cz, CPz, Pz, Oz; lateral: FP1, F3/4, F7/8, FC3/4, FT7/8, C3/4, T7/8, CP3/4, TP7/8, P4/5, P7/8, and O1/2. Recordings were referenced online to the left mastoid and re-referenced offline to the average of the left and right mastoids. The electro-oculogram (EOG) was recorded at four electrode sites; vertical EOG was recorded from electrodes placed above and below the left eye and the horizontal EOG was recorded from electrodes situated at the outer canthus of each eye. Electrode impedances were kept below 5 kΩ. The EEG and EOG recordings were amplified and digitized online at 1 kHz with a bandpass filter of 0.15–100 Hz. ## ERP Data Analysis All trials were evaluated individually for EOG or other artifacts. Trials contaminated by artifacts were excluded from the averaging procedure. This affected 9.5% of experimental trials. A digital 40 Hz low-pass filter was used on all data to reduce high-frequency noise. Event-related potentials were computed separately for each participant and each condition for the 1000 ms window after the onset of the target word relative to a 100 ms pre-stimulus baseline. Analyses focused on 18 electrodes that could be evenly distributed across the topographic factors of interest: F3, FZ, F4, FC3, FCZ, FC4, C3, CZ, C4, CP3, CPZ, CP4, P3, PZ, P4, O1, OZ, and O2. Statistical analyses on average voltage amplitudes were conducted in R separately for two time windows selected based on existing literature on the N400 component and visual inspection: 300–400 ms for the N400, and 600–800 ms for later differences. We conducted Type II SS omnibus repeated measures ANOVAs that fully crossed repetition (old vs. new) and predictability (expected vs. unexpected) with anteriority (anterior vs. central vs. posterior) and laterality (left vs. midline vs. right). Electrodes were distributed across the topographic factors as follows: left-anterior: F3, FC3; midline-anterior: FZ, FCZ; right-anterior: F4, FC4; left-central: C3, CP3; midline-central: CZ, CPZ; right-central: C4, CP4; left-posterior: P3, O1; midline-posterior: PZ, OZ; right-posterior: P4, O2. Univariate *F*-tests with more than one degree of freedom in the numerator were adjusted by means of the Greenhouse-Geisser correction. Further, in order to examine the potential interaction (or the lack thereof) of the effects of repetition and predictability on the N400, we compared two linear mixed-effects models, one with and one without an interaction term between predictability and repetition. We asked whether the model with an interaction term provided a better fit for the N400 data as compared to a model without it. Using the *lme4* package we fitted two linear mixed-effects models to the 300–400 ms time-window averages in the midline-posterior region. Both models had by-subject and by-item random intercepts, but the simpler model only had repetition and predictability as fixed effects while the more complex model included an additional repetition-by-predictability interaction term. We then conducted a likelihood ratio test to determine if the more complex model provided a better fit to the data. ## Behavioral Data Analysis Participants’ performance on the sentence comprehension and memory recognition tasks was measured. Comprehension accuracy was analyzed using mixed logit models with repetition, predictability and their interaction as fixed effects and by- item and by-subject random intercepts. D-prime (d’) scores were computed to examine participants’ performance on the recognition task. The log-linear correction method described in was used to avoid the appearance of non-finite values in the case of extreme false alarm or hit rates. We conducted a one- sample Wilcoxon test to examine if their average d’ score was above chance. # Results ## Behavioral Results Participants performed well on both the sentence comprehension questions and the recognition memory task. They answered the comprehension questions with a mean accuracy of 90.8% (old-expected: 91%; old-unexpected: 90.2%; new-expected: 92.6%; new-unexpected: 89.3%). Mixed logit models revealed that comprehension accuracy was higher for sentences in the expected conditions than in the unexpected conditions (expected = 91.8%; unexpected = 89.8%; β = −0.15, *p*(Wald) \<.05) but it was not impacted by the repetition status of the target word. The average d’ score on the memory test was 1.83 (SD = 0.82), which was significantly above chance (*p*\<0.001). Overall, the behavioral data show that participants performed adequately on both the comprehension and the memory components of the current task. ## Event-related Potentials (ERPs) during Reading Comprehension shows the grand average ERPs at PZ to target words in all four conditions and the topographic distribution of the repetition effect (new minus old) separately for the unexpected and expected conditions in the 300–400 ms time interval. and show the grand average ERPs across all scalp sites in the expected conditions (old vs. new) and unexpected (old vs. new) conditions respectively. Visual inspection indicates that both experimental factors had a clear effect on the N400: the amplitude of the N400 was reduced for old relative to new target words, and it was also reduced for expected relative to unexpected target words. shows the results of the statistical analyses in both time-windows. We report statistics for significant main effects and interactions below. Consistent with these observations, omnibus repeated measures ANOVA in the 300–400 ms interval revealed significant main effects of both repetition and predictability. The main effect of repetition was driven by ERPs being less negative in the old than in the new conditions (*F*(1,23)  = 14.94, *p*\<.01). The main effect of predictability was driven by ERPs being less negative in the expected than in the unexpected conditions (*F*(1,23)  = 11.51, *p*\<.01). Both effects were broadly distributed across the scalp, as shown by the lack of significant interactions with topographic factors. Crucially, no significant interaction between these two factors was obtained (all *F*s\<2). This additive pattern is displayed in, which shows the average ERP amplitude in the 300–400 ms interval in the midline posterior region. The main effect of predictability and repetition and the absence of an interaction between them also held in two alternative time-windows that have been previously used to assess N400 effects (200–400 ms and 300–500 ms). In order to further examine the additivity of the effects of predictability and repetition, two linear mixed-effects models were fitted to the N400 data in the midline posterior region. Both models had predictability and repetition as fixed effects and by-subject and by-item random intercepts; one of them included an additional interaction term between predictability and repetition. A likelihood ratio test revealed that the more complex model did not provide a significantly better fit of the data than the simpler model (*X²*₍₁₎  = .21; *p* = .64), and thus the removal of the interaction term from the model was statistically justified. In fact, the near-zero *X²* value shows that including an interaction term in the model hardly improves its fit of the data at all. Even though both analyses used a null hypothesis significance testing approach and thus neither allowed us to confirm the null hypothesis, the results from both analyses are consistent in showing that repetition and predictability modulated the size of the N400 in an additive fashion. In the 600–800 ms interval, omnibus repeated measures ANOVA revealed a marginally significant main effect of repetition (*F*(1,23)  = 3.35, *p* = .08) and a marginal repetition × anteriority interaction (*F*(2,46)  = 3.2, *p* = .07). These effects were not followed up further as they failed to reach statistical significance and they were not predicted by any of the hypotheses examined in the current study. # Discussion The aim of the present study was to investigate how recent exposure to a word interacts with the processing of contextual information during sentence comprehension. We devised a paradigm that allowed us to orthogonally manipulate a word’s repetition status and its predictability in a sentence context. In line with previous studies that have examined the effects of lexical repetition and predictability separately,, both factors led to a reduction in the N400 amplitude in the present study. In addition, we show evidence that their effects are additive, such that the N400 response to a recently encountered word is reduced by a similar amount regardless of the word’s predictability in context. Our observation that word repetition reduces the N400 response to both expected and unexpected words differs from the results of Besson and colleagues, who found that repetition reduced the N400 response to unexpected words only. We attribute this discrepancy to two primary differences in the experimental paradigm and materials used across studies. First, while all of the expected words in Besson et al.’s study had high cloze probability, in the current study only expected words of intermediate cloze probability were used. Since highly predictable words elicit small N400 responses, they might result in floor effects that could have led to the absence of a repetition effect in the expected condition in Besson et al.’s study (cfr.). Further, unlike Besson et al.’s sentence repetition task, in the current paradigm the target words were first studied in a word list, and they were presented in a sentence context only during the reading comprehension phase. This not only minimized potential encoding differences between expected and unexpected words during familiarization, but also eliminated potential concerns about the effects of context repetition. Therefore, the present paradigm permitted a truly orthogonal manipulation of a word’s repetition status and its predictability in context, and we believe that it was these methodological improvements that allowed us to observe repetition effects on the N400 response to expected as well as unexpected words. The current findings of additive effects of repetition and predictability on N400 amplitude are consistent with two hypotheses. First, memory for recent words and contextual information may impact lexical semantic processing via distinct mechanisms that independently modulate a word’s activation level. Specifically, we propose that the ease of lexical semantic processing is modulated by (i) bottom-up, exposure-driven changes to the word’s residual activation level in long-term semantic memory, and (ii) top-down, pre-activation of the word’s semantic features as a result of the semantic interpretation of the preceding context. Repetition facilitates lexical semantic processing because recent exposure to a word increases its activation level in memory; lexical semantic processing is facilitated for predictable words because comprehenders incrementally compute a semantic interpretation of the sentence context and integrate it with their world knowledge to anticipate and pre- activate likely upcoming words. Under this view, previous exposure to a word and linguistic predictions act on the same representations stored in long-term semantic memory, but they exert their effects through distinct and independent mechanisms. Alternatively, contextual information and recent exposure to a word may both impact lexical semantic activation via a predictive mechanism. Crucially, the facilitative effect of repetition on lexical semantic activation may not be fully attributable to residual activation of previously encountered words. Instead, lexical semantic activation may be facilitated upon repetition because having recently encountered a word directly strengthens comprehenders’ expectations for that word to appear again; this is essentially the assumption also made by cache-based natural language processing models. Under this view, lexical repetition and predictability modulate lexical semantic activation via a shared neurocognitive mechanism: previous exposure to a word as well as sentential interpretation are incorporated into predictive computations, which in turn facilitate word recognition by pre-activating lexical semantic representations in memory. Some evidence for such a view has been demonstrated in the domain of face recognition , where the facilitative effect of repetition is modulated by the likelihood that repetitions occur in the experiment. If this view is correct, then the additivity of these factors (or the lack thereof) provides information about how different types of evidence are combined in generating predictions about upcoming words, rather than indicating that these factors impact lexical semantic activation through different and independent mechanisms. In sum, it is important to distinguish between a view that holds that sentential context and prior exposure to a word modulate lexical semantic activation via two distinct mechanisms (prediction and residual activation from recent exposure, respectively), and the possibility that both factors modulate lexical semantic activation via the same mechanism (prediction). Future research will be needed to address this question. Meanwhile, the current findings provide no support for the other two proposals outlined in the Introduction, both of which predicted a significant interaction between lexical repetition and context predictability. First, the current results are inconsistent with the proposal that prior exposure facilitates lexical semantic processing during comprehension primarily by strengthening predictions that are already licensed by the sentence context. This account predicted a larger effect of repetition for expected than unexpected words. However, the present results suggest that word repetition facilitates lexical semantic activation of target words both when their occurrence is predictable by contextual information and when it is not. The present results also do not support the proposal that lexical semantic activation during comprehension is primarily modulated by the predictions afforded by the sentence context, with lexical factors exerting an influence only when words are not predictable by context. This proposal draws on the close relationship between repetition and semantic priming, and the previous observation that semantic priming has a larger effect on the N400 amplitude in words lists or in less constraining sentence contexts than in more constraining sentence contexts. The current findings, however, suggest that repetition priming displays a different profile than what has been previously reported for semantic priming, as it affects both expected and unexpected words alike. It is important to note that differences between repetition and semantic priming have previously been noted in behavioral studies. For example, while repetition effects are known to occur over relatively long intervals, semantic priming effects tend to be short-lived. In addition, repetition effects are reliably obtained across different tasks, whereas semantic priming effects are reduced when tasks are changed, for example, from lexical decision to naming. These differences have led to the proposal that repetition and semantic priming might reflect different underlying processes. For example, it has been suggested that while repetition effects might be due to an automatic lexical activation mechanism, semantic priming might partly be the result of top-down prediction,. However, as the current study did not directly contrast interactions with repetition priming and with semantic priming, future work will be needed to examine the extent to which this discrepancy might be explained by differences in experimental paradigms across studies. Importantly, one assumption in the present study is that the effect of predictability on the N400 reflects the behavior of a predictive mechanism that makes use of contextual information to pre-activate words with higher cloze probability, facilitating lexical semantic processing. This is in line with the view that comprehenders make probabilistic predictions about likely upcoming words, and the finding that the N400 is sensitive to a word’s predictability across the full range of possible cloze probability values,. However, like most previous studies that examined the effects of cloze probability (e.g., –; c.f. ), the current study was not designed to determine exactly how the expected target words (or their semantic features) were pre-activated by sentence context. While target words may have been pre-activated based on the message- level information of the sentence context, they may also have been pre-activated partly due to their lexical relationship with other words or phrases in the context. Therefore, in the current study, a word may have been more expected on the basis of its relation to the sentence context as a whole, to words and phrases in the sentence context, or both. Finally, the methodology used in the current study has several important limitations. First, although rarely discussed in this light, it is possible that repetition also modulates the P3 component, which is sensitive to the detection of improbable stimuli (for review see –), and this effect may not have been distinguishable from the N400 effect if these two components had overlapped in time. In the current study, we cannot rule out the possibility that repetition modulated the P3 in addition to the N400 response to the target words during sentence comprehension, but we consider it unlikely given some crucial differences between our paradigm and those that have been used to elicit P3 effects in the past. First, previous studies have shown that P3 effects are strongly attenuated when a participant’s attention is directed away from the task in which the targets are embedded, and in the current study the studied words were not task-relevant during the sentence comprehension phase. In fact, participants were explicitly instructed to base their recognition decisions on whether the words had been presented in the familiarization phase and to ignore whether the words appeared in the sentences. Furthermore, unlike most studies on the P3 component, which have used tasks that required participants to track one (or two) very infrequent targets embedded in a stream of stimuli, the present study presented eight targets in each block, and each target had a 50% chance of occurring in a sentence during the reading comprehension phase. Therefore, we believe that the N400 effect of repetition in the current study is unlikely to be due to an overlapping P3 effect. Second, the repetition manipulation in the present paradigm was not naturalistic, in the sense that participants did not encounter the repeated words in passages, but instead studied them before the sentence-reading phase. This was done in order to orthogonally manipulate the effects of repetition and predictability, and also to allow a more direct comparison with previous studies, which also used a memory paradigm. We believe that this paradigm provides a first step towards examining how lexical and contextual variables work in tandem to facilitate the processing of words in sentences; future research will need to develop more naturalistic methods for examining their interaction. In summary, the current study presents a new experimental paradigm for examining the combined effects of lexical repetition and predictability. We demonstrate that contextual information and memory for recently encountered words have qualitatively similar and additive effects on the N400 response during sentence comprehension. A better understanding of how these variables work in tandem will inform theories about how language is processed in everyday language situations, where sentential context and previous occurrences of a word are expected to jointly aid in the comprehension of discourse. Therefore, by allowing a truly orthogonal manipulation of different factors that are known to impact lexical semantic activation, the current paradigm provides a useful tool for future research on word recognition during language comprehension. The authors thank Cybelle Smith for her help in stimulus creation and data collection. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: WYC SL SB GM DP EL. Performed the experiments: WYC SL SB GM DP. Analyzed the data: WYC SL. Contributed reagents/materials/analysis tools: WYC SL SB GM DP EL. Wrote the paper: WYC SL EL.

# Introduction Early, high-quality cardiopulmonary resuscitation (CPR) is a key contributor to maximizing survival from out-of-hospital cardiac arrest (OHCA). The immediate initiation of CPR can double or quadruple survival. Current resuscitation guidelines indicate compression targets for rate, depth, chest recoil, and chest compression fraction. However, providing chest compressions in adherence to recommendations is generally difficult in the field even for well-trained rescuers. The most frequent non-compliances are long pauses in chest compressions and high compression rates resulting in low compression depths, all these decreasing the likelihood of restoration of spontaneous circulation and survival. Monitoring the compression technique to provide real-time guidance to rescuers contributes to improving CPR quality. Principally used by the advanced life support (ALS), CPR feedback usually relies on advanced systems based on accelerometers, force sensors or magnetic induction which can be connected to the monitor-defibrillator or used as stand-alone devices. Unfortunately, these advanced systems are not generally available for basic life support (BLS) bystanders and first responders using automated external defibrillators (AEDs). Most current AEDs acquire the transthoracic impedance (TI) signal together with the ECG through defibrillation pads. TI represents the resistance of the thorax to electrical current flow. It is used to check for proper electrode contact and adjust the defibrillation energy. Oscillations in the TI signal are correlated with ventilations, chest compressions or presence of circulation. In particular, chest compression activity can be observed in the TI signal in the form of fluctuations around the patient’s baseline impedance synchronized with each compression. Methods for the automated detection of chest compressions have been proposed using several TI features derived from the time domain, the frequency domain and the combination of both. In fact, some defibrillators implement proprietary algorithms to provide CPR feedback based on the TI. These methods were generally tested with signal segments extracted from OHCA episodes collected by the ALS using monitor-defibrillators with CPR feedback. In this context, we wanted to design a simple algorithm for detecting compression activity and computing compression rate in real-time that could be used in BLS settings. The algorithm had to be robust to account for the large dispersion in compression rates observed in the field. Additionally, it had to be simple enough to be implemented in low-cost AEDs with low computational power. We hypothesized that the autocorrelation of the TI signal could provide the necessary information for that purpose. The method was optimized and evaluated retrospectively with AED recordings from OHCA treated by a BLS ambulance service. # Materials and methods ## Data collection Data for the study came from the BLS ambulance system of Emergentziak- Osakidetza, the emergency medical services system in the Basque Country (Spain). Emergentziak-Osakidetza serves a population of approximately 2 200 000 inhabitants, with a density of about 300 inhabitants per km² (4 429 per km² in the urban area). We gained access to all LIFEPAK^® 1000 (Stryker, formerly Physio-Control, USA) AED recordings collected from consecutive patients from 2013 through 2014 that were available in the regional emergency services central office. Recordings belonged to the out-of-hospital cardiac arrest registry (OHSCAR). The regional prospective collection was approved by the Ethical Committee of Clinical Research of the Basque Country (CEIC-E). No patient private information was included in the database. LIFEPAK 1000 AEDs stored ECG and TI signals in their internal memory. Prior to their storage, the ECG was band-pass filtered to suppress direct current and high frequency noise, and the TI was high-pass filtered to suppress patient’s baseline impedance. We extracted ECG and AED signals from the AED recordings and exported them to Matlab^® (Mathworks, USA) format. Then, signals were resampled to a common sampling frequency of 250 Hz. ## Data annotation Signal intervals corresponding to shock administration, disconnections of defibrillation pads and file storage errors were discarded from further analysis. ECG and TI signals were reviewed to annotate the beginning and end of each chest compression interval within the episodes. Each individual chest compression was identified by its corresponding fluctuation in the TI signal. We developed a custom-made software with a graphical user interface for annotation tasks. Experts with large experience in analyzing AED signals were in charge of the annotation process. They jointly reviewed and annotated a subset of the episodes in order to define the annotation rules. The remaining episodes were randomly split in three parts, and each part was annotated by a single reviewer. The quality of the TI signal was generally sufficient for the reliable identification of individual chest compressions, but in some cases the ECG was used for confirmation. shows two signal segments, with chest compression series delimited by red lines. Each individual compression is depicted with a grey dotted line. In example (A), compressions are clearly distinguishable both in the ECG and in the TI. In example (B), however, compressions are more easily identifiable in the ECG. ## Algorithm description We developed an algorithm to estimate compression rates in the range 60–250 cpm. This way, the algorithm could detect and correct severe rate disagreements with respect to current guidelines recommendation. For that purpose, it processes the TI signal in consecutive non-overlapped 2-s analysis windows. The window size was chosen to ensure that a window comprised at least two compressions even at the slowest rate to be detected (60 cpm). The algorithm was designed to provide an estimate of the chest compression rate every 2 s if compression activity is detected, or a zero value, otherwise. The first step involves low-pass filtering the 2-s TI signal interval to enhance the fundamental component of the fluctuation caused by chest compressions. Then, the algorithm computes a biased estimate of the autocorrelation of the filtered TI. This provides a measurement of the similarity of the signal with its delayed version. When the TI waveform shows regular fluctuations, the autocorrelation presents a maximum at the time-lag corresponding to one period of the fluctuation. The position of the maximum represents the average time between consecutive chest compressions. Its inverse is an estimate of the average compression rate in the analyzed 2-s window. We designed the algorithm to locate a peak exceeding a predefined threshold in the lag range from 0.24 to 1.0 s. This corresponds to the range of compression rates to be detected (60–250 cpm). During a chest compression pause, the autocorrelation of the TI will have a disorganized waveform not presenting a prominent maximum. If there is no peak satisfying the amplitude condition in the fixed lag range, the output of the algorithm for compression rate is zero, i.e. the analyzed window is classified as “no chest compressions”. Further details on the algorithm and some graphical examples are provided in the supplementary materials. ## Performance evaluation Episodes were randomly split into training (15% of the recordings) and validation subsets. The algorithm was designed and optimized with the training subset and evaluated with the validation subset. We launched the algorithm from the beginning to the end of the TI signal available per episode. This way we evaluated the algorithm as closest as possible to its real performance when implemented in an AED. First, the method was evaluated in terms of its ability to detect compression activity. All the outputs of the algorithm (one value every 2 s) were classified in one of the following four categories: - TP (true positive): the algorithm reported a non-zero compression rate value for a window with annotated chest compressions. - TN (true negative): the algorithm reported a zero value for a window with no chest compressions. - FP (false positive): the algorithm reported a non-zero compression rate value for an analysis window with no chest compressions. - FN (false negative): the algorithm reported a zero value for a window with annotated chest compressions. Then, according to this categorization, the algorithm was evaluated in terms of its ability to detect chest compression activity through the following figures of merit: - Sensitivity (Se): percentage of annotated compression windows that were correctly detected (i.e. $100 \cdot \frac{TP}{TP + FN}$). - Positive predictive value (PPV): percentage of detected compression windows that actually contained compressions (i.e. $100 \cdot \frac{TP}{TP + FP}$). - Specificity (Sp): percentage of annotated no-compression windows that were correctly detected (i.e. $100 \cdot \frac{TN}{TN + FP}$). - Negative predictive value (NPV): percentage of detected no-compression windows that actually did not contain compressions (i.e. $100 \cdot \frac{TN}{TN + FN}$). These figures of merit were computed both globally, that is, by jointly analyzing all the 2-s windows in all the episodes of the validation subset, and per episode, that is, for the 2-s windows comprised by each episode separately. Second, the method was evaluated in terms of its ability to reliably estimate compression rate. The reference for compression rate was computed from the annotations as the inverse of the mean time between the compression instances manually annotated in the 2-s analysis window (see Data annotation). Error in compression rate estimation was computed as the difference between the reference values and the algorithm estimates. Third, the method was evaluated in terms of its ability to reliably estimate chest compression fraction (CCF). The reference CCF per episode was computed as the percentage of 2-s windows with annotated chest compressions. Error in CCF estimation was computed as the difference between the reference value and the algorithm estimate. ## Statistical analysis Distributions were reported as median (IQR) as they did not pass Lilliefors normality test. Global and per episode figures of merit with their 95% confidence intervals (CI) were computed. # Results From a database of 242 AED recordings (one per patient) containing continuous and concurrent ECG and TI signals, 237 were included in the study. Two episodes were discarded because the ECG was compatible with spontaneous circulation and no chest compressions were observed. Three episodes were discarded because the signals were corrupted by noise impeding reliable annotation of chest compression instances. The training set contained 30 episodes and the validation set 207 episodes. A total of 2766.6 min were reviewed and annotated: 449.7 min in the training subset and 2316.9 min in the validation subset. A total of 303,548 chest compressions were annotated: 47,937 (training) and 255,611 (validation). Finally, a total of 66,610 analysis windows were processed by the algorithm: 10,682 (training) and 55,478 (validation). summarizes the characteristics of the episodes included in the study. shows the distributions of compression rate (A) and CCF (B) in the AED database. Distributions are reported separately for the training and validation subsets and also jointly (global distributions). Considering independence between the diagnoses of the 55,478 2-s windows in the validation subset, global algorithm performance was: Se/PPV 98.7%/98.7%, Sp/NPV 97.1%/97.1%. In the per episode analysis, mean performance values and their CI were: Se 98.6% (95% CI: 98.2–99.1), PPV 98.5% (95% CI: 98.1–98.9), Sp 97.1% (95% CI: 96.6–97.6), and NPV 96.9% (95% CI: 96.1–97.8). shows the distributions per episode. The median unsigned error in the estimation of the chest compression rate per episode was 1.7 (1.3–2.9) cpm. depicts the error in the estimated rate as a function of the reference. The error was below ±10 cpm in 97.7% of the analyzed windows containing chest compressions. The unsigned error in the estimation of CCF was 2.9% (1.7–4.4). shows the distributions of CCF per episode calculated from the reference and by the algorithm, and the distribution of the unsigned error. # Discussion The use of monitoring and feedback devices that guide CPR quality during resuscitation contributes to meeting quality recommendations. These devices are mostly used in ALS settings, since most monitor-defibrillators already incorporate advanced technology for measuring chest compression quality (generally based on accelerometers or on electromagnetic fields). However, this is still a challenge in BLS settings, as many AEDs do not incorporate any type of CPR feedback. Fitting all AEDs (even low-cost devices) with CPR feedback capabilities to guide rescuers on compression performance could be a significant step forward to achieve the goal of early, high-quality CPR provided by bystanders and first responders. Minimizing interruptions in chest compressions and providing adequate compression rates are two critical CPR quality components. Interruptions in chest compressions compromise blood flow to the heart and brain and decrease defibrillation success and survival. Guidelines recommend a tight regulation of compression rate to the target range of 100–120 cpm; at lower rates adequate forward flow may not be generated, while too high rates may decrease compression depth and impede proper heart refilling, reducing the effectiveness of chest compressions. In this context, we proposed an algorithm to detect chest compression activity and estimate compression rate to be used for feedback in BLS settings. It is based solely on the TI signal available in most commercial AEDs. The algorithm can operate in real-time computing a value for compression rate every 2-s that could be displayed to the rescuer during the resuscitation attempt with convenient timing to provide adequate guidance. It can also be launched after the event for debriefing purposes. The algorithm was accurate in the detection of compression activity, reporting Se and PPV above 98% and Sp and NPV above 97%. These results are comparable to those reported by previous methods relying on the TI signal with more complex approaches. Alonso et al. reported a Se/PPV of 97%/97% using the amplitudes and durations of each TI oscillation (time-domain approach). González-Otero et al. reported a Se/PPV of 96%/97% with 2-s analysis windows and TI features computed in the frequency domain. Kwok et al. reduced the analysis window to 1-s, and used TI features computed in both the time and frequency domains, and a Hidden Markov model to improve accuracy. Reported Se and PPV were 99% and 98%, respectively. More recently, Coult et al. combined TI features from the time and frequency domain computed in 5-s TI segments and reported Se and Sp above 98%. We have demonstrated that the information conveyed in the TI autocorrelation was equally reliable with a simpler approach. Furthermore, all the cited studies used OHCA recordings collected by the ALS using monitor-defibrillators. We used BLS AED recordings since our method was intended to be applied in that setting. Accurate online detection of presence/absence of compression activity is relevant for several reasons, particularly in BLS settings with lay people in the field: first, alerting rescuers on long pauses in chest compressions could contribute to reducing no-flow time. Second, the ability to detect programmed compression pauses (for example, pauses for ventilation in the 30:2 protocol or pauses during rescuers switch) could be used to launch automated analyses of the ECG. According to current guidelines, CPR has to be interrupted every 2 min for rhythm analysis, as chest compressions induce artifact in the ECG that could lead to an incorrect diagnosis by the AED algorithm. Some manufacturers have tried to avoid these interruptions by designing a shock advise algorithm that can diagnose the corrupted ECG during chest compressions. Our method has the advantage of using the standard shock advice algorithm during chest compression pauses, which could reduce the implementation challenge and the time to market. Furthermore, AEDs could have the capability of detecting circulation through the automated analysis of the ECG and the TI recorded during pauses for ventilation, as proposed in reference. In this reference, the presence of QRS complexes in the ECG and a circulation component in the TI during the pause are used for automated circulation assessment. For that purpose, a robust and accurate method like the one we describe in the present study is pivotal. Our algorithm was also accurate in the estimation of chest compression rate, and could help rescuers to adhere to the recommended 100–120 cpm range during resuscitation. We reported a median error of 1.7 cpm, comparable to other studies (1.2 cpm, 1.8 cpm, and 1.8 cpm). Our algorithm was designed to detect chest compression rates between 60 and 250 cpm, a range much wider than current recommendations. Thus, the algorithm could be useful to bystanders with no experience in cardiopulmonary resuscitation. Finally, the algorithm could be a powerful tool for the retrospective analysis of CPR quality in debriefing sessions. Post-event assessment of CCF, mean and instantaneous compression rate, location and duration of the compression pauses are key in reporting the quality of CPR. The main advantages of the proposed algorithm are its high accuracy in the detection of the presence of chest compressions and in the estimation of compression rate. The algorithm relies on a single feature computed from the TI autocorrelation, i.e., the method is simple and of low computational cost. Consequently, the algorithm could be incorporated in commercial AEDs with minimal software modifications. Finally, the algorithm is robust, since it reliably performs in a wide range of compression rate conditions. However, integrating the method in AEDs and evaluating how well it performs in a BLS system would require further investigation. ## Limitations Data used in the study came from a single AED model, and results may be dependent on the specific acquisition characteristics of the defibrillator. Generalizability of the results would benefit if the algorithm were tested with a larger database. Another limitation is that compressions were manually annotated in the TI signal used for designing the method. Therefore the algorithm replicated the annotations made by the experts. However, no other references such as acceleration or force signals are usually available in AED recordings. Compression rates in our dataset were much higher than guidelines recommendations and showed a high dispersion. This favored the design of a robust algorithm but it is unusual to find BLS recordings with such high rates. When some of the BLS crew members were questioned about the matter, they admitted their tendency to compress the chest too fast at the time the study database was collected. They recognized the importance of training, and the use of metronome and feedback systems in the field. In the past years, CPR quality training programs have been developed in the Basque Country BLS system. # Conclusion Monitoring of chest compression activity and compression rate based on the autocorrelation of the TI signal is reliable. Our algorithm is simple and easy to implement in current AEDs to guide rescuers during CPR and could also be used for debriefing purposes. Since the TI signal is routinely acquired by most AEDs through defibrillation pads, incorporating this functionality in AEDs could potentially improve the quality of chest compressions provided by bystanders and first responders. # Supporting information The authors thank the BLS providers of Emergentziak-Osakidetza for collecting the data used in this study. [^1]: Author Digna María González-Otero is employed by Bexen Cardio, a Spanish medical device manufacturer. Bexen Cardio had no additional role in study funding, or study design, data collection and analysis, decision to publish, or preparation of the manuscript. This does not alter our adherence to PLOS ONE policies on sharing data and materials. [^2]: ‡ DA, CC and JFU also contributed equally to this work.

# Introduction The development of biotechnology is contributing to the rapid growth of the biological literature. For example, PubMed (<http://www.ncbi.nlm.nih.gov/pubmed/>.), a free resource that is developed and maintained by National Center for Biotechnology Information (NCBI), contains more than 20 million citations of biomedical literature from MEDLINE, life science journals, and online books. The enormous volume of biological literature available provide a massive data resource for researchers, but it also a challenge for mining new information and discovering new knowledge, which has become a very important research subject. Biological named entity recognition can be regarded as a sequence segmentation problem where each token in a sequence is assigned a biological name label (e.g. PROTEIN, DNA, RNA, CELL-LINE, CELL-TYPE,), which can be used to identify specified biological terms in text, or label OTHER which represents the term isn’t a predefined type of biological one. Biological named entity recognition has a key role in biological text mining. It is fundamental for biological information extraction and mining techniques, such as biological relation extraction. However, it is difficult to correctly identify biological terms in text because they use alphabets, digits, hyphens, and other characters,. Arbitrarily referring to biological terms makes it even harder to conduct automatic recognition using computers. In biological text, biological named entities are usually multi-word phrases and some have prefixes and/or suffixes, which makes it harder to determine the boundaries of terms. Biological terms are also affected by their context. In some cases, a biological term has different meaning among species. As a result, it is difficult for computers to recognize biological terms automatically. Identifying biological terms from text is very important in bioinformatics. In this study, we propose a novel approach for biological named entity recognition. ## Related Work Biological term recognition is one of the hottest research areas. Many researchers are interesting in mining biomedical terms from text, which is a key step in extracting of knowledge with an overall aim of identifying specific terms, such as genes, proteins, diseases and drugs –. In general, several methods are used for biological named entity recognition, i.e., dictionary-based approaches, rule-based approaches, and machine learning- based approaches. However, dictionary-based approaches tend to miss undefined terms that are not mentioned in the dictionary. The overall results of dictionary-based approaches rely heavily on a predefined dictionary. There is an enormous number of biological terms and new terms are constantly emerging, so it is impossible to produce a complete dictionary containing all biomedical terms. Therefore, the use of a dictionary can provide the highest precision, but we can also miss many terms. In rule-based biological term recognition systems, the rules used for identifying terms are critical, but there are generally no recognition rules that cover all cases. Machine learning-based approaches train models using a training data set and the models can identify predefined types of terms. Machine learning approaches are now a mainstream method of named entity recognition. Many algorithms are widely used, such as Bayesian approaches, Hidden Markov Model (HMM), Support Vector Machines (SVM), Conditional Random Fields (CRFs), and Maximum Entropy (ME). For example, AbGene developed by Tanabe *et al.* has an 85.7% precision rate, 66.7% recall rate, and 76.2% F1 rate when using the Bayesian method with manual post-processing. An HMM-based system designed and implemented by Zhou *et al.* can recognize protein, DNA, RNA, cell- type, and cell-lines from text. Their system has a 72.55% F1 rate. Kazama *et al.* used SVMs to identify protein, DNA, cell-type, cell-line, and lipid, with a 73.6% F1 rate. Tsai *et al.* developed a CRF system to find protein mentions, achieving a 78.4% F1 rate. Lin *et al.* used ME to recognize 23 categories of biological terms with a 72% F1 rate. However, many methods that perform well in general text do not work as well as expected, – because there are many obstacles in biological term recognition. First, a biomedical term may have several different written forms, e.g., epilepsy and falling sickness refer to the same disease, which is a disorder of the central nervous system that is characterized by loss of consciousness and convulsions. Second, an entity can be represented using different types, e.g., cancer can be used to represent a disease as well as a genus of crabs in the family Cancridae. Third, abbreviations of terms, especially arbitrarily referred abbreviations, cause even more ambiguity problems. For example, PC may refer to prostate cancer, phosphatidyl choline, or even a personal computer. Fourth, many biomedical terms are phrases or compound words, or they may have a suffix or prefix. All of these factors make it more difficult for computers to identify biomedical terms automatically. Researchers have applied many methods to improve the performance of machine learning approaches, such as combining different approaches and proposing a hybrid approach, conducting post-processing after machine learning, and adding biomedical domain knowledge to machine leaning-based term identification systems. In this paper, we combined all these methods to raise the precision and recall rate. # Results We used SVM, Stanford CRFs and two SVM-CRF hybrid approaches to identify biological terms from text. One SVM-CRF hybrid approach used SVM to separate biological terms from non-biological terms before using Stanford CRFs to identify the type of the biological term, while the other used SVM-CRFs to recognize biological terms before applying our proposed algorithms to improve the prediction results. The parameters for the SVM and Stanford CRFs used in the tests are listed in and. In the first round, we tested four approaches using data from the GENIA corpus. The F1 score for the SVM-CRFs combined approach with amendment was better than the other three approaches in five classes and it was close to the best in the remaining classes. Its macro-F1 score was greater than those of the other three approaches. The detailed testing results are shown in. The macro-precision, macro-recall, and macro-F1 rates for the results are shown in. In the second round, we tested four approaches using data from JNLPBA04. The F1 scores for the two SVM-CRF approaches were better than those of the other approaches. The SVM-CRFs combined approach with amendment had the highest macro-F1 score. The detailed results are shown in. The macro-precision, macro- recall, and macro-F1 rate results are shown in. # Discussion The results showed that the SVM-CRFs hybrid approach could identify biological terms from text well and they performed better than conventional SVM and CRFs approaches. We found in some cases, that SVM had higher precision but it tended to miss terms and unstable when trained with a small-sized data set. If the positive data are much less than the negative one, its optimal hyper plane will be biased to negative. Moreover, when the number of feature dimensions is much higher than the size of training set, over-fitting is very likely to happen. For example, monocyte macrophage lineage associated surface antigen is a protein term. However, the result by SVM is not correctwhere the word associated should be tagged as IG#protein. This error is caused because the number of positive samples of the word “associated” as IG#protein is much less than that of negative ones. The results showed that although the performance of CRFs was medium, they maintained a balance between precision and recall rate, indicating that this was a stable approach. All the results suggested that combining SVM and CRFs can provide better performance because this hybrid technique was complementary. The basic idea of our approach was to make full use of the power of SVMs as a binary-class classifier, which facilitates data labeling with CRFs. However, SVM and CRFs are the two very different algorithms, so simply combining them could cause inconsistencies. The proposed amendment algorithms were designed to correct any inconsistencies and promote their performance. # Materials and Methods ## Materials There are many benchmark corpuses for biological named entity recognition, such as the GENIA data set, JNLPBA04 shared task data set, GENETAG data set, and MEDSTRACT data set. The GENIA corpus was developed for applying natural language processing technology to biological text mining. It contains 2,000 MEDLINE abstracts with more than 400,000 words and almost 100,000 annotations of biological terms. JNLPBA04 has several shared tasks for natural language processing in biomedicine and its application. Bio-entity recognition is one of the tasks of JNLPBA04. The JNLPBA04 data set is often used as a benchmark data set for evaluation methods. In the first round of testing, we divided data from the GENIA corpus into two parts, i.e., one part for training and the other for testing. We randomly picked 2000 DNA terms, 683 RNA terms, 2000 protein terms, 2000 cell line terms, 2000 cell type terms, and 2000 other types of terms for training. We then selected 400 DNA terms, 166 RNA terms, 400 protein terms, 400 cell line terms, 400 cell type terms, and 400 other types of terms for testing. In the second round of testing, we randomly selected 2000 DNA terms, 950 RNA terms, 2000 protein terms, 2000 cell line terms, 2000 cell type terms, and 2000 other types of terms from JNLPBA04. We then picked 400 DNA terms, 118 RNA terms, 400 protein terms, 400 cell line terms, 400 cell type terms, and 400 other types of terms for testing. ## SVM Terms Identifier SVM performs well in solving small sample size, nonlinear, and high-dimensional pattern recognition problems and other machine learning problems. Assume that we are given data where is either 1 or −1, indicating the class of. In our previous experiment, we used SVM to identify biological terms from text. We used word, word shape, part-of-speech, and morphology as features for identification, as shown in. The results were good. SVM uses a line or surface to separate the data. Thus, SVM is suitable for binary classification problems but not multiple-class problems where there are more than two candidate objective classes. In most cases, name entity recognition is a multiple-class task. As a result, the initial binary SVM is not fit for most name entity recognition tasks. We can use two main types of approaches to solve multiple-class problems. One is to update an SVM kernel function that can merge the multiple classification surface problems into an optimization so as to solve multiple class classification in one pass. The alternative is to apply multiple binary classifiers until they finish the job. ## CRFs Terms Identifier CRFs are often used for the labeling or parsing of sequential data, such as natural language text or biological sequences. CRFs work well in named entity recognition tasks. Many features can be used in CRFs. For example, term appearance (e.g., capitalization, affixes, etc.) and orthographic features (e.g., alphanumeric characters, dashes, Roman numeral characters, etc.) are used frequently. However, CRFs have many drawbacks. First, CRFs use a limited size of context rather than the whole text because of computational limitation, thereby limiting the contextual information. Second, splitting the context of the whole text into small pieces of context will generally separate inherent relationships among them, and simply combining these pieces of context again cannot reproduce the original context due to the loss of relationships during splitting. For example, a CRF biological term identifier uses a two-word context. The whole text could be split into many pieces of two-word contexts. As a result, the same term in the different places of the text could be tagged with different results due to the variation in the context. However, SVM deals with the whole text so it does not have such restrictions. Third, CRFs are affected by the data distribution. If we want to achieve better results, the data should have an exponential distribution. However, biological terms in texts generally do not meet this data distribution prerequisite. ## SVM-CRFs Combined Biological Name Entity Recognition One of the new research areas in machine learning is combining useful algorithms together to provide better performance or for achieving smooth and stable performance. SVM and CRFs are two conventional algorithms that can deal with named entity recognition tasks well. As stated earlier, the feature context used by SVM is global and it does not have the same constraints as CRFs. SVM is initially the best fit for binary-class tasks and it does not perform well on multiple-class tasks. CRFs generally require more computational time and space than SVMs. Thus, although CRFs have many drawbacks, they are very good at sequential data tagging tasks, which is a typical problem in name entity recognition. Thus, we combined SVM and CRFs because they can complement and facilitate each other. In our approach, biological named entity recognition was regarded as a two-step task. The first step was to determine whether a candidate term was a biological one. If it was a biological term, we determine its class of entity. The first step was a binary classification task where the result was either yes or no, before we could fully use SVM to complete the task. We then used CRFs to infer the type of biological term. Finally, we merged the results returned by SVM and CRFs, before performing an amendment process. ## Inconsistency Removal In this paper, we used a BIO pattern for the resulting tags: tag that started with the character B began a term; tags starting with the character I represented the intermediate words of a term; while tags starting with the character O indicated that the word was not a biological term. For example, the tag BG#protein shows that the word is the starting word of a protein, while the tag IG#protein is an intermediate word for a protein. Thus, the following words with tagscan be composed as a complete protein term: *NOTCH1 ankyrin repeat region*. Given the statement above, we propose a phased approach (Algorithm 1) for determining whether a term is a biological term, as shown in Algorithm 1. Algorithm 1 determined whether a term was a biological one. The input was the word set of all terms. The output was words with the tag *Bio* showing that the word was part of a biological term or the tag *O* showing that the word was not a biological term. Words tagged with *Bio* are further processed by CRFs to determine their biological classes. However, SVM and CRFs are two different algorithms. Simply merging the results returned by SVM and CRFs could cause inconsistency. For example, the term *CsA treated cell* is a cell line mention. Its correct tag should be The SVM identifier predicted the word *CsA* and word *cells* as biological words, but the word *treated* was predicted as a non-biological term. The final results of the SVM and CRFs are Therefore, we needed to amend any inconsistencies to improve the results. Before the amendment, we determined which terms were inconsistent. We use the following two rules to identify inconsistent terms: - Rule 1: If the precursor and the successor of a word are both middle words of a biological term, the word should be also a middle word of the term. - Rule 2: A term begins with a word tagged with a start tag. Rule 1 and Rule 2 removed any inconsistencies caused by shifts in context. We used Algorithm 2 to carry out the term consistency analysis, as shown as follows. Algorithm 2 determined word inconsistency of a term by merging the results of SVM and CRFs, and returning a pending inconsistent terms list. ## Term Length Maximizing Using Rule 1 and Rule 2, we can identify and eliminate inconsistencies. In the example, the prediction results for the term *CsA treated cell*will be treated as correct, although the results are not exactly the best fit. Thus, we propose a new rule to address this type of inconsistency. - Rule 3: The length of a biological term is expected to be as long as possible. According to Rule 3, biological terms should be as long as possible. Using our approach, we extend a term from left to right to validate whether the extended terms are biological terms. Thus, given, if is tagged as a biological term, we have to check: If any of the extended terms are in a biological term list, it is definitely a biological term. However, it is impossible to produce a complete biological term dictionary. Therefore, we need to make some deductions to predict the tags of the extended word. We used a maximal forward and backward probability squeezing approach to extend the term. The maximal forward probability approach determines each forward output probability of state t on the basis of state t−1, while the maximal backward probability determines each backward output probability of state t on the basis of the state t+1. Our approach identifies the output with the maximal product result for the forward probability and the backward probability. We assume an output sequence and a hidden state sequence. Let be the transfer probability from state t−1 to state t, while is the probability of observing all of the given data up to state t−1. At state t−1, given an output sequence and a hidden state, we can find the forward output using the following equations. Let, be the output probability from state t to state t +1 and be the probability of all future data from state t +1 to state t. At state t+1, given output sequence and hidden state, we can conduct inference and find the backward output using the following equations. The final result maximizes the product of the result returned by forward inference and backward inference, as shown in the following equation. An illustration of maximal forward and backward probability squeezing is shown in. The maximal bidirectional probability squeezing method that uses the forward probability and backward probability to predict the outputs of intermediate states tends to lead to bias when dealing with states that are rare. Thus, we add positive gain to rare event cases to reinforce their probability and avoid bias, as shown in Algorithm 3. Algorithm 3 adds positive gain to rare cases to reinforce their probability and avoid bias. We also maintain the context window as large as possible, so the output has the maximal positive gain, as shown in Algorithm 4. Algorithm 4 is maximal bidirectional probability squeezing, which uses the forward probability and backward probability to predict the output. Algorithm 4 also maintains a maximal context window so the output has the maximal positive gain. When we use Rule 3 to maximize the term length, we gradually extend the context window size. We initially set the context window size for the tag as 3. The sequence piece of the context window will then be, while the pending sequence is extended to. We take the piece and use Algorithm 4 to infer the resulting tag. We then judge whether it is correct using Algorithm 2. If correct, the output of the sequence will be revised, but otherwise the context window will be extended left one step and right one step, making it. The pending sequence will also be extended to. We then determine the state of using Algorithm 4 with the context window. This is conducted iteratively until the predictive tag result is correct according to Algorithm 2 or we still cannot find the correct result after the whole output sequence has been treated. The amendment of the output sequence in various contexts is performed using Algorithm 5 Algorithm 5 ensures that the results in context will be adaptively extended gradually. ## Performance Evaluation We evaluate the results in terms of precision, recall rate, and F₁ rate. Precision, recall rate, and F₁ are given by the following equations. For example, when we identify a protein term, the definition of true positive, false positive, true negative, and false negative are regarded as: True positive: protein term correctly identified as protein. False positive: non-protein term incorrectly identified as protein. True negative: non-protein term correctly identified as non-protein. False negative: protein term incorrectly identified as non-protein. We also used macro-precision, macro-recall and macro-F₁, to evaluate the overall performance of the identifiers. Their definitions are as follows : ## Conclusions The vast biological literatures provide a highly reliable information source for biological research. Mining information and finding new knowledge is a very important new subject, where the identification of biological terms is fundamental. We propose a novel machine learning approach to achieve biological named entity recognition. This approach used an SVM to determine whether the term is a biological term, before CRFs were used to infer the type of a biological term. We then judged whether the merged result was consistent in the new global context and applied an amendment approach that used maximal bidirectional squeezing with positive gain in an adaptive context algorithm for correcting inconsistent terms. The results showed that our approach could achieve biological named entity recognition and it performed better than CRFs and SVM alone. LIBSVM and the Stanford Named Entity Recognizer used in this work are highly appreciated. [^1]: Conceived and designed the experiments: BS. Performed the experiments: FZ. Analyzed the data: FZ. Contributed reagents/materials/analysis tools: FZ BS. Wrote the paper: FZ BS. [^2]: The authors have declared that no competing interests exist.

# Introduction Tumorigenesis is a multistep process, likely including genetic mutations in oncogenes and tumor suppressor genes. The mutator hypothesis states that mutations leading to enhanced genomic instability (termed “mutator mutations”) drive cancer pathogenesis by accelerating the acquisition of oncogenic mutations. Originally formulated around DNA polymerases and repair enzymes, the mutator hypothesis has been broadened to include microsatellite instability, chromosomal instability, and deficits in checkpoint activation. Although mutator mutations have been found in the germline in certain familial cancer syndromes, the generalized mutator hypothesis focuses on *somatic* mutator mutations occurring as a step in the evolution of somatic cells towards malignancy. On the contrary, it has been argued that mutator mutations (MM) are unnecessary for cancer development, and that the observed incidence rates of cancer may be explained by mutations occurring at the normal rate in conjunction with multiple rounds of lineage expansion and selection. The debate concerning the relevance of the mutator hypothesis has centered around whether mutator mechanisms are required to explain the appearance of a single cancer cell within a human lifetime. A novel approach was recently suggested, based on the wider perspective that all potential mechanisms of tumorigenesis are in play, but those which produce malignant lineages most efficiently are most likely to contribute to clinical cancers. *Efficiency* is defined as the expected number of malignant lineages generated up to and including a reference timepoint by any particular tumorigenic mechanism. This shifts the issue from analyzing the waiting time to a single cancer cell, and fitting it to epidemiologic data, to the evaluation of the *relative efficiencies of mutator and non-mutator pathways in cancer lineage production*. In this framework, mutator mutations, lineage expansion, and selection are not mutually exclusive and could all simultaneously contribute to tumorigenesis. It is also noted that the conversion rate of normal cells to cancer cells (“cancer lineage birth rate”) likely far outstrips the number of clinically observed cancers, due to numerous malignant and premalignant lineages being eliminated by immune surveillance, failure to establish a blood supply, or competition from other premalignant lineages. Thus models which match the cancer lineage birth rate to clinical cancer incidence may have inherent limitations. It may be more relevant to evaluate the potential contribution of mutator mutations to the efficiency of tumorigenesis, as opposed to whether mutator mutations are necessary to explain a rate of cancer lineage birth rate equal to that of clinically observed cancer incidence. Furthermore, attempts to compare absolute theoretical cancer rates to absolute observed cancer rates are very sensitive to the underlying parameters and other assumptions, leading to variability in conclusions, whereas in the calculation of relative efficiencies, many parameters cancel in the ratios, minimizing the danger of overfitting of models and providing the potential for more robust conclusions. It is assumed in these models that any given malignant lineage has an approximately constant and low probability of developing into a clinical cancer. An analysis of the relative efficiency of tumorigenesis with and without a somatic mutator mutation, in the absence of lineage expansion (LE), demonstrated that mutator mutations enhance tumorigenic efficiency under many realistic scenarios, despite the need for an extra mutation step to acquire the mutator mutation itself. Mutator mutations generally do not enhance efficiency for cancers whose pathogenesis requires only two genetic alterations, but increase dramatically in importance as the number of steps in tumorigenesis increases. However, as the model did not explicitly include lineage selection and expansion, the question of the contribution of mutator mutations to tumorigenic efficiency in the presence of lineage selection and expansion remained open. Mutator lineages are also more likely to suffer deleterious mutations that reduce their fitness and potentially lead to extinction. This effect has been termed negative clonal selection (NCS). To date, no analysis of this effect integrated with simultaneous genetic evolution of the tumor has been performed. In order to account for the effects of selection and expansion of fitter lineages, as well as negative clonal selection, I systematically consider the *fitness landscape*, or the multidimensional space representing cellular fitness, as a function of cellular genetic makeup within an environmental context. Pathways through this fitness landscape are termed *fitness trajectories*. Trajectories of special interest for tumorigenesis are those which begin with a normal cell and end with a transformed malignant cell. This paper presents mathematical models which represent the general case of tumorigenesis across a variety of fitness trajectories, including multiple situations where the mutator lineage suffers reduced fitness (NCS), or achieves increased fitness leading to lineage expansion (LE). Four cases (“fitness trajectories”), which differ in the fitness of intermediate lineages in the tumorigenic process, are considered for both mutator and non-mutator pathways with respect to the production or birth rate of new malignant lineages. In order to become a malignant lineage, a normal lineage must accumulate a fixed number of oncogenic mutations (hits). Lineages with less than the full complement of oncogenic mutations may still expand their relative numbers, or risk extinction, according to their relative fitness, on a continuous basis throughout the process. The cases differ with respect to the assumed fitness trajectory, i.e. the relative change in fitness with each successive oncogenic mutation. In case 1, the *incremental* lineage expansion case, the lineages acquire a fixed increment in fitness with each successive oncogenic mutation, finally achieving their maximum fitness when they have acquired a full complement of oncogenic mutations. Given that most of the increased tumorigenic efficiency due to a mutator mutation can be captured by the case in which the mutator mutation is an initial step, we focus in the mutator pathway analysis for case 1 on mutator mutations occurring as an initial step. In case 2, the *cooperative* lineage expansion case with *early* mutator mutation, there is no increase in fitness until a subset of oncogenic mutations have been acquired, at which point the fitness increases rapidly and lineage expansion begins. Additional oncogenic mutations may then be required to achieve the fully malignant phenotype. The mutator mutation occurs early, e.g. at any point before the lineage expansion. In case 3, the *cooperative* lineage expansion case with *late* mutator mutation, the situation is analogous to case 2 except that the mutator mutation occurs after the onset of lineage expansion, during the period when additional oncogenic mutations are occurring towards reaching a fully malignant phenotype. Since cases 2 and 3 are alternate mutually exclusive subsets of the same fitness trajectory, their relative efficiencies (compared to non-mutator pathways) are additive. In case 4, the mutator and wild type lineages are subject to negative clonal selection. The lineages have a subset of loci (reduced fitness or “RF” loci), mutation of which *may* lead to reduction in fitness, depending on the genetic and environmental context. When a reduced fitness locus is mutated, the lineage is at risk for fitness reduction. Lineages with fitness reduction become extinct, thus potentially limiting the advantage conferred by a mutator mutation. While cases can be proposed that are mixtures of these four cases, it should be possible to infer their properties once these four archetypal fitness trajectories are analyzed. Thus, based on analysis of these pathways in combination (together with the constant fitness pathway previously analyzed), any conceivable fitness landscape could be analyzed. Using these models, I evaluate the relative contribution of mutator mechanisms to tumorigenesis, considering in a quantitative fashion those issues which have historically been raised as counterarguments to the mutator hypothesis, and demonstrating predominance of mutator pathways in most instances. In addition, in the presence of negative clonal selection, I find an optimal mutation rate for tumor evolution, which appears to differ from that for species evolution. The models are focused on enhanced single base substitution rates, and it would be of interest to specifically model other forms of genetic instability that might lead to deletions or to chromosomal instability. The analysis raises several provocative questions: 1. As mutator pathways appear to predominate in most instances, can the diversity and complexity of tumors be addressed by current therapeutic strategies? 2. Can tumor diversity and genetic instability be used to stratify patients for prognosis and therapy? 3. Can therapy be designed to increase the mutation rate in tumors beyond the optimum derived in this paper, resulting in lethal mutagenesis? 4. Can the onset of tumors be delayed to beyond the human lifetime, and therefore prevented, by small decreases in the mutation rate? 5. What are the underlying reasons for quantitative differences between tumor and species evolution? # Results ## Model outputs The results for the four cases below are presented in terms of two key model outputs: (1) relative tumorigenic efficiency of mutator vs. non-mutator pathways, N_rel, and (2) the minimum fold increase in mutation rate required from a mutator mutation before the mutator pathway has a relative tumorigenic efficiency greater than or equal to 1, termed α_50%. N_rel is the ratio of malignant lineages produced by mutator and non- mutator pathways under the specified conditions. The fraction of clinical cancers arising by mutator pathways is given by N_rel/(1+N_rel), and mutator mechanisms predominate if N_rel\>1. In non-lineage expansion models, in which progression to a malignant lineage is a rare event, N_rel can be expressed as a ratio of probabilities, P_rel. α is the multiplicative factor by which a mutator mutation increases the somatic mutation rate per cell generation, e.g. the magnitude of the genetic instability. In the lineage expansion (cases 1–3) and constant fitness models, there is a minimum value of α, which we term α_50%, at which mutator pathways are expected to contribute to half of clinical cancers, and above which mutator pathways predominate. The fraction of total cancers caused by a mutator pathway with a given α when compared to an alternative non-mutator pathway is given by α^C/(α^C+α_50%^C), where C is the number of oncogenic mutations required for malignant transformation. Thus, a mutator mutation must confer a minimum level of genetic instability to be relevant. In evaluating the importance of mutator pathways in a particular model, we need to determine if α_50% is within a range commonly seen in known mutator mutations. Mutations in base selection and proofreading generally increase mutation rates 10–100 fold, and increased random mutation frequencies of up to 500-fold have recently been observed in human tumors. In evaluating the results below, mutator pathways are expected to predominate when α_50% is at or below commonly observed values of α (ca. 10–500). For the negative clonal selection model (case 4), an additional key output is an optimal value of the fold increase in mutation rate, α_optimal, which maximizes the importance of mutator pathways. This corresponds to an optimal mutation rate, k_mut-optimal. ## Model inputs The results depend on the properties of the tumor under consideration, which in turn define the inputs to the models. The key input parameters for all the models are: C, the number of oncogenic mutations required for transformation to the malignant phenotype; R≥0, the natural logarithm of the relative fitness of a malignant cell compared to wild type (meaning that with each successive generation the relative numbers of the malignant lineage increase by a factor e^R); 0≤R_p≤R, the component of R which is due to enhanced proliferation (the remainder would be due to decreased apoptosis); T, the time (in cell generations) to malignancy; N_ML, the number of genomic “mutator loci”, in nucleotides, mutation of which leads to genetic instability; and k_mut, the mutation rate per nucleotide base per cell generation in wild type cells. In the *cooperative lineage expansion* models (cases 2 and 3), an additional input parameter is introduced: D, the number of oncogenic mutations required for an increase in fitness. In the *negative clonal selection* model (case 4), I introduce the input parameter N_RFLN-D (N_{reduced fitness loci net-dominant}), which is an indicator of the vulnerability of the genome to mutations which may reduce cellular fitness. It consists of the number of loci, in base pairs, single copy mutation of which may reduce fitness of the lineage, where the loci are divided into subclasses, and the number in each subclass is multiplied by the probability that a mutation of it will lead to a fitness reduction as a function of genetic and environmental context. Key input parameters and the ranges over which they have been varied in the calculations, as well as key model outputs, are summarized in. The remainder of this section describes selected results and their dependence on input parameters. Further detailed results, not shown in the Figures, are given in Supplementary Tables. In the calculations, I assume N_ML is 100, a very conservative assumption. If N_ML = 1000, α_50% would decrease by a factor of 10^1/C (relative to the same case with N_ML = 100), further enhancing the potential role of mutator pathways. Using the equations in, one may rapidly explore a wide variety of other questions and input parameters. ## Case 1: incremental lineage expansion In analyzing this case, we assume that C oncogenic mutations are required for transformation to the malignant phenotype (C generally varying between 2 and 12 based on epidemiologic data of cancer incidence as a function of age), that a malignant cell has increased fitness R relative to wild type (meaning that with each successive generation the relative numbers of the malignant lineage increase by a factor e^R), and each successive oncogenic mutation leads to an incremental increase in fitness R/C. Based on the previous finding that the major component of efficiency in mutator pathways is due to initial mutator mutations, we evaluate the mutator pathway assuming the mutator mutation occurs first. We find in this case that multiple lineages are simultaneously expanding at different exponential rates, corresponding to lineages with 1, 2, … C−1 oncogenic mutations, and therefore incrementally different fitness. Thus the full expression for the number of malignant cells generated by either pathway is the sum of exponentials. If we approximate these expressions by the highest order term (i.e. the most rapidly growing exponential), representing the pool of cells with C−1 oncogenic mutations from which the new malignant lineages are drawn by one more mutation, we obtain several results (Supplementary and). α_50% is calculated using equations\[11–12\], and N_rel by equation \[13\], in. Firstly, mutator mechanisms predominate in most instances, although the value of α_50% increases slightly compared to the constant fitness case (Supplementary). For most parameter values in case 1, α_50% remains within the range of α for commonly observed mutator mutations, indicating that mutator pathways will have a significant role in tumorigenesis. The relative importance of mutator pathways in tumorigenesis increases as the number, C, of oncogenic mutations required to generate a malignant phenotype increases, as judged by α_50% values. When 2 or fewer oncogenic mutations are required, non-mutator pathways predominate. When 4 or more oncogenic mutations are required, mutator pathways predominate. When 3 oncogenic mutations are required, the results depend on the parameter values (Supplementary). Mutation of both copies of a recessive oncogene would count as 2 oncogenic mutations. When compared to the constant fitness case, incremental lineage expansion limits the importance of mutator pathways when two oncogenic mutations are required for cancer (C = 2) and to some extent at C = 3 with a low wild type mutation rate (k_mut), but for higher values of C, α_50% continues to be well within commonly observed ranges. For example, when three oncogenic mutations are required for cancer (C = 3), the wild type mutation rate is low (k_mut = 10⁻¹¹), and the relative fitness advantage e^R of malignant cells relative to wild type is 2, a 770-fold increase in the mutation rate would be required for mutator pathways to be observed in 50% of the cancers (Supplementary). This increase is at or beyond the upper range of increase in mutation rate due to common mutator mutations. In contrast, when six oncogenic mutations are required for cancer (C = 6), α_50% ranges from 11–30 (Supplementary), in the lower range of commonly observed values of mutation rate increase due to mutator mutations, suggesting a predominance of mutator pathways, in that most mutator mutations would then correspond to α\>α_50%. Note in Supplementary that this result for C = 6 is unchanged for all combinations of wild type mutation rate, cell generations to cancer, and degree of fitness increase within the explored parameter values. When judged by relative efficiency N_{rel 1∶0}, the importance of mutator pathways is reduced relative to the constant fitness case to a greater degree than one would judge based on α_50%. This is because the value of N_{rel 1∶0} is very sensitive to small changes in α, and therefore a relatively small increase in α is required to compensate for the effect of incremental lineage expansion on N_{rel 1∶0}. Based on the analytical model, the relative efficiency N_{rel 1∶0} is reduced in the case of incremental lineage expansion by a factor of RT (C−1)/\[(C+1)C\] compared to the constant fitness case. However, α_50% would need to increase by a factor of only {RT (C−1)/\[(C+1)C\]}^1/C to compensate for this. For example, with the relative fitness of malignant cells e^R = 2, the number of cell generations T = 5000, and the number of required oncogenic mutations C = 6, the relative tumorigenic efficiency N_{rel 1∶0} is reduced over 400-fold relative to the constant fitness case. But only a 2.7 fold increase in α can restore the same relative importance of mutator pathways under these circumstances. The analytical model (equations\[11–13\]) shows that the relative importance of mutator pathways N_{rel 1∶0} increases with increasing wild type mutation rate k_mut and increasing fold-increase in mutation rate, α, similar to the constant fitness case. Very large relative fitness advantages for malignant cells e^R somewhat further reduce the importance of mutator pathways (, Supplementary). For example, when three oncogenic mutations are required for cancer (C = 3), the number of cell generations to cancer T = 5,000, and the wild type mutation rate is k_mut = 10⁻¹¹, α_50% is 510 when the relative fitness advantage e^R of malignant cells relative to wild type is 1.2, 770 when the relative fitness advantage is 2, and 1100 when the relative advantage is 7.4 (Supplementary). Finally, the analytical model (equations\[11\] and \[13\]) shows that the relative importance of mutator pathways N_{rel 1∶0} is approximately independent of the number of cell generations T, in contrast to the constant fitness case, where it is proportional to T. Different cancer types are thought to typically arise after different numbers of cell generations T. The relative importance of mutator pathways in these different cancer types may thus depend on the fitness landscapes experienced by cells with less than the full complement of oncogenic mutations. ## Case 2: cooperative lineage expansion, early mutator mutation In this circumstance, D oncogenic mutations occur leading to a sudden cooperative increase in fitness. At some time during the acquisition of these initial D mutations, a somatic mutator mutation may occur. After the acquisition of the first D oncogenic mutations, and consequent increase in fitness, an additional C–D oncogenic mutations must occur to complete the transformation to a malignant lineage. α_50% is calculated using equation \[16\], and N_rel by equation \[13\], in. In this case also, mutator mechanisms predominate. The results with regard to α_50% are depicted in. As in the incremental lineage expansion case, the calculations show a slight increase in α_50% relative to the constant fitness case, while still generally indicating a predominance of mutator pathways. When judged by relative efficiency N_rel, the importance of mutator pathways is again reduced relative to the constant fitness case to a greater degree than one would judge based on α_50%, again due to the high sensitivity of N_rel to the value of α, but a small change in α can compensate. When compared to the constant fitness case, cooperative lineage expansion with early mutator mutation limits the importance of mutator pathways when few oncogenic mutations are required for cancer (C = 2) and to some extent at C = 3 with a low wild type mutation rate (k_mut), but for higher values of C, α_50% continues to be well within commonly observed ranges. For example, when three oncogenic mutations are required for cancer (C = 3), the relative fitness of malignant cells e^R = 2, and the wild type mutation rate is low (k_mut = 10⁻¹¹), an 880-fold increase in the mutation rate would be required for mutator pathways to be observed in 50% of the cancers (see and Supplementary). This increase is at or beyond the upper range of increase in mutation rate due to common mutator mutations. In contrast, when six oncogenic mutations are required for cancer (C = 6), α_50% ranges from 11–30 (Supplementary), in the lower range of commonly observed values of mutation rate increase due to mutator mutations, suggesting a predominance of mutator pathways. The analytic models (reference and equations \[13–14\] and \[16\] in this paper) show that mutator pathways are more likely if they occur early within this window, and also for higher wild type mutation rate k_mut, and fold change α in mutation rate due to a mutator mutation, similar to the results for the constant fitness and cooperative lineage expansion with late mutator mutation cases. Very large relative fitness advantages for malignant cells e^R somewhat further reduce the importance of mutator pathways (, Supplementary). For example, when three oncogenic mutations are required for cancer (C = 3), the number of cell generations to cancer T = 5,000, and the wild type mutation rate is k_mut = 10⁻¹¹, α_50% is 590 when the relative fitness advantage e^R of malignant cells relative to wild type is 1.2, 880 when the relative fitness advantage is 2, and 1260 when the relative advantage is 7.4 (Supplementary). However, in contrast to the constant fitness and cooperative lineage expansion with late mutator mutation cases, the analytic model (equations \[13\] and \[16\]) shows that the relative contribution of mutator pathways is independent of number of cell generations T. Finally, the relative contribution of mutator pathways is shown by the analytic model (equations \[13\] and \[16\]) to be independent of the number of oncogenic mutations required for an increase in fitness, D, in contrast to the cooperative lineage expansion with *late* mutator mutation case. ## Case 3: cooperative lineage expansion, late mutator mutation In this circumstance, D oncogenic mutations occur leading to a sudden cooperative increase in fitness. After this occurs, an additional C–D oncogenic mutations must occur to complete the transformation to a malignant lineage. During this latter period, a somatic mutator mutation may occur. For both mutator and non-mutator pathways, the lineages will have greater numbers of cells due to their increased fitness. However, this increased fitness is constant for both types of pathways, and remains constant during the period in which the possible occurrence of a mutator mechanism is being considered. Thus, the *ratio* N_rel of malignant cell lineages produced by mutator and non-mutator pathways will be nearly equivalent to the probability ratio P_rel previously derived for the constant fitness case, except that the parameter C (number of oncogenic mutations required for cancer) is now replaced by C–D (the number of oncogenic mutations required for cancer after the original fitness increase), and a factor representing more rapid acquisition of the mutator mutation due to more rapid proliferation multiplies N_rel (if the increased fitness includes more rapid proliferation). α_50% is calculated using equation \[20\], and N_rel by equation \[21\], in. Mutator mechanisms predominate in most instances as long as C–D≥3 (see and Supplementary), as judged by the values of α_50%. In the case of C–D = 3, for example, α_50% ranges from 12 to 252, depending on various parameter values (Supplementary). This range is well within that seen with known mutator mutations. As the number of oncogenic mutations required for cancer after the original fitness increase (C–D) increases further, greater predominance of mutator pathways is expected. For cooperative lineage expansion with C–D\<3, non-mutator pathways, or mutator pathways with early mutator mutations, are more likely pathogenic mechanisms. Importantly, the analytic results (equations \[20–21\]) imply that the importance of this pathway may depend on the number of oncogenic mutations required for increased fitness, D, when other parameters, including the number of oncogenic mutations to cancer C, are held constant. In this case, the fewer oncogenic mutations are required for increased fitness, the greater the relative predominance of this mutator pathway with late mutator mutations. The dependence of the results on D is illustrated in and documented for other parameter values in Supplementary. As in the constant fitness case, mutator pathways are also more likely if they occur early within this window, and for higher wild type mutation rate k_mut, fold increase α in mutation rate due to a mutator mutation, and number of cell generations T. For the cooperative lineage expansion case with *late* mutator mutation, the relative importance of mutator pathways is somewhat further *increased* at very large relative fitness advantages for malignant cells, e^R (Supplementary), in contrast to the incremental lineage expansion case (case 1) and the cooperative lineage expansion case with *early* mutator mutation (case 2). In the cooperative lineage expansion case with late mutator mutation, a greater fitness advantage increases the pool of cells which may acquire a late mutator mutation. For example, when the number of oncogenic mutations required for cancer C = 6, the number of oncogenic mutations required for the cooperative fitness increase D = 2, the number of cell generations to cancer T = 5,000. and the wild type mutation rate k_mut = 10⁻¹¹, α_50% is 29 when the relative fitness advantage e^R of malignant cells relative to wild type is 1.2, 27 when the relative fitness advantage is 2, and 22 when the relative advantage is 7.4 (Supplementary). ## Case 4: negative clonal selection In this model, lineages have a constant risk per cell per cell generation of suffering a reduction in their fitness. Lineages with fitness reduction are assumed to eventually become extinct (the probability of this occurring is very high in large cell populations). This phenomenon was previously studied in isolation, and termed negative clonal selection (NCS). In the current model, those lineages which do not become extinct are at the same time continuously and progressively acquiring oncogenic mutations. The instantaneous risk of fitness reduction is the product of the mutation rate per nucleotide base per cell generation k_mut (or αk_mut after a mutator mutation) and N_RFLN-D (N_{reduced fitness loci net-dominant}), an indicator of the vulnerability of the genome to mutation, consisting of the number of loci, in base pairs, single copy mutation of which may reduce fitness of the lineage, where the loci are divided into subclasses, and the number in each subclass is multiplied by the probability that a mutation of it will lead to a fitness reduction as a function of genetic and environmental context. As in the constant fitness case, and the incremental lineage expansion case (case 1 above), we approximate mutator pathways by considering mutator mutations occurring as an initial step in tumorigenesis. N_rel (which in this case is equal to the relative probability P_rel of mutator versus non-mutator pathways) is calculated using equations \[28–30\] in. In contrast to the other cases, the relative efficiency N_rel of mutator to non-mutator pathways does not continue to increase with greater fold increases α in the mutation rate. Increased mutation rates speed the acquisition of oncogenic mutations, but at the same time increase the risk of fitness reduction and extinction. In this type of fitness landscape, the relative efficiency N_rel of mutator compared to non-mutator pathways increases with greater fold increases α in mutation rate, until an optimum at which the growth of the malignant lineage begins to be limited by negative clonal selection. An approximate optimum for mutation rate k_mut can be estimated for this circumstance from the theoretical treatment: Note that the treatment focuses on dominant reduced fitness mutations only. Recessive reduced fitness mutations, requiring mutation of both alleles, were found to be quantitatively insignificant. Within the parameter ranges considered within this paper, the optimal mutation rate for tumor evolution, k_{mut optimal}, varies from 2.1×10⁻¹⁰ to 3.6×10⁻⁶ per nucleotide base per cell generation. This optimum is generally higher than estimated mutation rates of wild type embryonic stem cells or somatic cells, 10⁻¹¹ to 10⁻⁹. An approximately optimal value of the fold increase α in mutation rate due to a mutator mutation is therefore given by α_optimal = k_{mut optimal}/k_mut. In the presence of an anti-apoptotic mutation, the vulnerability of the genome N_RFLN-D would be reduced to a very low value, further raising the optimal mutation rate and diminishing the potential effect of negative clonal selection. Thus, mutator pathways with α\>α_optimal would be more efficient if they occurred after an anti- apoptotic mutation. The relative importance of mutator pathways increases with increasing number of required oncogenic mutations for malignant transformation, but in contrast to the other cases, the minimal number of oncogenic mutations at which mutator pathways are favored \[log (N_{rel 1:0})\>0\] varies depending on the strength of negative clonal selection, as shown in for a fold increase in mutation rate α = 100, a wild type mutation rate k_mut = 10⁻¹¹, and a number of cell generations to cancer T = 5,000. At maximal negative clonal selection, there must be at least 5 oncogenic mutations required for malignant transformation before mutator pathways are favored. Additional more detailed results are given in Supplementary. In general mutator pathways are favored when 5 or more oncogenic mutations are required for malignant transformation and not favored when 2 or fewer oncogenic mutations are required. Results when 3 or 4 oncogenic mutations are required depend on parameter values. Whereas in the absence of negative clonal selection, higher values of the fold increase α in mutation rate due to a mutator mutation, number of cell generations T, and wild type mutation rate k_mut generally favor mutator mutations, in the presence of negative clonal selection the relative importance of mutator pathways depends on the parameter values in a complex way. depicts the relative prevalence of mutator pathways with initial mutator mutations N_{rel 1∶0} as a function of α, for the highest levels of negative clonal selection, with the number of oncogenic mutations required for cancer C = 5, the number of cell generations to cancer T = 5,000, and the wild type mutation rate k_mut = 10⁻¹¹, illustrating the decrease in the relative importance of mutator pathways beyond an optimum. Supplementary shows the relative probability of a mutator pathway with an initial mutator mutation compared to no mutator pathway, in the presence of negative clonal selection (N_{rel 1∶0, NCS}) for numerous combinations of parameter values not shown in the Figures. # Discussion This paper highlights a new approach to evaluation of the importance of somatic mutator mutations in tumorigenesis. The relative efficiencies of mutator and non-mutator pathways are considered, thus circumventing the need for comparisons of absolute rates of tumorigenesis with epidemiologic data, with their inherent limitations such as the assumption that every malignant cell becomes a clinical cancer. It was previously shown that mutator pathways are generally more efficient in the setting of constant fitness despite requiring an additional step for acquisition of the mutator mutation. More efficient mechanisms are likely to play a proportionately larger role in tumorigenesis. However, tumorigenesis involves changes in fitness including decreased fitness leading to extinction (negative clonal selection) and increased fitness leading to lineage selection and expansion. These could have profound effects on the relative importance of mutator pathways in tumorigenesis. Hence, in the present work I quantitatively consider the variations in fitness that have historically been raised as counterarguments to the mutator hypothesis. This analysis confirms the importance of mutator pathways in tumorigenesis across most representative fitness landscapes. Importantly, the results do not negate the important role for simultaneous lineage expansion and selection. In all cases, mutator pathways are favored when there are more oncogenic mutations required for cancer. Depending on the situation, mutator pathways will be favored when the number of required oncogenic mutations exceeds the range of 3–5. Genome wide sequencing of solid tumors suggests tumors may harbor between 14–20 oncogenic mutations. The functional dependence of the importance of mutators on other key parameters varies depending on the different cases considered. For example, in some cases, increasing the number of cell generations T enhances the importance of mutators, in others it has no effect, and in still others there is an optimum beyond which further increases in T decrease the importance of mutators. The case which best matches clinical phenomena needs to be determined and may vary for different cancer types. For constant fitness models, negative clonal selection models, and incremental lineage expansion models, mutator mutations are more efficient if they occur early. However, for cooperative lineage expansion models, whether it is more efficient for the mutator mutations to occur before (case 2) or after (case 3) lineage expansion onset depends on the parameters, and can be determined in any particular circumstance by comparing N_rel, the relative efficiency of mutator compared to non-mutator pathways, for case 2 and case 3. While the models generally predict that mutator mutations are an early event, this may be difficult to verify experimentally. Mutator mutations may occur in only a minority of premalignant lesions, but the mutator premalignant lesions are predicted to overwhelmingly be the ones that actually develop into cancers. The model allows the calculation of the fraction of pre-malignant lesions with a given number of oncogenic mutations that harbor mutator mutations in exactly the same way that it allows calculation of the fraction of malignant tumors which develop by the mutator pathway as a function of the number of oncogenic mutations. It therefore follows that premalignant lesions with two or less oncogenic mutations will be less likely to harbor a mutator mutation, but that those few mutator early premalignant lesions will ultimately produce the majority of advanced premalignant lesions and cancers. Since the mutator lesions become progressively enriched with increasing progress towards malignancy, any experimental technique which looks at average mutation frequency across many premalignant lesions might falsely conclude that a mutator mutation is a “late event”. One would need to be able to look at the presence or absence of a mutator mutation in large numbers of premalignant lesions and/or cells individually to verify this prediction. The mutator lesions and/or cells may remain in the minority until lesions with 3–4 or more oncogenic mutations are surveyed. The analysis predicts that for genetic DNA repair syndromes such as xeroderma pigmentosum (XP) and hereditary non-polyposis colon cancer (HNPCC), simple relationships will obtain between the observed degree of increased risk, the number of oncogenic mutations required for malignant transformation, the time to cancer onset, and the increased mutation rate as a result of the disorder. These parameters are all measurable, with the exception of the number of oncogenic mutations required for malignant transformation, which can be estimated based on evolving molecular knowledge in experimental systems. The model also predicts that in syndromes with inherited mutator mutations such as HNPCC, premalignant lesions such as polyps will be more efficiently converted to cancer. Because of this very efficient acquisition of additional oncogenic mutations, HNPCC can result in a higher number of cancers without a higher number of polyps. The greater enhancement of number of cancers compared to the enhancement of the number of polyps can be anticipated based on the greater number of oncogenic mutations required to produce the former. In contrast, in a genetic cancer syndrome which is not based on a mutator mutation, familial adenomatous polyposis (FAP), we see an increase in the number of polyps at least equivalent to the increase in the number of cancers. The model predicts, again correctly, that the mutator mutation accessible through HNPCC will result in an increased risk of colon cancer but not of embryonal carcinomas. Thus, in HNPCC, in which a single further mutation confers a mutator phenotype, there is no increased risk of embryonal cancers, which require fewer oncogenic mutations in their formation, despite an increased colon cancer risk corresponding to the larger number of oncogenic mutations in the pathogenesis of colon cancer. In retinoblastoma, few oncogenic mutations are required, but as the mutator mutation is also the pathognomonic oncogenic mutation, its occurrence is favored. In the case of negative clonal selection, tumorigenic efficiency does not continue to increase with increased mutation rate, and actually decreases beyond an optimum rate. The optimal mutation rate for tumor evolution is consistently higher than one published estimate of the mutation rate of embryonic stem cells, and is generally higher than the estimates for mutation rates of somatic cells. If we assume the estimate of the stem cell mutation rate is correct, and that it has evolved to be optimal for species evolution, the comparison emphasizes that while tumor and species evolution may be partially analogous, they may differ in important quantitative and even qualitative features. Thus, mutator mutations may be important for the evolution of tumors without being important in species evolution. The optimum mutation rate for tumor evolution estimated by this treatment is analogous to the optimum mutation rate for viruses under selection pressure from antiviral therapy, which is the reciprocal of the viral genome length. The present work differs in that it assumes multiple steps are required for malignant transformation (hence the factor C), that not all sites when mutated lead to reduced cellular fitness (hence N_RFLN-D rather than the full genome length), that either member of a diploid gene pair can suffer a reduced fitness mutation (hence the factor of 2), and that multiple cell generations, rather than 1, need to be considered (hence the factor T). The increased optimal mutation rate calculated here for tumors is consistent with the several experimental demonstrations of increased tumor incidence in mice genetically engineered to have mutator mutations –, as well as the several hundred fold higher random mutation frequency in human tumors observed experimentally compared to surrounding normal tissues, when mutations are measured at a genetically neutral site. Note these mutations in human tumors may be present in only a very small minority of cells within a tumor mass, and can be detected only by PCR amplification from single copies. While the observed random mutation frequency also depends on the number of cell generations, it is unlikely the cell generation number increase in tumors is several hundred fold relative to normal tissues. In colon cancer for example, the normal tissue arises from a highly proliferative stem cell compartment which has been proliferating since birth. Finally, the derivation of an optimal mutation rate greater than wild type is consistent with studies of bacteria competing for survival in culture, which show the winners have an increased, but not excessively increased, mutation frequency \[Loh E, Salk JS, Loeb LA, unpublished\]. I will now briefly consider the more speculative questions raised in the introduction. 1. As mutator pathways appear to predominate in most instances, can the diversity and complexity of tumors be addressed by current therapeutic strategies? Clinically, mutator lineages will lead to enhanced heterogeneity within tumors, enhancing the probability that a sub-population of tumor cells manifests pre-existing resistance to therapy. In addition, sensitive cells within mutator lineages will evolve to resistance more quickly. Both pre- existing and acquired resistance emphasize the need for therapeutic combinations. 2. Can tumor diversity and genetic instability be used to stratify patients for prognosis and therapy? The question of how many agents to give in combination, especially when limited by toxicity, may depend on measures of underlying tumor diversity and plasticity. 3. Can therapy be designed to increase the mutation rate in tumors beyond the optimum derived in this paper, resulting in lethal mutagenesis? The proposed anti-cancer strategy termed “lethal mutagenesis” involves increasing the mutation rate to the point where negative clonal selection threatens survival of the malignant lineage. The mathematical model of negative clonal selection can potentially be adapted to allow estimation of the tumor mutation rates which need to be reached to achieve this effect, as well as the mutational “therapeutic window” between tumor and normal tissue. Current inhibitors of cell cycle checkpoint kinases, under preclinical and clinical development as chemotherapy and radiation sensitizers, exemplify methods to further enhance the mutagenic effect of therapy, by bypassing pauses for repair of DNA damage. These agents may be selective based on the absence of a functional p53 checkpoint in tumors. 4. Can the onset of tumors be delayed to beyond the human lifetime, and therefore prevented, by small decreases in the mutation rate? As cancer is generally a disease of the elderly, only a modest delay in its onset is required to reduce its importance compared to other causes of mortality. As the efficiency of tumorigenesis is a function of the mutation rate raised to the power of the number of oncogenic mutations required for malignant transformation, only a very modest reduction in mutation rate would be required to delay the onset of cancer. This in turn suggests that prevention of cancer through public health and/or pharmacologic measures aimed at reducing the mutation rate could be effective. 5. What are the underlying reasons for quantitative differences between tumor and species evolution? The genome may be more tolerant of mutation in the context of tumorigenesis, in which homeostasis need not be maintained in the whole organism, than in the evolution of species, in which this homeostatic constraint must be obeyed. This may result in a higher optimum mutation rate for tumor evolution. In summary, this paper provides a general quantitative evaluation of the relative importance of mutator pathways compared to non-mutator pathways in tumorigenesis, accounting for fitness variation and selection, thus directly addressing the major historical criticisms of the mutator hypothesis. Mutator pathways predominate in most but not all instances. The optimal mutation rate is higher for tumor evolution than for species evolution. # Methods The present work builds on methods previously published concerning both tumor evolution at constant fitness and negative clonal selection in the absence of tumor evolution. However, it significantly extends this work by providing a general analytic solution for determining the relative efficiency of mutator pathways, compared to non-mutator pathways, for fitness pathways involving increased fitness of malignant lineages and their precursors, as well as providing an analysis of negative clonal selection in the presence of tumor evolution. All the models are “multi-hit models”, in that malignant lineages, with C oncogenic mutations, arise by mutation from precursor lineages with C−1 mutations, which in turn arise from precursor lineages with C−2 mutations, etc. In the constant fitness case, the probability of having C oncogenic mutations at time T is derived for the non-mutator pathway based on minimal assumptions. The corresponding quantity for the mutator pathway is then the probability weighted integral of all the possible pathways with an additional somatic mutator mutation occurring at any time between 0 and T. In the original negative clonal selection work, the number of viable lineages surviving decays exponentially with a time constant β equal to the product of the mutation rate per base per cell generation, k_mut, and a general indicator of the vulnerability of cellular fitness to dominant mutation, N_RFLN-D, the net number of loci in bases, single copy mutation of which will reduce fitness, adjusted for the probability of fitness reduction associated with each locus, summed over possible genetic and environmental contexts. The asssumption that all lineages with reduced fitness will become extinct is nearly true for large populations, and dominant reduced fitness mutations are of greater quantitative significance than recessive ones. The models do not assume that different lineages are competing for an ecologic niche of a fixed size, as is often the case in evolutionary game theory and related techniques. It is assumed that wild type lineages maintain their numbers and the size of their ecological niche, lineages with reduced fitness become extinct, and lineages with increased fitness, including malignant lineages and their precursors, increase exponentially at rates determined by their relative fitness, potentially breaking anatomic barriers to form malignant or benign tumors respectively within expanded niches. These assumptions, intuitively aligned with the occurrence of benign premalignant lesions in solid tumors, enable the derivation of complete analytic solutions incorporating both multi- hit tumorigenesis and arbitrarily varied fitness landscapes. The exponential growth rates are determined by relative fitness R. This follows from the assumption that the change in relative numbers of a lineage per unit time is proportional to its relative fitness, i.e.: where N is the number of cells, dN/N is the change in their relative number, R is the relative fitness, and dt is an instant of time measured in cell generations. Integrating from 0 to T cell generations leads to an exponential growth equation depending on relative fitness, i.e. Exponential growth is thought to characterize nascent tumors until they reach a limiting size. Note a relative fitness of zero corresponds to wild type fitness and results in constant numbers over time. The lineage expansion models (cases 1–3) are formulated in terms of “expectation values,” or the mean number of malignant cell lineages generated by a particular tumorigenic mechanism at a reference timepoint. This is in contrast to the constant fitness model in which a single cell lineage experiences a limited number of cell divisions, so that mutation in any single nucleotide locus during the lifetime of a single cell and its progeny is a rare event. Probabilities of rare events joined as “or” approximately add (P(A or B)≈P(A)+P(B)), and the constant fitness model is expressed in terms of probabilities. In the presence of lineage expansion, a large population of N cells may result from the lineage of a single cell, where N is equal to or greater than the reciprocal of the per nucleotide mutation rate per cell generation times the number of cell generations. In that instance, the probability that at least one cell in this population harbors a mutation at a particular nucleotide locus may approach 1, and probabilities of individual events joined as “or” do not simply add. However, expectation values are still additive under an “or” operation, simplifying the theoretical treatment. Biologically, this corresponds to the postulate that not every malignant lineage leads to a clinical cancer, and the mechanisms which produce the most malignant lineages are most likely to contribute to tumorigenesis. The models are designed for large cell populations. In scenarios where there are small clusters of cells, such as intestinal crypts, the models may be thought of as reflecting the average behavior of a population of crypts, including the fact that benign tumors or polyps may arise in some. Other key assumptions, parameters, and parameter values have been previously reviewed. Key input and output parameters and their values are also given in the “Results” section and in. Below we give the equations used to calculate the results in this paper and the general strategy for deriving them. Full derivations are available upon request. These full derivations also include demonstrations that all cases are identical to the previously derived constant fitness case in the limit where the natural log fitness advantage R (cases 1–3) or susceptibility constant β for negative clonal selection (case 4) approach 0, serving as a check on the mathematics. ## Case 1: Incremental Lineage Expansion Given the assumption that any fully malignant lineage with C oncogenic mutations has an equal chance of becoming a clinical cancer, N_rel, the relative number of clinical cancers due to mutator compared to non-mutator pathways at time T or earlier, is given by: where N_Ci,C-mut***(T)*** and N_Ci,C***(T)*** are the number of new lineages initiated with C oncogenic mutations, and C oncogenic mutations are required for malignant transformation, up to and including time T, for mutator and non-mutator pathways, respectively. The model is a multi-hit model wherein C oncogenic mutations are required for tumorigenesis, and is described by a network of ordinary differential and integral equations. Lineages with no oncogenic mutations do not increase their numbers (denoted N₀), but lineages with n oncogenic mutations increase in number by a factor of e^nR/C each *wild type* cell generation. Hyperproliferative mutations which decrease the generation time are appropriately factored into the value of R. Lineages with n mutations also increase their numbers by initiation events: i.e., mutations from the lineages with n−1 oncogenic mutations. The rate of initiation of new lineages with n oncogenic mutations at any instant t, dN_ni,C ***(t)***, is given by the product of the number of cells with n−1 oncogenic mutations at that instant, N_n−1,C ***(t)***; the mutation rate, k_mut; the number of remaining unmutated oncogene loci, N_OL-n+1 (where N_OL is the number of oncogenic loci available for mutation in a wild type cell); and a factor to account for more cell generations per unit time if there is a hyper-proliferative mutation, e^Rp: The negligible effect of mutation in decreasing precursor populations is ignored, simplifying the treatment. The total number of lineages with n mutations initiated by time T is the integral of the instantaneous initiation rate from 0 to T: The number of cells in a given class of lineage by time T is given by the integral of the instantaneous initiation rate at time t, multiplied by the lineage expansion from time t to T for that lineage class, over the interval t = \[0,T\]. In particular, malignant initiation events occur from lineages with C−1 oncogenic mutations, as they acquire their Cth oncogenic mutation: The total number of malignant initiation events by time T is simply the integral of this instantaneous initiation rate from 0 to T: In summary, the instantaneous initiation rate of lineages with one oncogenic mutation is given by a first order differential equation. Based on the instantaneous initiation rate of lineages with one oncogenic mutation, and their lineage expansion, one can derive an expression for the number of cells with one oncogenic mutation as a function of time. The instantaneous rate of initiation of lineages with two oncogenic mutations is proportional to the number of cells with one oncogenic mutation at any given time. In turn, based on the instantaneous initiation rate of lineages with two oncogenic mutations, and their lineage expansion, we derive an expression for the number of cells with two oncogenic mutations at any point in time. The instantaneous initiation rate of cells with 3 oncogenic mutations is then proportional to the number of cells with 2 oncogenic mutations, and so on. Expressions for the number of cells and lineage initiations were derived for several values of n (number of oncogenic mutations) and C (number of oncogenic mutations required for malignant transformation), and based on the algebraic details, general expressions were hypothesized for all n and C, verified for the cases explicitly derived, and proven for all n and C by mathematical induction. For mutator pathways, the mutator mutation is assumed to occur first, based on previous work suggesting that mutator pathways are more efficient if the mutator mutation occurs early. The rate of formation of cells with the original mutator mutation, dN_0,C-mut ***(t)***, is given by the product of the initial number of cells N₀, the number of mutator loci N_ML, and the mutation rate: Once a lineage with a mutator mutation is formed, analysis of its progress parallels that of the non-mutator pathway, except for the increase in the mutation rate constant by the factor α. The numbers of malignant cell initiations by both mutator and non-mutator pathways is the sum of exponentials representing lineages with 0 to C−1 oncogenic mutations. α_50%, the minimum fold increase in mutation rate at which mutator pathways account for 50% of observed cancers (derived by setting N_rel = 1), is approximated by considering only the most rapidly growing exponential: The exact expression, considering all exponentials, is: where the number of combinations, or ways “a” oncogenes can be selected from a set of C−1 oncogenes, without regard to the order of selection; A and B are proportional to the number of initiation events by time T for non- mutator and mutator pathways (with α = 1), respectively. The approximate expression is within 1% of the exact expressions \[12a–h\] for the vast majority of cases, and within 5% for all cases examined. N_rel, the relative efficiency of mutator compared to non-mutator pathways, is given by the following expression when the mutator mutations increase the mutation rate by a factor of α: ## Case 2: Cooperative Lineage Expansion, Early Mutator Mutation In this case, there is no change in fitness until D≤C oncogenic mutations have occurred. These first D steps, with or without a mutator mutation, can therefore be analyzed using the strategy previously outlined for the constant fitness case. In this case, for the mutator pathway, the mutator mutation is allowed to occur anywhere up to (but not including) the point where D oncogenic mutations have occurred, and these D possible time intervals are summed. Once D oncogenic mutations have occurred, the lineage has a natural log fitness advantage R, and expands by a factor of e^R per *wild type* cell generation (see equations \[2–3\] above). During this period, the fitness is also constant, although greater than it was prior to the D oncogenic mutations. The total number of cells with C−1 oncogenic mutations (by non-mutator or mutator pathways) at time T is given by the integral over t from 0 to T of the product of: the number of cells N_D−1,C ***(t)*** or N_D−1,C-mut ***(t)*** with D−1 oncogenic mutations (without or with a mutator mutation, respectively) at time t; the instantaneous mutation rate per locus for conversion to cells with D oncogenic mutations (k_mut for non-mutators, α k_mut for mutators); the number of unmutated oncogenic loci at the time of conversion to cells with D oncogenic loci (N_OL−D+1), the exponential lineage expansion of cells with D oncogenic mutations from time t to time T (e^R(T-t)); and the probability that any progeny of this expanded lineage would have acquired the final C−D−1 oncogenic mutations in time T−t (P_C−1\|D,C***(T−t)*** or P_C-1\|D,C-mut***(T−t)***) for non-mutators and mutators, respectively, adjusted for the absence or presence of mutator mutations and the increased number of cell generations per unit time, if the fitness increase includes an increase in proliferation rate: The probability of the expanded lineage acquiring C−D−1 oncogenic mutations in time T-t is the product of the number of ways of selecting C−D−1 oncogenic mutations from N_OL-D remaining unmutated oncogenic loci, ; and the single step mutation probability per locus (k_mut(T−t) for non- mutators, α k_mut(T−t) for mutators, adjusted by a factor of to account for increased number of cell generations per unit time if there is a hyperproliferative mutation), raised to the (C−D−1)st power: As in case 1, we use the number of cells with C−1 oncogenic mutations to calculate the instantaneous rate of malignant initiation events at any time, integrating that from 0 to T to obtain the number of malignant initiation events at or prior to T. N_rel, the relative number of clinical cancers due to mutator compared to non-mutator pathways at time T or earlier, is given as before by the ratio of total number of malignant initiation events at or before time T for mutator divided by non-mutator pathways (see equation \[4\]). α_50% is again derived by setting N_rel = 1. An approximate expression for α_50%, in the limit of significant lineage expansion, and increasingly accurate as the fitness advantage R and the fold mutation rate increase α get larger, is: N_rel is again given by. ## Case 3: Cooperative Lineage Expansion, Late Mutator Mutation In this case, there is no change in fitness and no mutator mutation until D≤C oncogenic mutations have occurred. These first D steps can therefore be analyzed using the strategy previously outlined for the non-mutator pathway in the constant fitness case. Once D oncogenic mutations have occurred, the lineage has a natural log fitness advantage R, and expands by a factor of e^R per *wild type* cell generation. During this time, an additional C−D oncogenic mutations will also occur, with a mutator mutation occurring anywhere from the 1^st to (C−D)th step in the process. During this period, the fitness is also constant, although greater than it was prior to the D oncogenic mutations. The total number of cells with C−1 oncogenic mutations at time T by a mutator mechanism, N_C−1,C-mut***(T)***, is given by the sum, over the possible steps k at which a mutator mutation can occur, of double integrals. These double integrals from t equals 0 to T are of the product of the number of cells with D−1 oncogenic mutations at time t, N_D−1,C-mut***(t)***; the instantaneous rate of conversion of these cells to cells with D oncogenic mutations, (N_OL−D+1) k_mutdt; the lineage expansion from time t to T, e^R(T−t); and an internal integral representing the likelihood of subsequent acquisition of the remaining C−D−1 oncogenic mutations and a mutator mutation in time T-t. This internal integral is over t' equals 0 to T-t, and the integrand is the product of the probability of having k−1 additional oncogenic mutations between time t and time t+t' before the mutator mutation occurs, P_k−1,t'; the instantaneous rate of occurrence of the mutator mutation (adjusted to account for the reduced cell generation time in the presence of a hyperproliferative mutation) at time t', e^Rp N_ML k_mut dt'; and the probability that the remaining C−D−k oncogenic mutations will occur in the remaining T−t−t' cell generations, P_{C−1,k−1,t',t}. In analogy with previous arguments, where the first term represents the number of possible combinations of k−1 oncogenic mutations from N_OL-D unmutated oncogenes, is the probability of one oncogenic locus being mutated in the time t' (given that any hyperproliferative mutation increases the mutation rate per *wild type* cell generation by a factor), and is the probability of k−1 oncogenic loci being independently mutated in this time, and The number of cells with C−1 oncogenic mutations at time T by non-mutator pathways is the same as in case 2. As in cases 1 and 2, we use the number of cells with C−1 oncogenic mutations to calculate the instantaneous rate of malignant initiation events at any time, integrating that from 0 to T to obtain the number of initiation events at or prior to T. N_rel, the relative number of clinical cancers due to mutator compared to non-mutator pathways at time T or earlier, is again given by the ratio of total number of malignant initiation events at or before time T for mutator divided by non-mutator pathways (see equation \[4\]). α_50% is again derived by setting N_rel = 1. An approximate expression for α_50%, in the limit of significant lineage expansion, and increasingly accurate as the fitness advantage R and the fold mutation rate increase α get larger, is: N_rel is given by: In the case of cooperative lineage expansion, we can determine whether an early or late mutator mutation is more efficient by comparing the respective values of N_rel. A late mutator (case 3) is more efficient (and therefore more likely) than an early mutator (case 2) if and only if: For α≫1, is well approximated by the simple expression: We see that early mutator mutations (i.e. before the fitness increase) are much more likely in the cooperative case for larger values of D. ## Case 4: Negative Clonal Selection In this case, lineages are subject to negative clonal selection (NCS), or random dominant reduced fitness (RF) mutations that are deleterious with a certain probability proportional to N_RFLN-D. Lineages with reduced fitness become extinct. The loss of fitness is described by a first order differential equation, leading to exponential decay of surviving lineages (P_S is the probability of survival), with exponent given by minus the product of the susceptibility constant β, the number 2 (given diploid cells), α (for mutator lineages only) and the number of cell generations T. In turn, β is the product of the mutation rate constant k_mut and the net number of dominant RF loci N_RFLN-D : As this case does not involve lineage expansion, the model can be expressed in terms of probabilities rather than expectation values. The probability of a malignant lineage initiation by a non-mutator pathway by time T, P_{cancer, 0, NCS}, is the product of the probability of surviving negative clonal selection for T cell generations (equation \[24a\]) and the probability of having C oncogenic mutations at time T : The maximum probability of malignancy as a function of underlying mutation rate can be found by differentiating this expression with respect to the mutation rate k_mut (bearing in mind equation \[24c\] for β), and setting this derivative equal to zero, leading to equation from the “Results” section. The optimal mutation rate for carcinogenesis calculated in this way is generally significantly higher than the wild type mutation rate. The derivation is analogous to that derived from quasispecies theory, except it considers the need to acquire C mutations rather than 1, T cell generations rather than 1, N_RFLN-D rather than the full genome length as the size of the target which can mutate to reduced fitness, and the factor of 2 to account for a diploid genome. The mutator pathway probability of carcinogenesis is evaluated assuming the mutator mutation occurs first. Loss of lineages due to NCS is more rapid after a mutator mutation, but so is the acquisition of oncogenic mutations. The probability of malignant lineage initiation with a mutator mutation as step 1, P_{cancer, 1, NCS}, is the integral from t equals 0 to T of the product of the instantaneous rate of occurrence of the mutator mutation at time t (N_ML k_mut dt); the probability of surviving negative clonal selection until time t without a mutator mutation, P_{0, t, NCS}; and the probability of acquiring C oncogenic mutations while surviving negative clonal selection in time T–t after enhancement of mutation rate by an mutator mutation, P_{C-mut, 0, t, NCS}: The probability of surviving negative clonal selection until time t, P_{0, t, NCS}, is given by \[24a\] and \[24c\] with T = t. The probability of acquiring C oncogenic mutations while surviving negative clonal selection in time T–t after enhancement of mutation rate by a mutator mutation, P_{C-mut, 0, t, NCS}, is given by the product of: the probability of surviving negative clonal selection for T-t cell generations given a mutator mutation, e^{−2αβ(T−t)}; the number of ways to choose C oncogenes from a set of N_OL oncogenes, ; and the probability of C oncogenic mutations occuring as independent events in time T−t, \[αk_mut(T−t)\]^C: The relative efficiency or probability (N_rel or P_rel) of a mutator pathway with an initial mutator mutation to that of a non-mutator pathway in the presence of negative clonal selection is the ratio of the malignant initiation probabilities for mutator vs. non-mutator pathways, and is given by: where For (α−1)βT≪1, an alternative expression for Z must be used to maintain adequate computational precision: ## Lemmas Required to Reproduce the Derivations in Cases 1–3 To reproduce the derivations above, the following identities involving factorials are required. Proofs of these identities are available on request. # Supporting Information I thank Dr. Lawrence Loeb for critical review of the manuscript, and Dr. Alfred Knudson for helpful discussions. [^1]: Conceived and designed the experiments: RAB. Performed the experiments: RAB. Analyzed the data: RAB. Contributed reagents/materials/analysis tools: RAB. Wrote the paper: RAB. [^2]: Robert A. Beckman is a stockholder in Merck & Co., Inc. and Johnson & Johnson, Inc.

# Introduction Direct reprogramming of somatic cells into induced pluripotent stem (iPS) cells by forced expression of a small number of defined factors (e.g., Oct3/4, Sox2, Klf4 and c-Myc) has great potential for tissue-specific regenerative therapies, avoiding ethical issues surrounding the use of embryonic stem (ES) cells and problems with rejection following implantation of non-autologous cells. The iPS cells have been generated from a variety of mammalian species including mice, monkeys, dogs, pigs and humans. Mouse iPS cells have been generated from cells of all three embryonic germ layers, including mesodermal fibroblasts and B lymphocytes, endodermal hepatocytes, gastric epithelial cells and pancreatic cells, and ectodermal keratinocytes. The reprogramming process appears to be highly inefficient and is likely affected by many factors, including the age, type and origin of the cells used. Recently, a “stochastic model” predicted that most or all cells are competent for reprogramming. However, the kinetics of reprogramming appear to vary when target populations from different tissues are used. Mouse hepatocytes and gastric epithelial cells appear to be more easily reprogrammed and require less retroviral integration than fibroblasts. Dermal papilla cells, which endogenously express high levels of Sox2 and c-Myc, have been reported to be reprogrammed more efficiently than skin and embryonic fibroblasts. Although the mechanisms underlying differences in reprogramming efficiency are not yet clear, some cell types might be more easily reprogrammed using specific exogenous factors than others. Importantly, the use of cell types with a high reprogramming efficiency could reduce the number of transduced factors needed, decreasing the chance of retroviral insertional mutagenesis and increasing the likelihood of ultimately replacing the remaining factors with small molecules. For future clinical application, it is therefore crucial to identify cell types that can be more easily reprogrammed; ideally, these cells should also be derived from a feasible and accessible source tissue to permit autologous use. From the standpoint of accessibility, the oral mucosa is one of the most convenient tissues for biopsy. Indeed, gingival tissues are routinely resected during general dental treatments, such as tooth extraction, periodontal surgery and dental implantation, and are generally treated as biomedical waste. Interestingly, clinical observations and experimental animal studies consistently indicate that wound healing in the oral mucosa has better outcomes than in the skin, although the healing process and sequence are similar. Therefore, it has been postulated that oral mucosal cells possess distinctive characteristics promoting accelerated wound closure. The oral mucosa is composed of a thin keratinocyte layer with underlying connective tissue. Gingival fibroblasts (GFs), which are the major constituents of the gingival connective tissue, play an important role in oral wound healing, and are phenotypically and functionally different from skin fibroblasts. The establishment of primary GF cultures is relatively simple because GFs adhere and spread well on culture plates, and proliferate well without requiring specific culture conditions. Stem cell-based therapies using bone marrow aspirates have been successfully used in dentistry to regenerate maxillary/mandibular bones and periodontal tissue –; however, bone marrow aspiration from iliac crest is not an easy operation for dentists because of limitations of the dental license and specialty. Efficient reprogramming of GFs could make the gingiva an ideal source for iPS cells that could be used for autologous cell therapy and drug screening applications, especially in dentistry. In addition, gingiva normally discarded as biomedical waste would be an ideal source of donor cells from healthy volunteers to establish an iPS cell bank for a wide range of medical applications. We hypothesized that iPS cells could be produced from fibroblasts derived from gingival tissue. Such cells could be used as a valuable experimental tool for investigating the basis of cellular reprogramming and pluripotency, with possible future clinical applications. # Results ## Generation of Mouse GF-Derived iPS Cells Mouse GF cultures were established from either palatal mucosal tissues (pGFs) or mandibular tissues (mGFs) obtained from adult male mice. After four-factor transduction, several small-cell colonies were detected in pGFs and mGFs cultures (5 passages) under phase contrast microscopy within 14 days (10 days on feeders). More than 100 colonies were obtained in each 60-mm dish of pGF (day 17) and mGF (day 21) cultures. GFP expression in the cultures was monitored under fluorescence microscopy because the correct generation of iPS cells requires the silencing of the retroviral transgenes. Ten colonies from each dish (total 20 colonies) showing lower GFP expression were picked up for further expansion in ES medium. After expansion, 5 clones (3 from pGF cultures and 2 from mGF cultures:) displaying proliferation and morphology characteristic of ES cells were selected. For three-factor transduction, a small number (approximately 50 in a 100-mm dish) of ES cell-like colonies emerged in both the pGF and mGF cultures with few background cells within 50 days after transduction. Twenty colonies in the mGF culture were then mechanically picked for expansion. Most colonies were expandable, and ten colonies were finally selected for clonal iPS cell cultures. The colonies in the selected clone cultures grew in a tight and round shape. Transmission electron microscopy (TEM) revealed tight and continual cell membrane contacts, and showed a large nuclear to cytoplasmic ratio and prominent nucleoli, representing the typical ultrastructure of mouse ES and iPS cells. We refer to these ES-like cells as pGF-iPS-4F-1, -2, -3 and mGF-iPS-4F-1, -2 cells (four-factor transduction), and mGF-iPS-3F-1 to -10 cells (three-factor transduction). ## Characteristics of Mouse GF-iPS Cells All generated GF-iPS-3F and -4F cell colonies showed robust staining for alkaline phosphatase (ALP). To confirm ES cell-like characteristics, the expression of undifferentiated ES marker genes was determined by reverse transcription-polymerase chain reaction (RT-PCR). All GF-iPS-4F cell clones expressed various markers for undifferentiated ES cells, including Nanog, ERas, Rex1 (Zfp42), and Oct3/4 (endogenous), to various extents but at lower levels than in mouse ES cells. All GF-iPS-3F cell clones expressed these ES cell marker genes at levels comparable to those in ES cells. In contrast, these genes were not expressed in parental GFs and SNLP feeder cells. Bisulfite genomic sequencing was performed to evaluate the methylation status of cytosine guanine dinucleotides (CpGs) in the promoter regions of the pluripotency-associated genes, Nanog and Oct3/4. The methylation analysis revealed the percentage methylation of CpGs in the Nanog promoter regions of mGF, GF-iPS-4F (average of 2 clones), GF-iPS-3F (average of 2 clones) and ES cells to be 73.3%, 30%, 0% and 3.3%, respectively. The respective percentages for Oct3/4 in mGF, GF-iPS-4F (average of 2 clones), GF-iPS-3F (average of 2 clones) and ES cells were 85.3%, 30.7%, 4.7% and 1.3%. These results suggest that the highly methylated CpGs in Nanog and Oct3/4 promoters of parental GF cells were demethylated, and that these promoters became active during iPS cell induction. ## Differentiation of Mouse GF-iPS Cells After 3 days of floating cultivation, the GF-iPS cells and mouse ES cells formed ball-shaped structures and embryoid bodies (EBs). Ten days after the expansion of EB-like structures from pGF-iPS-4F cells on gelatin-coated plates, the attached cells showed various morphologies , resembling neuronal cells, cobblestone-like cells, and epithelial cells. By twenty days after expansion, some clumps of cells had started pulsating, suggesting that they had differentiated into cardiomyocytes ( **and**). Thirty days after expansion, von Kossa staining revealed osteogenic cells with mineralized nodule formation. Immunocytochemistry revealed positive staining for β-III tubulin (a marker of ectoderm), α1-fetoprotein (AFP) (endoderm) and α-smooth muscle actin (α-SMA) (mesoderm) in pGF-iPS-4F-1 and mGF-iPS-3F-2 cell cultures. Other clones of GF- iPS-4F and -3F cells also showed positive staining for these proteins. These data demonstrate that GF-iPS-4F and GF-iPS-3F cells could differentiate into cells from all three germ layers *in vitro*. ## Teratoma Formation by Mouse GF-Derived iPS Cells Apparent tumor formation was observed in mGF-iPS-3F- and pGF-iPS-4F-injected mice at weeks seven and ten after injection, respectively. Extracted tumors containing GF-iPS-3F cells were larger than those containing GF-iPS-4F cells. Histological examinations showed that the tumors contained various tissues, including keratin-containing epidermal tissues (ectoderm), neural tissues (ectoderm), striated muscle (mesoderm), cartilage (mesoderm) and gut-like epithelial tissues (endoderm). These data demonstrate that the GF-iPS-3F and -4F cells generated in our study were capable of differentiating into tissues representative of the three germ layers *in vivo*. ## Germline Chimeras from GF-iPS-3F Cells GF-iPS-3F cells (C57BL/6J black mouse-derived) were microinjected into Jcl:MCH white mouse-derived blastocysts, which were then transplanted into the uteri of pseudo-pregnant Jcl:ICR white mice. This yielded 13 out of 52 (25%), 9 out of 22 (40.9%), 21 out of 34 (61.8%), 19 out of 22 (86.4%) and 16 out of 27 (59.3%) adult chimeric mice from mGF-iPS-3F-1, -2 -3, -4 and -6 cells, respectively, as determined by the coat color. Chimeric mice from the mGF-iPS-3F-3 clone were then mated with Jcl:MCH white females to verify germline transmission, and one pup obtained from the mating was derived from mGF-iPS-3F cells, as revealed by coat color in black. Taken together, these data demonstrate that GF-iPS-3F cells possess *in vivo* developmental potential comparable to that of ES cells. ## Reprogramming Efficiency of Mouse GF-Derived iPS Cells To compare the reprogramming efficiency between mouse GFs and tail-tip fibroblasts (TTFs), pGF, mGF and TTF cultures were established from the same individual mouse. The reprogramming efficiency of the pGFs, mGFs and TTFs at passage 4 was 1.2%, 0.6% and 0.1%, respectively. During the experimental period (4–10 passages), the reprogramming efficiency was the highest in the pGFs, followed by the mGFs, and then the TTFs. No ES cell-like colonies had emerged from the TTF cultures transduced after 7 passages, whereas pGFs transduced after 10 passages were still amendable to reprogramming, at a rate of 0.6%. Cell proliferation assays performed on cells at passage 5 showed that the number of pGFs and mGFs on day 8 was significantly higher than that of TTFs (*P*\<0.01). The proliferative capacity of pGFs was maintained for at least 20 passages, while that of TTFs decreased significantly after 10 passages (data not shown). Real-time RT-PCR showed that the expression level of telomerase reverse transcriptase (Tert) mRNA in pGFs and mGFs was maintained for 6 passages, while that in TTFs decreased as the passage number increased. Similar expression levels were detected for c-Myc, Klf4, Sox2, p53 and p21 genes among the pGFs, mGFs and TTFs, although the expression levels of klf4, p53 and p21 in GFs were slightly higher than in TTFs at passage 6. Expression of Oct3/4 mRNA was not detected in the primary GFs and TTFs. ## Induction of Human Gingival Fibroblast (hGF)-Derived iPS Cells When gingival tissues from the patient were cultured on a gelatin-coated dish, fibroblasts and epithelial cells proliferated out of the tissues. Homogeneous fibroblast culture was established in serum- and calcium- containing media. Four-factor-transduced cells yielded iPS cell-like colonies; these colonies were picked mechanically and five clone cultures (hGF-iPS-547A-1 to -5) were established. The colonies could be expanded and displayed the same morphology and growth characteristics as colonies of iPS cells obtained from human dermal fibroblasts and human ES cells. ## Characteristics and Differentiation of the hGF-iPS Cells The colonies of all generated hGF-iPS cell clones stained positively for ALP activity. RT-PCR analysis showed that all hGF-iPS cell clones expressed ES cell specific genes, including NANOG, REX1, TERT, endogenous OCT3/4 and SOX2, at levels comparable to those in H9 and KhES3 human ES cell lines. In contrast, these genes were not expressed in parental hGFs or SNLP feeder cells. Bisulfite genomic sequencing revealed the percentage methylation of CpGs in the NANOG promoter regions of parental hGF, hGF-iPS (average of 2 clones) and H9 human ES cells to be 32.8%, 3.2% and 6.3%, respectively. The respective percentages for OCT3/4 in hGF, hGF-iPS (average of 2 clones) and H9 human ES cells were 64.6%, 5.2% and 7.3%. Cloned hGF-iPS-547A-2 and -3 cells formed teratomas after injection into SCID mouse testes. Histological examination of the teratomas at week nine after injection revealed representative tissues originating from the three embryonic germ layers, including ectodermal neural tissues, mesodermal cartilage and endodermal gut-like epithelial tissues. Therefore, like human ES cells, the hGF- iPS cells generated in this study had the capacity to differentiate into tissues representative of the three germ layers *in vivo*. # Discussion Tissue engineering become a new frontier in dentistry for, among other applications, regeneration of missing oral tissues. However, engineering applications for tooth, jawbone, temporomandibular joint cartilage and periodontal tissues await the establishment of a stem cell source that allows easy collection by dentists. In view of the potential clinical applications of iPS cells, various types of discarded or easily obtainable normal human tissues, especially those that can be obtained with minimal patient discomfort such as peripheral blood, should be considered as potential sources of iPS cells. From a practical standpoint, the complicated cell isolation process, low numbers of isolated cells and slow proliferation necessitate a long-term *ex vivo* expansion step for obtaining sufficient cells for iPS induction. Such a step is costly, time-consuming, and increases the risk of cell contamination and loss. Several sources have shown more efficient iPS cell generation, such as human keratinocytes from hair follicles or epidermal biopsies as well as mesenchymal stem cells of dental origin from dental pulp, and impacted third molars. However, there are several practical limitations to using these cells for tissue engineering. For example, expansion of keratinocytes requires serum-free low- calcium medium, which is relatively costly, to prevent terminal differentiation. Isolation of the dental stem cells may not be sufficiently convenient to allow easy harvesting whenever the cells are needed because it requires tooth or pulp extraction surgery, and the missing tissues can not regenerate. In addition, culture of these non-terminally differentiated cells requires a high level of skill to adequately maintain cellular homogeneity. Similar limitations apply to many other possible sources of iPS cells. An ideal autologous source should thus allow easy collection of a large number of cells that can be grown in a simple culture system, and that can quickly be cultured to quantities sufficient to obviate extensive and long-term expansion. Oral gingival tissue may represent such a source. In this study, GFs from both mouse and human gingival tissues were easily established. The mandibular mucosa of a 10-week-old mouse is too small to extract only gingival tissues. Therefore, muscle and bone tissues around the mandibular gingival tissues were carefully removed, and mGFs were obtained from the outgrowth of fibroblastic cells from the remaining tissues. This technical limitation may have resulted in the possible contamination of mGFs with some myoblasts or bone marrow stromal fibroblasts. On the other hand, the mouse palatal mucosa was easily extracted en bloc without any contamination by surrounding tissues. Therefore, the pGFs used in this study may have been more homogeneous than the mGFs. Nevertheless, both types of adult mouse primary GFs proliferated well and could be successfully reprogrammed into iPS cells that differentiated into cells and tissues representing all three germ layers *in vitro* and *in vivo*. All reprogramming approaches investigated to date seem to involve epigenomic modification. GF-iPS-3F showed a greater decrease in the CpG methylation ratio in the promoter regions of Nanog and Oct3/4 in comparison to GF-iPS-4F cells that resulted in a methylation pattern similar to that of mouse ES cells. Nakagawa *et al.* reported that the omission of c-Myc transduction resulted in the generation of high-quality iPS cells from mouse TTFs, in which Nanog is strongly activated and the retroviruses are silenced. Consistently, our results showed that GF-iPS-3F cell clones highly expressed ES cell marker genes, including Nanog and endogenous Oct3/4. On the other hand, GF-iPS-4F cell clones expressed these genes at lower levels than in mouse ES cells and showed partial DNA methylation in restricted areas of the promoters. These results may in part be due to incomplete reprogramming effects of four-factor transduction during GF-iPS-4F cell induction in our system, which did not utilize a specific system for reprogrammed cell selection (e.g., drug selection and mice genetically modified for Nanog expression). Because incompletely reprogrammed GF-iPS-4F cells at least possess multipotency, they might be used in some tissue regenerative approaches; however, concerns remain that reactivation of the c-Myc oncogene in iPS cells could increase tumorigenicity, thereby hindering potential clinical applications. c-Myc has recently been shown to be dispensable for direct reprogramming. It is conceivable that one major function of c-Myc is to enhance proliferation, thereby accelerating the reprogramming process, possibly by increasing the speed of stochastic events that lead to the formation of iPS cells. Consistently, the three factors without c-Myc were able to initiate a slower reprogramming process that was sufficient to fully reprogram mouse GFs after a longer time period. Indeed, GF-iPS-3F cells demonstrated almost complete DNA demethylation in the promoter regions of Nanog and Oct3/4. In addition to the specific induction, the long reprogramming time course in GF-iPS-3F cells may be responsible for the low level of methylation in the promoters. All five tested GF-iPS-3F clones readily produced viable chimeric newborn and adult mice. Moreover, mGF-iPS-3F-3 cells contributed to germline transmission, which definitively demonstrates that they were pluripotent and functionally indistinguishable from ES cells. So far, 4 out of the 78 mGF-iPS-3F cell chimeras (5%) produced in the study died within six months; however, apparent tumors were not observed in the dead mice. These results suggest that high-quality iPS cells can be generated from adult mouse GFs by transduction of the three factors (Oct3/4, Sox2 and Klf4) without any specific system for the selection of reprogrammed cells. However, because Klf4, the remaining oncogenic factor, or insertional mutagenesis due to retroviral transduction itself might also cause tumor formation, it will be important to investigate the possibility of using recombinant proteins (and small molecules) to reduce the number of genetically transduced factors required for iPS cell induction, or even to entirely obviate the need for viral gene delivery. TTFs were the first type of adult cells to be reprogrammed into iPS cells. Since then, other adult cell types with the potential for easier reprogramming have been tested. Our results show that pGFs can be reprogrammed to pluripotency at least 7-fold more efficiently than TTFs at the same passage number derived from the same mouse. In addition, GFs, especially pGFs, maintained their high reprogramming efficiency for at least ten passages. On the other hand, no-ES cell-like colonies emerged from TTFs transduced after 7 passages, possibly due to replicative senescence. The lower reprogramming efficiency of mGFs compared to pGFs in our system may have been due to the establishment of a heterogeneous cell population in the mGFs. It was initially speculated that the high efficiency of iPS cell generation from GFs might have been due to high endogenous expression levels of at least one of the four defined pluripotency-inducing factors, or due to reduced activation of the p53 pathway. However, no significant differences in the endogenous expression of the four factor genes or of p53 or p21 were detected. On the other hand, GFs showed significantly greater proliferation in comparison to TTFs. Additionally, pGFs proliferated well for at least 20 passages, while the proliferation of TTFs decreased after 10 passages (data not shown). Moreover, GFs consistently showed higher expression of Tert mRNA compared to TTFs. Tert is one of the major subunits in the telomerase complex, and the transcription of Tert gene correlates with telomerase activity in most cells. The high expression of the Tert gene in GFs may therefore explain their high proliferative capacity. The high proliferative capacity of the GFs should be advantageous for retroviral integration due to increased likelihood of cell division during transduction. The higher efficiency of GF reprogramming, therefore, may at least partially be due to the high proliferation rate of the GFs. Wounds in the oral mucosa show faster closure with less scar formation than skin wounds, partly because oral GFs differentially express early wound closure-related genes, such as FGFR1OP2/wit3.0. It is therefore possible that intrinsic differences in gene expression patterns between the GFs and TTFs may also underlie differences in reprogramming efficiency. We also demonstrated that iPS cells could be generated from human gingival tissues, which underscores the potential value of this promising cell source for human applications. When we transduced human dermal fibroblasts (HDF1388) in parallel with human GFs under the same experimental setting and infection protocols, fewer ES cell-like colonies emerged from the dermal fibroblasts, suggesting that human GFs might be more readily reprogrammed into iPS cells. Yan *et al.* recently reported that dental stem cells can be efficiently reprogrammed into iPS cells by lentiviral transduction of LIN28/NANOG/OCT4/SOX2 and by retroviral transduction of c-MYC/KLF4/OCT4/SOX2. However, their protocol did not generate ES cell-like colonies from human gingival fibroblasts and foreskin fibroblasts. They also indicated that dental stem cells express a number of ES cell-associated genes, thus suggesting that these stem cells have epigenetic advantages for reprogramming. Therefore, the reprogramming efficiency of terminally differentiated GFs may be inferior to that of undifferentiated stem cells. It is unknown at present whether iPS cells derived from different types of cells behave in the same manner. Specifically, iPS cells from different cell types may differ in their ability to undergo guided differentiation. Therefore, GF-iPS cells should be further characterized and compared to ES cells and iPS cells derived from other sources. Potential differences in the reprogramming efficiency between cells isolated from humans and mice also remain to be elucidated. Nonetheless, the high replication capacity of GFs should permit not only the generation of sufficient cells for iPS cell induction, but also the efficient generation of iPS cells from multiple expanded cell cultures. The intrinsic features of GFs from easily obtainable gingival tissues could be of benefit for regenerative medicine and drug screening, especially in dentistry, as it is easy for dental associates to establish primary cell cultures with minimal patient discomfort. Additionally, establishment of iPS cell banks with various human leukocyte antigen (HLA) types should be useful for general regenerative medicine, as the establishment of clinical-grade iPS cell lines from individual patients would require much time and high cost. Collection of gingiva considered until now to be biomedical waste from healthy volunteers and efficient iPS cell generation from this tissue may allow the development of a cell banking system for a wide range of medical applications. In conclusion, the efficient reprogramming of mouse gingival fibroblasts to pluripotency is expected to provide a valuable experimental model for investigating the basis of cell source-dependent cellular reprogramming and pluripotency, which may thus lead to a practical alternative for the generation of patient- and disease-specific pluripotent stem cells. # Materials and Methods ## Ethics Statement All animal experiments in this study strictly followed a protocol approved by the Institutional Animal Care and Use Committee of Osaka University Graduate School of Dentistry (approval number: 20-009). Written approval for human gingival tissue collection and subsequent iPS cell generation and genome/gene analyses performed in this study was obtained from the Institutional Review Board at Osaka University Graduate School of Dentistry (approval number: H21-E7) and the Ethics Committee for Human Genome/Gene Analysis Research at Osaka University (approval number: 233), and written informed consent was obtained from each individual participant. ## Cell Culture Mouse GF cultures were established from either pGFs or mGFs obtained from 10-week-old adult male C57BL/6J mice. To establish TTF cultures, tails from mice were peeled and minced into 1-cm pieces. The extracted palatal and molar mucosal tissues or minced tails were placed on a 0.1% gelatin-coated 30-mm tissue culture dish and maintained in MF-start medium (Toyobo, Osaka, Japan) at 37°C with 5% CO₂. When fibroblasts migrated out of the tissues, the tissues were removed. When the cells reached subconfluence, they were harvested and transferred to a gelatin-coated 60-mm tissue culture dish (Passage 1) and cultured in “fibroblasts and Platinum-E (FP) medium”, which consists of Dulbecco's Modified Eagle's Medium (DMEM without sodium pyruvate; Nacalai Tesque, Kyoto, Japan), 10% fetal bovine serum (FBS; Sigma, St. Louis, MO), 50 units/ml penicillin, and 50 µg/ml streptomycin (Nacalai Tesque). Culture in serum- and calcium-containing media favorably selects fibroblasts from heterogeneous populations of migrating epithelial cells and fibroblasts in the mucosal tissues. Primary hGF cultures were established according to the previously described protocol from healthy gingival tissues discarded during surgery on a 24-year-old man. SNLP76.7-4 feeder cells and the mouse ES cell line (AB2.2) were graciously supplied by Dr. Allan Bradley of the Sanger Institute (London, UK). Platinum-E packaging cells for retrovirus production were graciously supplied by Dr. Toshio Kitamura (University of Tokyo, Japan). Human ES cell line H9 was obtained from WiCell™ Research Institute (Wilmington, MA), and human ES cell line KhES-3 and human dermal fibroblast-derived iPS cells were obtained from Institute for Frontier Medical Sciences, Kyoto University. The human ES cells were treated according to the guidelines for utilization of human ES cells established by the Ministry of Education, Culture, Sports, Science and Technology, Japan. ## Retrovirus Production Moloney murine leukemia virus (MMLV)-based retroviral vectors (pMXs-IRES-puro) containing murine and human c-Myc, Oct3/4, Sox2 or Klf4 cDNA were obtained from Addgene (Cambridge, MA), and the pMX-GFP retroviral vector was purchased from Cell Biolabs (San Diego, CA). Nine micrograms of each plasmid vector were separately added to tubes containing Opti-MEM-I mediun (Invitrogen) and FuGENE 6 transfection reagent (Roche, Basel, Switzerland); each plasmid was then separately transfected into 100-mm dishes containing Platinum-E packaging cells. The transfection efficiency was monitored by evaluation of GFP expression under a fluorescence microscope. The efficiency of transfection into Plat-E cells was typically \>60%, as indicated by GFP expression. The next day, the culture medium was exchanged for fresh FP medium. After 24 hours, the virus-containing supernatants were mixed together and used for retroviral transduction. ## Induction of iPS Cells Twenty-four hours before transduction, mouse pGFs and mGFs (5 passages) were seeded at 5×10⁵ cells per 100-mm dish in FP medium containing 3 ng/ml bFGF (Peprotech, London, UK). For the four-factor transduction, supernatants with retroviruses coding c-Myc, Oct3/4, Sox2, Klf4 and GFP were mixed at a ratio of 1∶1∶1∶1∶3. When the mGFs were transduced with only three factors, the supernatants containing retroviruses coding Oct3/4, Sox2, Klf4 and GFP were mixed at a ratio of 1∶1∶1∶3. The cells were incubated overnight in the virus/polybrene (4 µg/ml)/bFGF (10 ng/ml)-containing supernatants. On days one and three after transduction, the culture medium was exchanged for fresh FP medium containing 3 ng/ml bFGF. At four days after transduction, the cells in the culture dishes were re-seeded onto 60-mm dishes at 0.1–1×10³ cells/cm² for the four- factor transduction, and onto 100-mm dishes at 0.7–1×10⁴ cells/cm² for three-factor transduction; mitomycin C-inactivated SNLP76.7-4 cells were used as a feeder layer. The next day, the culture medium was exchanged for “ES medium”, which consisted of DMEM, 15% FBS, 2 mM L-glutamine, 1×10⁻⁴ M nonessential amino acids, 1×10⁻⁴ M 2-mercaptoethanol, 50 U penicillin, and 50 µg/ml streptomycin. The medium was changed every day. The three-factor-infected GFs were harvested with deteriorated feeder cells at 30 days after transduction and then were re-seeded onto a new feeder layer. The colonies demonstrating minimal GFP expression were identified and mechanically picked for expansion. After expansion, clonal colonies showing ES cell-like proliferation and morphology, including a round shape, large nucleoli and scant cytoplasm, were selected for establishing clonal iPS cell cultures. To demonstrate the expression of the ES cell marker ALP in GF-derived iPS cell colonies, a standard ALP staining protocol was used. TEM was used to determine the structure of individual cells from the GF-derived iPS cell colonies. The colonies were fixed first with 1% paraformaldehyde and 1.25% glutaraldehyde in phosphate-buffered saline (PBS), and second with 2% osmium tetroxide. The fixed colonies were then embedded in epoxy resin and stained with methylene blue, followed by 1-µm sectioning of the resin for optical microscopy. Additionally, 70-nm sections were stained with 2% uranyl acetate and Sato's lead stain, and examined using a JEM 1200EX (JOEL, Tokyo, Japan) operated at 80 kV for TEM. Induction of iPS cells from primary hGFs (7 passages) via introduction of four factors was performed using a previously described protocol. According to this protocol, hGFs were first infected with a lentivirus to express the mouse Slc7a1 gene (by a plasmid vector from Addgene) and then infected with retroviruses coding human c-MYC, OCT3/4, SOX2 and KLF4 genes for iPS cell induction. ## RT-PCR Analysis Total RNA derived from mouse or human clonal GF-derived iPS or ES cell colonies was used for RT-PCR analysis. Total RNA was extracted with an RNeasy Mini Kit (QIAGEN, Hilden, Germany). After DNase I treatment (Ambion, Austin, TX), cDNA was synthesized from 1 µg of total RNA using Super Script III reverse transcriptase (Invitogen, Carlsbad, CA). The cDNA target was amplified by PCR using Taq DNA polymerase (Promega, Madison, WI) following the manufacturer's recommendations. The primer pairs used are given in. PCR products were subjected to 1.5% agarose gel electrophoresis with ethidium bromide staining and visualized under ultraviolet light illumination. The expression of glyceraldehyde-3-phosphate dehydrogenase (GAPDH) mRNA was used as an internal control. ## *In vitro* Differentiation of Mouse GF-Derived iPS Cells To determine the differentiation ability of GF-derived iPS cells *in vitro*, we used floating cultivation to form EBs. For EB formation, mouse GF-derived iPS cells were harvested by trypsinization and transferred to low-attachment bacterial culture dishes in the ES medium. After 3 days of floating cultivation to form EBs, aggregated cells were plated onto gelatin-coated 8-well glass chamber slides (Nalge Nunc International, Naperville, IL) or 12-well tissue culture plates, and incubated in ES medium. The culture medium was changed twice a week. For immunocytochemistry, cells cultured in glass chamber slides for 3 to 10 days after expansion were fixed in 10% buffered formalin phosphate (Wako, Osaka, Japan) and incubated in 1% bovine serum albumin and 0.1% Triton-X100 in PBS for 20 min. After two washes in PBS, the cells were incubated with a mouse anti- human α-SMA monoclonal antibody (0.05 mol/L; clone 1A4, Dako, Glostrup, Denmark) or rabbit anti-human AFP polyclonal antibody (0.05 mol/L; Dako) for 30 min at room temperature, or a mouse anti-human β-III tubulin monoclonal antibody (0.5 µg/ml; clone TU-20, Millipore, Temecula, CA) or control IgG (0.5 µg/ml; mouse IgG whole molecules: Santa Cruz Biotechnology, Santa Cruz, CA) overnight at 4°C. The cells were then washed and incubated for 30 min at 37°C with Alexa 488 (green dye) or 568 (red dye) conjugated to goat anti-mouse or anti-rabbit IgG (1∶500; Molecular Probes, Eugene, OR), followed by 4,6′-diamidino-2-phenylindole (DAPI: Roche) nuclear staining. For osteogenic cell detection, EBs cultured in gelatin-coated 12-well tissue culture plates in ES medium for 30 days were stained using a standard von Kossa staining method to demonstrate the extent of nodule mineralization. ## Bisulfite Genomic Sequencing Genomic DNA was isolated from the aggregated mouse GF-derived iPS cells and ES cells floating for 3 days or from the attached human GF-derived iPS cells and H9 ES cells harvested by CTK solution. Information about the promoter regions and CpG loci of Nanog and Oct3/4 was obtained from a previous study and the Data Base of Transcriptional Start Sites (DBTSS Ver. 7.0: <http://dbtss.hgc.jp/>). Bisulfite treatment was performed using the EpiTect Bisulfite kit (Qiagen) according to the manufacturer's recommendations. Bisulfite PCR primers, are listed in. Amplified products were cloned into the pGEM-T Easy Vector (Promega). Five to eight randomly selected clones were sequenced with the SP6 forward and reverse primers for each gene. ## Teratoma Formation and Histological Analysis Eight-week-old immunodeficient mice (C.B-17 SCID; Clea Japan, Tokyo, Japan) were anesthetized with diethyl ether and an intraperitoneal injection (0.1 ml per 100 g body weight) of a 10× dilution of Nembutal (Dainippon Sumitomo Pharmaceutical, Osaka, Japan). Twenty microliters of a mouse or human GF-derived cell suspension (0.2–0.5×10⁶ cells/testis) in cold Hank's balanced salt solution or DMEM/F12 (Invitrogen) were injected into the medulla of mouse testes using a Hamilton syringe. The mice were thereafter housed with free access to water and food under specific pathogen-free conditions. After 7–10 weeks, the teratomas were excised after perfusion with PBS followed by a fixative solution containing 1% paraformaldehyde and 1.25% glutaraldehyde, and subjected to histological analysis. Specimens were embedded in paraffin, and sectioned at 3 µm for hematoxylin and eosin (H&E) staining. ## Chimera Formation Superovulation \[intraperitoneal administration of 5 I.U. pregnant mare serum gonadotropin (PMSG) followed after 48 hr by 5 I.U. human chorionic gonadotropin (hCG)\] was induced in eight-week-old female mice \[Jcl:MCH (ICR), CREA Japan\], which were then mated with Jcl:MCH (ICR) males. Embryos at the 2-cell stage were collected at day 1.5 after vaginal plug observation and flushed in M2 medium (Sigma). Embryos were then cultured in KSOM culture medium (Chemicon) in the incubator (37°C, 5% CO₂ in air) until they became blastocysts. Mouse GF-derived iPS cells were harvested using 0.25% trypsin to obtain a single cell suspension. Single cells were then transferred into the micromanipulation chamber in a drop of DMEM medium containing 10% fetal calf serum and 15 mM HEPES. Groups of 20 to 25 cells were injected into each single blastocyst. The injected embryos were then transplanted into 2.5 dpc pseudopregnant Jcl:ICR recipient females. Chimeric male mice were mated with female mice \[Jcl:MCH (ICR)\] to validate germline transmission. ## Determination of Reprogramming Efficiency TTF, pGF and mGF cultures were established from the same individual mouse (10 weeks of age). Four-factor transduction (without GFP) was performed using cell cultures with identical passage numbers. The transduction of each cell type was performed simultaneously using the same virus-containing supernatants. The cell cultures used for the comparisons were between passage numbers four and ten. Each cell culture was then seeded in 6-well tissue culture plates (1×10⁴ cells/well) with the feeder cells. iPS cell colonies were identified based on ES cell-like morphology, and ALP staining was used to facilitate their identification. The reprogramming efficiency was calculated as the number of iPS colonies formed per number of transduced cells seeded. A cell proliferation assay was performed on pGF, mGF and TTF cultures with identical passage numbers. The cells were seeded in 96-well tissue culture plates (2×10³ cells per well) and maintained in FP medium. The culture medium was renewed every other day. The number of cells was evaluated using the WST-1 cell counting assay (Dojindo Laboratories, Kumamoto, Japan) as described previously. The endogenous mRNA expression of Oct3/4, Sox2, Klf4, c-Myc, p53, p21 and Tert in pGFs, mGFs and TTFs at passages 4 to 6 was determined by real-time RT-PCR analysis. TaqMan primer and probe sets used are Mm00488369_sl (Sox2), Mm00516105_gl (Klf4), Mm00487804_ml (c-Myc), Mm00441964_g1 (p53), Mm00432448_m1 (p21), Mm00436931_m1 (Tert) and 4352339E (GAPDH). # Supporting Information We thank Dr. Jiro Miura and Shinya Uraguchi (Osaka University School of Dentistry) for technical assistance and Dr. Devang Thakor (Harvard Medical School) for scientific comments. We are also grateful to Dr. Yasuhiko Tabata (Institute for Frontier Medical Sciences, Kyoto University) and Dr. Tetsuya Ishii (CiRA, Kyoto University) for their valuable support to initiate this work. [^1]: Conceived and designed the experiments: HE KO SY. Performed the experiments: HE KO HK GY SF. Analyzed the data: HE. Contributed reagents/materials/analysis tools: HE KO MS TM SY. Wrote the paper: HE. Administrative support: HY. [^2]: The authors have declared that no competing interests exist.

# Introduction One of the key insights of cultural evolutionary theory is that cumulative culture crucially depends on demography. Indeed, a wide variety of models of cultural transmission has reproduced the result that changes in population size may drive cultural change: increases in the former favour cumulation, while decreases may occasion cultural loss (–, but see). These model-theoretical findings are used to explain particular cultural transitions, e.g., the loss of culture in Holocene Tasmania, the Upper Paleolithic transition, or the growth of scientific knowledge since the Industrial Revolution. Interestingly, although said models express cumulation in terms of complexity increases, they differ considerably in how they construe the latter term (i.e. complexity). Characterizations have been given in terms of change in fitness ; in terms of transmission inaccuracy –; and in terms of the number of elements a cultural trait consists of. This may be an asset rather than a drawback: if there is convergence of results, demographic change may be offered as a possible explanation of a broad suite of patterns of cultural change, viz. all patterns that can be plausibly construed as cumulative under one of the various characterizations on offer. This is especially pertinent in case empirical evidence is not sufficiently abundant to prefer one particular construal of cultural change. Consider for instance the Upper Paleolithic transition, which according to Powell et al is characterized by "substantial increase in technological and cultural complexity, including the first consistent presence of symbolic behavior, such as abstract and realistic art and body decoration \[…\], systematically produced microlithic stone tools \[…\], functional and ritual bone, antler, and ivory artifacts, grinding and pounding stone tools, improved hunting and trapping technology \[…\], an increase in the long-distance transfer of raw materials, and musical instruments." As intuitively plausible as this may seem, it still needs to be established that the Late Pleistocene cultural complexity referred to by Powell and colleagues really is adequately captured by the characterization of complexity assumed in current models (including their own). For example, does the emergence and consistent presence of symbolic behaviour demonstrate that cultural skills become more complex in the sense of becoming harder to transmit faithfully? That is, was Upper Paleolithic symbolic behaviour actually more error-prone than previous (non-symbolic) behaviour? Or is this transition better understood in terms of an increase in the number of elements cultural behaviours encompassed, with symbolic behaviours encompassing more elements than previous (non-symbolic) behaviours? Does the same apply for long-distance transfer of raw materials? For hunting technologies? Currently available evidence does not afford conclusive answers to these questions; arguably, the answers are underdetermined by any conceivable evidence. In this epistemic situation, the diversity of characterizations of complexity—and of model assumptions more generally—may save the day: the larger the set of intuitively plausible definitions of complexity in the family of demographic models, the more likely it is that at least one empirically valid characterization of the considered pattern of change is included in the set. That is, the more robust the relation between population size and cumulative cultural change is under variations of characterizations of cumulation and of modelling assumptions, the more credible or widely applicable are demographic explanations of cultural change. Conversely, if the dependence fails to obtain under some characterizations or auxiliary assumptions, an episode of cultural change can only be justifiably attributed to demographic change in case empirical evidence speaks against these unfavourable characterizations or assumptions. In this paper, we offer a cultural-evolutionary model that is based on yet another characterization of complexity, and we examine whether it can be safely added to the family of models that show a relation between cultural and demographic change. If it can, the dependence of cumulative culture on demography would stand even firmer. If it can not, patterns of cultural change that are cumulative in our newly introduced terms have not yet been shown to be susceptible to demographic explanation. We characterize cumulation, like some existing contributions to the literature, in terms of increasing complexity. Yet rather than characterizing transmission accuracy or the sheer number of elements in a cultural trait, we follow Herbert Simon in taking complexity to consist in the density of *interaction* between the elements of a trait: cultural change is cumulative in case the transmission of cultural traits sustains ever more intricate interdependencies between the components or elements of these traits. To illustrate the plausibility of the assumption, consider the production of early stone tools. presents action hierarchies for Oldowan and Late Acheulean flake detachment (after ; see also references therein). The latter is more complex than the former not primarily because it has more constitutive elements, but rather because these elements are organized in a more elaborate hierarchical structure that comprises more nested levels: the addition of platform preparation to the superordinate goal of percussion in Late Acheulean flake detachment introduces four nested levels, so that the method contains six nested levels in total (versus four in Oldowan flake detachment). The success of the superordinate level (i.e. percussion) is thus contingent on four elements (rather than three, as in Oldowan production), namely position core, hammerstone grip, strike *plus* platform preparation, and the success of the latter is itself contingent on the interplay of a whole set of lower-level actions. Action hierarchies for multi-component blade technologies would be even more intricate. Here the success of percussion would, for example, depend on stringent selection/importation of raw materials – and on the properties of other components (such as the haft). The more intricately these elements are interrelated, the more difficult it becomes to predict how changes in one place will affect elements elsewhere in the hierarchy. Even a very small error introduced during transmission in one element may have profound repercussions on the performance of other elements, and thus on success overall (which in turn makes it difficult to predict expected changes through time, see ). Therefore, in cases like these, complexity defined as transmission inaccuracy and complexity as defined by Simon (let us call it S-complexity) need to be carefully distinguished. The literature on cultural evolution has not overlooked S-complexity. Most importantly, the intricate interaction between the material components of technologies has been offered as evidence that cumulative cultural changes can not have been the result of individual learning: "\[Several technologies\] are very complicated artifacts with multiple interacting parts made of many different materials. (…) Determining the best design is, in effect, a high dimensional optimization problem that is usually beyond individual cognitive capacities…". Similarly, Mesoudi and O'Brien have introduced a two-peaked fitness landscape in their experimental study of the transmission of projectile- point design. Despite this supporting role of S-complexity in cultural- evolutionary theory, it has not yet been implemented in cultural-evolutionary models. There is at least one reason to expect that doing so has strong, and negative, repercussions for demographic explanations. The intuition behind the dependence of cumulative cultural change on demography is strength in numbers: larger populations can sustain a more complex culture simply because they are more likely to contain individuals whose cultural traits are at least as good as those of the individuals they imitated from previous generations. However, under the assumption that mentor selection is imperfect, populations as a whole profit from such outstanding individuals only insofar as these can be detected by others as suitable objects of imitation—and here larger populations are clearly at a *dis*advantage. At high levels of S-complexity, this disadvantage may become insurmountable. For under these conditions, only very sizable populations will contain any individuals who have been able to avoid the slight transmission inaccuracy in cultural transmission that would have interfered with overall success of the complex cultural trait. These individuals might not in turn be able to transmit their traits, however, since potential students might be unable to find them in a population of this size. # Materials and Methods We devised a model with successive generations of agents, each of whom learns from a parent of the previous generation through oblique, pay-off based transmission—much as in. Adjustments to these models, required to implement S-complexity, were based on Stuart Kauffman's so-called *NK*-logic. In line with Herbert Simon's proposal, here stands for the number of components, whereas expresses the level of interaction between components. Let us explain the model in more detail. The agent-based model (implemented in Netlogo, code available from the authors) contains a population of agents, each of whom exhibits a variant of a cultural trait (say, a technology or technological skill). The configuration of any variant is given by a binary string of elements. For example, if, the string *01101* would refer to a variant which differs in the second element from a variant characterized by *00101*. We follow Kauffman in assuming that each element can be only in two states (*0* or *1*). Although it is in principle possible to extend the model assuming any number of possible states, working with binary values is intuitive enough. For example, consider percussion in Late Acheulean flake detachment, which consists of four elements: platform preparation, positioning of the core, holding the hammerstone, and striking. For each of these actions we could assign a *1* when the action is executed in one way, and a *0* when executed in another way. Percussion on a prepared platform would then be represented by the string *1111*, percussion on the ground by the string *0111*, and so forth. Evidently, one could increase the level of precision by adding more elements. For instance, one could characterize platform preparation with a three-element string, stating values for hammerstone selection, positioning of the core and light percussion. To any variant is assigned a skill or fitness value, which is defined as the average of the *contributing values* of each of the individual elements; the contributing value of each element is in turn determined by its own state (*0* or *1*), and by the state of other elements. In case of, changing the state of one element (e.g., from *0* to *1*), will only affect the value of the element itself. In case of, the contributing value of an element may change directly or as a result of a change in the single other element influencing it. The interaction parameter can be any natural number between *0* and. By changing the parameters *N* and *K* we can tune the complexity of cultural traits. The least complex trait is one for which and, per definition. Complexity increases with increasing ; further, for any, maximal complexity is obtained when. Under the latter circumstances, a change in one element affects the fitness of all other elements. In line with previous research, we use transmission inaccuracy as an additional measure of complexity. However, whereas transmission inaccuracy has been previously expressed only as the magnitude of the error naïve individuals make, in our paper it is determined by a copy error rate, i.e. a number expressing the *probability* that a transmission error will be made, where an error consists in unfaithful replication of an element of the copied sequence. Multiple errors (i.e. elements changing states) may occur in the same transmission attempt. The *impact* of these errors can, like, be expressed as a real number between 0 and 1. illustrates the *NK*-logic, as well as transmission over successive generations, for a population size. On the left is represented a trait with three elements without interdependencies between elements. The variant of the trait in Generation 1 is *000*, with an average fitness of 0.4. Now the offspring in Generation 2 will try to imitate the cultural parent in Generation 1. Note that in case, offspring will get the opportunity to select a cultural parent, but more on this below. Imitation happens with copy error rate, which is in our model defined as the probability of an element changing state in a transmission attempt. So for, any element has a 1% chance of changing state (from *0* to *1* or from *1* to *0*). In the worked-out example, we assume deterministic inaccuracy for the sake of clarity: in each generation, exactly one element changes state. From Generation 1 to Generation 2, the third element is copied inaccurately, and thus receives a new contributing value (from 0.3 to 0.6, the magnitude of the error being 0.3). Since, this mutation does not affect any other element in the string. Compare this with the right-hand side of, where, which means that each element interacts with one other element in the string. In this example, these interactions are one-sided influences, of the first element on the second element, of the second on the third, and of the third element on the first. The interactions in this example are regular, but this need not to be the case. For if, for example, an element has three links coming in, but an *average* of links coming out. Now, for example, if the third element changes state, as when going from Generation 1 to Generation 2, the value of that element changes (from 0.3 to 0.6) and so does that of the first element (from 0.7 to 0.3). In this case, the improvement with regard to the third element does not lead to an increase in overall skill or performance value, since this improvement interferes negatively with the contributing value of the first element. When that element changes state, as it happens from Generation 2 and Generation 3, its value increases, but now the second element is maladjusted (its contributing value drops to 0.0; note that our model also allows for positive interferences). Even incremental innovation, characterized by deliberately changing one component or constitutive action at a time, is therefore a very delicate matter; although one element may contribute positively to overall performance, there is a chance that it does do so at the price of lowering the contribution of several other elements. Consequently, for higher and, it gets increasingly harder for agents to find a configuration which outperforms its predecessor, and even small copy error rates may have large detrimental consequences. As a result, only very sizable populations will be lucky enough to contain an individual who does better than its parent. Simulations proceeded as follows. We first generated *NK*-tables. An *NK*-table lists all string combinations for *N* elements, each with a corresponding, predefined contributing value (initially being drawn from a Uniform \[0,1\]; later we also considered Normal \[0,1\] and Gumbel \[0,1\] distributions). The string combinations with corresponding overall values on the right-hand side of could be interpreted as representing *part* of a predefined *NK*-table for. If an agent would try out a certain configuration, say *101*, its overall performance would be simply given by the average fitness value given in the *101*-row of the *NK*-table (i.e. in the example). We generated 200 *NK*-tables for each combination and, which amounts to 6,000 *NK*-tables in total. Populations of size (and for a selected range of parameter settings of size ; more on this in the Results section) had to "explore" these 6,000 *NK*-tables. Simulations were initialized at step 1 by assigning to each agent of Generation 1 the same string of size, with a configuration and fitness randomly drawn from the *NK*-table under consideration. Each one of the next steps of the simulation consisted of the following sub-steps: 1. a new generation of agents is introduced; 2. each individual of the new generation selects a cultural parent from the previous generation, and this depending on the parent's fitness; 3. the individual copies the selected parent's trait with copy error rate ; 4. each individual receives a fitness based on its acquired trait; 5. the new generation replaces the parent generation and the average and maximum fitness of the population, and, are measured. Pay-off biases are thus implemented in the second sub-step. We considered two implementations, one in which parent selection is perfect, i.e. the single best parent is selected by each offspring individual (BEST); in the other, parents are selected proportionally to their fitness (WEIGHTED). After 100 steps (i.e. after 100 generations), simulations were stopped, and three measures were computed. First, the maximum fitness of the last generation, or where and refers to the last generation. Second, the average cumulation between the first and later generations (as), or where refers to the last generation, to the first generation, and to the generation. Finally, third, the cumulation between the first and last generation (as in), given by In order to compare the performance of populations of varying sizes, we applied, for each parameter combination, a Wilcoxon signed-rank test, comparing the sample of 200 observations (corresponding to the 200 *NK*-tables explored for each parameter combination) obtained for the populations under test. This pairwise Wilcoxon comparison is appropriate, since we let populations of varying sizes always explore the same *NK*-tables *and* let them start with identical initial strings/fitnesses. Note that our model does not allow complexity and population size to evolve. This means we are able to compare only how populations of a fixed size are able to sustain a technology of a given complexity. Yet, we follow our benchmark studies, here, and assume that from such comparisons can be inferred causal claims (i.e. claims about the extent to which demographic change may *favour* cultural change). Although we believe that this inferential step needs extra argument, we thus take it to be unproblematic here. Importantly, this does not undermine a negative result of our study: if it demonstrates the comparative advantage of larger over smaller populations to be non-robust, demographic explanations are compromised, regardless of whether or not the causal inference can be justified. # Results compares populations of and, assuming Uniform distributions and WEIGHTED pay-off bias, with red dots plotting the *p*-values resulting from the two-sided Wilcoxon signed-rank tests for the parameter combinations marked by the black dots below. Here the null hypothesis is that populations of sizes and produce: no significantly different maximum fitness (upper part, *p*-values for); no significantly different average cumulation between the first and later generations (middle part, *p*-values for); and no significantly different cumulation between the last and first generation (lower part, *p*-values for). So red dots under the black dashed line indicate parameter combinations for which the null hypothesis should be rejected for a significance level of 0.05. For those combinations where we observed a significant difference, we checked whether it was in favour of (so outperforming) by means of one-sided tests. Since this turned out always to be the case, we do not explicitly refer to the results of these one-sided tests in the remainder. The graphs exhibit several patterns. Let us start with the *p*-values for and. Generally, larger populations outperform smaller populations (so red dots fall under the 0.05 threshold) as long as complexity, expressed either as, or is low. When complexity increases, the larger populations of still produce higher maxima than the smaller populations of (as evident from the -values for), but the former lose their consistent advantage over the latter for two reasons. First, by lowering -values, higher values of and/or lead to *invisibility* of good parents, i.e. they contrast less with lower-skilled individuals. Consequently, given WEIGHTED pay-off bias, the contribution of inferior parents to transmission increases. Importantly, this holds for small and large populations alike. Second, higher values of and result in *instability*, in the following sense. For cumulation to occur, successive generations must be able to build on previous achievements; populations thus must be able to transmit a relatively stable knowledge base. That high values of undermine this may be self-evident, but high values of have a similar effect. Consider for instance the case where and. In this case, there is a 39.5% chance of at least one element changing state during transmission, an error which affects the fitness of other elements. Now if is high, say, a good innovation is very easily lost; a transmission error in one element leads to 50 new draws, which, for a Uniform distribution, will average out close to 0.5. So even if the transmission error is beneficial (i.e. it leads to a higher contributing value for the element itself), it will be largely undone by the new values drawn for the element's interdependent elements. More colloquially, excellent traits can be expected to deteriorate dramatically in transmission if even one of their elements would change state. Interestingly, although the qualitative results for and are the same, large populations are, in sustaining traits of higher complexity, slower to lose their advantage with respect to the former. An explanation for this is that the variance of is larger than that of for, in particular in cases where a population effect is found for but not for. So, even if under these conditions the mean and median values of and are similar, the variance of will be too sizable to yield a significant difference in the test for it. This evidently leaves the question why would exhibit a larger variance in the first place. Here the explanation is that, when and/or are sufficiently but not exceedingly high, small populations, due to their size, go through repetitive, quick episodes of substantial loss and cumulation. In case of, these fluctuations are averaged out by averaging over, resulting in variances lower than those of. Note that further increasing population size does not solve said issues of invisibility and instability, as can be gleaned from. That table gives the *p*-values of Wilcoxon signed-rank tests for versus populations sized, and this for the two first parameter combinations for which didn't outperform. It appears that under these conditions even populations of 10,000 individuals are not significantly better at sustaining highly complex cultural traits than populations of only 10 individuals. Further, comparisons between populations of size and populations of size support the idea that size effects are transitive. That is, for these smaller population sizes, no effects were observed that were not also present when comparing and. Finally, trends observed for versus for Uniform distributions also occur under Normal \[0,1\] and Gumbel \[0,1\] distributions. Trends are different under the assumption of BEST pay-off bias, where offspring is able to identify and imitate the single best individual in the parent generation. As can be seen in, demography now more generally makes a difference in sustaining cumulative culture. Only when transmission is highly erroneous or, population size contributes little to accumulation. Note that the results for BEST pay-off bias reinforce our earlier argument concerning invisibility. BEST pay-off bias by construction removes the invisibility constraint: however small the contrast between the best cultural parent and the lesser-skilled members of her generation, the BEST condition guarantees that she will be identified by all offspring. Under these circumstances, retains its advantage over for higher 's and 's, except when is high. # Discussion This study examined the robustness of a regularity suggested by previous modelling efforts, namely a strong dependence of cumulative culture on demography. More particularly, the aim was to verify whether that link was independent from—rather than an artifact of—previous models' assumptions about cumulation. To that effect, we added a measure of complexity to those already implemented in cultural-evolutionary models, and we adapted existing models so that cumulative cultural change could be expressed in terms of what we called 'S-complexity' (after Herbert Simon). This complexity is a function not only of a trait's number of components, but also of the number of interactions between these components. Our hypothesis was that in the face of increasing S-complexity, the link between demographic change and cumulative culture would collapse. The results of the simulations reported here lend support to our hypothesis: except under the highly optimistic assumption of BEST pay-off bias, large populations tend to lose their advantage in sustaining cumulative cultural change when cultural traits get too intricate. We identified two reasons for this. The first is that high S-complexity weighs heavily on social learners' ability to stand out under WEIGHTED pay-off bias. That is, except when pay-off biased selection is perfect and offspring is able to identify the single best individual in the parent generation, offspring is very often imitating inferior parents whose pay-offs are insufficiently different from even the best individuals in the population. The second reason is that cumulative culture demands stability or continuity, which is undermined not only by high copy error rate, but also by high values of. When is high, even a slight change in a trait's set-up will have a profound impact on the trait's overall value. Thus, the slightest error in transmission has the potential to completely destroy a previous achievement; the latter may be haphazardly reinvented on a later date, but not due to a cumulative process of building improvements on improvements. These results add to the suspicion that the dependence of cumulative culture on demography is not general, but applies to a specific range of cases (for empirical evidence questioning this dependence, ; but see). Previously, it has been shown to obtain only under a limited number of assumptions concerning learning biases, ; here it has been shown to obtain only insofar as previous assumptions about complexity are not violated and one makes the additional, highly optimistic assumption that naïve individuals are always able to identify and get the opportunity to learn from the single best parent in the population. How does this bear on explanations of particular episodes of cultural change? Since assumptions about complexity couldn't be discounted by means of robustness analysis, the only option seems to attempt to discount them on empirical grounds. If it would turn out that the Upper Paleolithic transition (for instance) didn't correspond to increases of S-complexity, Powell et al's explanation would stand firm. Conversely, it would compromise Powell et al's account if the transition were marked by the emergence of more intricate innovations, with increasing interdependencies among components (e.g., between procuring, transporting, preprocessing, and processing materials). We take it that as regards the Upper Paleolithic transition the choice between Powell et al's and our assumptions about complexity are underdetermined by the available evidence; so that currently neither their demographic explanation nor its negation can be discarded. So contrary to Powell et al's claims, it still may very well be that increased cognitive capacity (e.g., increased causal understanding of the interdependencies between components) gave rise to the Late Pleistocene emergence of modern human behaviour; or that some other factor or combination of factors made us modern. More generally, this study shows the importance and usefulness of robustness analysis. Besides sorting out claims which hold independently from the simplifying assumptions of the models they are based on, robustness analysis usefully guides data gathering: it tells for which assumptions we still need empirical confirmation (i.e. those assumptions which it cannot discount) and for which we can remain blissfully, or at least safely, ignorant (i.e. those assumptions which are inessential to the phenomenon of interest). Robustness analysis therefore is and should be an integral part of model building and assessment. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Analyzed the data: AQ KV. Contributed reagents/materials/analysis tools: AQ KV WH. Contributed to the writing of the manuscript: KV WH AQ.

# Introduction Optical highlighters comprise a class of fluorescent proteins which either turn on (Photo-Activation, PA) or change (Photo-Conversion, PC) their emission wave length in response to photo-stimulation with Ultra-Violet light. Among the most popular are the monomeric *Anthozoa* derived Green-to-Red photo-convertible proteins (mEOS2, Dendra2 and mKikGR), which irreversibly photo-convert from a green to red fluorescent state upon irradiation with UV light. This property has afforded biologists the ability to selectively label sub-populations of tagged- proteins and to track their sub-cellular migrations in real-time, significantly enhancing the understanding of complex biological processes. A typical PC experiment consists of defining a Region of Interest (ROI) in the green channel and photo-converting the ROI to red using a short laser pulse. The movement of the PC protein is then monitored by time-lapse microscopy, revealing novel protein trafficking destinations and migratory patterns. Typical analysis of PC data requires the extraction of fluorescence intensity values within the ROIs, widely handled by commercial microscope software control packages in conjunction with the open source project, ImageJ, and its associated plugins before using spreadsheet software to manually normalize and plot intensity values from different ROIs. Nevertheless, this process can be very time consuming and prone to error, prompting a demand for a new software enabling the automated analysis of PC datasets. While software packages are readily available for Fluorescence Recovery After Photobleaching (FRAP) datasets (e.g. Virtual FRAP, easyFRAP, FRAPCalc), key experimental differences between FRAP and PC protocols (e.g. one color vs. two color time lapse microscopy) make these packages ill-suited for analysis of PC datasets. In particular, PC experiments employ dual color time-lapse protocols in order to track the migration of a newly generated PC signal throughout the entire cell. As such, tracking of the PC signal relies upon appropriate extraction of signal information from two channels, as well as efficient normalization and quantification of fluorescent signals within multiple ROIs simultaneously. Increasingly, PC proteins are applied to investigate the dynamics of proteins residing in a-membranous cellular organelles (e.g. Nucleoli) or transient supra-molecular assemblies (e.g. Splicing Speckles or Stress Granules). However a poor Signal to Noise Ratio (SNR) can mask valuable information on protein residency and migration in these small cellular sub-compartments, as the fluorescent molecules undergoing PC include only a limited proportion of the total cellular population. Issues also arise when handling large volumes of 2D images generated from live cell imaging studies, and which contain rapid changes in protein dynamics. Hence, a more dedicated analysis package with tailored noise filtering and segmentation algorithms is required in order to successfully quantify and retain the low intensity, high frequency fluorescent signals obtained from PC experiments. Here, we provide users with a new convenient toolkit, which can be easily incorporated into the image analysis workflow and significantly accelerates the process of determining trafficking patterns of Green-to-Red photo-convertible fusion proteins. We introduce MATtrack, a quantitative analytical tool, which is tailored towards processing datasets obtained from dual-color, multi-dimensional (x,y,t) live cell imaging studies using photo-convertible proteins, and which was developed in the technical computing language, MATLAB. Importantly, MATtrack comprises a simple user interface and its implementation requires no specialist programming knowledge. # Materials and Methods ## Plasmids pDendra2-C and pDendra2-Fibrillarin expression vectors were a generous gift from Dr. Konstantin Lukyanov (Institute of Bioorganic Chemistry, Russian Academy of Sciences, Moscow, Russia). Dendra2UBC9 was obtained by PCR amplification of the cDNA encoding UBC9 from pCDNA3-UBC9-SV5, obtained from Dr. Ronald Hay (University of St. Andrews, St. Andrews, UK) and cloning in frame into the EcoR1 sites of the Dendra2-C expression vector. ## Cell Culture HeLa cells (ATCC CCL-2) were maintained in Dulbecco’s Modified Eagles Medium (DMEM) supplemented with 10% Foetal Calf Serum (FCS) at 37°C in a humidified atmosphere of 5% CO₂. 24 hours before imaging, HeLa cells grown in μ-slides (IBIDI) were transfected with 50 ng of Dendra2, Dendra2Fibrillarin or Dendra2UBC9 complexed with 0.15 μl Fugene HD (Roche). Prior to photo-conversion experiments, DMEM was exchanged for HEPES buffered DMEM without phenol red supplemented with 10% FCS.. ## Image Acquisition Image acquisition was carried out as described previously. Briefly, HeLa cells transiently expressing Dendra2 fusion proteins (Dendra2, Dendra2-Fibrillarin and Dendra2UBC9) were imaged on a Nikon Eclipse Ti E Spinning Disk Confocal Microscope, at 37°C in a humidified atmosphere. Transfected cells were identified by exciting the Dendra2 signal (Green) with a 488nm laser at 1–8.5% laser power, and detecting emission at 512/518 nm. Next, a sub-cellular region was targeted using a FRAP-PA unit (Andor) and photo-converted with a short (1000μs/ pixel) pulse of a 405nm diode laser administered at 25% laser power. Photo-converted Dendra2 (Red) signal was obtain with a 561 nm diode laser at 25% laser power, and emission was detected at 624/40 nm. Protein trafficking events were monitored by recording a time series in the red channel. The MATtrack software, including data pre-processing, image analysis and the user interface, were implemented in MATLAB ver 7 (R14) but is compatible with more recent versions and has been tested up to MATLAB 2014b with Image Processing Toolbox. ## Image Analysis in Andor IQ2 Comparative analysis of Dendra2 protein trafficking was performed using AndorIQ2 software (Andor). The mean fluorescence intensity was quantified in user defined ROIs and exported to excel for further analysis. The data was normalized by subtracting the background signal and subsequently quantified over time. Movies of Dendra2 protein trafficking were generated using IMARIS (Bitplane). # Results and Discussion An overview of the MATtrack image analysis workflow is provided in. MATtrack contains multiple processing algorithms to firstly separate a dual color time- lapse multi-tiff data set into its red and green components, before subjecting the red images to noise filtering, normalization, contrast stretching and temporal smoothing in order to improve detection of the PC signal. Next, MATtrack employs a Seeded Region Growing Algorithm to accurately delineate user selected sub-cellular ROIs, and subsequently detects protein trafficking events by automatically quantifying time-dependent variation in fluorescence intensity within the selected regions. Finally, MATtrack builds a “Migration Map” from a statistical model of the data set, providing a convenient means to visualize protein trafficking patterns and areas of significant accumulation. We tested MATtrack using image sequences derived from PC time-lapse experiments of HeLa cells transiently expressing different Dendra2-fusion proteins. Dendra2 alone as well as Dendra2 fusion proteins consisting of the nucleolar protein Fibrillarin (Dendra2-Fibrllarin) or the SUMO-conjugating enzyme UBC9 (Dendra2UBC9) were employed to determine whether our image processing and analytical algorithms could discriminate between purely diffusive as well as directed trafficking behaviors, mediated by Nuclear Localization Sequences (NLS), protein-protein interactions or RNA binding motifs. ## Image Processing ### Red/green image separation Color information from dual color time-lapse protocols may not always be retained in the output file type (.tif). Consequently, the red and green component images in the image sequence require automatic separation. Here, the variance in the image is used to systematically distinguish between red and green images. Indeed, green images in the dataset capture the pre-PC signal from Dendra2 and contain more data and consistently higher variance than red images, even when a PC event has occurred prior to implementation of the acquisition protocol. Hence, MATtrack compares the variance of fluorescence in the initial red frame, which represents the pre-conversion signal, with the variance of fluorescence in the initial green frame, which contains the full data. Since the pre-conversion state contains little or no signal in the red frame, this should have a very low variance while the green image will be very varied. Thus, MATtrack can clearly distinguish and separate the red and green frames. ### Noise filtering Images in the test data sets were acquired using a Spinning Disk Confocal Microscope (SDCM), generating an image sequence of high spatial and temporal resolution containing a low intensity, high frequency PC signal of interest. Potentially, this signal could be obscured by noise introduced during image capture or subsequent digitization and thus requiring a dedicated noise reduction algorithm in order to preserve the signal of interest. Noise is identified by examining the images in both the spatial domain and the time direction, where sharp spatial variations in local neighborhoods as well as sharp local temporal changes (i.e. frame-to-frame) were classified as noise. Comparatively individual elements of the image should remain relatively homogeneous, while true temporal changes in signal should be relatively smoother and maintained over several frames. A 3x3 median filter, which examines the intensities in the local neighborhood of each pixel and sets the new pixel value to be the median (centre value) of the neighbors, is applied to reduce spatial noise and preserve image features without attenuating the high frequency signal of interest. The data set is then temporally smoothed by applying a Gaussian filter in the time direction, which sets each new pixel value to a weighted sum of its neighbors, thereby minimizing local frame-to-frame variation. A Gaussian filter was defined in the time (t) direction according to : $$G(t) = \frac{1}{\sqrt{2\pi}\sigma}e^{- \frac{t^{2}}{2\sigma^{2}}}$$ Where σ is the standard deviation of the filter. With σ = 1 and using a width of 5 (stretching across five frames), this subsequently translates to a 1-D mask with: $$\text{G}\mspace{2mu} = \ \left\lbrack {0.054\ 0.24\ 0.4\ 0.24\ 0.054} \right\rbrack$$ This is passed across the image sequence and centered on each frame, transforming each pixel to the weighted sum of its temporal neighbors. Importantly, this smoothing is curtailed at the start and end of the image sequence, to prevent extension of the filter beyond the length of the image sequence. ### Contrast Enhancement An initial background extraction method was implemented in order to enable more accurate identification of ROIs during image analysis. Here, the first red frame in the image sequence (pre-conversion) is classified as background. This is subtracted from each subsequent frame, leaving the area of interest strongly contrasted with the background region. In order to maximize the contrast and increase the dynamic range of the image, the image sequence is contrast stretched and normalized, setting the maximum intensity in the whole sequence to 1 and the minimum to 0. The maximum and minimum intensity values across the entire image sequence are determined (*I*_*min* and *I*_*max*, respectively) and the intensity of each pixel in the sequence, *I(x*,*y*,*t)* is scaled using the linear transform: $$I\left( {x,y,t} \right) = \frac{I\left( {x,y,t} \right) - I_{\text{min}}}{I_{\text{max}} - I_{\text{min}}}$$ Of note, contrast stretching is applied across the entire sequence at once, as stretching frame by frame would lead to apparent fluorescence changes in the time domain. ## Image Analysis ### Statistical Model Analytical protocols are implemented to identify sub-cellular destinations and map the trajectories of the protein of interest. A statistical model of the processed image sequence is built by calculating the mean image, **M**, and standard deviation image, **S**, determined from the mean, *M(x*,*y)*, and standard deviation, *S(x*,*y)*, at each pixel location across the entire image sequence (Eqs and): $$M(x,y) = \frac{1}{N}{\sum\limits_{t = 1}^{N}{I\left( {x,y,t} \right)}}$$ $$S(x,y) = \sqrt{\frac{1}{N}{\sum\limits_{t = 1}^{N}\left( {I\left( {x,y,t} \right) - M\left( {x,y} \right)} \right)^{2}}}$$ where *N* is the number of images in the sequence and *I(x*,*y*,*t)* is the intensity of each pixel. The mean image represents the average intensity of each pixel across the sequence, providing an improved quality image of the entire cell, while the standard deviation image indicates which pixels exhibit high temporal variation, thus highlighting cellular areas of highest protein trafficking. ### Migration Map To generate a map to visualize protein trafficking patterns and regions of accumulation, we designed a classification algorithm, which automatically examines the variation of each pixel during the image sequence. This approach was preferable to employing a machine learning algorithm as these systems require significant human intervention during the training phase, and the resultant systems can be unable to handle irregular images. In contrast, the algorithm proposed here is advantageous as classification is performed by direct examination of the image sequence, requiring no human intervention in the classification and requiring no training, while allowing for adaptability to unusual images. Here, the region encompassing the cell is identified by locating the areas of highest variation (i.e. highest standard deviation). By analyzing the standard deviation image, **S**, and taking a threshold–defined as the mean of **S** plus one standard deviation of **S**–the cellular region is defined as: $$A\left( {x,y} \right) = \left\{ \begin{array}{l} {1\mspace{23mu}\text{if}\mspace{27mu}\text{S}\left( {x,y} \right) > \left( {\overline{S} + \sigma_{S}} \right)} \\ {0\mspace{23mu}\text{if}\mspace{27mu}\text{S}\left( {x,y} \right) < \left( {\overline{S} + \sigma_{S}} \right)} \\ \end{array} \right.$$ The region formed is then processed using morphological opening to remove spurious blobs. Within this region, the pixels are then classified by observing the changes over time of each pixel in the region, generating a “Migration Map” of the photo-convertible protein intra-cellular movement throughout the time series. Here, if a cellular region is bright but dims as the image sequence progresses, it is classified as the original PC location. The centroid of this section is taken as the approximate PC point. Conversely–dim becoming bright–indicates a region to which the protein is migrating. Extreme brightness, either relative to the rest of the image or in absolute terms (i.e. saturation), suggests regions of high accumulation. Finally, regions that remain relatively dark with minimal variation are classified as background. (The unprocessed raw data is presented as S1 Movie and S2 Movie). ## Application To test our classification algorithm, image sequences derived from PC time lapse experiments of HeLa cells transiently expressing different Dendra2-fusion proteins were analyzed. Firstly, Dendra2 was employed as a model of diffusive behavior in a cellular environment, as it lacks interaction with cellular sub-compartments, and localizes throughout the nucleoplasm and cytoplasm (left panel). After PC of a nucleoplasmic region, the corresponding region on the Migration Map is colored dark grey and the cytoplasm light grey, indicating diffusion of Dendra2 away from the PC region and across the nuclear envelope into the cytoplasm (right panel). Of note, no white areas can be observed in the Migration Map, indicating Dendra2 does not preferentially accumulate in specific sub-compartments. Next, the classification algorithm was examined to determine if it could accurately identify regions of protein accumulation using the fusion proteins Dendra2-Fibrillarin and Dendra2UBC9, which are known to accumulate in the nucleolus and PML nuclear bodies, respectively (left panel). In both case, a nucleoplasmic region was photo-converted and the migration map was used to describe the resulting trafficking behavior. Remarkably, the Migration Map correctly outlined how Dendra2-Fibrillarin or Dendra2UBC9 diffused through the nucleoplasm and specifically interacted with the nucleolus or PML nuclear bodies respectively (right panel; White regions). Hence, our algorithm correctly classified sites of protein accumulation even when they resided in very small sized sub-nuclear structures (PML nuclear bodies) which could be problematic for the user implementing manual ROI delineation alone ### Region Selection A simple user interface allows Regions of Interest (ROIs) to be selected and the corresponding fluorescent signal to be analyzed and outputted automatically in graphical format. Here, the user selects single points on the mean image, **M**, of the image sequence and a Seeded Region Growing Algorithm is used to define a ROI based on a set of corresponding nearby points. By automatically identifying a set of similar points, a regional average is calculated and tracked through the sequence. If only the selected point was tracked, this may cause issues as the selected point may be an outlier. The region selection also overcomes the issue of low signal to noise ratio, whereas an individually tracked pixel may vary dramatically due to noise. To define ROIs, we implemented a Seeded Region Growing Algorithm introduced by Adams & Bischoff. In this technique, the user selected point is taken as a seed and its neighbors are searched to identify up to 50 local equivalent pixels, bypassing any outliers. Here, equivalence is defined as Similar Mean Intensity, but alternative equivalence measures can be incorporated including: Standard Deviation, Region classification or distance from the selected point. Importantly, this algorithm abrogates the need for time-consuming manual delineation of the ROI. Furthermore, the algorithm employed here enables more accurate determination of ROIs when compared to other ROI selection algorithms where the ROI is defined as a fixed area around the selected point. A comparison of the Seeded Region Growing algorithm with a standard ROI selection algorithm is shown in. Here, selection of a small distinct nuclear region, such as a nucleolus occasioned the incorporation of unwanted nucleoplasm in the ROI, potentially resulting in erroneous quantification of the nucleolar signal. In contrast, the Region Growing Algorithm employed here results in incorporation of pixels only belonging to the selected pixels region (e.g. cytoplasm, nucleoplasm, etc.). The ROI selection method allows tracking of irregularly shaped areas which would make it compatible with other trafficking behavior including the trafficking of proteins associated with cytoplasmic membrane, the endoplasmic reticulum, the Golgi apparatus, endosomes, and lytic compartments. Nevertheless these remain to be tested and validated. ### Graphical Output After ROI selection, the average fluorescence intensity of the ROIs over the image sequence as well as the background are quantified and plotted against time providing an integrated view of protein re-localization to different cellular regions. For comparison the Dendra2 and Dendra2UBC9 datasets are provided as S1 Movie and S2 Movie respectively. This much-used method of identifying protein re-localization events by constructing a graph of mean fluorescence intensity over time from individual user-selected regions is seen in. The user can also utilize the migration map to identify points or regions of interest based on their trafficking behavior. The size of the selected regions and number of image points can be specified by the user or will revert to pre-defined defaults (a maximum of ten points selected, with each region extending to a maximum of 250 pixels). ### Accounting for Photobleaching Because photo-bleaching can lead to erroneous trafficking behaviors in PC experiments, the total cell fluorescence must be examined to determine whether any reduction in fluorescence has occurred throughout the sequence. Any drop in fluorescence is reported as a percentage to the user so that low quality datasets may be discarded. Moreover, photobleaching is automatically corrected by normalizing the signal in the tracked ROIs against the total cellular fluorescence at each time point. ### MATtrack validation compared to AndorIQ2: Tracking Dendra2UBC9 sub-cellular migration To validate our software accuracy in tracking PC protein re-localization events, we performed a comparative analysis of Dendra2UBC9 sub-cellular trafficking in HeLa cells using AndorIQ2 software. Here, the photo-converted nucleoplasmic region (P1), 2 additional nucleoplasmic regions (P2 and P3) and one additional cytoplasmic region (P4) were delineated manually using the SDCM acquisition software AndorIQ2, which automatically quantified the MFI in the defined ROIs over the time series. The data was then exported to excel, where the ROI MFI was normalized to the background by the user before being plotted against time, which was completed in 10 minutes. In parallel, we employed automated analysis of the same ROIs using MATtrack, which was completed in approx. 2 minutes. Moreover, MATtrack enabled automatic monitoring of the total photo-bleaching in the dataset, which was determined to be 15.098%. Comparison of the AndorIQ2 and MATtrack output results showed highly similar trafficking profiles. Dendra2UBC9 trafficked from the PC nucleoplasm and migrated to the other nucleoplasmic regions (P2 and P3). In parallel we observed a lower signal in the cytoplasm (P4), indicating a small proportion of Dendra2UBC9 was re-localizing to the cytoplasm. Close inspection revealed sharp frame-to-frame variation in signal processed using AndorIQ2, which had been removed by MATtrack’s integrated noise filtering algorithms. In addition, we observed a steady decline in Dendra2UBC9 signal in ROIs P1 and P2 in the AndorIQ2 dataset, which was absent from the MATtrack dataset. Potentially, this signal represented “true” Dendra2UBC9 trafficking away from these regions. Alternatively, this decline was an artifact of photo-bleaching. To distinguish between these possibilities, we re-analyzed the dataset in MATtrack without photo-bleaching correction, resulting in a decrease in signal in ROIs P1 and P2 and indicating that photo-bleaching was responsible for the different outputs between the two methods. Importantly, this highlighted that AndorIQ2 and MATtrack performed comparably for analyzing PC protein trafficking, however the automatic integration of ROI selection, image processing and analytical algorithms in MATtrack resulted in significantly faster result output than AndorIQ2. We anticipate that MATtrack would extend the same advantages over other commercial software packages with integrated image acquisition and analytical protocols (Olympus, as well as the open source software ImageJ, which requires time consuming manual ROI delineation and data export to additional statistical software packages). Hence, MATtrack streamlines the image analysis workflow and enables more rapid dataset analysis compared to other software packages, increasing the rate of result output, giving the user a more comprehensive view of protein trafficking at the single cell level in a shorter time frame. # Conclusion MATtrack is a new software tool that automates the analysis of PC datasets in order to accurately monitor and characterize live protein trafficking *in vivo*. MATtrack implements multiple algorithms to process the raw data, alleviating the burden of identifying appropriate processing techniques which can pose significant challenges to the non-expert user. MATtrack significantly streamlines the image analysis workflow to produce meaningful, quantitative results. Finally, its open-source code enables other programmers to refine existing algorithms and contribute new algorithms to expand MATtrack to suit their individual needs. This work was supported by DIT’s School of Electrical & Electronic Engineering and the UCD National Virus Reference Laboratory (NVRL). The authors wish to acknowledge access to and use of the UCD Conway Imaging Core Technologies, Conway Institute for Biomolecular and Biomedical Research, University College Dublin. In particular, we wish to acknowledge the UCD Conway Imaging Core Facility and Katarzyna Welzel for their technical assistance with image acquisition and PC on the SDCM, the latter being funded by Science Foundation Ireland (SFI). Finally, the authors wish to acknowledge Dr.Konstantin Lukyanov for his extremely generous gifts of the pDendra2c and pDendra2-Fibrillarin vectors (Institute of Bioorganic Chemistry, Russian Academy of Sciences, Moscow, Russia). [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JC EW VWG. Performed the experiments: EW. Analyzed the data: JC. Contributed reagents/materials/analysis tools: WWH DS. Wrote the paper: JC EW VWG.

# Introduction The maize silk is functionally equivalent to the stigma and style of a typical pistil. It is a specialized elongated tissue that begins to senesce about 8–10 days after it emerges from the husks. In normal conditions, pollination is completed within 1–2 days after silk emergence. Thus, a period of 8–10 days of silk viability is normally sufficient for seed setting. However, in hybrid seed production where plants are emasculated, a female parent with a long period of silk viability is critical for a good seed set. Thus, elucidating the molecular mechanisms of silk viability is necessary for the production of elite inbred lines and hybrid selection, which in turn contribute to maize hybrid seed production and field production. Before losing viability, the maize silk can receive viable pollen, promotes pollen germination, guides pollen tube navigation, and aids pollen–ovule interactions. Thus, silk viability has wide-ranging implications other than the specific function of supporting pollen germination through to fertilization. Under conducive conditions, compatible pollen grains hydrate and germinate pollen tubes. The tubes elongate via tip growth, penetrate the cell layers of the stigma, and enter the trichome, navigating within the transmitting tracts of the silk. Usually, only a single pollen tube grows through the micropyle and eventually reaches the ovule where fertilization occurs. All of these biological processes are critical for reproduction. Thus, several studies have analyzed the molecular mechanisms of pollen adhesion, pollen germination, pollen tube guidance, and pollen–ovule interactions. In pistil-interacting pollen tubes, genes involved in signal transduction, transcription, and pollen tube growth are highly expressed. Cysteine-rich peptides (CRPs) for cell–cell communications were shown to be important for pollen–pistil interactions in several studies. Genes involved in amino acid and lipid transport were shown to be important for, and unique to, reproductive processes in maize silks. The cytosolic free Ca²⁺ concentration is another important factor in normal pollen tube growth and morphology. A Ca²⁺ channel in pollen is formed by glutamate receptor-like proteins, which are regulated by D-serine in the pistil. Nitric oxide, which has negative chemotropic activity in *lily* pollen tubes and is involved in pollen tube guidance in *Arabidopsis*, might function in tip- growth downstream of Ca²⁺ signaling. For normal pollen tube growth, lipid transfer protein 5 (related to lily stigma cysteine-rich adhesin) and cysteine-rich receptor-like kinases were shown to be important for stigma- mediated reproductive processes in *Arabidopsis* and maize, respectively. Normal pH and K⁺ homeostasis in the pollen tube are important for guidance of the pollen tube to the ovule. cAMP was shown to play a second messenger role in regulating pollen tube growth and reorientation. Receptor-like kinase proteins that maintain pollen tube integrity and pollen tube attractants secreted by synergid cells are also required to guide pollen tube growth to the embryo sacs to complete fertilization. A defensin-like cysteine-rich peptide protein encoded by *ZmES4* was shown to cause pollen tube burst in mature maize synergid cells by opening K⁺ channels. Although many studies have identified key genes and molecules in the mechanisms of pollen–pistil interactions, pollen tube guidance, and pollen–ovule interactions in maize, few have focused on the molecular basis of silk viability, especially the genes and proteins related to silk viability before pollination. In practice, silks of maize hybrids always have much longer period time to accept pollen, complete double fertilization, and obtain seeds than their parental lines. Such as, silk viability of the improved lines of the local germplasm TangSPT that extensive used in maize breeding in China only has 5–6 days, while, 7–8 days silk viability were kept for their hybrid combinations with other lines from different heterotic group. However, the molecular mechanism of silk viability and its heterosis remain largely unstudied. The aims of this study were to: (1) identify key proteins related to silk viability at different silk developmental stages using a proteomics approach using three inbred lines, (2) illuminate the potential molecular mechanism of silk viability heterosis using two different hybrid combinations and its corresponding inbred lines, and (3) identify the common factors regulating both silk viability and its heterosis. # Results ## Silk viability evaluated by seed setting rate To evaluate differences in silk viability, the seed setting rate in the ear mid- base region (5–15 rounds from the ear base) was analyzed in three inbred lines and two hybrids. At the five sampling stages, D₄, D₆, D₈, D₁₀, and D₁₂, the average seed setting rate of three biological replications for the inbred lines Xun928 and Lx9801 was 99.3%, 97.4%, 95.0%, 26.0%, 13.5%, and 96.3%, 89.3%, 78.8%, 16.3%, 1.9%, respectively. The inbred line Zong3 sustained a high seed setting rate of 100% for all sampling stages. The ANOVA results showed that the difference of seed setting rate was significant between the parental inbred lines and their corresponding hybrids (*P* \< 0.05), except for D₄ between the inbred line Xun928 (*P* = 0.079), Zong3, and their hybrid combination Xun928×Zong3. The different sampling stages of Xun928, Lx9801, Xun928×Zong3, and Lx9801×Zong3 also showed significant differences (*P* \< 0.01). Thus, least-significant difference (LSD) multiple comparisons were performed and showed that the seed setting rate between each sampling stage was significantly different both for Xun928 and Lx9801. However, non-significant difference was detected between D₄ and D₆ both for the two hybrids Xun928×Zong3 and Lx9801×Zong3. The seed setting rate significantly decreased from D₈ to D₁₀ in both Xun928 and Lx9801 and the two hybrids. However, the decrease in the seed setting rate was slower in the hybrids than in the two inbred lines because of heterosis in the hybrids. Compared with those of the parental inbred lines, the seed setting rate of the hybrids fell between the mid-parent and high-parent values; i. e, seed setting rate showed partial dominant heterosis. Based on the phenotype of seed setting rate and heterostic degree, only the sampling stages with a significant difference at the 0.01 level were used for the proteomic analysis. Thus, the hybrids at D₈, D₁₀, and D₁₂ were used in the proteomic analysis of heterosis, and the inbred lines Xun928, Lx9801, and Zong3 at stages D₆, D₈, D₁₀, and D₁₂ were used in the proteomic analysis of silk viability. ## Differentially accumulated proteins For the 2-DE analysis, only protein spots that showed the same trend in the three biological replicates were retrieved ( and Figs). After normalization and ANOVA, only 3, 7, and 16 differentially accumulated protein spots were obtained for the inbred lines Xun928, Lx9801, and Zong3, respectively. These protein spots, which showed maximum changes more than 1.5-fold (*P* \< 0.05) during the four sampling stages, were manually excised and analyzed by MS. Among the 26 differentially accumulated protein spots, 17 and 7 protein spots showed the lowest and the highest levels at D₆, respectively. Meanwhile, 8 and 4 out of the 14 protein spots corresponding to D₆ showed the highest and the lowest levels at D₁₂, respectively. Among them, protein spot 46 gradually accumulated during silk development, while protein spot 61 gradually diminished. For the heterosis analysis, 46, 47, and 37 protein spots with maximum changes of more than two-fold (*P* \< 0.05) between the hybrid Xun928×Zong3 and its corresponding parents were retrieved at D₈, D₁₀, and D₁₂, respectively. The corresponding numbers of protein spots with more than two-fold changes between the hybrid Lx9801×Zong3 and its two parents were 24, 37, and 24, respectively. Out of the 215 differentially accumulated proteins, about 57% (122 protein spots) were additively accumulated and 43% (93 protein spots) were non-additively accumulated in the two hybrids. Among the non-additively accumulated proteins, five interaction patterns were observed; “−”, “− −”, “+”, “+ −”, and “+ +”, accounting for about 34% (32 protein spots), 3% (3 protein spots), 42% (39 protein spots), 12% (11 protein spots), and 9% (8 protein spots) of the non-additively accumulated proteins, respectively. The “− −” pattern was only detected in the hybrid Xun928×Zong3 at D₁₂, and the “+ +” pattern was only detected at D₁₀ and D₁₂ in the two hybrids. The “+ −” pattern was found in the hybrid Xun928×Zong3 at D₈ and D₁₂ and the hybrid Lx9801×Zong3 at D₁₀ and D₁₂. The other two major non-additive accumulation patterns “+” and “−” were well distributed across the three sampling stages in each hybrid. ## Differentially accumulated proteins identified as important for silk viability and its heterosis Three proteins differentially accumulated during silk development were identified, including gi\|413944345 (protein spot 63 in Zong3), gi\|414869037 (protein spot 8 in Xun928), and gi\|195635735 (protein spot 16 in Xun928). These three proteins also differentially regulated the heterosis of silk viability in the two hybrids (Tables). gi\|413944345, which differentially regulated silk development in the common paternal line Zong3, showed differential accumulation in the two hybrids Xun928×Zong3 and Lx9801×Zong3 at almost all sampling stages. gi\|414869037 and gi\|195635735, which were specific for silk development in the inbred line Xun928, contributed to the heterosis of silk viability only in the hybrid Xun928×Zong3 (protein spots 155 and 239 at D₁₀ and D₁₂; protein spot 96 at D₈), and not in Lx9801×Zong3. The functional category analysis showed that these three proteins were involved in anthocyanin biosynthesis, methionine metabolism, and suberin biosynthesis. Additionally, the proteins gi\|195643366 (APx2, cytosolic ascorbate peroxidase; protein spots 74, 190, 268), gi\|413942605 (6-phosphogluconate dehydrogenase isoenzyme B, protein spots 104, 176, 247), and gi\|413920184 (O-methyltransferase ZRP4, protein spots 89,150, 272) regulated the heterosis of silk viability at all sampling stages only in the hybrid Xun928×Zong3. The proteins gi\|414878829 (glutathione S-transferase 4, protein spots 127, 216, 298) and gi\|414881303 (anthocyaninless1, protein spots 138, 224, 285), affected the heterosis of silk viability at all sampling stages only in the hybrid Lx9801×Zong3. Three out of these five proteins were involved in secondary metabolism in the phenylpropanoid pathway. The proteins gi\|195629642 (lichenase-2 precursor, protein spots 267, 290) and gi\|413920639 (chitinase 1, protein spots 255, 275) contributed to silk viability heterosis only at the late silk developmental stage (D₁₂) of the two hybrids. ## Functional category and KEGG pathway enrichment classifications of differentially accumulated proteins The differentially accumulated proteins associated with silk viability and its heterosis was in similar functional categories. Unknown proteins comprised a large proportion of the differentially accumulated proteins, accounting for 38% and 40% of the proteins related to silk viability and its heterosis, respectively. The proteins involved in metabolism group accounted for the largest proportion of the differentially accumulated proteins, accounting for 43% of proteins related to silk viability and 42% of proteins related to the heterosis. Proteins involved in protein biosynthesis and folding, including transcription, translation, folding, sorting and degradation, were the second most abundant group and were specific to the heterosis of silk viability. Other important categories, based on protein abundance, were stress and defense response, plant hormone biosynthesis and signal transduction, and cellular processes. In the largest category, metabolism, there were six and eight subcategories of proteins involved in silk viability and its heterosis, respectively. Among them, methionine metabolism and flavonoid metabolism were important for both silk viability and its heterosis (Tables and), and lipid metabolism and energy metabolism were specific to the heterosis of silk viability. The protein–protein interaction networks involved in silk viability and its heterosis were analyzed by searching the String database. Three proteins were implicated in silk viability and its heterosis: gi\|413944345 (KOG1192), gi\|414869037 (KOG2263) and gi\|195635735 (NOG293481). These proteins had only one or two interacting proteins and were distributed at a remote node in the network. Some reductases or dehydrogenases were located at the interaction nodes and played an important role in the protein–protein interaction networks both for silk viability and its heterosis; for example, KOG1502-KOG2450-KOG0022 in the interaction network for silk viability and KOG1502-KOG1577-KOG2450-KOG0022-KOG0725 in the interaction network for the heterosis of silk viability. The two important branches for the heterosis of silk viability were ATP energy production (KOG1758-KOG1353-KOG1350-KOG1626-) and protein metabolism (KOG0177-KOG0179-KOG0863-KOG1688-). Additionally, two glutathione S-transferases (gi\|195619648 KOG0406, gi\|162460516 KOG0867) might play a crucial role in providing energy and proteins for the entire protein–protein interaction network for the heterosis of silk viability. Meanwhile, Kinases (KOG1367, KOG2440), enolase (KOG2670), and isomerase (KOG1643) increased the complexity of the protein–protein interaction network for the heterosis of silk viability compared with the network for silk viability, which was complicated by cell cytoskeleton proteins. # Discussion ## Comparison of protein categories related to phenotypes of the inbred lines and their corresponding hybrids The results in this study revealed that several functional categories of proteins corresponded to the seed setting rate phenotype of the inbred lines and its corresponding hybrids. For the inbred lines, proteins involved in flavonoid metabolism, methionine metabolism and cytokinin signaling, made the highest contributions to contributed the highest silk viability in the inbred line Zong3. Compared to the inbred line Lx9801, the stronger silk viability of the inbred line Xun928 was attributed to proteins involved in methionine metabolism, and these proteins also contributed to the heterosis of silk viability in the hybrids, but the proportions of their contributions differed. Proteins involved in flavonoid metabolism were important for heterosis of silk viability at all sampling stages in the two hybrids. However, proteins involved in methionine metabolism contributed differently to the heterosis of silk viability at different developmental stages of silks in the two hybrids: at D₁₀ and D₁₂ for Xun928×Zong3, and at D₈ and D₁₀ for Lx9801×Zong3. These results implied that proteins contributing to silk viability were not always as important for the heterosis of silk viability in the hybrids. Compared with the hybrid Lx9801×Zong3 at the three sampling stages, the hybrid Xun928×Zong3 accumulated more proteins involved in protein biosynthesis and folding, stress and defense responses, signal transduction and cell detoxification in response to genetic and environmental changes. Thus, the hybrid Xun928×Zong3 showed stronger heterosis than the hybrid Lx9801×Zong3. ## Potential regulation networks revealed by differentially accumulated proteins related to silk viability and its heterosis ### Methionine metabolism and salvage cycle Nutrient supply is a basic requirement for successful fertilization. The nutrients in pollen, however, can support only about 2 cm of tube growth in the maize silk. Thus, the maize silk must provide enough nutrients to support pollen tube growth over a longer distance. Consistent with this, many differentially accumulated proteins were related to cysteine and methionine metabolism (Tables). Proteins involved in methionine supply were important for silk viability. Methionine functions not only as a building block for protein synthesis, but also as a signaling molecule in communicating intracellular metabolic events to receptors on the cell surface. Therefore, methionine could supply appropriate signals to support pollen tube growth and guidance in the maize silk. In plants, methionine synthase (MeSe EC 2.1.1.12; protein spots 8, 121, 155, 239) catalyzes the terminal step of the methionine synthesis pathway by transferring a methyl group to homocysteine (Hcy), producing methionine. However, this *de novo* synthesis is energetically expensive and highly tissue-specific. To save energy consumption, about 80% of the methionine is recycled. S-Adenosylmethionine (AdoMet)-dependent transferase (protein spots 148, 194, 238) plays a critical role in methionine recycling by transferring the methyl group from AdoMet to S-adenosylhomocysteine (AdoHcy). This is not the only methionine recycling pathway in plants. Methylthioribose-1-phosphate isomerase (protein spot 64), which catalyzes the phosphorylated methylthioribose (MTR) to methylthioribulose-1-P, is the first and ubiquitous enzyme for methionine recycling. The accumulation of adenine (Ade), a by-product of MTR formation, inhibits methionine recycling. On the other hand, Ade is also a substrate for phosphoribosyl Transferase 1 (APT1) (protein spot 46), which regulates cytokinin levels by converting active cytokinin forms to inactive ones. Loss of APT1 activity leads to excess accumulation of cytokinins, inducing a myriad of cytokinin-regulated responses, such as delayed leaf senescence, anthocyanin accumulation, and downstream gene expression. AdoMet, as the major product of methionine metabolism, is an important cofactor that modulates various biological activities. As the major methyl-group donor, AdoMet can regulate transmethylation reactions at the levels of DNA metabolism, RNA metabolism, and protein post-translational modifications. AdoMet is also involved in metabolic and developmental regulation, since it is a substrate for thesynthesis of nicotianamine, ethylene (1-aminocyclopropane-1-carboxylate synthase), and polyamines. AdoMet metabolism is complicated by its interaction with plant growth hormones such as cytokinins and auxins. Thus, AdoMet is involved in regulating plant developmental by fine-tuning gene transcription, cell proliferation, and the production of secondary metabolites. In this study, positive regulators of nutrients production were identified in the inbred lines Zong3 (protein spots 46, 64) and Xun928 (protein spot 8), but not in the inbred line Lx9801. In the hybrid combinations, relatively more positive regulators (protein spots 148, 155, 238, 239) were identified at the late developmental stages (D₁₀ and D₁₂) in Xun928×Zong3. However, only two positive regulators (protein spots 121, 194) were identified in hybrid Lx9801×Zong3 and differentially accumulated at D₈ and D₁₀. These results were consistent with the stronger silk viability of the inbred lines Zong3 and Xun928 than that of Lx9801, as well as the high seed setting rates (84.6% for D₁₀ and 80.2% for D₁₂) and mid-parent heterotic degrees (34.4% for D₁₀ and 41.4% for D₁₂; data not shown) during the late sampling stages in the hybrid Xun928×Zong3. For comparison, Lx9801×Zong3 showed seed setting rates of 70.4% and 66.9% at D₁₀ and D₁₂; and mid-parent heterotic degrees of 21.1% and 31.3% at D₁₀ and D₁₂ (Tables –). ### Photosystem and energy metabolism Photosynthesis provides fuel for plant growth by converting light energy into chemical energy. Ribulose-1,5-bisphosphate carboxylase/oxygenase (RuBisCO), catalyzes the first major step of the Calvin cycle (carbon fixation) to produce energy-rich carbohydrates. This reaction uses ATP as an energy source and NADPH as reducing power, and is often the rate-limiting step in photosynthesis. In all eukaryotes, Rubisco is an oligomer consisting of eight large subunits bound to eight small subunits. The large subunits (protein spots 67, 106, 178) contain the enzymatically active substrate binding sites and are synthesized in the chloroplast. The small subunits are synthesized in precursor form by cytoplasmic ribosomes. Assisted by the RuBisCO large subunit-binding protein (protein spot 40), mature small subunits assemble with large subunits to form the oligomeric holoenzyme in the stroma. Several studies have shown that increased expression levels of RuBisCO subunits could increase photosynthetic efficiency by increasing catalytic activity and/or by decreasing the oxygenation rate. ATP synthase (EC 3.6.3.14), a key enzyme in energy metabolism, is widely involved in oxidative and photosynthetic phosphorylation and plays an important role in many processes in plants. It consists of two rotary motors: the membrane-integrated CF_o and the hydrophilic CF₁. CF_o mainly participates in proton transport through thylakoids, whereas CF₁ contains the nucleotide binding, catalytic, and regulatory sites of the ATP complex. CF₁ contains five subunits: α (protein spot 199), β, γ, δ (protein spot 182), and ε. The gene encoding the CF₁ α subunit, *atpA*, was shown to be related to cold resistance, and the transcript level of *atpA* was positively correlated with ATP synthase activity. Mutation of the *atpA* gene in a cytoplasmic male sterile line caused an energy supply shortage during flower development, resulting in abnormal microspore development compared with its maintainer. In this study, proteins involved in energy metabolism differentially accumulated in the inbred line Zong3 (protein spot 40) and the hybrid Xun928×Zong3 at D₈ and D₁₀ (protein spots 67, 106, 178, 182). A sufficient energy supply may be important to support stronger silk viability of Zong3, compared with those of Xun928 and Lx9801, and the stronger heterosis of silk viability in Xun928×Zong3 than in Lx9801×Zong3. ### Protein metabolism and cell senescence Proteins have a vast array of functions within living organisms, including catalyzing metabolic reactions, replicating DNA, responding to stimuli, and transporting molecules from one location to another. Many elaborate regulation mechanisms are involved in converting DNA sequences into functional proteins. Translation, the assembly of proteins by ribosomes, is an essential part of the protein biosynthetic pathway and requires initiation and elongation complexes. Eukaryotic translation initiation factor 5A (eIF-5A) (protein spots 72, 146) not only regulates protein synthesis but also acts as an important determinant of cell proliferation and senescence. In dividing and dying cells, different isoforms of eIF-5A execute its biological switching function in response to physiological and environmental cues. The elongation factor-1 (EF1) complex (protein spots 112, 161, 251) is responsible for the enzymatic delivery of aminoacyl tRNAs to the ribosome. EF1A is responsible for the selection and binding of the cognate aminoacyl-tRNA to the acceptor site of the ribosome. EF1 delta (protein spots 112, 251), functions as a guanine nucleotide exchange factor in regenerating active EF1A-GTP from inactive EF1A-GDP. During and after protein synthesis, polypeptide chains often fold into their native secondary and tertiary structures, whether they are used in the cell or secreted. To achieve their final correct states, cellular and secreted proteins require the help of several other folding proteins or chaperones. Protein disulfide isomerase (protein spot 252) catalyzes protein-folding, allowing proteins to reach their final correctly folded state without enzymatic disulfide shuffling. Unneeded or damaged proteins are transferred to proteasomes, an active complex composed of α subunits and β subunits (protein spots 229, 288), to be degraded into amino acids that are used to synthesize new proteins. At all sampling stages, many differentially accumulated proteins involved in protein biosynthesis and correct folding were identified in the hybrid Xun928×Zong3. However, more proteins involved in proteasomes differentially accumulated during the late sampling stages (D₁₀ and D₁₂) in the hybrid Lx9801×Zong3. These results implied that the hybrid Lx9801×Zong3 might consume more resources during normal metabolism, which weakened its silk viability, especially at the late silk developmental stages. During normal plant development, the insoluble polyesters suberin and cutin form extracellular lipophilic barriers to prevent membrane leakiness. However, membranes become leaky when the cell begins to senescence. This process is usually accompanied by the accumulation of proteins involved in lipid metabolism. In this study, O-methyltransferase (protein spots 89, 150, 163, 171, 236, 243, 272), the first rate-limiting enzyme in suberin synthesis, and GDSL- motif lipase/hydrolase (protein spot 96), an enzyme involved in the hydrolysis and transfer of activated monomers in cutin synthesis, differentially accumulated in the hybrid Xun928×Zong3. Proteins involved in lipid metabolism only differentially accumulated in the hybrid Lx9801×Zong3. These results implied that membrane leakiness might occur earlier in Lx9801×Zong3 than in Xun928×Zong3 at the late silk developmental stages. Thus, silk viability was lost earlier in Lx9801×Zong3 than in Xun928×Zong3. This pattern of protein accumulation might also explain the faster decrease in the seed setting rate and the weaker heterotic degree in the hybrid Lx9801×Zong3 at the late silk developmental stages. ### Phenylpropanoid metabolism and plant hormones regulation The plant hormone auxin regulates cell elongation, division, differentiation, and morphogenesis. Many proteins involved in elaborate temporal and spatial regulation of auxin metabolism, transport, and signaling have been identified. Auxin-binding protein 1 (ABP1: protein spot 193) mediates cell elongation and, directly or indirectly, cell division. In previous studies, ectopic and inducible expression of ABP1 conferred auxin-dependent cell expansion in tobacco cells that normally lack auxin responsiveness and antisense suppression of ABP1 eliminated auxin-induced cell elongation and reduces cell division. A homozygous null mutation of ABP1 was embryo-lethal in *Arabidopsis*. Auxin-induced swelling of proteoplasts and intact guard cells can also be attributed to ABP1. Pyrophosphate-energized vacuolar membrane proton pump 1 (protein spot 69) facilitates auxin transport and regulates auxin-mediated developmental processes by modulating apoplastic pH. Flavonoids in the phenylpropanoid pathway are another regulator of active auxin and have species-specific roles in nodulation, fertility, defense, and ultraviolet protection. Flavonols have been shown to negatively regulate the polar auxin transport (PAT) by competing for free auxin with auxin efflux carriers such as PIN and ABCB (PGP proteins) in vivo. Dihydroflavonol 4-reductase (DFR4, EC1.1.1.219: protein spot 53) and UDP- glucoside: flavonoid glucosyltransferase (EC 2.4.1.115: protein spots 63, 76, 79, 99, 117, 119, 122, 123, 138, 152, 159, 188, 197, 200, 201, 202, 212, 224, 231, 241, 285) are the first and last enzymes in the anthocyanin biosynthetic pathway, respectively. Their differential accumulation may be related to competition for the dihydroflavonol substrate with the flavonol branch, and thus, could indirectly affect PAT. The relatively higher contents of anthocyanin biosynthetic enzymes corresponded to higher levels of glutathione S-transferase- like proteins (protein spots 127, 216, 217, 298), which transport anthocyanins from the ER to the vacuole in the hybrid Lx9801×Zong3 at the three sampling stages. Cytokinin is another important hormone that regulates cell proliferation and differentiation. The two active forms of cytokinins are the isopentenyl adenine (iP)-type and the zeatin-type. The metabolic regulation of cytokinin includes biosynthesis, interconversion, inactivation and degradation. β-glucosidase (EC 3.2.1.21, protein spot 61) encoded by *Zm-p60*.*1* catalyzes the release of active cytokinins from their inactive storage and transport forms (cytokinin-O- glucosides). Over-expression of *Zm-p60*.*1* disrupted zeatin homeostasis in intact transgenic plants, rendering them hypersensitive to exogenous zeatin. Inactive cytokinin can also be generated by cytokinin-O-glucosyltransferase (protein spot 107). Studies of maize transformants harboring zeatin O-glucosyltransferase have shown that zeatin O-glucosylation affects root formation, leaf development, chlorophyll content, senescence, and male flower differentiation through developmental modifications. Proteins involved in the regulation of hormone levels (ABP1, pyrophosphate- energized vacuolar membrane proton pump 1; anthocyanin biosynthesis pathway; and cytokinin-O-glucosyltransferase) were identified as being important for both silk viability and its heterosis. These proteins differentially accumulated in the hybrid Xun928×Zong3, whereas only those involved in anthocyanin biosynthesis differentially accumulated in the hybrid Lx9801×Zong3. The flexibility of the systems regulating hormone levels may explain the increase in silk viability in the hybrid Xun928×Zong3, resulting in the high seed setting rate and strong heterosis. In summary, proteins gi\|413944345, gi\|414869037, and gi\|195635735 were attractive and might be related with silk viability as well as its heterosis. Significant correlation (*r* = 0.827^\* for gi\|413944345; *r* = -0.365^\* for gi\|414869037; *r* = 0.556^\* for gi\|195635735) was detected between these protein spots accumulation level and seed setting rate. Thus, we could propose the following hypotheses regarding proteins related to silk viability and its heterosis: methionine salvage, protein synthesis, and ATP supply function as positive regulators of silk viability, and therefore, contribute to strong silk viability and its heterosis in Zong3 and Xun928×Zong3, respectively. Active fatty acid metabolism, a signal for cell wall degradation, and anthocyanins, which negatively regulate local hormone accumulation, were related to weaker silk viability in Lx9801 and Lx9801×Zong3. The metabolism of cutin and suberin, which were derived from phenylpropanoid precursors, might confer stronger silk viability (Zong3 and Xun928) and stronger heterosis of silk viability in hybrids by slowing the silk aging process, especially during the late stages of silk development. # Conclusions In this study, the heterosis of silk viability could be mainly attributed to additive accumulation of differentially regulated proteins, although proteins that accumulated in a non-additive manner made a similar contribution. Simple additive and dominant effects at a single locus, as well as complex epistatic interactions of metabolic pathway genes at two or more loci, resulted in partially dominant silk viability heterosis in the hybrids. For silk viability, most important differentially accumulated proteins were those involved in methionine metabolism for nutrient supply, phenylpropanoid metabolism for hormone homeostasis, protein biosynthesis and metabolism for genetic information processing, and carbon fixation for energy generation. # Materials and Methods ## Plant materials Three typical inbred lines, Zong3, Xun928, and Lx9801, with different silk viability were used in this study. Among more than one hundred inbred lines, the silk viability of the inbred line Zong3 was extremely high. The inbred lines Xun928 and Lx9801 had relatively weak silk viability. To assay the heterosis of silk viability in different genetic backgrounds, two hybrids, Xun928×Zong3 and Lx9801×Zong3 were created in this study. The two hybrids and the three inbred lines were planted on the farm of Henan Agricultural University (Zhengzhou, China; E 113°42′, N 34°48′) in summer of 2013, when the daily average temperature was 14.3°C. The annual average rainfall is 640.9 mm in this region. Each plot consisted of ten 5-m-long rows, with 20 cm of in-row spacing and 67 cm of inter-row spacing. Only the middle rows were sampled to avoid edge effects. Before the silks emerged from the husk, ear shoots were totally covered with bags to avoid pollen contamination. To evaluate the silking time accurately, the silking time of each ear of the materials was recorded in the field. Day 1 (D₁) was marked as the day that the silks emerged above the ligule of the outer leaf of the husk. Silks were removed from the mid-base region of each ear at D₄, D₆, D₈, D₁₀, and D₁₂ and immediately frozen in liquid nitrogen in the field. Each sample was collected with three biological replications and 10 ears were mixed for each replication. At the same time, the ear for each sample was saturation- pollinated by hand on the sampling day to measure the seed setting rate. Pollination was completed between 9 and 10 a.m. (below 37°C) to ensure consistent pollination efficiency. The ears were harvested at physiological maturity, and only the seeds at the mid-base (5–15 rounds from the base) were used to calculate the seed setting rate of the cob according to silk development characteristics. The seed setting rate was calculated by dividing the total number of spikelets by the number of fully grown seeds. ## Protein extraction and MS identification Each genotype was assayed with three biological replications corresponding to each sampling stage. Frozen silks (1 cm, approx. 1.0 g) of each biological replication (mixture of silks from ten different plants) were fully ground in liquid nitrogen and then extracted in 10 mL pre-cooled trichloroacetate (TCA) buffer (10% w/v TCA in acetone with 0.07% β-mercaptoethanol) with vortexing for 2 h at 20°C. After centrifugation at 15,000 × g for 30 min, the supernatant was discarded and the precipitate was rinsed with 10 mL chilled buffer (80% acetone with 0.07% β-mercaptoethanol) four times by centrifuging for 10 min at 15,000 × g. The final cleaned precipitate was freeze-dried under a vacuum. The dried protein pellet per 1 mg was resuspended in 20 μL buffer (8 M urea, 2 M thiourea, 4% (w/v) CHAPS and 40 mM dithiothreitol (all from Solarbio)). The protein was quantified using a Bio-Rad protein assay with bovine serum albumin as a standard and used for two-dimensional gel electrophoresis (2-DE). Three technical replications were assayed for each biological replication. For each technical replication, equal amounts of total protein extract (800 μg) were used for isoelectric focusing (IEF). Immobilized dry strips (24 cm, Imobiline drystrips, Bio Rad, Hercules, CA, USA) with a linear gradient of pH 4–7 were rehydrated for 16 h at 50 V. The IEF conditions for separating proteins were as followes: slow 250 V for 30 min, rapid 250 V for 2 h, rapid 500 V for 2 h, rapid 1,000 V for 2 h, linear 9,000 V for 5 h, rapid 10,000 V for 10 h, and a constant 500 V for the final 12 h at 20°C. Strips were immediately equilibrated in 10 mL of two types of SDS equilibration buffer for 15 min each. Buffer 1 contained 0.375 M Tris-HCl pH 8.8, 6 M urea, 20% glycerol, 4% SDS, and 2% DTT and buffer 2 contained 0.375 M Tris-HCl pH 8.8, 6 M urea, 20% glycerol, 4% SDS, and 2.5% iodoacetamide. IPG gel strips with the proteins were embedded into the top of a polyacrylamide gel (12%) after equilibration and separated at a constant voltage of 50 V for 30 min. Then, a constant voltage of 200 V was maintained until the electrophoresis was finished. Digital images of the gels stained with Coomassie brilliant blue G250 were obtained with a scanner (UMAX Power Look 2100 XL). Spot detection and matching was performed with the default parameters using the “spot detection wizard” function in PDQuest 8.0 software. The “find spot centers” function was used with default auto-noise smoothing and background subtraction. A Gaussian model was selected to generate a master gel for each image file. All the gels were matched to the reference master gels selected and normalized in automated mode followed by manual group correction. The normalization parameters were “total quantity in valid spots”, “total density in gel image”, “mean of log ratios”, and “local regression model”. After normalization, ANOVA was used to calculate the significance of differences in the relative abundance of protein in individual spot features among the developmental stages of a certain inbred line, as well as among hybrids and their corresponding inbred lines at each developmental stage. For protein spots further assayed by MS, the maximum intensity variation criterion was set to ≥ 1.5-fold and ≥ 2-fold (*P* \< 0.05) among different sampling stages for each inbred line, and between hybrid and parental inbred lines at each sampling stage, respectively. The selected proteins were excised manually from gels, subjected to in-gel digestion with trypsin, and then destained using 25 mM ammonium bicarbonate in 50% (v/v) acetonitrile for 15 min at room temperature. The discolored spots were vacuum-dried and incubated with modified porcine trypsin at 37°C overnight. After centrifugation, the supernatant was collected and vacuum-dried, and then the precipitate was re-dissolved in 60% acrylonitrile/0.1% trifluoroacetic acid (TFA) (100 μL) for 15 min to obtain the peptides. Then, a 0.3 mL peptide sample and 0.3 mL matrix consisting of 10 mg/mL α-cyano-4-hydroxycinnamic acid in 50% acetonitrile and 0.1% TFA was analyzed on a matrix-assisted laser desorption ionization-time of flight (MALDI-TOF) mass spectrometry (MS). The parameters of the MS were set with 4000 Series Explorer software (Applied Biosystems). The lists of theoretical peptide MS from each peptide-map-fingerprinting (PMF) combined with MS/MS were used to search the NCBI (National Center for Biotechnology Information) database without repetition for homologous sequences using MASCOT 2.2 software ([www.matrixscience.com](http://www.matrixscience.com/)). The search criteria were as follows: 1) peptide mass tolerance of 100 ppm; 2) maximum of a single missed tryptic cleavage; 3) fragment mass tolerance of 0.4 Da; and 4) carbamidomethylation by cysteine residues as fixed modifications and oxidation by methionine residues as dynamic modifications. Only proteins with a MASCOT score \> 60 with 95% confidence and at least two matched peptides were accepted. Gene Ontology (GO) annotations and the theoretical Mr/pI for the identified proteins were retrieved from <http://www.geneontology.org/> and <http://www.expasy.ch/tools/pi_tools.html>, respectively. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis was carried out using the blast function in BLAST2GO. The protein–protein interaction network was analyzed by the publicly available program STRING (<http://string-db.org/>). Clusters of Orthologous Groups (COG) of proteins functions were used to construct the networks. Only an interaction networks with a high confidence (0.700 for silk viability or 0.900 for heterosis of silk viability) and no more than five interactors were retained. The eukaryotic orthologous groups (KOGs) were considered prime selections for a single protein spot. ## Data analysis Protein spots that had no significant difference in average spot intensity from the mid-parent value at the 0.05 level were considered additively accumulated (A). The accumulation pattern of each non-additive protein was classified as described by Hoecker et al.. Average spot intensities of proteins that deviated significantly from the mid-parent value of the parental lines at *P* \< 0.05 level were defined as non-additive proteins. “+” and “−” were used to indicate that the protein spot intensity identified in the F₁ hybrid was similar to the high parent and low parent values, respectively. “+ +” and “− −” indicated that the protein spot intensity identified in the F₁ hybrid was significantly different from the high parent and low parent values, respectively. “+ −” indicated the protein spot intensity identified in the F₁ hybrid fell in between the mid-parent and high parent or mid- parent and low parent values. ANOVA, LSD, and correlation analysis were performed with the corresponding function in Excel 2007. # Supporting Information CRPs cysteine-rich peptides Hcy homocysteine AdoMet S-adenosylmethionine MTA methylthioadenosine APT1 Adenine Phosphoribosyl Transferase 1 RuBisCO Ribulose-1,5-bisphosphate carboxylase/oxygenase ABP Auxin-binding protein 1 AdoHcy S-adenosylhomocysteine Ade adenine MS mass spectrometry IEF isoelectri focusing MALDI-TOF matrix-assisted laser desorption ionization-time of flight PMF peptide-map-fingerprinting GO Gene Ontology KEGG Kyoto Encyclopedia of Genes and Genomes COG Clusters of Orthologous Groups KOGs eukaryotic orthologous groups LSD least-significant difference [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: ZF JT. Performed the experiments: ZM YQ XZ YW. Analyzed the data: ZF ZM. Contributed reagents/materials/analysis tools: FZ YW. Wrote the paper: ZF.

# 1. Introduction In order to adapt to the dynamic, competitive and complex market environment, team operation is favored. With intensified environmental uncertainty and changing demands, it is increasingly difficult to rely solely on the wisdom and management methods of team leaders to avoid organizational risks and improve organizational effectiveness. Therefore, it becomes more relevant for the team to actively express their work-related views to achieve continuous team optimization in the incessant iterative process. Positively team expression represents a new perspective on the behavior of suggestions, which is called the team voice behavior. Van Dyne and LePine introduced voice behavior into organizational behavior. Frazier and Bowler clearly defined team voice behavior as "the performance of all team members as a collated voice behavior". In China, research on team voice behavior is increasing day by day. In the available studies, Liu and Liao defined team prohibitive voice as "a working process or behavior that all team members express a common sense to the leader, that possibly impairs the team". In retrospect of existing global literature, team voice behavior can provide organizations with various ideas, opinions, and suggestions to help leaders become more informed of the work processes and problems, promote organization decision-making, and avoid error checking. Team voice behavior can help prevent organizational risks, improve organizational effectiveness, and predict team performance. As the core managers in the team, leaders can influence the team voice behavior with their seemly leading styles and behaviors. Leaders can better manage the team with appropriate cognitive thinking. Paradoxical thinking acts as a positive way of thinking that deals with contradictory things to discover new opportunities and experience new kinetic energy. It conduces to integrate the two extremes of the contradiction, regards the tension and conflict caused by the paradox as an opportunity and challenge, and resolves disputes between contradictory needs, interests, ideas, and opinions. Consequently, the negative and tense working state of the team can be improved, and the positive expression can be stimulated, thus promoting team voice behavior. With the popularity of paradoxical leaders and gray management, the paradoxical thinking between leaders and employees in organizations will become the mainstream thinking mode advocated by enterprises. However, most research focuses on the individual level of employees and leaders but rarely involves the team level Similarly, most research on voice behavior focuses on employees. Research involving team voice only takes team characteristics as antecedent variables, without the influence of the thinking style of leaders on team voice behavior. Previous studies have found that paradoxical thinking of leaders can improve team competitive advantage and the good development of employees and voice behavior. However, the research of leadership paradoxical thinking on voice behavior at the team level is still relatively lacking. Therefore, in order to fill the gap in the research on the influence of paradoxical thinking of leaders on the team and to clarify whether the team voice behavior is affected by the leaders thinking, this study attempts to enrich and perfect the research related to paradoxical thinking and team voice behavior by analyzing the mechanism of paradoxical thinking of leaders affecting team voice behavior. In addition, social exchange theory suggests that after receiving benefits from others, individuals will provide relative benefits or help in return to maintain social exchange relationships. When team leaders or organizations offer support, care, and understanding to team members, team members perceive more responsibilities, obligations, and missions to repay the leader or the team, thereby expressing more positively and showing more voice behaviors. Leaders with a paradoxical thinking are adept in identifying, accepting, and proactively tolerating contrary tasks and requirements. Faced with the paradoxical work requirements of accomplishing high-performance work and improving innovation capabilities, leaders with paradoxical thinking treat this contradictory work goal with an inclusive view. They will be tolerant of team members by offering help even if the members do not complete their work. With the understanding and acceptance of the leaders, the team members show more communication, interaction, and cooperation. Team members work together to accomplish paradoxical works through mutual support and cooperation. Team cooperation and team voice behavior are also affected by the team atmosphere. An enabling team atmosphere promotes team cooperation to a certain extent, encourages team members to work towards common goals, and hearten employees to deliver suggestions and stimulate their work enthusiasm. The core of social exchange theory is reciprocity. In a team forgiveness climate, mistakes and failures of team members are tolerated in the face of paradoxical work. In such a climate, to repay the tolerance and understanding, employees complete their work with mutual understanding, help, and cooperation by showing more proactive communication and expression. Therefore, based on the social exchange theory, this research explores the mechanism of the paradoxical thinking of leaders on team voice behavior by adopting team cooperation as a mediating variable and team forgiveness climate as a moderating variable. First, this research starts with the thinking characteristics of team leaders and reveals the role of the paradoxical thinking features of leaders on team voice behavior. Second, from the perspective of social exchange theory, whether team cooperation is a bridge between paradoxical thinking of leaders and team voice behavior was tested, and the path by which the paradoxical thinking features of leaders influence on team voice behavior was revealed. Finally, from the perspective of mutual benefit, the moderating effect of a forgiveness climate on the relationship between the paradoxical thinking of leaders and team cooperation was explored in consideration of the contextual factors of team climate. Then, whether the contextual factors of the team forgiveness climate affect the paradoxical thinking of leaders was unveiled through the effect path of team cooperation. The above research is how this paper enriches the theoretical research on the formation of team voice behavior. # 2. Research theory and hypothesis ## 2.1 Paradoxical thinking of leaders and team voice behavior Team voice behavior refer to a concentrated expression of the suggestion behavior of team members. This behavior includes interaction between group members, rather than a simple superposition of the suggestions of individual employees. Frazier and Bowler clearly defined it as "the performance of all team members as a collated voice behavior". Team voice behavior helps the team avoid risks and improve organizational effectiveness and team performance. In a work team, the team leader represents the core of the team. The personal thinking characteristics of a leader exert a huge impact on the team. As a stable trait tendency of the individual, the thinking mode can show continuous personal motivation and behavioral focus. Leaders with paradoxical thinking are adept at proactively responding to and managing contradictions and conflicts. They tend to understand and accept team mistakes and failures using the connections between contradictory things. According to the social exchange theory, pointed out that when the benefits provided by others are obtained, individuals will give relative benefits or help in return to maintain social exchange relationships. Team members who feel supported and encouraged by the tolerance and encouragement of leaders conduct more communication and discussion and become more proactively express their advice to repay the favorable behavior of leaders. Thus, the lack of information and wrong decisions of leaders in the management work can be avoided. First, while realizing organizational needs, leaders with paradoxical thinking also consider the individual needs of their teams. Allowing members to give full play to their strengths and abilities and carry out personalized authorization and decentralization can increase the work autonomy and diversity of the team members. Employees will also give more suggestions in return to repay the leaders. Second, in an ever-changing organizational environment, leaders with paradoxical thinking can well-balance the competing needs in the organization. They are able to show their characteristics of openness, tolerance, and flexibility to help members feel an open and supportive environment where employees are more likely to deliver suggestions. Finally, as the core of the team, the leader represents the role model for employees. Leaders with paradoxical thinking show the team how to deal with contradictory work in a complex environment. Faced with the possible risks of suggestions, team members will also imitate their leaders to consider the pros and cons of expressing suggestions and then choose the right time to do so. Hence, this research proposes Hypothesis 1: 1. **Hypothesis 1:** *The paradoxical thinking of leaders positively affects team voice behavior*. ## 2.2 The mediating role of team cooperation The basis of team cooperation is the mutually perceived common goal among team members. Team cooperation means that team members present a behavioral state of mutual dependence and support for the common goal within the team. Specifically, to achieve team goals, team members transform the output into results with certain cognitions and behaviors. Thomas believes that team members have a win- win mentality for cooperation and that team cooperation makes it easier to achieve their own and team goals. Studies have shown that when working cooperatively, team members are more willing to improve the performance of the team and the organization through mutual help. When completing tasks cooperatively, they will show greater decision-making performance. A team is more likely to succeed when its members cooperate with each other. Team cooperation behavior is influenced by the leader and the team members themselves. Leaders act as role models in team cooperation, and their characteristic differences of thinking can affect the way team members work. First, leaders with paradoxical thinking characteristics integrate the two extremes of contradiction, showing more openness, support and guidance. According to the core view of social exchange theory—Mutual benefit and reciprocity, when leaders give more instructions for team work, team members will receive more support and guidance, thus deepening mutual understanding and cooperation among team members. Second, according to the social exchange theory, improve the relationship between leaders and employees, and between employees. When team members experience the paradoxical thinking characteristics of the leader, it will be helpful to integrate conflicts and promote mutual understanding and help among team members, thus showing closer cooperation between them. Finally, leaders with paradoxical thinking are more likely to understand team failures or setbacks and display higher tolerance for team faults. According to the principle of mutual benefit and reciprocity in the basic social exchange theory, the understanding and tolerance of leaders enable team members to show more mutual communication, help, and cooperation to repay this favorable factor. The work mode of team members lays the foundation of team voice behavior. First of all, team cooperation promotes close contact and communication between team members. Thus, voice behavior emerges when members express themselves proactively, facilitating the achievement of team organizational goals. Secondly, leaders with paradoxical thinking characteristics set an example of understanding and tolerance. In addition, team members will imitate and show tolerance and understanding to others, prompting them to work toward common goals and actively offer advice and suggestions. Finally, team cooperation behaviors with clear team goals can promote better team development, strengthen the active expression of team members, and motivate team voice behavior. Therefore, this research proposes Hypothesis 2: 1. **Hypothesis 2:** *Team cooperation plays a mediating role in the paradoxical thinking of leaders and team voice behavior*. ## 2.3 The moderating role of forgiveness climate Team forgiveness climate refers to the perception of team members being supported by the team when exhibiting benevolent and altruistic responses in the face of conflicts and failure. It is mainly reflected in the tolerant attitude towards individuals who make mistakes or fail. Studies have shown that a forgiveness climate can promote the problem-solving efficiency of the work team, ease and improve the working relationship between team members; it can also increase the productivity of employees, reduce turnover and improve team performance. The team climate is an important situational factor in team cooperation. In different team climates, even the same thinking characteristics or management style of the leader can produce various team behaviors. According to the social exchange theory, if an individual has previously been rewarded with a certain stimulus behavior, then when a similar one appears again, the individual may adopt the same behavior to reward others. With the support and guidance given by the leader of paradoxical thinking, the team strives to achieve the team goals in the form of cooperation. At the same time, When the forgiveness climate releases a signal similar to the inclusion of failure, team members conduct more communication and cooperation to repay the forgiveness by teams, resulting in more proactive behaviors and expressions. Moreover, the team forgiveness climate is conducive to moderating the relationship between leaders and team members, improving the interpersonal relationship between team members, and forming a more friendly and trustworthy relationship. According to the social exchange theory, when faced with exchange risks, people will maintain new exchange relationships based on the trust established by previous exchange experience. Therefore, when leaders with paradoxical thinking show understanding and acceptance in an inclusive and considerate team climate, team members will trust the leaders and team and cooperate more actively to complete team tasks and goals. Even if a short-term relationship crisis occurs between the team and the leader, they will tolerate and cooperate based on the original trust relationship. In addition, in a forgiveness team climate, team members tend to maintain an optimistic mood, no longer complain or blame others for their mistakes, at the same time, the team leader guides the work by integrating the thinking mode of contradiction. Promote staff to complete team work with good communication and cooperation. On the contrary, members in a team with a low forgiveness climate, often regard failure as a shame, with intolerance of mistakes at work, shirk the work responsibilities. Even though the team leader with paradoxical thinking can integrate conflicts and contradictions, it is difficult for team members to communicate and cooperate to complete the work. Thus, this research proposes Hypothesis 3: 1. **Hypothesis 3:** *The team forgiveness climate plays a positive role in the paradoxical thinking of leaders and team cooperation relationship*: *in a stronger forgiveness climate*, *the positive relationship between the paradoxical thinking of leaders and team cooperation is stronger*. *While in a weaker forgiveness climate*, *the positive relationship is weaker*. After integrating the mediating role in Hypothesis 2 and the moderating role in Hypothesis 3, this study proposes a moderating mediation model. Team cooperation plays a mediating role between the paradoxical thinking of the leader and team voice behavior. However, this mediating role is affected by the degree of the team forgiveness climate. Specifically, when the team forgiveness climate is relatively high, the influence of the paradoxical thinking of leaders on team voice behavior can be transmitted through team cooperation. The team is infected by the forgiveness climate, and the acceptance and understanding of the paradoxical thinking of leaders play a role, enabling the team to cooperate and express themselves actively. Therefore, this research puts forward Hypothesis 4: 1. **Hypothesis 4:** *Team forgiveness climate positively moderates the indirect effect of paradoxical thinking of leaders through team cooperation that affects team voice behavior*: *a stronger team forgiveness climate indicates that this indirect effect is better*. In summary, based on the social exchange theory, this research takes Chinese mainland companies as research samples to explore the influence of the paradoxical thinking of leaders on team voice behavior. Moreover, based on theories and existing research findings, four hypotheses are put forward by introducing team cooperation as the mediating variable and the team forgiveness climate as the moderating variable. The specific research model diagram is shown in. # 3. Research design methods and variable measurement ## 3.1 Research samples and procedures This study adopts questionnaire surveys to obtain reliable and realistic first- hand research data. The sample data is mainly acquired from enterprises in Shanghai, Guangdong, Zhejiang, Jiangsu, Guizhou, and Sichuan province in China. Data were obtained by a team of three people working jointly to distribute questionnaires, collect questionnaires, and unify the survey. Before the survey, leaders and team members in this study were informed that there was no right or wrong answer, with promised anonymity and confidentiality of the questionnaire. The "leader-employee" matching survey method was adopted. Data on paradoxical thinking, team forgiveness climate, and team voice behavior were answered by the leader, and team cooperation was answers by the team members. Moreover, in order to avoid common method deviations, this research was divided into three-time points to distribute questionnaires, with an interval of one-month, and a duration of 3 months (February to May 2021). The leader completed the first survey, including paradoxical thinking of leaders and a team forgiveness climate. A total of 150 questionnaires were issued, with 148 questionnaires returned. Team members completed the second survey, including team cooperation. A total of 560 employee questionnaires were distributed to 148 teams, with 510 employee questionnaires from 130 teams collected. In the third survey, leaders scored the team voice behavior. A total of 130 leader questionnaires were distributed, and 118 were returned. In addition, with the subjects’ permission, our research team added common control variables, including gender, ages, education level, working years, and team size. After the completion of the survey, the last 4 digits of the mobile phone numbers were used as the matching basis for “leader-employee”. Finally, 477 valid questionnaires were obtained from 101 teams, with an effective team recovery rate of 67.33%. In the leader questionnaire, 62 are males, accounting for 61.4% of the survey, and 39 are females, occupying 38.6%; there are 19 people with college degrees, accounting for 18.8% of the survey, and 41 people with bachelor’ degrees, occupying 40.6%; 39 people have master’ degrees, with a proportion of 38.6%, and 2 people have doctor degrees, with a proportion of 2%. The average education level is at the undergraduate level, and the average ages are 40.09 years. The average working years are 11.78 years, and the average team size is 8.07. In the employee questionnaire, there are 217 males and 260 females, accounting for 45.5% and 54.5% of the survey, respectively. The average education level is at the college level, the average ages are 35.24 years, and the average working years with leaders are 5.54. ## 3.2 Measurement of variables For the reliability and validity of the questionnaire, this study draws on the existing mature scales. Before the survey, according to a standard translation and back-translation procedure and double-checking with the questionnaire distribution team, the scale was finally accurately translated into Chinese. This study adopts the Likert 5-point scale (1 to 5 in the questionnaire represent "strongly disagree" to "strongly agree", respectively). ### Paradoxical thinking The paradox-style scale compiled by Miron-Spektor et al. was used to measure the paradoxical mindset of leaders. There are 9 items in the questionnaire. For examples, when dealing with conflicting views, I have a better understanding of the problem; "when trying to solve conflicting problems, I will be full of vitality." The Cronbach’s α coefficient of the scale in this study is 0.862. ### Team forgiveness climate With the forgiveness climate scale developed by Cox, there are 4 items in total. For examples, "in the team, we can tolerate the faults and mistakes of the team members; in the team, we do not hold grudges". The Cronbach’s α coefficient of the scale in this study is 0.845. ### Team cooperation Using the team cooperation scale compiled by Chatman and Flynn, there are 4 items in total. For examples, "It is important to maintain harmony within our team; there is a high degree of cooperation among our team members". The Cronbach’s α coefficient of the scale in this study is 0.838. ### Team voice behavior Based on the method suggested by the measurement team of Walumbwa et al., Van Dyne and LePine’s scale was adapted to measure individual-level suggestions, with a total of 6 items. For examples, "the team member comes up with and puts forward opinions and suggestions on problems that affect the work team; team members communicate with each other about their work, even if they have different opinions." The Cronbach’s α coefficient of the scale in this study is 0.801. ### Control variables Demographic variables have been found to influence voice behavior. Team voice behavior is affected by gender, age, length of service, and education level of the leader. In addition, team size may impact the results of the study. In order to verify the model more accurately, this study measures gender, age, length of service, education level of leaders, and team size as control variables. # 4. Results ## 4.1 Common method bias analysis Although this study adopts a multi-stage and multi-source research method to avoid homologous deviation in operation, paradoxical thinking, team forgiveness climate, and team voice behavior are all evaluated and answered by leaders. Therefore, a common method bias is necessary for the part of the questionnaire filled by leaders. Harman single factor test was used for factor analysis of all problems. It is found that the variance explanation of the first factor is 34.29%, less than the recommended value of 50%. Therefore, the relationship between variables is credible without obvious homology deviation. The overall research results are not be seriously affected. ## 4.2 Data aggregation test When data at the individual level is aggregated to the group level, a consistency test of scoring is required. Since this research focuses on the concept of team cooperation at a team level, the aggregation principle of “consistency coefficient R_wg not less than 0.70” is followed. The results show that the average R_wg of team cooperation consistency is 0.938 (median R_wg is 0.963, minimum R_wg is 0.78, and maximum R_wg is 1.00). In addition, scholars have further pointed out that only when the intraclass correlation coefficient (ICC (1)) is greater than 0.12 with ICC (2) over 0.60 can individual-level data be aggregated into group- level data. The result shows that the ICC (1) value of team cooperation is 0.143, and the ICC (2) value is 0.611. By combining the indicators of R_wg and ICC the score consistency test in this study meets the standard for data aggregation. ## 4.3 Discriminant validity test In order to test the discriminative validity of the variables involved in this study, the structural equation model was used to carry out factor analysis to test the variables in the study. The team cooperation variables are aggregated to the team level after employees fill in, while paradoxical thinking, team forgiveness climate, and team voice behaviors are filled out by leaders. Team cooperation has been distinguished from other variables in the operation means of the research method. Therefore, this study only needs to verify the variables answered by leaders: the discriminative validity of paradoxical thinking, team forgiveness climate, and team voice behavior. Confirmatory factor analysis was used to test the discriminative validity of each variable. Therefore, the fitting index was selected to judge the fitting degree of the model. The chi- square difference must reach the significant level, the root mean square of approximate error (RMSEA) must be less than 0.08, and the comparative fitness index (CFI) and Tuck-Lewis index (TLI) must be greater than 0.9. According to, the three-factor model is better than other factor models in the fit of the sample data (χ2 = 187.07, df = 120, RMSEA = 0.07, SRMR = 0.08, CFI = 0.93, TLI = 0.91). It is indicated that the discriminative validity of the questionnaire design in this study is sound, and three factors represent three different constructs that can be used for regression analysis. ## 4.4 Descriptive statistics and correlation analysis To further clarify the relationship between the paradoxical thinking of leaders, team forgiveness climate, team cooperation, and team voice behavior, this research conducted a correlation analysis on the relationship between various variables. The results in indicate that the correlation coefficients between variables in the hypothetical relationship are significant. This result also supports the validation of research hypotheses, but further verification is required. ## 4.5 Regression analysis This study uses multiple linear regression and the process method to analyze the relationship among the paradoxical thinking of leaders, team forgiveness climate, team cooperation and team voice behavior. Among them, gender, age, and length of service of leaders, and team size are used as control variables. The results are shown in. Analysis of control variables: According to the results of Model 1 in, compared with female-led teams, male-led teams are more inclined to team cooperation (B = -0.119, SE = 0.055, P \< 0.05). In addition, in larger teams, members are more inclined to team cooperation (B = 0.027, SE = 0.007, P \< 0.001). The results of Model 4 in reveal that members show more team voice behavior in female-led teams (B = 0.211, SE = 0.091, P \< 0.05) compared with that in male-led teams. Additionally, in larger teams, members are more apt to team voice behavior (B = 0.045, SE = 0.012, P \< 0.001). Analysis of Hypotheses 1 and 2: The results of Model 5 in indicate that after controlling the natural attributes of the leader and team size, the paradoxical thinking of the leader positively affects the team voice behavior (B = 0.209, SE = 0.092, P \< 0.05). Therefore, Hypothesis 1 holds. From the results of Model 6 in, it can be known that after controlling the natural attributes of the leader and team size, team cooperation has a mediating effect on the relationship between the paradoxical thinking of the leader and team voice behavior (B = 0.561, SE = 0.176, P \< 0.01). In order to further clarify the above effect, the Bootstrap method was used for testing, and the results are shown in. After controlling the natural attributes of leaders and the team size, the indirect effect of team cooperation is significant with 0 excluded in the 95% confidence interval; the direct effect of the paradoxical thinking of leaders on team voice behavior is not significant, and the confidence interval of 95% is over 0. It is indicated that team cooperation plays a complete mediating role in the relationship between the paradoxical thinking of leaders and team voice behavior. Therefore, Hypothesis 2 is established. Analysis of Hypotheses 3 and 4: The results of Model 3 in indicate that the interaction between the paradoxical thinking of the leader and team forgiveness climate is significant after controlling the natural attributes of the leader and team size (B = 0.214, SE = 0.078, P \< 0.01). This interaction term produces ΔR² = 0.266 (P \< 0.001) based on the control and manipulated variables. Therefore, the team forgiveness climate moderates the relationship between the paradoxical thinking of leaders and team cooperation. In order to explain this significant moderating effect, the method proposed by Aiken and West was used to adjust the levels of moderating variables to determine the mean (plus or minus) by one standard deviation (±1SD). As shown in, the result indicates that the high team forgiveness climate means that the positive influence of the paradoxical thinking of leaders on team cooperation is relatively strong. Therefore, Hypothesis 3 holds. In order to further test the moderated mediating effect, this paper uses the Bootstrap method to test the moderated mediating model based on existing research. The specific analysis results are shown in. The indirect effect value in a high forgiveness climate is 0.135, with its figure \[.006.277\] in the 95% confidence interval. However, it is -0.005 in a low forgiveness climate, with its figure \[-.115.061\] in the 95% confidence interval. Thus, the forgiveness climate only moderates the mediating role of team cooperation in the relationship between the paradoxical thinking of leaders and team voice behavior to a relatively high degree. Therefore, Hypothesis 4 holds. # 5. Research conclusions and discussions ## 5.1 Research conclusion Based on the social exchange theory, this study examines the mediating role of team cooperation between the paradoxical thinking of leaders and team voice behavior from leaders’ thinking characteristics. The moderating effect of a team forgiveness climate is investigated to reveal the influence mechanism of leaders’ paradoxical thinking on team voice behavior. Based on the "leader- employee" matching questionnaire of 477 employees in 101 teams, conclusions are drawn as follows: 1. The paradoxical thinking of leaders predicted team voice behavior; 2. Team cooperation mediated the relationship between the paradoxical thinking of leaders and team voice behavior; 3.Moreover, the positive relationship between the paradoxical thinking of leaders and team cooperation is stronger, when the forgiveness climate is stronger. In contrast, the positive relationship between the paradoxical thinking of leaders and team cooperation is weaker, when the forgiveness climate is lower; 4. Team forgiveness climate also exerts a positive role in moderating the indirect effect of the paradoxical thinking of leaders on team voice behavior through team cooperation: in a stronger team forgiveness climate, the indirect effect is stronger. ## 5.2 Theoretical significance First, this research starts with the characteristics of the paradoxical thinking of leaders and expands the antecedent variables of team voice behavior from the perspective that personal thinking characteristics of leaders influence team. The paradoxical thinking of leaders has a positive predictive effect on the team voice behavior. Although scholars have done relevant studies on paradoxical leaders, most studies are based on the impact of paradoxical leaders on employee voice. This research can enrich the research concerning the influence of the paradoxical thinking of leaders on team voice from the team level. Second, team cooperation plays a mediating and explanatory role in the influence of the paradoxical thinking of leaders on team voice. From the perspective of the influence of leaders’ thinking characteristics on team members and the communication and cooperation of team members due to leaders’ acceptance and understanding, this study reveals the effect path of the paradoxical thinking characteristics of leaders on their team voice behavior. Additionally, whether team cooperation is the bridge between the paradoxical thinking of leaders and team voice behavior is tested. Finally, this study verifies that the team forgiveness climate positively moderates the relationship between the paradoxical thinking of leaders and team cooperation. Moreover, the mediating role of team cooperation in the influence of the paradoxical thinking of leaders on team voice is verified. When the forgiveness climate releases a signal similar to embrace failure, team members conduct more communication and cooperation to repay the team’s forgiveness, resulting in more proactive behaviors and positive expressions. This conclusion supports the mutually beneficial view in social exchange theory. In a team forgiveness climate, team members are more willing to communicate and cooperate and are more proactive in expressing their opinions, with the team showing more voice behaviors. This paper expands the boundary condition research on the relationship between the paradoxical thinking of leaders and team voice behavior. It also reveals whether the contextual factors of team forgiveness climate influence the effect path of the paradoxical thinking of leaders on the team voice behavior through team cooperation. That is how this work can enrich the theoretical research of team voice formation. ## 5.3 Practical significance This study provides the following inspiration for companies to stimulate team voice behavior. First, the cultivation of the paradoxical thinking that shapes managers: 1. The correct identification of the thinking characteristics of leaders to guide them to maintain moderate paradoxical thinking. 2. The cultivation of managers’ paradoxical thinking needs to focus on their training to develop holistic mindsets. Their understanding of two aspects of contradictions in the team should be enhanced, and two extremes of contradictions should be integrated, to promote a more proactive expression of teams. Second, the guidance of team cooperation: Since team cooperation plays a mediating role in the relationship between the paradoxical thinking of leaders and team voice behavior, business managers should guide and encourage employees to complete team cooperation goals cooperatively in daily work. Additionally, managers also should provide resources that promote team cooperation and create a team culture atmosphere of win-win cooperation. Finally, the creation of a team forgiveness climate: Team forgiveness climate positively moderates the mediating role of team cooperation in the influence of the paradoxical thinking of leaders on team voice behavior. Thus, the enterprise should establish a team forgiveness climate and dispel concerns of team members on the possible negative effects of voice behavior. In addition, in a team with a forgiveness climate, the paradoxical thinking of leaders is more likely to positively predict the team voice behavior, which is conducive to exerting the role models of leaders. Enterprise managers should pay attention to the construction of team culture. They can encourage the team to offer suggestions by establishing a good relationship with employees, understanding their work, and allowing different opinions. # 6. Research limitations and future research directions This research is exploratory research on team voice with Chinese mainland corporate teams as the object. Although the research has achieved some valuable conclusions, the following limitations exist. First, although this study adopts survey data from multiple sources at multiple stages and time points, the team voice behavior may fail to conduct an accurate evaluation because team voice is evaluated by the leaders of team members. Therefore, the following research is suggested to adopt the multi-sources (team members—team leaders—leaders of team leaders) and multi-stage investigation methods. Moreover, the outcome variable of team voice behavior should be filled out by leaders of team leaders. Second, the too-much-of-a-good-thing (TMGT) effect in paradoxical thinking is not considered. Studies have found that excessive paradoxical thinking can cause individuals to spend much time integrating contradictory elements and neglect non-logical thinking such as imagination and intuition. The paradoxical thinking of leaders is too low to understand and integrate different knowledge and ideas, thus reducing the voice behavior of the team. Therefore, the organization should be concerned about the paradoxical thinking level of the leader at a medium level, which is conducive to the positive development of the organization. Similarly, if the forgiveness climate is too low, team members will feel harsh and severe, which makes the social exchange relationship worse and the guidance and support provided by the team less, affecting team cooperation. On the contrary, when the team has a high forgiveness climate, it is easy to be interpreted as not completing tasks and not being responsible, thus affecting the development of the team and the organization. Therefore, the forgiveness climate is conducive to the development of the organization and the team in the middle degree, and the forgiveness climate in the high and low degree has a certain negative impact on the staff and the team. According to the above, our team suggests that in future research, we should consider the too-much-of-a- good-thing effect of paradoxical thinking and forgiveness climate, which may provide a new perspective for future research. Of course, future research should also consider the pros and cons and the degree of paradoxical thinking that has the greatest positive effect. Finally, the data in this study are collected from Shanghai, Guangdong, Zhejiang, Jiangsu, Guizhou and Sichuan province in Chinese mainland. Therefore, whether the relationship between the variables found in this study is universal for companies in other areas needs more research. **Note:** A sample of the questionnaire to provide complete transparency to the context and framing that led to the generated data. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Vascular endothelial growth factor-A (VEGF-A) is a key player in physiologic as well as pathologic angiogenesis. A variety of diseases ranging from tumor growth, through asthma to exudative age-related macular degeneration are associated with a deregulation of this angiogenic factor. As a consequence, anti-VEGF therapeutics have been developed to target pathologic angiogenesis in various contexts: Anti-VEGF drugs alone or in combination improved survival rates or progression-free survival rate in certain cancer types. Anti-VEGF drugs also have been investigated in ophthalmology in randomized clinical trials and have been found to improve treatment outcomes. Along with the increasing clinical role of anti-VEGF compounds, VEGF-A itself has been investigated intensively as a potential diagnostic biomarker or as a predictive marker for treatment response. In addition, antagonizing VEGF-A has been found in some studies to be associated with a significantly increased risk of myocardial infarction, hypertension and stroke. Even small amounts of intravitreally injected VEGF-binding agents were found to alter systemic VEGF-A levels. As a consequence, determination of systemic VEGF-A levels is part of the systemic safety analysis in various studies investigating VEGF-binding proteins. However, due to the use of different protocols in quantifying systemic VEGF-A levels, comparison of the results remains difficult. The aim of this study was therefore to investigate the independent impact of various pre-analytical parameters on the assessment of VEGF-A levels in blood samples from healthy volunteers. Identifying these sources of bias and noise will improve the reliability of future VEGF-A readings and will help to clarify the role of angiogenic factors in disease and response to therapy. # Methods ## Sample acquisition and distribution Blood samples from six healthy volunteers (three male, three female, aged between 20 and 40 years) were taken for the initial data set at three centers, two at each center (one male, one female). For additional experiments, blood samples were taken from eight participants (four female and four male). All procedures for obtaining blood samples were standardized following a detailed protocol: blood was first taken from the male participant, then from the female participant in a lying position. Blood sampling had to be scheduled between 7:30 and 10 a.m. Participants had to come to the site in a fasting status with last meal eaten before midnight the previous day and with no extraordinary stress exposure (e.g. excessive sports) the day before. Blood was taken first with a neonatal cannula and then with a butterfly needle. In order to avoid long stasis, the tourniquet was released directly after venous puncture and before blood was collected. The first drops of blood were discarded. After centrifugation of blood samples and before analysis, all aliquots were stored at -20°C and shipped on dry ice if applicable. Sample acquisition was repeated one week later in order to investigate intrapersonal variations in VEGF-A levels over time. Each center deliberately altered predefined pre-analytical parameters to investigate this parameter's impact on VEGF-A measurements. One aliquot from each sample was measured at the site of blood sampling, one aliquot from the same sample was sent to a second independent measuring center and a third aliquot was sent to a central laboratory. By this approach, each sample was measured independently at three different centers (acquisition center, measuring center and central laboratory) in order to investigate whether the parameter "center" was an independent confounder of VEGF-A measurements. The process of sample acquisition and distribution is illustrated in. ## Parameters investigated All parameters investigated as potential confounders of systemic VEGF-A levels are listed in. ## Analyzing kits for measured parameters The measurement of VEGF-A was done at all centers with the same batch of ELISA kit (R&D Cat No. DVE00). At the central laboratory VEGF-A was additionally measured with a Luminex assay (R&D Systems, Human VEGF High Sensitivity Kit Cat No. LHSCM293). Platelet-factor-4 (PF-4) was measured using an ELISA kit from R&D Systems (Cat. No. DPF40). IGF-1 was measured using an ELISA kit from R&D Systems (Cat. No. DG100). ## Missing values and values under the limit of detection Values that were missing from one center because this center did not have this sample were omitted. Paired analysis was not performed in those cases and samples were excluded from the linear regression model. In total, 6 out of 404 aliquots were excluded due to missing values. Values under the limit of detection (provided by the manufacturer of each kit as minimal detectable dose (MDD)) were uniformly replaced by half the MDD value. For the VEGF-A ELISA assay, MDD was given as 9 pg/ml by the manufacturer, for the VEGF-A Luminex assay MDD was 1,74 pg/ml, for the IGF-1 ELISA MDD was 22,4 ng/ml and for the PF-4 ELISA MDD was 4,63 pg/ml. ## Statistical Analysis Multiple regression analysis was performed using a linear regression model to assess the independent parameters affecting VEGF-A plasma measurements. Computations were performed using R. Significance of statistical difference is indicated by asterisks \*p\<0.05, \*\*p\<0.01, \*\*\*p\<0.001; ns: not statistically significant. In addition to the multiple regression analysis we performed descriptive paired analyses using Spearman’s correlation to measure coefficients of determination (R²) where appropriate. Data are presented as Tukey box plots and bar charts with standard deviation. ## Ethics This study was approved by the Ethics Committee at the University of Freiburg, Germany and all participating centers. Participants provided their written informed consent to participate in this study. # Results All VEGF-A levels reported in this section were measured by ELISA unless otherwise stated. The central laboratory samples were measured in single aliquots. At all other centers, samples were measured as technical duplicates and mean values were used for further statistical analysis. ## Multiple regression analysis Results from the multiple regression analysis are shown in. The following parameters were identified as statistically significant independent confounders of VEGF-A levels: *analyzing center*, *anticoagulant*, *type of centrifuge*, *measuring method* and *sex of the healthy volunteer*. The linear regression model estimates VEGF-A levels measured at the various centers to be between 32 pg/ml lower (center D) and 35 pg/ml higher (center E) compared to the central laboratory that was set as reference. Multiplex bead array measurements are estimated to be on average 38 pg/ml lower compared to ELISA measurements. Female participants had on average 24 pg/ml higher VEGF-A levels compared to male participants in our data set. Both, PECT anticoagulant (-34 pg/ml) as well as in particular CTAD anticoagulant (-73 pg/ml) resulted in significantly lower VEGF-A measurements compared to EDTA. Using a centrifuge with a swing-out rotor yielded on average of 23 pg/ml lower VEGF-A values compared to a fixed-angle centrifuge. The following parameters were NOT identified as independent confounders of VEGF-A levels in our data set: *intrapersonal variation over one week* (-3 pg/ml at second vs. first time point), *cannula* (+5 pg/ml for neonatal vs. butterfly), *time before centrifugation* (-8 pg/ml if waiting up to 2 hours), *time after centrifugation* (+8 pg/ml if waiting up to 6 hours) and *filling level of collection tube* (+7 pg/ml for half-filled vs. completely filled tubes). ## Paired analysis In addition to the linear regression model, we analyzed the following parameters in a descriptive analysis in order to better visualize their impact on mean values and range of measured values. ### Monovariante center comparison The box-plot analysis and paired plot comparisons shown in clearly illustrate the variability of results obtained at different centers by measuring aliquots that were taken from the same participants, processed in the same manner and analyzed using the same batch of an ELISA kit. In some cases, a good correlation between the centers and the central laboratory is observed, e.g. center E and F (R² = 0.98 and 0.95, respectively). In contrast, high deviations were found for other centers, for example when comparing center A (R² = 0.79) and center C (R² = 0.5) to the central laboratory. depicts results for all samples that were taken at center A with aliquots being sent to center B and the central laboratory for ELISA measurements. It is obvious that measured VEGF-A values from identical aliquots differed depending on the center at which the measurements were performed. A similar effect was observed for samples that were collected at center C with identical aliquots measured at centers C, D and the central laboratory. shows good correlation between samples obtained at center E and measured at center E, F and the central laboratory. ### Anticoagulants: EDTA, PECT and CTAD VEGF-A in the systemic circulation is stored in thrombocytes to a great extent. Thrombolysis therefore represents a potent confounder for measurement of free VEGF-A plasma levels and must consequently be avoided by using appropriate anticoagulants. EDTA is the most widely used anticoagulant in clinical routine, but both PECT and CTAD anticoagulants have been suggested as alternatives with potential superior effects in preventing thrombocyte activation. For both PECT and CTAD, lower amounts of free VEGF-A were measured compared to EDTA plasma in samples from the same participants taken at the same time point using the same method. This may in part be explained by higher thrombocyte activation in EDTA samples compared to CTAD or PECT samples as reflected by higher readings for PF-4, a factor usually stored in thrombocytes and released upon thrombocyte activation. When put into relation to values of VEGF-A levels obtained with EDTA, VEGF-A levels obtained with CTAD were lower than those obtained with PECT, indicating that unwanted release of VEGF-A by thrombocytes may even be lower when CTAD is used compared to PECT. There is no general effect of the anticoagulant on protein biomarkers because IGF-1 determination was not affected by the choice of anticoagulant. In order to evaluate whether the lower VEGF levels in the CTAD and PECT samples could be due to the two anticoagulants interfering with the ELISA assay, we added known concentrations of recombinant VEGF-A to PBS containing either EDTA, PECT or CTAD. There was no difference in the recovery rate of spiked-in VEGF-A, confirming that CTAD and PECT do not interfere with the VEGF ELISA assay. However, it was notable that older aliquots of recombinant VEGF-A (stored over one year at -80°C) yields significantly lower recovery rates in all samples tested. Furthermore, it was confirmed that both assays measure only free VEGF-A. ### Type of centrifuge: fixed angle vs. swing-out rotor The type of centrifuge could be an important pre-analytic parameter with impact on VEGF-A readings since pellet formation and pellet density in fixed angle vs. swing-out rotor centrifuges may lead to different degrees of contamination of plasma samples with pellet components such as cell debris containing VEGF-A from intracellular sources. In direct comparison, slightly higher VEGF-A levels were detected when samples were centrifuged with a fixed angle vs. a swing-out rotor centrifuge. This difference was identified as an independent confounder of VEGF-A readings in the linear regression analysis. Interestingly, high PF-4 levels, indicative of thrombocyte activation, were mainly observed in samples centrifuged with a fixed-angle rotor. IGF-1 levels were measured as an additional independent biomarker but were not affected by different centrifugation conditions. ### Measuring method The most widely used methods to measure VEGF-A are ELISA and multiplex bead arrays. Therefore, the amount of VEGF-A was determined in all samples using these two methods. As shown in, absolute readings from the two assays cannot be compared. In our hands, multiplex bead array measurements resulted in lower values compared to ELISA measurements: Median VEGF-A level measured by ELISA was 36 pg/ml (IQR: 19–63 pg/ml), whereas median VEGF-A level measured by multiplex bead array was 10 pg/ml (IQR: 7–18 pg/ml). This influence of analytical method was identified as an independent confounder in the linear regression analysis. The scatter plot and the Bland-Altmann plot indicate a good correlation between the two methods, although absolute values were not identical. In order to further evaluate the differences between the two measuring methods, we measured standard curves produced with the VEGF calibrator protein supplied with both kits in both assays. The results showed that the VEGF-A provided with the Luminex kit gives higher signals for identical protein concentration (as specified by the manufacturer) irrespective of the assay used. This means that measured sample concentrations, which are back-calculated using a standard curve produced with the calibrator protein from the Luminex kit, result in lower concentrations of target VEGF independent of the assay used. ### Extreme values The ability of a method to correctly retrieve known values at the upper and lower assay range is an important parameter for ensuring data consistency for samples containing very high or low analyte concentrations. Therefore, samples with known VEGF-A levels were either diluted or spiked with fixed amounts of VEGF-A (spike-in recovery), and it was investigated if the expected levels of VEGF-A could be determined correctly by ELISA or multiplex bead array. For diluted samples measured with ELISA, we found values close to the expected analyte concentrations, if a previously measured sample was re-measured in a 1:10 dilution. Samples re-measured by ELISA after a 1:5 dilution tended to result in slightly higher values than mathematically expected. The multiplex bead array method also measured slightly higher values than expected both after a 1:5 as well as after a 1:10 dilution. When samples with spiked-in VEGF-A were measured by ELISA, observed recovery rates were approximately 80%. These experiments were performed only at our central laboratory to test basic assay performance and were therefore not included in the linear regression model. ### Time before and after centrifugation We investigated whether storing blood samples before or after centrifugation would affect VEGF-A readings. When EDTA plasma samples were stored at room temperature for up to two hours before centrifugation, VEGF-A levels were only increased in a few samples. PF-4 readings also tended to be slightly higher in samples with longer incubation times and correlated with higher VEGF-A readings. This was, however, not found to be a significant independent confounder of VEGF-A readings in the linear regression model. In an additional experiment, we evaluated whether even longer waiting times (up to 24 hours before centrifugation) would affect VEGF-A readings. While long-term storage before centrifugation had a significant impact on VEGF-A readings from EDTA samples, there was no change in PECT or CTAD samples. A subset of EDTA- and PECT plasma samples was stored at room temperature or at 4°C, respectively, for up to 6 hours after centrifugation before the plasma was separated from the cell pellet. In some of the EDTA plasma samples, higher VEGF-A levels were measured. This was, however, not significant in the linear regression model. Importantly, none of the samples collected in PECT anticoagulant showed a detectable change in measured VEGF-A levels over time. In support of the notion that PECT anticoagulant very potently inhibits thrombocyte activation, we found considerably higher PF-4 values in EDTA samples compared to PECT samples (note the different y-axis in the two PF-4 graphs). ### Butterfly vs. neonatal cannula Variations in blood sample acquisition can significantly affect measurement results. For example, this is well established for potassium sampling when release of intracellular potassium leads to erroneous high plasma readings after partial cell lysis during sample acquisition. Likewise, thrombolysis occurring during sample acquisition might result in incorrect high VEGF-A levels. Since VEGF-A measurements have become increasingly important also in neonatal patients with the advent of anti-VEGF treatment in infants suffering from retinopathy of prematurity (ROP), VEGF-A levels were determined in samples obtained with neonatal cannulas or with standard butterfly needles (all taken from the same adult participants). The scatter plot in indicates a minor trend for higher VEGF-A readings in some samples that were obtained with neonatal cannulas. Many values, however, are located near the bisecting line indicating comparable measurement results. The kind of cannula was not a statistically significant independent confounder in the linear regression analysis. ### Intrapersonal fluctuations over one week In order to investigate whether VEGF-A levels vary in healthy participants over time, systemic VEGF-A levels were re-measured in new samples obtained from the same individuals after one week. Slight variations were observed, but these were not identified as independent confounders in the linear regression model. ### Filling level of the collection tubes In clinical routine, sample tubes are not always filled completely. This leads to altered ratios of sample volume to anticoagulant. shows the scatter plot for completely filled vs. half-filled tubes. There was no difference found (most values are located near the bisecting line) and accordingly, filling level was not identified as an independent confounder in the linear regression analysis. # Discussion Systemic VEGF-A levels are measured as biomarkers in different medical fields, such as oncology, pneumology and ophthalmology \[, –\]. Absolute values for VEGF-A levels vary noticeably between studies (e.g. plasma VEGF-A levels of healthy probands were reported to be between 81 pg/ml and 180 pg/ml) and therefore, cross-study comparisons are difficult. This is due to the fact that the methodology for VEGF-A measurements from human blood differs and that various confounders, e.g. source (plasma or serum), blood sampling techniques and anticoagulants (heparin plasma, CTAD plasma, EDTA plasma or others), can considerably affect measurable VEGF-A levels. In this study, we identified a number of parameters that can all independently lead to a bias of measured VEGF-A values from human blood samples. These parameters are: type of anticoagulant, assay method, center performing the measurement, type of centrifuge, sex of the participant. These parameters are extremely important to consider when systemic VEGF-A is measured for example as part of a safety evaluation after anti-VEGF therapy or in other fields, where systemic VEGF-A levels are of importance for example in evaluating clinical outcome or in providing risk estimates. It is important to emphasize that for this type of study the number of individual probands is not the determining factor (we had only 12 probands in our study providing the blood samples). It is rather the number of *aliquots* retrieved from these samples that determines the validity of such a study. For this purpose, we analyzed a total of 476 aliquots that were all processed in an exactly predefined manner, standardizing all pre-analytic steps but the one which was to be investigated. This allowed us to investigate the individual pre- analytical variable's impact on VEGF-A measurements independent from any inter- individual variations in VEGF-A levels. There are various explanations why certain pre-analytic steps can affect VEGF-A measurements. For example, it can be speculated that individual factors play a certain role as reflected by the parameter “measurement center”. In our study, all centers used the same batch of ELISA kits, so lot variability can be ruled out as an interfering parameter in our study. Beyond the human factor, different types of spectrophotometers for ELISA read-out may play a role. We therefore recommend using one central laboratory experienced in the applied measurement method. In order to investigate whether the type of assay applied for analysis has an independent effect on VEGF-A level measurements, all samples were analyzed by multiplex bead array in addition to ELISA. We detected significantly lower VEGF-A values using the multiplex bead array technique compared to the ELISA technique. The Bland-Altmann plot in, however, displays a horizontal line (except for three outliers). This demonstrates that relative values are comparable, but that the multiplex bead array retrieved consistently lower absolute values at a relatively fixed fraction from ELISA readings in our hands. This phenomenon was described previously for a comparison of ELISA and multiplex bead array for other cytokines, like IL-1β, TNF-α or IFN-γ. The difference in absolute values was detected despite using internal standard curves in both assays. One possible explanation could be that the VEGF-A used for generating the standard curve differs between the two techniques. Indeed, the calibrator protein provided with the VEGF ELISA kit resulted in lower signal levels in both the ELISA and Luminex system compared to the calibrator protein from the Luminex kit (shown). This explains the lower absolute VEGF-A levels measured in the multiplex bead assay. We therefore conclude that both methods can be used for human plasma VEGF-A measurements but that absolute values cannot be compared between the two techniques. In contrast to “free” recombinant VEGF-A used for generating standard curves, VEGF-A in plasma samples can be bound to proteins and might therefore escape detection. Both assays used in our study measure only free VEGF as confirmed by addition of aflibercept. One very obvious parameter with significant impact on VEGF-A readings is the anticoagulant used for plasma sampling. Since we were interested in measuring free VEGF-A and since it is well established that high levels of VEGF-A are stored in thrombocytes, serum measurements were ruled out for our comparative analysis. However, even between the different plasma anticoagulants, we found significant differences with regard to their potency to reliably suppress thrombocyte activation. As reflected by PF-4 measurements, the highest degree of thrombocyte activation (and therefore the highest risk of plasma being contaminated by VEGF-A released from thrombocytes) was present in the EDTA samples (both in completely filled and half-filled tubes). This is important, since most centers use EDTA as their standard anticoagulant for plasma sampling. However, for reliable measurement of VEGF-A as a biomarker EDTA does not appear to be the ideal plasma anticoagulant. In both PECT and CTAD buffers, considerably less PF-4 was detected which is indicative of less thrombocyte activation. As a consequence, VEGF-A levels measured in PECT and CTAD buffer showed less variability and were significantly lower than values measured from EDTA plasma. These lower VEGF-A levels in PECT and CTAD buffers do more likely reflect the true free plasma VEGF-A levels, while values from EDTA buffers may be higher due to varying degrees of additional VEGF-A released from thrombocytes. In addition, storing of centrifuged samples *before plasma separation* from the pellet did not affect measurement of VEGF-A when PECT was used as anticoagulant. Similarly, sample storage up to 24 hours *before centrifugation* did not result in different VEGF-A readings when CTAD or PECT samples were used, while EDTA samples resulted in a wide variability of measured VEGF-A levels when samples were incubated over 24 hours before centrifugation. While PECT buffer is currently not commercially available, CTAD tubes can be obtained as standard clinical equipment. For these reasons and since our experiments of VEGF-spiking into PBS containing EDTA, PECT or CTAD did not show any interference of any of the anticoagulants with the ELISA assay, we recommend the use of CTAD plasma for VEGF-A measurements, which is in accordance with the observations from Zimmermann et al. Starlinger et al. explored in their study, among other parameters, the influence of a Venflon^® versus a Vacutainer^® butterfly cannula on the measured VEGF-A levels and did not find an effect on measured VEGF-A levels, but on PF-4 levels. Since our study was in part conducted in preparation for the clinical study CARE-ROP, a study for investigating the efficacy and safety of different ranibizumab dosages for the treatment of retinopathy of prematurity (NCT02134457), we compared neonatal with butterfly cannulas. Similar to the results from Starlinger et al., we did not find a significant effect of these two types of cannulas (butterfly vs. neonatal) on measured VEGF-A levels. Neonatal cannulas can therefore be used for VEGF-A sampling. Svendsen et al. observed that measured VEGF-A levels are highly dependent on the applied centrifugation force: the higher the centrifugation force, the lower the measured VEGF-A levels, possibly due to better separation of thrombocytes from plasma. In our study, we compared two types of centrifuges (swing-out rotor versus fixed angle rotor) using equivalent settings. Even with identical centrifugal forces, VEGF-A levels differed significantly between the two types of centrifuges, with values from swing-out rotor centrifuges being lower than values from fixed-angle centrifuges. PF-4 values tended also to be lower in the swing-out rotor centrifuge, indicating that less thrombocyte activation occurs and therefore less contamination of free plasma VEGF-A with VEGF-A released from thrombocytes. Consequently, we recommend using the same type of centrifuge with identical settings for all samples acquired during the course of a study, preferentially a swing-out rotor centrifuge. Intrapersonal VEGF-A levels did not significantly vary over one week. This is in line with Svendsen et al. who also did not find a variation in VEGF-A levels within two weeks, but found variations over longer time periods. The fact that VEGF-A levels from the same person measured twice within one week were highly reproducible in our study, demonstrates not only the reliability of our sampling and measuring method but further emphasizes how strong the confounding effects of the interfering parameters are. The fact that IGF-1 measurements were not affected by the investigated pre-analytic parameters further confirms that these identified parameters are specific confounders for VEGF-A measurements. In conclusion, our study identified potent confounders of VEGF-A measurements that need to be standardized in order to obtain reliable and reproducible results in clinical studies. Based on our results we recommend the use of CTAD as anticoagulant, a standardized type of centrifuge where possible and one central laboratory for measurements of free VEGF-A from human plasma. Either an ELISA or multiplex bead array method can be used but absolute values are not interchangeable between these two assays. # Supporting Information The authors thank Ulrike Hagemann (University of Tuebingen), Susanne Denning (University of Ulm), Marc Leinweber (University of Freiburg), Serap Luick (University of Kiel) and Gerburg Nettels-Hackert (University of Muenster) for their excellent technical assistance. [^1]: This study was supported through a research grant from Novartis, Germany. J. M. Walz is a part time employee at Novartis; H. Deissler has received consultancy honoraria and research support from Novartis and travel grants from Bayer HealthCare and Novartis; L. Faerber is an employee at Novartis; P. Heiduschka has received research support from Novartis, Bayer HealthCare and Affiris; A. Klettner has received consultancy honoraria, speaker fees and research support from Novartis; T. U. Krohne has received consultancy honoraria and speaker fees from Bayer HealthCare, Heidelberg Engineering and Novartis, research grants from Alcon and Novartis and travel grants from Bayer HealthCare, Heidelberg Engineering and Novartis; F. Ziemssen has received consultancy honoraria from Alimera, Allergan, Bayer HealthCare and Novartis, speaker fees from Alcon, Alimera, Allergan, Bayer HealthCare, Heidelberg Engineering and Novartis and research support from Allergan and Novartis; A. Stahl has received research support from Novartis and speaker honoraria or consultancy honoraria from Boehringer Ingelheim, Novartis and Zeiss; D. Boehringer, J. C. Goepfert, S. M. Kleeberger and N. Schneiderhan-Marra have no conflict of interest. There are no patents, products in development or marketed products to declare. This does not alter the authors’ adherence to PLOS ONE policies on sharing data and materials, as detailed online in the guide for authors. [^2]: Conceived and designed the experiments: HD PH AK TK FZ AS. Performed the experiments: JW HD PH AK TK JCG SMK NS-M FZ AS. Analyzed the data: DB HD PH AK TK NS-M FZ AS. Contributed reagents/materials/analysis tools: DB HD PH AK TK NS-M FZ AS. Wrote the paper: JW DB HD LF PH AK TK NS-M FZ AS.

# Introduction The time spent in sedentary behaviour is increasing in modern societies, and it currently represents an important public health burden. Sedentary lifestyle is a risk factor for all-cause mortality, metabolic syndrome or type 2 diabetes, obesity, and even cancer, independently of physical activity levels. To reduce sedentary time in the general population, the simplest, most effective, and most accessible method is to decrease lying and sitting time. Although it is commonly believed that lying, sitting, and standing require a different energy expenditure (EE), there is still controversy. Whereas some studies reported no EE differences between lying and sitting, others have shown that EE was 20% higher in lying than in sitting. On the other hand, replacing sitting with standing is recommended to decrease sedentary time and increase the daily energy expenditure, but the difference in EE between sitting and standing also remains controversial. When comparing sitting vs. standing, studies show EE differences ranging from 10 to 100%. These contradictory findings might be partially explained by the lack of a rigorous control in the experimental design, data collection (i.e. different gas collection system), and/or data analysis. Other variables, such as anthropometric characteristics, body composition, age, or sex, may have also contributed to these discrepancies, but their role is largely unknown. There is a lack of studies showing the EE differences in the three positions, lying, sitting, and standing, and it would be of public health interest to better understand the EE changes across positions. The objectives of the present study were (i) to compare differences on EE across three positions: sitting, lying, and standing, and (ii) to determine the associations between the change on EE across these postures (lying vs. Sitting, lying vs. Standing, and sitting vs. Standing) with anthropometric and body composition parameters in young healthy adults. # Material and methods ## Study participants The participants were enrolled in the ACTIBATE study (ClinicalTrials.gov ID: NCT02365129) and met the following inclusion criteria: (i) being non-smokers, (ii) not taking any medication, (iii) not having an acute or chronic illness, and (iv) not being pregnant. The study was approved by the Human Research Ethics Committee of both the University of Granada (n°924) and Servicio Andaluz de Salud (Centro de Granada, CEI-Granada), and complied with the revised ethical guidelines of the Declaration of Helsinki (revision of 2013). All participants signed the written informed consent before their enrolment. A total of 84 young healthy sedentary adults aged between 18 to 25 years old were recruited for the current study. We had some technical problems in data collection with specific participants (i.e. calibration of the gas analyzer), and we excluded some participants because they did not strictly meet the standardized previous conditions (i.e. fasting time, physical activity, drugs or supplements intake etc) or because they talking or moving during the data collection. Thirty-one individuals were excluded and, therefore a total of 53 participants were included in the analysis. ## Procedures The study was conducted between April and June 2017. The participants were instructed to refrain from any moderate or vigorous physical activity within 24 and 48 hours, respectively, before the testing day, and not to consume caffeine and/or dietary supplements in the 24 hours prior to testing. The participants arrived to the laboratory by car or by bus (avoiding any physical activity after waking up) and in a fasted state (between 5 and 6 hours). The experimental design can be seen in. The EE measurements were performed by indirect calorimetry following the current recommendations. Briefly, all the measurements were carried out in the same room and by the same trained staff. Before being evaluated, all the participants confirmed that they had met the previous study conditions. Then, they lay on a reclined stretcher in a supine position during the 5–10 previous minutes to the indirect calorimetry measurement. They were instructed to breathe normally, and not to talk, fidget, or sleep. The same position and instructions were maintained during the next 15 minutes, when the indirect calorimetry measurements were performed. Afterwards, the participants sat on a supported chair with flat back placed near the bed during 10 minutes following the same instructions mentioned previously. Finally, the participants were asked to stand-up slowly and avoid unnecessary movements. In this static position, indirect calorimetry was measured for another 10 minutes. The transitions from lying to sitting, and from sitting to standing took a maximum of 5 minutes which were removed from the analysis. Indirect calorimetry measurements were performed with the CPX Ultima CardiO2 (Medical Graphics Corp, St Paul, USA), using an oronasal mask (model 7400, Hans Rudolph Inc, Kansas City, MO, USA), equipped with a prevent metabolic flow sensor (Medgraphics Corp, Minnesota, USA). Flow calibration was performed using a 3-L calibration syringe at the beginning of every testing day, and the gas analyser was calibrated using two standard gas concentrations following the manufacturer’s instructions before each EE measurement i.e. lying, sitting, and standing. The data obtained from the indirect calorimetry assessment included: EE (calculated by the Weir abbreviated equation \[assuming negligible protein oxidation\] and expressed as Kcal/day: EE = \[3.9 (VO2) + 1.1 (VCO2)\] \* 1.44), respiratory quotient (RQ), minute ventilation (VE), and respiratory rate (RR). Heart rate (HR) was recorded with a heart-rate monitor (Polar RS800CX, Polar Electro, Kempele, Finland) in all measurements. The weight (±10 g) and height (±5 mm) were measured without shoes and with light clothing, using a digital integrating scale (SECA 760, Hamburg, Germany), and a stadiometer (SECA 220, Hamburg, Germany). Body mass index (BMI) was calculated as weight (kg)/height (m²), and body composition (lean body mass and fat body mass) was determined by Dual Energy X-ray Absorptiometry (HOLOGIC, Discovery Wi). A detailed explanation of body composition procedures can be found elsewhere. Lean and fat body mass percentage was calculated as lean body mass (kg)\*100/weight (kg), and fat body mass (kg)\*100/weight (kg), respectively. ## Data analysis The gas exchange parameters were averaged every minute with the Breeze Suite (version 8.1.0.54 SP7, MGC Diagnostic) software. Later, we discarded the first 5 minutes record in each position and averaged the obtained data from the 6^th to the 10^th minute, which has been previously showed to be a valid option for indirect calorimetry data analysis. To determine the interindividual variability in response to the change on EE of changing from one position to another (lying vs. sitting, lying vs. standing, and sitting vs. standing), the participants were also categorised as spenders and savers. ‘Spender’ refers to a participant with a rise in EE \>5% between two positions which is maintained during the entire assessment period, and ‘saver’ refers to those who showed little or no change in EE (a rise in EE \<5%) between two positions. ## Statistical analysis We performed a sample size calculation based on a minimum predicted change of 5% in EE between lying and standing positions, and an SD for this change of 10%. A sample size of 40 participants was predicted to provide a statistical power of 80% considering a type I error of 0.05. Therefore, a sample size of 53 participants was enough to test our hypothesis. We conducted repeated measures analyses of variance (ANOVA) to compare EE, RQ, VE, RR, and HR across positions (lying, sitting, and standing), using Bonferroni post-hoc comparisons. The differences between variables were computed in all cases as: (i) standing—lying, (ii) standing—sitting, and (iii) sitting—lying. Linear regressions were conducted to examine if the EE differences between two positions could be explained by anthropometric or body composition parameters. We also conducted t-student unpaired-samples test to study the differences between spenders and savers. The analyses were conducted using the Statistical Package for Social Sciences (SPSS, v. 21.0, IBM SPSS Statistics, IBM Corporation), and the level of significance was set at P\<0.05. # Results shows the descriptive characteristics of the study participants. No interaction by sex was observed in EE, RQ, VE, RR, and HR changes across all positions studied (all P interaction≥0.477). shows the mean values of EE, RQ, VE, and RR in lying, sitting, and standing positions. The EE was significantly higher standing than in lying and sitting positions (mean difference: 0.121±0.292 and 0.125±0.241 kcal/min, respectively; change percentage: 9.7±10.9 and 10.2±11.6%, respectively; all P\<0.001), and no differences were observed between lying and sitting positions (P = 1.000). The RQ was higher lying than in sitting and standing positions (mean difference: 0.04±0.06 and 0.04±0.05, respectively; change percentage: 3.7±6.8 and 4.0±6.0%, respectively; all P\<0.001). The VE was higher standing than in lying and sitting positions (mean difference: 1.22±0.88 and 1.27±0.94 L/min, respectively; change percentage: 13.5±9.2 and 13.9±9.7%, respectively; all P\<0.001), yet no differences were found between lying and sitting positions. Furthermore, there were differences in RR between lying and sitting positions, and between sitting and standing positions (mean difference: 1.88±2.65 and 1.18±2.02 breath/min, respectively; change percentage: -14.93±20.65 and 6.90±13.77%, respectively; all P\<0.001); yet no differences were observed between lying and standing positions (P = 0.295). Moreover, The HR was significantly higher standing than in sitting (mean difference: 16±8 beats per minute; change percentage: 16.6±9.1%; P\<0.001) and lying positions (mean difference: 25±15 beats per minute; change percentage: 26.0±17.8; P\<0.001). The HR was also higher sitting than lying (mean difference: 9.2±6.4 beats per minute; change percentage: 11.3±8.7; P\<0.001). shows the association between the change on EE of changing from one position to another in all different position pairs (i.e. lying vs. sitting; sitting vs. standing; sitting vs. standing) and the anthropometric and body composition parameters. There was a significant negative association between the EE differences in sitting vs. standing position and lean body mass (R² = 0.078, P = 0.048), yet no significant associations between EE differences with the rest of the anthropometric and body composition parameters were observed in each position pair studied (all P\>0.321). We also noticed no significant associations between the RQ differences in all different position pairs and the anthropometric and body composition parameters (all P\>0.321). According to the EE change from lying to sitting, 71.7% of participants (n = 38) were classified as savers, and only 28.3% (n = 15) were spenders (EE change in savers and spenders: -3.2±5.5% vs. 11.9±7.1%, respectively; P\<0.001;). By definition, the EE change from lying to standing was higher in the spenders group (spenders: 71.7%, n = 38; mean change -2.6±6.5% and 15.3±9.0%, savers and spenders, respectively, P\<0.001;). Furthermore, the EE change from sitting to standing was also higher in the spenders group (spenders: 18.9%, n = 10; -0.8±3.6; vs. 8.5±3.7%, savers and spenders, respectively, P\<0.001;). There were no significant differences between savers and spenders in RQ in any of the three pairs of positions (all P\>0.248). We also studied the anthropometric and body composition parameters as well as the EE lying, sitting, and standing in savers and spenders, and in the three pairs of positions studied: (i) lying vs. sitting, (ii) lying vs. standing, and (iii) sitting vs. standing. There were no significant differences in age, BMI, lean body mass and fat body mass, except when comparing saver vs. spender in sitting vs. standing position in term of lean body mass (42.65±8.60 vs. 37.50±6.32 kg respectively, P\<0.05). We found significant differences in EE between all of the three position pairs established between saver and spender. All of these findings persisted after controlling for the time of the day when the test was performed and for the menstrual cycle period in women (data not shown). Interestingly, we also found that EE relative to lean body mass was higher in women compared with men in lying, sitting and standing position (0.032±0.008, 0.032±0.007, 0.035±0.009 vs. 0.025±0.006, 0.024±0.00, 0.026±0.006, respectively, all P\>0.01). # Discussion The main findings of this study showed that standing increases EE above sitting and lying values (\~10%), while sitting and lying paradoxically seems to represent similar EE. Taken together, these findings suggest that decreasing lying and sitting times could be a simple strategy to slightly increase energy expenditure. Our results concur with those of others who also showed higher EE when standing than when sitting. However, Monnard et al. reported that a multi-ethnic male cohort had the same EE when comparing the sitting and standing positions. Our results also indicate that EE is similar in sitting compared to lying in young healthy adults, which concurs with other studies. Similarly, although no differences were found in EE between lying and sitting, we found lower RQ values when the participants were sitting compared with when they were lying. The changes (decrease) in RQ from lying to sitting, and from lying to standing had been previously found, but it remains necessary to thoroughly investigate the physiological mechanisms that could explain such changes. Our results suggest that the lack of differences in EE between lying and sitting, and higher values in RQ in the former could be due to an extra activation of ventilatory muscles (e.g. diaphragm) and to a loss of ventilatory efficiency in lying vs. sitting, which would also be reflected in higher RR in lying position. Postural maintenance needs a higher muscle activation sitting than lying. Therefore, a higher EE would be expected when sitting than when lying. We hypothesise that gravity-induced displacement of viscera weight against the diaphragm would cause and over-load to ventilatory muscles which ultimately contributes to a higher EE, compensating the relaxed postural muscles. This could be reflected in higher RR while lying, despite similar VE (i.e. decrease in ventilatory efficiency). When the ventilatory muscles are slightly over-loaded, they would need to recruit fast fibers (which involve carbohydrate oxidation metabolism) in higher proportion than postural muscles, which would explain why we found higher RQ but similar energy expenditure in lying than in sitting positions. However, it is well-known that RQ is heavily influenced by VE and that hyperventilation leads to more clearance of CO2 obtaining a higher RQ. This issue is a reflection of breathing pattern but not a change in substrate utilization. Therefore, these findings should be deeply investigated in future studies. Paradoxically, our results presented a large inter-individual variability showing lower EE during standing compared to lying in a number of participants (n = 7). A possible explanation for this finding could be that the initial lying assessment had not reached a steady-state, thereby overestimating the measurement. We showed that a total of 14 participants did not meet the steady state criteria, but after sensivity analysis excluding those individuals that did not attain the steady state, the results persisted. Further studies are needed to deeply investigate this concern. Interestingly, most participants in our study showed a small or even no rise in the EE when standing compared to sitting, being categorised as savers a total of 81% of the study participants, which concurs with other studies. The mechanism by which the large majority of participants appear to be spenders remains to be elucidated, but anthropometric and body composition variables may, in part, be related to these phenotypes. We did not find any association between EE differences and almost every anthropometric, or body composition variables when compared lying vs. sitting, and lying vs. standing position. However, despite the inter-individual variability among participants in terms of EE, this study indicates that when compared sitting vs. standing position, EE differences could be explained in part by lean body mass. Miles-Chan et al. compared energy cost in standing vs. sitting positions, and no associations were found between the EE differences and anthropometry (body weight or height) in 22 young adults with normal BMI, which concurs with our findings. However, a recent study showed that the body weight and the leg length might contribute to the inter-individual variability in the EE (R² = 0.548; P = 0.001, and R² = 0.460; P = 0.006, respectively) considering sitting vs. standing positions, but it was not taken into account body composition parameters in the regression analysis. We found a significant negative correlation between the EE differences in sitting vs. standing position and lean body mass. Our results could be explained by the fact that lean body mass is positively correlated with efficiency in term of EE. Therefore, the present study supports the argument that individuals with lower lean body mass have lower EE in resting condition (i.e. lying position), but show higher EE differences considering sitting vs. standing position. There are clinical and research implications derived from these findings. Firstly, in order to develop several strategies related to fighting against metabolic diseases related to energy balance, it is important to consider that the EE is higher when standing than when lying and sitting. Secondly, the EE can be accurately determined in both lying and sitting positions, but it is partially incorrect when the aim is to describe the nutrient oxidation rate, because RQ was not comparable between positions. Therefore, it is tempting to speculate that lying with a specific bed inclination could be the best approach to determine the “real” EE, since this position would help to avoid extra activation of ventilatory muscles increasing ventilatory efficiency, and also contributing to a minimum recruitment of postural muscles. The results of this study should be considered with caution as there are some limitations. Although we standardised the protocol test, the order of the positions was not randomised and a drag effect may have been produced. We carefully controlled the fasting time (5–6 hours) prior to the test, but the best practice guidelines suggest to established at least 7 hours. Moreover, the composition of the previous meal was not standardised, and this fact may have influenced the RQ measurement. The shorter supine resting period prior to the test protocol may have affected subsequent measurements during the lying position and thus the lying vs. sitting comparison. In addition, although the majority of the participants of the current study met the steady state criteria in lying (73.6%), sitting (58.5%), and standing (62.3%), there was a number of individuals that did not meet the above-mentioned criteria causing a potential overestimation of EE in all postures. In conclusion, our findings support that increasing the time spent standing could be a simple strategy to increase the EE. In fact, it is clear that reducing sitting time should be encouraged according to estimations indicating that substituting 6 hours of sitting per day with standing results in 45 additional kcal in daily energy expenditure. Therefore, our findings have important clinical implications including a better monitoring, characterizing, and promoting countermeasures to sedentariness through low-level physical activities. # Supporting information We are grateful to Ms. Carmen Sainz-Quinn for assistance with the English language. This study is part of a Ph.D. Thesis conducted in the Biomedicine Doctoral Studies of the University of Granada, Spain. The authors declare no competing financial interests. [^1]: The authors have declared that no competing interests exist.

# Introduction Urogenital chlamydia is the most prevalent, curable bacterial sexually transmitted infection (STI) worldwide, with a significant public health burden, especially in young women. The causative bacterium, *Chlamydia trachomatis* (Ct) causes a high rate of asymptomatic infections and is associated with adverse outcomes like infertility, ectopic pregnancy and pelvic inflammatory disease (PID). To reduce transmission and late complications, active case finding and early treatment are critical strategies. The standard diagnostics are Nucleic Acid Amplification Tests (NAAT), but they are expensive and require sophisticated laboratory conditions. This makes NAAT unsuitable for the detection of Ct for most low-resource settings. Therefore the World Health organization (WHO) has launched a priority program that is designated to develop affordable and reliable point-of-care (POC) tests for STIs that are predominant in low resource countries \[<http://www.who.int/std_diagnostics>\]. In this program, WHO has formulated the ASSURED criteria that POC tests have to meet: Affordable, Sensitive, Specific, User-friendly, Robust and rapid, Equipment- free, Deliverable to those who need them. The POC test result should be readily available, while the patient waits, to ensure prompt treatment. This is especially important where patient return for treatment is low. It is estimated that a POC test of moderate sensitivity (63%) combined with immediate treatment on-site may lead to the treatment of more infected individuals than an ultra- sensitive and specific NAAT alone when patient return is low. Moreover, counselling messages are most efficient when a diagnosis can be communicated during the same consultation. These advantages are relevant for industrialized countries as well, even if POC tests have a lower sensitivity than standard NAAT. To date, POC tests for urogenital chlamydia show disappointing test characteristics, especially low sensitivity. In a recent evaluation, three POC tests for urogenital chlamydia, currently on the market, showed poor sensitivity between 12% and 17% in a non–manufacturer-sponsored clinical study. In contrast, one POC test for urogenital chlamydia (Diagnostics for the Real World, Cambridge, UK) especially developed for low-resource countries has an asserted sensitivity of over 80%. A manufacturer-sponsored diagnostic field study in the Philippines revealed sensitivities of 71% and 87% among women at high risk and low risk for STI, respectively. Suriname, South America, is a low-resource country and affordable and reliable diagnostics to detect Ct are urgently needed. Therefore, we aimed to evaluate the performance of this promising POC test in two outpatient clinics in Suriname, with the objective to use this test for intervention of the chlamydia epidemic. # Methods ## Study sites and population The study was approved by the ethical committee of the Ministry of Health of the Republic of Suriname (VG010-2007) and the ethical committee of the Academic Medical Centre, University of Amsterdam, the Netherlands (MEC07/127). Patients were recruited at two sites in Paramaribo, Suriname: 1. The Dermatological Service, an integrated outpatient clinic that offers free-of-charge examination and treatment of STIs and infectious skin diseases like leprosy and leishmaniasis. All consecutive women who visited for an STI check-up were asked to participate in the study and were considered to be at high-risk for STI. 2. The Lobi Foundation is a center for birth control and sexual health. As women who visit this clinic do not attend primarily to be checked for STI, these participants were considered to be at low risk for STI. Recruitment took place between July 2009 and February 2010. Exclusion criteria were: use of antibiotics in the past 7 days, age younger than 18 years and previous participation. After written informed consent, patients were given a unique code to participate anonymously. Participants were interviewed about demographic characteristics, including self-reported ethnicity as Suriname is a multiethnic society, with many ethnic groups such as Creoles and Maroons (both descendants from the African diaspora due to slave trade), Hindustani, Javanese, and Chinese (all descendants from labor immigrants), Caucasians (descendants from Dutch farmers), indigenous Amerindian people and Mixed race persons. Moreover, participants were asked about willingness to wait for POC test results, although in our study participants did not receive the results from POC, and if they used any products for vaginal hygiene like douches, herbs, or other home products, and if so, in what frequency. Data were entered into an MS Access database. ## Specimen collection and testing procedures Nurse-collected vaginal swabs were obtained blindly for the Chlamydia Rapid Test (CRT) (Diagnostics for the Real World (Europe), Cambridge, UK) and NAAT (Aptima, Gen-Probe, San Diego, USA) testing using a cross-over model. This means that in the first half of the included women the swab for the CRT was taken first and the second of the included women NAAT was taken first. Nurses were trained to collect the swabs before routine speculum examination was performed. A minimum period of 10 times for CRT and 10 seconds for NAAT of contact between the tip of the swab and the vaginal wall in a rotating motion was ensured. CRT was immediately performed according to the manufacturer's instructions on-site in the laboratory. All technicians that performed the CRT were trained with proficiency panels as provided and instructed by the manufacturer. Technicians did not receive information about the participant. The test results were interpreted and recorded by two laboratory technicians separately. CRT results were defined as indeterminate when the laboratory technicians reported discordant results or when CRT failed (i.e. control line did not appear). The samples for NAAT testing were collected according to the manufacturer's instructions, and shipped to the Public Health Laboratory in Amsterdam where they were tested within 50 days after collection. NAAT test results were communicated with the two recruitment sites in Suriname and participants with a positive-Ct NAAT were treated with doxycycline 100 mg bid for 7 days at Lobi Foundation and 10 days bid at the Dermatological Service or, in case of (possible) pregnancy, with a single 1000 mg oral dose of azithromycin. ## Chlamydia Rapid Test The CRT was performed as described previously. Version 6.1 of the *Chlamydia* Rapid Test (Professional use) (P/N 1200-20) instructions for use (C03-0008) was used. Shortly, each swab was subjected to extraction by sequential addition of 400 µl of reagent 1, 300 µl of reagent 2, and 100 µl of reagent 3 to the swab in a tapered sample preparation tube, with gentle mixing between additions. The sample preparation reagents were administered with unit dose pipettes. The extraction tube was then capped and used as a dropper to deliver 5 drops (approximately 100 µl) of the extracted sample to a tube containing the lyophilized amplification and detection reagents. The resulting mixture was agitated gently until a clear pink solution was obtained, after which the test strip, coated with a monoclonal antibody to chlamydial lipopolysaccharide (LPS) and including a procedural control, was added to the solution and allowed to stand for 25 minutes before the result was read. Each swab was subjected to one extraction. The test strip was used in the interpretation of the result; a clearly visible test line indicated a positive result, provided that the control line was also visible on the test strip. ## NAAT testing For NAAT testing, the monospecific Aptima chlamydia assay for the detection of *Chlamydia trachomatis* rRNA (Gen-Probe Inc., San Diego, USA) was used with the accompanying vaginal swab specimen collection kit. The protocols described in the package inserts were followed. Technicians performing NAAT were blinded to the results of the POC-Ct and did not receive clinical information. This NAAT is an FDA-approved commercial test and was used to estimate the Ct prevalence at both study sites. ## Quantitation of Ct load and HLA Quantitative Ct load was determined for samples with a discrepant test result between CRT and NAAT, and for samples that tested positive for CRT as well as for NAAT using a real-time PCR targeting the cryptic plasmid. Ct load was expressed as inclusion forming units (IFU) based on defined serial dilutions of Ct cultured in human cells with over \>90% infected HeLa cells of 100 IFU to 0.001 IFU taking into account also DNA from non-viable Ct particles. The human cell load was assessed by determination of human HLA copies in combination with a defined serial dilution of quantified human DNA using the following primer probe combination: HLA-F 5′-TTG-TAC-CAG-TTT-TAC-GGT-CCC-3′ HLA-R 5′- TGG-TAG- CAG-CGG- TAG-AGT-TG,-3 and HLA-Probe 5′-FAM- TTC TAC GTG GAC CTG GAG AGG AAG GAG -BHQ1-3′. By using a chlamydial and a human target, the average chlamydial/human cell ratio, and IFU/swab were calculated. ## Statistical analysis To evaluate the performance of CRT compared to NAAT sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were calculated using standard methods. Specimens with indeterminate results by CRT were excluded. An independent t-test was used to compare log-transformed Ct loads between true-positive and false-negative CRT results. Analyses were performed with SPSS package version 19.0 (SPSS Inc., Chicago, IL). The study has been reported according to the STARD checklist for the reporting of studies of diagnostic accuracy. # Results ## Study population and specimens In total, 1019 women were asked to participate in the study, of whom 917 were included and 102 did not meet the inclusion criteria or declined to participate. Five women were excluded from the CRT performance evaluation due to either discrepancy in CRT result between two lab technicians (n = 3) or failure of CRT (n = 2). General characteristics of the 912 women included in the CRT performance evaluation are shown in. Their median age was 30 years (IQR 25–36), 336 (36.9%) were of Creole/Maroon ethnicity and 229 (25.1%) were of Hindustani ethnicity. Twenty-one (2.3%) women reported having had sex for money or goods. Almost all women 900 (98.7%) would wait for the CRT test result if the test were a standard offering in their clinic. Of these women, 660 (73.3%) would be willing to wait for a maximum of half an hour to receive the results, the other 240 (26.7%) would be willing to wait for at least an hour. ## Ct prevalence and CRT performance results Ct prevalence was 20.8% in the high-risk population (visiting the Dermatological Service) and 9.2% in the low-risk population (visiting Lobi Foundation). Combining the results of the two sites, the sensitivity and specificity of the CRT in identifying Ct compared to NAAT were 41.2% (95% CI, 31.9%–50.9%) and 96.4% (95% CI, 95.0%–97.5%), respectively. PPV of the CRT was 59.2% (95% CI, 47.5%–70.1%) and NPV was 92.9% (95% CI, 91.0%–94.5%). Sensitivity and specificity of CRT compared to NAAT were comparable for the high-risk population (39.4% and 94.4%) and the low-risk population (42.0% and 96.8%). ## Quantitative load measurements Quantitative Ct bacterial load and human HLA were assessed for the samples that showed discrepant results for CRT and NAAT (n = 89) and for samples that were CRT and NAAT positive (n = 42). Ct bacterial load could be detected in 99/131 samples and human HLA in 126/131 samples. Of the 42 samples that tested positive for CRT and NAAT, quantitative Ct bacterial load was detected in all 42 samples and human HLA in 39 samples. Of the 60 samples that tested CRT negative and NAAT positive, quantitative Ct bacterial load was detected in 55 samples and human HLA in all 60 samples. Of the 29 samples that tested CRT positive and NAAT negative, quantitative Ct bacterial load was detected in 2 samples and human HLA in 27 samples. Quantitative Ct bacterial load was 73 times higher in NAAT-positive/CRT-positive samples (geometric mean 120 IFU) compared to NAAT-positive/CRT-negative samples (geometric mean 1.64 IFU, p\<0.001). Human DNA concentration did not differ between the true-positive and false-negative CRT results (p = 0.835). The average chlamydial/human cell load ratio (Ct concentration) was 60 times higher in NAAT-positive samples where CRT detected Ct infection (geometric mean 0.32 IFU/human cell) compared to loads that CRT did not detect (geometric mean 0.0053 IFU/human cell, p\<0.001). Quantitative HLA load was comparable for NAAT- positive/CRT-positive samples (geometric mean 344 cells) compared to NAAT- negative/CRT-positive samples (geometric mean 451 cells, p = 0.424). Quantitative Ct loads were comparable for women reporting symptoms like vaginal discharge, irregular menstruation, pain during intercourse, lower abdominal pain or dysuria and women without the specific symptom (data not shown). Women visiting the high-risk STI clinic had comparable quantitative Ct loads with those visiting the low-risk clinic (p = 0.525). Sensitivity of the CRT was comparable for those who practiced any vaginal hygienic measures, 37.5% (95% CI, 23.6%–53.1%), compared to those who did not practice vaginal cleansing, 43.3% (95% CI, 31.3%–56.0%). When comparing women who practice vaginal cleansing frequently, at least once a week, with those who cleanse less than once weekly, sensitivity of CRT yields comparable results, 39.1% (95% CI, 21.1%–59.8%) and 27.3% (95% CI, 7.5%–57.8%), respectively. Based on the overall median Ct load, NAAT-positive samples were divided in two groups with either a low- (range 0.006–12.5 IFU) or high-grade quantitative bacterial Ct load (range 13.5–6470 IFU). In the low-grade bacterial load group, the CRT sensitivity was 12.5% (95% CI, 5.2%–24.2%), whereas in the high-grade Ct load group the sensitivity was 73.5% (95% CI, 59.9%–84.4%). # Discussion We found a disappointingly low clinical sensitivity of 42.0% and 39.4% of the CRT in low-risk and high-risk Surinamese women, respectively, compared to the sensitivity of 86.8% in low-risk women and 71% in high-risk women in the Philippines, reported earlier in a study supported by the manufacturer. The discrepancy might partly be explained by the use of a different reference test. Where we used Gen-Probe's Aptima platform as a reference test, in the Philippines study the Roche Amplicor (Roche Molecular Systems, Branchburg, NJ) was used. Although current generation NAATs have comparable sensitivities, NAAT could be more sensitive than Roche Amplicor. A somewhat lower sensitivity of CRT in our study could be expected with a more sensitive control test, but this does not explain the large difference in sensitivity found in the Philippine study and our results. Another explanation for the lower sensitivity we found could be attributed to a different wash-out period for antibiotic use between the two studies. We excluded women who used antibiotics in the last 7 days, while in the Philippines study women who used antibiotics in the previous month were excluded. Time to clearance of LPS antigen, which is targeted by the CRT, might be shorter after antibiotic use than time to clearance of Ct rRNA, which is targeted by NAAT. This could have caused the occurrence of false-positive NAAT samples, and consequently more false-negative CRT samples could be expected. Low sensitivity of the CRT due to inadequate collection resulting in a low sample yield could be ruled out since the human cell load in samples with true-positive and false- negative CRT results was comparable. The CRT had a 96.4% specificity. False- positive CRT results could have been caused by cross reactivity with *C. ptsittaci* or *C. pneumoniae* as described in the manufacturers manual. Yet infections with these organisms in the urogenital tract in humans are uncommon. As a false positive chlamydia diagnosis can have serious adverse social consequences a specificity of 96,4% is undesirable, especially in low prevalent settings. The CRT in our study had a few modifications compared to the study in the Philippines. We used unit dose pipettes instead of unit dose vials. Also, the nitrocellulose membrane was changed by the manufacturer and according to the manual, only one dipstick had to be used to interpret the results. However, when a test is renewed one might expect at least comparable diagnostic characteristics compared to the previous test. In the CRT evaluation study performed in the Philippines, the Ct prevalence was 6.3% in the low-risk group (women visiting an obstetrics-gynaecology clinic) and 17.9 to 32% in the high-risk group (female sex workers), which compares well with the prevalences found in our study, 9.2% and 20.8% respectively. The sensitivity figures found in our study were comparable for low-risk and high- risk women, 42.0% and 39.4% respectively. Quite surprisingly, in the Philippines study a much lower sensitivity was found in the high-risk group compared to the low-risk group. The authors explain this finding as a result of the use of vaginal creams and other feminine hygiene products, which can interfere with the CRT. In our study, the sensitivity of CRT was comparable for women who practiced any vaginal cleansing and those who did not. Although we consider the population recruited at Lobi Foundation a low risk group, with a prevalence of 9.2% this population would be considered high risk in many settings. Yet, with a prevalence of 20.8% as found at the Dermatological Service, the difference in prevalence between the two study sites is substantial. The sensitivity of CRT is higher in samples with a high bacterial load. The clinical relevance of organism load is still debated, but it is suggested that infections with high organism loads are more likely to lead to cervicitis or PID and are associated with multiple patient-reported symptoms. However, the association with patient-reported symptoms was only found with first-void urine and endocervical samples and not with self-collected vaginal samples. In our study, where nurse-collected vaginal swabs were used, quantitative Ct loads were not significantly different for asymptomatic women and women reporting one or multiple symptoms such as vaginal discharge or dysuria. The NAAT platform is a latest generation highly sensitive commercial diagnostic test for Ct. However no test is 100% accurate and a positive bacterial Ct load signal was detected in two samples that were NAAT negative and CRT positive. One sample had a Ct load of 62.9 IFU which might be explained by inhibition of high target load. The other sample had a very low load of 0.00261 IFU. Since the frequency of these discrepancies was extremely low, we do not consider that this finding significantly affects our test evaluation. A recent field study of the same CRT test but to detect ocular chlamydia infection (trachoma) found similar disappointingly low sensitivity (33.3%–67.9%) and specificity (92.4%–99.0%). Most commercially available and clinically evaluated POC tests for urogenital chlamydia show poor sensitivity results. Compared with the results found in our evaluation, the CRT of Diagnostics for the Real World outperforms some of the other commercially available products. Still, with a sensitivity of only 41.7%, this test performs under the minimally required sensitivity of 63% required for a POC test to treat more infected individuals than the standard NAAT in a setting with low patient return (\<65%). On the other hand, in situations where transmission during treatment delay and low return for treatment are considerable, even a POC test with a sensitivity below 63% could be beneficial in the prevention of ongoing STI transmission. A recent economic evaluation analysis using the same CRT as we evaluated in this study, showed that in the UK using NAAT is more cost-effective.. In that evaluation, a sensitivity between 73% and 85% for the CRT was assumed. POC tests available for systemic infections like HIV and syphilis are highly sensitive since they are based on the detection of serum antibodies. Infections caused by organisms like Ct (but also *N. gonorrhoeae)* are confined to mucosal tissue and normally invoke little to no production of antibodies. Therefore, the development of POC tests to diagnose mucosal Ct infections based on the detection of serum antibodies is, at least for now, not an option. Improved POC tests for Ct need to detect bacterial antigens or nucleic acids, even in cases with a low bacterial load. Promising steps have been made in the field of POC HIV-load NAAT using nanotechnology. Along the same lines, a POC test for urogenital chlamydia with sufficient sensitivity could be developed. Until reliable and affordable diagnostics are available, algorithms for syndromic management can be used for low-resource settings, although the success of algorithms for vaginal discharge varies between populations. In conclusion, the evaluated CRT of Diagnostics for the Real World has no added value in the management of Ct infections due to its low test performance. There is an urgent need for POC diagnostics for the detection of urogenital chlamydia meeting the ASSURED criteria, including adequate sensitivity. The authors would like to express their gratitude to all nurses and laboratory technicians of the Dermatological Service and the Lobi Foundation for data collection, and Susan T. Landry for editing the final manuscript. [^1]: Conceived and designed the experiments: JvdH LS SM AS HdV. Performed the experiments: JvdH AG SM AS. Analyzed the data: JvdH SM HdV. Contributed reagents/materials/analysis tools: LS AG SM AS HdV. Wrote the paper: JvdH LS AG SM HdV. [^2]: The authors have declared that no competing interests exist.

# Introduction Given the relationships of item response theory (IRT) models to confirmatory factor analysis (CFA) models, IRT model misspecifications might be detectable through model fit indices commonly used in categorical CFA. IRT models share many features with CFA models, and in some cases are equivalent to CFA models. Hence, IRT model misspecifications may be detectable through model fit indices commonly used in CFA. The Rasch model is mathematically equivalent to one- parameter IRT models, and it has been widely used to estimate item and ability parameters from measurement data. From the perspective of CFA, a unidimensional Rasch model can be considered as equivalent to a particular form of a unidimensional, categorical CFA model. Specifically, if the loadings of the unidimensional, categorical CFA model are constrained to be equal and fixed to one, the statistical expression of the model is equivalent to the unidimensional Rasch model. Previous research showed that weighted least squares means and variance adjusted (WLSMV) global model fit indices used in structural equating modeling practice are sensitive to person parameter estimate RMSE and item difficulty parameter estimate RMSE that results from local dependence in two-parameter logistic (2-PL) IRT models, particularly when conditioning on number of test items and sample size. In this study, I hypothesize that there is a clear relationship between WLSMV-based fit indices and LD-induced inaccurate parameter estimate when fitting models in which the discrimination parameter is fixed. That is, Rasch model applications will demonstrate a clearer relationship between these fit indices and inaccurate parameter estimate from LD than was found for 2-PL models in Huggins-Manley and Han. If this hypothesis is supported by the study, practitioners can use the fit indices that would be useful supplementary statistics to detect problematic LD in Rasch model applications. This would be a benefit to the measurement field as many current methods for detecting LD require complex computations, provide information at the item or item-pair level, and/or provide vague information with respect to how problematic the LD is for obtaining unbiased parameter estimates. Therefore, the purpose of this study is to evaluate the sensitivity of WLSMV-based global fit indices to violations of local independence in Rasch models, particularly violations that result in inaccurate person estimate and/or item parameter estimate. Additionally, to examine whether those fit indices are related to other widely used methodologies in applied Rasch modeling, the author compares the results to principal component on residual (PCAR) analysis. PCAR is the most popular method particularly in assessing local independence assumption in Rasch models. The use of the first eigenvalue from PCAR is used to check local dependency due to the presence of multidimemtionality. Several rules of thumb regarding the first eigenvalue from PCAR were suggested. Since the use of PCAR for local independence testing serves as a goodness of fit statistic as WLSMV- based fit indices does in this study, PCAR results are provided in the comparison of WLSMV-based fit indices. The Rasch model for binary responses is considered in this study. The specific research questions are: 1. Are CFI, TLI, and RMSEA indices from WLSMV estimation of binary Rasch models sensitive to LD that causes inaccurate estimates of item and person parameters? 2. Is the answer to Question 1 dependent on the number of observed variables, sample size, type of LD (trait or response dependence), and/or magnitude of LD? 3. Are the WLSMV-based fit indices useful in the comparison of PCAR results? ## Types of local dependence in rasch model Marais and Andrich distinguished two types of violations of LD. One, a violation of unidimensionality, is called trait dependence (TD). The other, a violation of statistical independence, is called response dependence (RD). ### Trait dependence Marais and Andrich pointed out that the TD can be considered a violation of unidimensionality. It is common that the scale used in social sciences tends to be constructed as a set of different subsets even though the instrument is intended to measure a single psychological trait. The instrument designed as a composition of the subsets has the advantage of obtaining measurement validity because the subsets design is capable of reaching out the various aspects of the trait. However, it has risk of violating the unidimensionality. Thus, if the subsets in the instrument measure a different latent trait, the instrument can no longer be considered a single instrument to measure a single latent trait. In this study, under the Rasch model, TD can be formularized as $$P\left( {\left. {X_{ip}^{(s)} = 1} \right|\theta_{p},\mspace{360mu} b_{i}} \right) = \frac{e^{({\theta_{ps} - b_{i}})}}{1 + e^{({\theta_{ps} - b_{i}})}},$$ where the subscript s represents subset s = 1, 2, …, S, *p* represents examinee, and *i* represents item. *θ*_*ps* is defined as the sum of primary traits among subsets and distinct traits by subset s. Specifically, let us consider a scale composed of s = 1, 2, …, S and let $$\theta_{ps} = \theta_{p} + \text{c}_{\text{s}}\theta_{ps}^{\prime}$$ where *θ*_*p* is the primary trait among subsets, and $\theta_{ps}^{\prime}$ is the distinct trait characterized by subset s. It is hypothesized that *θ*_*p* is not correlated with $\theta_{ps}^{\prime}$ and neither are any $\theta_{ps}^{\prime}$ among any subsets mutually correlated. This indicates that $\theta_{ps}^{\prime}$ is a trait unique to each subset. The value of c_s is the magnitude of each specific subset s and c_s \> 0. The correlation between the estimated traits of among any subsets s and t is given by $\rho_{st} = 1/\sqrt{1 + c_{s}}\sqrt{1 + c_{t}}$. Assuming *c*_*s* = *c*_*t* = *c* gives *ρ*_*st* = 1/(1 + *c*²) (Marais & Andrich, 2008). By setting the magnitude of trait dependence 0, 1, and 2, the correlation between estimated traits among subsets is 1.0, 0.5, and 0.2, respectively. The correlation of 1.0 between traits among subsets indicates no violation of unidimensionality in the Rasch model. On the other hand, traits correlation among subsets less than 1.0 indicates violation of unidimensionality depending on the magnitude of trait dependence term, c. ### Response dependence Marais and Andrich also suggested that RD can be considered as a violation of statistical independence. In other words, under RD, examinees’ response to an item depends on their response to a previous item. For example, if an examinee was able to give a correct answer for a previous item, he or she is more likely to get the next item correct. On the other hand, if the examinee gave an incorrect answer for the previous item, he or she is less likely to get the next item correct. This notion also generalizes to any set of binary, ordered response data (e.g., for an opinion-based item, 0 is disagree and 1 is agree). Under the Rasch model, RD can be formulized as following two equations: $$P\left( {\left. {X_{jp} = 1} \right|X_{ip} = 1;\theta_{p},\mspace{360mu} b_{i}} \right) = \frac{e^{({\theta_{p} - {({b_{j} - d})}})}}{1 + e^{({\theta_{p} - {({b_{j} - d})}})}},$$ and $$P\left( {\left. {X_{jp} = 1} \right|X_{ip} = 0;\theta_{p},\mspace{360mu} b_{i}} \right) = \frac{e^{({\theta_{p} - {({b_{j} + d})}})}}{1 + e^{({\theta_{p} - {({b_{j} + d})}})}},$$ where *p* represents the examinee, *i* represents the preceding item, and *j* represents the following item. This formula indicates that the response of an examinee to item *j* depends on the examinee’s response to item *i*. The value of d is adding or subtracting depending on whether the examinees’ response to preceding item was correct or endorsing. Specifically, if an examinee’s response to item *i* is correct or endorsing, the difficulty level of the following or dependent item *j*’ is made lower by the subtraction of the response dependence term (i.e., d). This makes the following or dependent item easier, which results in an increase in the examinee’s probability of getting the following or dependent item correct or endorsing. On the other hand, if an examinee’s response to item *i* is incorrect or not endorsing, the difficulty level of the following or dependent item *j*’ is made higher by the addition of the response dependence term (i.e., d). This makes the following or dependent item harder, which results in a decrease in the examinee’s probability of getting the following or dependent item correct or endorsing. ## Model fit indices for CFA with WLSMV estimation In order to deal with non-normality with categorical indicators, WLSMV methods have been recommended over the ML method. WLSMV is a limited information estimation method that utilize summary statistics (i.e., tetrachoric or polychoric correlations). As a limited information method, WLSMV is not only a robust method but also computationally fast, especially when the sample size and the number of dimensions are large. Thus, this study utilizes the WLSMV-based approach on CFI, TLI, and RMSEA indices. The calculations of those indices are presented below as Mplus, as is used in the study. Both CFI and TLI come from the *χ*² statistic obtained through the WLSMV estimation. They compare those statistics to the WLSMV *χ*² statistic obtained from a baseline modeling. The baseline model collectively places zero to all item parameters in the categorical data except thresholds. After obtaining *χ*² for both, the CFI and TLI are calculated as shown below. $$CFI = 1 - \frac{\text{max}\left( {\chi_{t}^{2} - df_{t},0} \right)}{\text{max}\left( {\chi_{t}^{2} - df_{t},\mspace{360mu}\mspace{360mu}\chi_{b}^{2} - df_{b},0} \right)},$$ and, $$TLI = \frac{\chi_{b}^{2} - \chi_{t}^{2}\left( \frac{df_{b}}{df_{t}} \right)}{\chi_{b}^{2} - df_{b}},$$ where *b* represents the baseline model, *t* represents the tested model, and *df* represents degrees of freedom. The RMSEA indices apply a fit statistic in quantifying the average discrepancies arising between expected and observed covariance within a specific data set in the defined latent variable model. The use of RMSEA in categorical data should be adjusted to account for the non-normality in the observed data. This is achieved by obtaining *χ*² from the WLSMV estimation, which enables the calculation of RMSEA as $$RMSEA = \sqrt{\frac{\chi^{2} - df}{df\left( {N - 1} \right)}},$$ where N represents the sample size, and *df* represents the degree of freedom. The RMSEA takes into account the model degree of freedom and sample size. The degree of freedom is considered a measure of model complexity. The rationale is that for more degrees of freedom, there is an increase in the number of variable relations in the model. Similarly, the relation between sample size and degrees of freedom is incorporated into the RMSEA value, which is standardized. In CFA practices, model fit indices are commonly used to gauge model misspecification. Since fit is a matter of degree, Hu and Bentler have provided recommended cutoff criteria for those indices (e.g., CFI/TLI \> 0.95, RMSEA \< 0.06), which are now widely used in CFA applications. However, those are maximum likelihood based fit indices for continuous variables and less is known about WLSMV-based fit indices for ordered categorical variables. Recent study showed that it is not appropriate to use Hu and Bentler’s conventional cutoff benchmarks (i.e., CFI, TLI, RMSEA) for ordered categorical variables with WLSMV estimation. Instead of using fixed cutoff criteria, they suggested WLSMV based fit indices would be used as diagnostics tools for model specification. ## Principal component analysis on residuals in rasch model The Rasch model assumes that the measurement model is unidimensional. The idea of PCAR is that there should be no meaningful pattern among the item residuals, after controlling for the single latent factor of the items by the Rasch model. The magnitude of the eigenvalue of the first component is considered to be an indicator of violation of unidimensionality. The rule of thumbs of the magnitude of the value depends on previous research. For example, Smith insisted the value of higher than 1.5 implies a lack of unidimensionality under 500 persons and 30 items. Smith & Miao suggested the value no less than 1.4 implies a lack of unidimensionality, while several studies suggested the value no less than 2.0 in assessing unidimensionality assumption. Although no fixed cut point would be applicable as Chou and Wang pointed out, PCAR has been widely used to examine dimensionality in Rasch models by many practitioners. # Methodology ## Design overview A simulation study was conducted in R 4.0.2, with batching of estimation to Mplus via Mplus Automation package. For PCAR, the *eRm* package and the psych package was used. A summary of the simulation method is presented here. The following subsections provide details for each part of the method. Factors manipulated in this study include type of LD (TD or RD), magnitude of LD, sample size, and number of items. Each simulation condition was replicated a thousand times. The item factor model constrains all loadings to be equal, which results in the same parameter estimation as the binary Rasch model. The Rasch model was fit to all data sets to estimate person and item parameters. Root mean square error (RMSE) were calculated as the evaluation criteria of parameter recovery. A series of correlational and graphic analyses was conducted to investigate the relationships between RMSE of parameter estimates and fit indices from WLSMV estimation. ## Data generation Simulated data were generated from the Marais and Andrich’s Rasch model for TD and RD, respectively (See Eqs through). For the violation of local independence in Rasch model due to RD, the simulated data was generated by Eqs and. ### Simulation conditions The basic design structures for both models are the same and consist of six subsets with dependence, TD or RD. All thetas including primary and subset specific thetas were drawn independently from standard normal distribution and item difficulties were drawn from -2 (i.e., easiest) to 2 (i.e., hardest) with equal intervals. The critical factor towards examining the research questions was varying the amount of LD in the generated data. The variation process on the amount of LD was varied for TD and RD. To align with Marais and Andrich (2008b), the constant c was varied from 0, 1, or 2 from the formula 1 for TD in the Rasch model; the constant d was varied from 0, 1, or 2 from the formula 3 and 4 for RD in the Rasch model. Under each manipulation of TD or RD, the number of items and sample size were varied in this study. To investigate the effect of test length factor, different number of items was considered. For smaller tests, the number of items was 30; to incorporate conditions associated with longer tests we also considered 60 items. The 30 items were distributed in six subsets of five items each, and the 60 items were distributed in twelve subsets of five items each. The first items of each subsets are the hardest ones as they are the preceding halo items that impact the following items of each subset. Based on the sample size condition from previous literatures, 250, 500, and 1,000 sample sizes were considered in this study. ### Data analysis and evaluation criteria The simulated data was fit to the Rasch model. The model calibrations were completed in Mplus with WLSMV estimation. The use of Mplus WLSMV estimations provided WLSMV-based fit indices such as CFI, TLI, and RMSEA in the Rasch model as earlier defined in Eqs through. After fitting the Rasch model and estimating ${\hat{\theta}}_{p}$, the RMSE index was calculated for $\hat{\theta}$ as shown below. $$\hat{\theta}\mspace{360mu}\text{RMSE} = \sqrt{\frac{\sum_{p = 1}^{N}{({\hat{\theta}}_{p} - \theta_{True,p})}^{2}}{N - 1}},$$ where ${\hat{\theta}}_{p}$ is the estimated ability for examinee p, *θ*_*True*,*p* is the known true ability value for examinee p, and N is the total number of examinees. Analogous RMSE values for $\hat{b}$ was calculated for each iteration of the simulation. $$\hat{b}\mspace{360mu}\text{RMSE} = \sqrt{\frac{\sum_{i = 1}^{I}{({\hat{b}}_{i} - b_{True,i})}^{2}}{I - 1}},$$ where ${\hat{b}}_{i}$ is the estimated difficulty for item i, *b*_*True*,*i* is the true item difficulty parameter for item i, and I is the total number of items. # Results Results are presented in three sections. Results from TD and RD are presented for the relationship between WLSMV-based fit indices (i.e., CFI, TLI, RMSEA) and RMSE of parameter estimates (i.e., ability and item difficulty), and comparison between those indices and PCAR, respectively. ## Trait dependence If the fit indices were helpful in detecting LD that results in RMSE in parameter estimates, there should be a relationship between WLSMV-based fit indices and RMSE of parameter estimates. Correlation analyses between those fit indices and such RMSE in the Rasch model are applied and results are presented in. shows that both RMSE of ability and difficulty estimates are strongly correlated with all the fit indices in the expected direction (i.e., fit indices worsen as RMSE increases). These results imply that WLSMV-based fit indices are potentially useful to detect LD cause inaccurate estimates in ability and difficulty parameter in the Rasch model when trait dependence exists. However, it is also shown that the relationships of each fit indices to RMSE of ability and difficulty estimates are dependent on simulation factors in this study (i.e., magnitude of trait dependence, number of items and sample size). The following sections will present the details. ### Relationship between CFI/TLI and RMSE of ability and difficulty estimates with test factors Since the correlation between CFI and TLI is very strong (r = 0.99), the results of those two fit indices are presented in the same section. The relationships between CFI/TLI and $\hat{\theta}$ RMSE are impacted by the interaction of two simulation factors—magnitude of trait dependence and number of items. As seen in, the fit indices worsen as the magnitude of TD increases, which means those indices work in detecting the presence of local dependence due to TD. Specifically, both CFI and TLI are mainly impacted by the magnitude of trait dependence. Notice that local independence assumption is met when c = 0 that shows good fit results above 0.95 even sample size is 250 in both CFI and TLI. When the trait dependence exists (i.e., c = 1 or c = 2), almost all fits are less than 0.95 but only 0.8% cases shows over 0.95 when c = 1 and no cases over 0.95 when c = 2. Test length is another factor that influences $\hat{\theta}$ RMSE but does not relate to CFI/TLI. More test items produce less $\hat{\theta}$ RMSE, and yet the length of the test is not related to CFI/TLI within all sample size condition. This result indicates that $\hat{\theta}$ RMSE results are more sensitive to number of items than are CFI/TLI fit results. For the relationships between CFI/TLI and $\hat{b}$ RMSE, when the local independence assumption is met (i.e., c = 0), both test sample size and test length are not factors that impact the relationship between CFI/TLI and $\hat{b}$ RMSE, and in such instances an acceptable CFI/TLI is aligned with low $\hat{b}$ RMSE. When a moderate magnitude of trait dependence exists (i.e., c = 1), number of items impacts the range of the resultant fit indices, but still only 0.8% cases shows over 0.95 (e.g., indicate less than good fit according to CFI/TLI). When a large magnitude of trait dependence exists (i.e., c = 2), there is a positive linear relationship between CFI/TLI and $\hat{b}$ RMSE, and more test items produce less $\hat{b}$ RMSE while having no particular influence on CFI/TLI. ### Relationship between RMSEA and RMSE of ability and difficulty estimates with test factors The relationships between RMSEA and $\hat{\theta}$ RMSE are impacted by the interaction of two simulation factors—magnitude of trait dependence and number of items. Considering the property of RMSEA that the smaller RMSEA the better fit, RMSEA works (i.e., increases) when the magnitude of trait dependence is increased. However, RMSEA is impacted not only by magnitude of trait dependence but also by number of items, unlike CFI/TLI. This is true for $\hat{b}$ RMSE as well presented in. Specifically, there are no major differences between RMSEA values across the simulation iterations and conditions in which the trait local independence assumption is met (i.e., c = 0); but when the trait local independence assumption is violated (i.e., c = 1 or c = 2), test length is a factor that influences both $\hat{\theta}$ /$\hat{b}$ RMSE and RMSEA. More test items produce less $\hat{\theta}$ /$\hat{b}$ RMSE and RMSEA values within all sample size condition. ## Response dependence For the response dependence scenario, the correlation results show a less strong relationship between fit indices and RMSE of parameter estimates as compared to the trait dependence scenario. shows that both RMSE of ability and difficulty estimates are small to moderately correlated with the fit indices though all the fit indices are significant and show expected directions. ### Relationship between CFI/TLI and RMSE of ability and difficulty estimates with test factors In both CFI/TLI demonstrate that the relationships between CFI/TLI and $\hat{\theta}$ RMSE are mainly impacted by number of items. In all RD simulation conditions, CFI/TLI results produce perfect fit except one condition (i.e., d = 2 & sample size = 250), which indicates that CFI/TLI seem not to work in detecting the presence of local dependence due to RD. Regarding parameter RMSE estimates, test length impacts $\hat{\theta}$ RMSE but not on $\hat{b}$ RMSE. Neither number of items nor sample size has an impact on the relationship between CFI/TLI and $\hat{b}$ RMSE, as can be seen in. $\hat{b}$ RMSE is only affected by the magnitude of response dependence and $\hat{\theta}$ RMSE is only affected by number of items when response dependence violation occurs. The relationships between RMSEA and $\hat{\theta}$ RMSE are impacted by sample size and the interaction of that effect with RD magnitude. In, RMSEA seem not to work in detecting the presence of local dependence due to RD as with CFI/TLI cases. Regarding parameter RMSE estimates, test length only impacts $\hat{\theta}$ RMSE but not $\hat{b}$ RMSE. Neither number of items nor sample size has an impact on the relationship between RMSEA and $\hat{b}$ RMSE, as can be seen in. $\hat{b}$ RMSE is only affected by the magnitude of response dependence and $\hat{\theta}$ RMSE is only affected by number of items when response dependence violation occurs. ## Comparison of PCAR and WLSMV fit indices In each simulation condition, PCAR was conducted to produce the eigenvalue of the first component. As those values are used to determine whether the data are unidimensional in the Rasch model, it would be beneficial to compare those to fit indices from WLSMV estimation. It is noted that PCAR is a method for checking unidimensionality that corresponds to local dependence due to TD in this study. As noted in response dependence section above, the fit indices from RD simulation condition produce nearly perfect fit regardless of magnitude of RD. Therefore, the result of comparison is presented mainly about TD with regards to the first eigenvalue of PCAR. The relationship between CFI/TLI/RMSEA and first eigenvalue of PCAR (V1), and mean of those values are presented in Figs, respectively. Within each sample size conditions, when c increases (i.e., the magnitude of TD increases) both CFI/TLI/RMSEA and V1 worsens. That is, it is much expected for those values to be varied as TD occurs. However, considering the test length factor, those values produce different scenarios. Specifically, CFI/TLI/RMSEA is very consistent within each test length condition (e.g., mean of CFI in c = 1 & sample of 250 is equal to 0.88 in both 30-item and 60-item condition), whereas V1 depends on test length condition (e.g., mean of V1 in c = 1 & sample of 250 is equal to 2.12 in 30-item and 3.10 in 60-item, respectively). The results align with Chou and Wang’s simulation study as they pointed out that V1 depends on sample size and test length. # Discussion This study utilizes Marais & Andrich’s violation of local independence assumption in the Rasch model for simulating data, and then applies binary Rasch models to estimate item and person parameters. The simulation allowed for a comparison of conditions that vary in magnitude of trait and response dependence violations, as well as other factors including sample size and number of test items. The idea stemmed from Huggins-Manley and Han’s previous research about 2-PL IRT models. They utilized multidimensional IRT models to simulate LD that counterparts to TD in this study; therefore, it is worthwhile to compare the results from TD to the results from their previous study. Huggins-Manley and Han’s found that test factors (i.e., number of items & sample size) were strongly related to RMSE in ability and difficulty parameters, respectively. However, the relationship between the fit indices and RMSE in item discrimination parameters was not as clear as the relationship between fit indices and RMSE in ability and difficulty parameters. For discrimination parameter, the relationship was associated with sample size as well as correlation among traits that is hard to control a priori in practical settings. Due to the complexity of discrimination parameters, it was expected that the Rasch model produces a more explicit relationship between fit WLSMV-based fit indices and LD-induced inaccurate parameter estimates. This study supported this claim for TD, but not for RD. The results of TD indicate that the WLSMV-based fit indices and RMSE of parameter estimates are closely related to each other. When TD occurs, the fit indices might be able to detect the violation because fits worsen as the magnitude of TD increases. It is worth noting that there are many instances in which either the fit indices or the parameter estimate RMSE is influenced by a factor (sample size or number of items), but the factor does not influence both. For example, it was shown that when TD increases to a large magnitude, RMSE of person ability and item difficulty also increase, and the increase is larger for shorter tests as compared to longer tests. However, the CFI/TLI values were the same across the different test lengths. Therefore while CFI/TLI can flag TD forms of LD, they were not sensitive to the same factors that resulted in LD-induced RMSE in person and item parameter estimates. This is one of many indications that the fit indices are useful for flagging LD concerns, but that any cutoff value used for CFI/TLI would not take into account the fact that LD-induced RMSE in parameter estimates is dependent on sample size. This finding aligns with Huggins-Manley and Han’s (2017) findings. On the other hand, fit indices are not able to detect the RD violations because all the fit results are constant as the magnitude of RD increases. Within the RD situation, the response patterns were highly distorted in cases where the true ability and item difficulty were exactly the same. This is because the responses of the rest of the items were dependent only on the response of the first item rather than the true ability and item difficulty. This does not impact on the global fit indices despite the distorted pattern due to RD. The estimated parameter with the discrepancy from the true parameter might show good fit statistics under the occurrence of RD. Is it useful to use WLSMV-based global fit indices in detecting LD for Rasch applications? The answer would be "yes" especially in the situation of TD occurs. Based on the results from the comparison of PCAR, WLSMV-based global fit indices have meaningful advantages over the use of V1. Both WLSMV-based global fit indices and V1 are expected to be varied as the magnitude of TD increases, however, WLSMV-based global fit indices are consistent regardless of sample size and test length while V1 is not. This result implies that the decision based on a fixed cut point of V1 with regard to assessing unidimensionality assumption in the Rasch models would not be appropriate, indicated by Chou and Wang as well. Regarding WLSMV-based global fit indices, those would possibly be used for the detection of TD. As Xia and Yang pointed out, however, that it is not appropriate to set cutoff values for WLSMV-based fit indices from ordered categorical data. Though this study does not aim to address the question of what the new cutoff values from WLSMV-based fit indices should be employed, we would know those indices are related to the violation of TD, which do not depend on sample size and test length. For example, a researcher fits the Rasch model that results in good fit results (e.g., CFI\>0.98, TLI\>0.98, RMSEA\<0.01), which indicates—at least—no violation of TD. However, given the relationships between fit indices and misspecifications due to RD, it is not feasible to detect RD using WLSMV-based global fit indices since those fits are always good regardless of RD. In educational or psychological setting, multidimensionality might occur in an assessment that results in LD. This type of violation of local independence assumption is defined as TD. Based on the findings, WLSMV-based global fits can be useful in detecting TD for the educational or psychological assessment types. In contrast, it is often found that RD occurs in health or physical-related field of research for the application of Rasch rating scale. On a rating scale measuring physical functioning, RD can be found when a subset of items have some features in common. For instance, measuring walking ability questions may have RD due to similarities in response format or item content. Marais presented several examples of RD occurrence, which is caused by types of questions (e.g., asking overall level of status following several other status asking items or an item is made negative of the other preceding item). Marais also pointed out that RD can be found where raters make judgements using an instrument consisting of a number of criteria (i.e., “halo effect” followed by different raters if they judge similar ratings on the different criteria than they would be if rated independently). Therefore, it should be noted that using WLSMV-based global fit indices might not be useful under the circumstances of application of those kinds of rating scales especially to have RD. It turns out that WLSMV-based global fit indices are much more sensitive to TD than RD, even though both are a form of LD. Therefore, focusing on Research Question 2 and test factors (i.e., number of items and sample size) in TD situation, both CFI and TLI are not impacted by test factors but only by trait dependence violation. But RMSEA is impacted by number of items especially when TD occurs that more items produce better RMSEA fit though the same magnitude of trait dependence is violated. This is an indication that parameter estimates in Rasch models become more robust to TD violations of local independence as the number of items increases. Also, this is an indication that cutoff scores should not be applied to the WLSMV-based global fit indices when fitting Rasch models because the appropriate cutoffs would vary by number of items. Based on all of the findings, some summarizing statements are provided here. First, WLSMV fit indices of CFI, TLI, and RMSEA are not sensitive to violations of local independence due to RD, and their use is not recommended for detecting RD. Second, WLSMV fit indices of CFI, TLI, and RMSEA are quite sensitive to violations of local independence due to TD. The larger the TD effect, the higher the RMSEA value becomes. The larger the TD effect, the lower the CFI and TLI value becomes. However, the magnitude of RMSE of parameter estimates that is associated with particular values of WLSMV-based CFI, TLI and RMSEA is dependent on number of items, cutoff values for these fit indices are not recommended. Generally speaking, longer tests are associated with lower RMSE of parameter estimates while length of test does not impact the magnitude of WLSMV CFI, TLI, and RMSEA. Finally, several limitations are addressed in this study. First, this study only considers dichotomous responses, but future research should include polytomous responses. Second, further investigation or justification are needed about the magnitude of TD or RD (i.e., c or d) that are varied as integers like 0, 1, 2 in this study. The c term is associated with the correlation between estimated traits among subsets (see footnote 1 on page 5), while the d term is just integer values adding or subtracting from item locations in the formula. It might be interesting to consider different levels of RD (e.g., d = 1.5), as is also the case with TD. Third, regarding the test design, this study applied a subtest design with a fixed number of items, but it is unknown whether the results will be the same with a different number of subtests or no subtest design. Also, sixty items, as well as one thousand examinees, may be considered quite large for some applications of Rasch in practice, such that future research may want to consider a fewer number of items or examinees for practical purposes. Lastly, this study focuses on assessing LD, especially global fit of the Rasch model using WLSMV-based fit indices. In this controlled study, when these WLSMV-based fit indices indicated misfit, this was due to TD, as TD was the source of misfit that was being manipulated. However, misfitting WLSMV-based fit indices will not always be due to TD as there are many other aspects to consider (e.g. individual item-fit, person fit, differential item functioning, targeting, etc.). As with any global fit indicator, the WLSMV-based fit indices will indicate an issue within the item set, but it will not reveal the nature of the issue, and therefore the item set would still require further investigation. To conclude, final answer statements with regard to research questions addressed in this study. Research question 1 asked CFI, TLI, and RMSEA indices from WLSMV estimation of binary Rasch models were sensitive to LD-induced bias in item and person parameter estimates. For each parameter estimates, bias from LD had significant correlated relationships to each fit indices in expected directions. It should be noted that the relationship between fit indices and parameter estimates are much strong in TD simulated condition. Research question 2 asked if the answer to question 1 was dependent on various simulated factors. For person parameters, the relationship between the fit indices and bias in person parameter estimates was affected by number of items. For item parameters, the relationship between the fit indices and bias in item parameter estimates was affected by the sample size. Research question 3 asked if the WLSMV-based fit indices useful in the comparison of PCAR. The fit indices were robust to various studied factors as the magnitude of LD (especially for TD) increases while the result from PCAR was not. The results reveal the usefulness of the method proposed in this study over the use of PCAR. [^1]: The author has declared that no competing interests exist.

# Introduction Breast cancer is considered as the most common cancer in women, accounting for 29% of estimated new cancer cases and 14% of estimated cancer-related deaths. Chemotherapy is one of the cornerstone treatments in patients with breast cancer, which overall improves breast cancer outcome by 5–10% in patients with node negative disease. However, its use is increasingly affected by chemotherapy resistance and lack of effective predictors. Recently, emerging evidences have suggested that carcinoma-associated fibroblasts (CAFs) could contribute to chemotherapy resistances in breast cancer treatment. As the most frequent component of stroma cells in tumor microenvironment, CAFs have been assumed to play an important role in the carcinogenesis and development of breast cancer. Moreover, Farmer et al. reported that increased stromal gene expression predicts resistance to preoperative chemotherapy with 5-fluorouracil, epirubicin and cyclophosphamide (FEC), suggesting that stroma activation could be involved in chemotherapy resistance. Additionally, it was shown that CAFs mediated resistance to chemotherapy by releasing collagen I. Loeffler et al. have developed a vaccine that could target CAFs and allow reversing resistance to chemotherapy. Considering the interaction between CAFs and chemotherapy resistance, it would be reasonable that chemotherapy-induced damage could have an impact on CAFs and change the expression of some relevant factors, which could participate in chemotherapy resistance. In our study, we have cultured CAFs which were derived from surgically resected primary breast cancers and compared the gene expression profiling of CAFs before and after chemotherapy. The goal will be to find candidate markers from tumor microenvironment that associate with breast cancer chemotherapy resistance and discuss the possibility that these markers could be used as predictors for chemotherapy efficiency and feasible for targeted therapy. # Materials and Methods ## 1. Ethics Statement The study was approved by the Institutional Review Board and Human Ethics Committee of Xuanwu Hospital, Capital Medical University. Written informed consent for using the samples for research purposes was obtained from all patients prior to surgery. ## 2. Cell Culture of CAFs and Breast Cancer Cell Line Tissues for primary cultures of CAFs were collected from 10 breast cancer patients who underwent complete surgical resection of their tumors at Xuanwu Hospital, Capital Medical University. Only tissues in excess of those required for clinical diagnoses were harvested for this study. Harvested tissues were placed in DMEM supplemented with 10% FBS and antibiotics (Invitrogen Corporation) for immediate transportation on ice to the laboratory. Tissues were minced into small pieces, washed with phosphate-buffered saline (PBS) three times and digested for 20 h at 37°C in prepared reagent containing collagenase type I and Hyaluronidase (Roche Molecular Biochemicals). The cell suspension was filtrated with 100 mesh screen and centrifuged at 1000 rpm for 5 min, and then the pellet was resuspended in the fresh DMEM containing 10% FBS. Cell counting was performed with the Beckman Coulter Cell and Particle Counter Z1. The population doubling was estimated based on the increase in cell number counted at each passage time. Moreover, MDA-MB-231 cells were cultured in DMEM supplemented with 10% FBS as the breast cancer cell line, according to the normal procedure (Sigma-Aldrich). ## 3. Immunohistochemistry (IHC) Primary antibodies for immunostaining included multi-cytokeratin (CK), Vimentin, α-smooth muscle actin (α-SMA), CD34, and TE-7 (anti-fibroblast antibody) (Labvision). CAFs were seeded in chamber slides and fixed in cold acetone. After antigen retrieval and blocking of endogenous peroxidase in 3% hydrogen peroxide, the cells were incubated with primary antibodies at room temperature in a moist chamber for 60 min. Specific signals were visualized by incubation with peroxidase-coupled secondary antibody for 60 min, followed by incubation with 3,3/-diaminobenzidine (DAB) used as a chromogen to create brown staining. Counterstaining was performed with hematoxylin for 5 min, and the slides were coverslipped. ## 4. Flow Cytometry (FCM) CAFs were collected and prepared as a single cell suspension by mechanical blowing with PBS at the concentration of 1×10⁵/ml. The expression of CD34 and CD45 (MACS) was detected using FCM (FACSC alibar; BD). ## 5. Cell Adhesion Assay The harvested MDA-MB-231 cells were diluted with DMEM at the concentration of 40000/ml, while CAFs were diluted at the concentration of 20000/ml. MDA-MB-231 cells were added to 24-well plates, which were divided into two groups. For one group based on co-culture assay, the filters were placed in 24 well plates and CAFs were added to each upper chamber (Transwell; Corning). For the other group, the filters and CAFs were not used. Both groups were then treated with 20 ng/ml Taxotere (Sanofi). Afterwards, the cells were incubated at 37°C in a humidified atmosphere of 5% CO₂, until confluent. Matrigel (BD) was equilibrated with serum-free DMEM by proportion of 1∶3 before coating, and then 100 µl matrigel was added to each well in new 24-well plates. Two groups of MDA-MB-231 cells were harvested and transferred to 24-well plates coated with matrigel. After incubation for 1 h, MDA-MB-231 cells were washed with PBS, fixed in 4% formaldehyde and stained with 5% crystal violet. The number of MDA-MB-231 cells that adhered to the bottom of coated wells was counted and the morphology was recorded with an inverted microscope (Olympus IX70). The assay was done twice, each in triplicate. ## 6. Invasion Assay The harvested MDA-MB-231 cells were diluted with DMEM at the concentration of 40000/ml, while CAFs were diluted at the concentration of 20000/ml. Matrigel was equilibrated with serum-free DMEM by proportion of 1∶3 before coating, and 50 µl/cm² matrigel was added to each filter. The filters were placed in 24-well plates, which were divided into two groups. For one group based on co- culture assays, MDA-MB-231 cells were added to each upper chamber, and CAFs were added to the lower chamber. For the other group, the filters and CAFs were not used. Both groups were then treated with 20 ng/ml Taxotere and incubated at 37°C in a humidified 5% CO2 incubator for 72 h. At the end of the incubation period, the cells on the upper surface of the filters were removed with a cotton swab, and the filters were fixed in 4% formaldehyde and stained with 5% crystal violet. The number of cells that migrated to the lower side of the filter was counted and the morphology was recorded with an inverted microscope (Olympus IX70). The assay was done twice, each in triplicate. ## 7. Proliferation Assay (MTT) The harvested MDA-MB-231 cells were diluted with DMEM at the concentration of 40000/ml, while CAFs were diluted at the concentration of 20000/ml. MDA-MB-231 cells were added to 24-well plates, which were divided into two groups. For one group based on co-culture assay, the filters were placed in 24 well plates and CAFs were added to each upper chamber. For the other group, the filters and CAFs were not used. Both groups were treated with 0 ng/ml, 4 ng/ml, 10 ng/ml, 20 ng/ml, 40 ng/ml Taxotere, respectively. Then the cells were incubated at 37°C in a humidified atmosphere of 5% CO₂ for 72 h. Assays were initiated by adding 100 µl MTT (2 mg/ml) to each well and incubating the cells for an additional 4 h at 37°C. Afterwards, the medium was removed and 1 ml dimethylsulphoxide (DMSO) was added to each well. Finally, the supernatants were transferred to 96-well plates in triplicate, which were read at a wavelength of 550 nm with a Thermo Scientific Multiskan® Spectrum. ## 8. mRNA Expression Profiling Totally 6 pairs of CAFs were prepared for microarray analysis. Each pair of CAFs were obtained from the same patient and classified into two groups. One group was treated with 20 ng/ml Taxotere for 24 h (regarded as after chemotherapy) while the other group was not processed with Taxotere (regarded as before chemotherapy). Total RNA was extracted from all cultured CAFs using the RNeasy kit (Qiagen) according to the manufacturer’s protocol. Microarray studies were performed by Capital Medical University Microarray Centre, using Illumina humanHT-12 v4 expression BeadChip based on Illumina BeadStation500. The biotinylated cRNA preparation, hybridization, and scanning of microarrays were done according to the manufacturer’s protocols. Biological replicates have been used to reduce errors. Illumina Gene Expression Beaderchips have internal control features to monitor data quality. The GenomeStudio software (Ilumina) calculates and reports a detection p-value, which determines whether a transcript on the array is called detected. In our study, a detection p-value below the threshold of 0.01 indicated that a gene could be considered as expressed. Differentially expressed genes in CAFs before chemotherapy vs. after chemotherapy were also identified and analyzed with GenomeStudio. The output was filtered to include genes whose expression was altered at least two-fold. The dataset of the microarray analysis has been deposited in ArrayExpress, with the accession number E-MTAB-1614. ## 9. Real-Time PCR Real-Time PCR was performed to confirm differential gene expression in cultured CAFs before and after chemotherapy (treated with 20 ng/ml Taxotere for 24 h), using BIO-RAD IQ5 Real-Time PCR System. cDNA was synthesized using 1 µg total RNA, oligo (dT), and Superscript™ III Reverse transcriptase (Invitrogen). Synthesis was done according to the manufacturer’s instructions. All the primers were designed with Primer Express software (Applied Biosystems) for thecandidate genes. Predicted PCR product sequences were verified by using BLAST for recognition of target and non-target sequences. ## 10. Statistical Analysis Statistical analysis was performed using SPSS 13.0 software (SPSS Inc). Student’s t test was used to test for statistical significance. Data were presented as the mean±standard error. *p*\<0.05 was considered to indicate a statistically significant difference. # Results ## 1. Characterization of Primary-cultured CAFs By using a study protocol approved by the Institutional Research Ethics Board, CAFs were cultured from 10 surgically resected primary breast cancers which were histologically confirmed. The cultured cells were morphologically characterized with flat spindle shape, rich cytoplasm and flat ovoid nuclear. With immunostaining, the primary-cultured CAFs showed positive expression of α-SMA, vimentin, and TE-7, but negative expression of CK and CD34. The morphological and immunohistochemical pictures of CAFs were represented in. Additionally, FCM showed negative expression of CD34 and CD45 in CAFs. ## 2. CAFs Promotes the Function of Breast Cancer Cells after Chemotherapy We assumed that chemotherapy-induced damage interacted with tumor microenvironment and hence compared the function of MDA-MB-231 cells after chemotherapy (treated with Taxotere) co-cultured with CAFs and that without CAFs. By using cell adhesion assay, invasion assay, and proliferation assay (MTT), it was observed that after chemotherapy, MDA-MB-231 cells co-cultured with CAFs displayed increasing adhesion, invasiveness and proliferation abilities, compared with MDA-MB-231 cells without CAFs. The representative pictures of cell functional studies were shown in. ## 3. Comparison of Gene Expression Profiling in CAFs before and after Chemotherapy Totally 24314 expressed genes were detected in the microarray assay and 35 differentially expressed genes were identified (absolute fold change \>2) between CAFs after chemotherapy (treated with 20 ng/ml Taxotere) and before chemotherapy, including 17 up-regulated genes and 18 down-regulated genes. The differentially expressed genes were summarized in and with clustering analysis. Moreover, Gene Ontology (GO) analysis revealed that these genes were mainly involved in nucleotide binding, actin binding, cytoskeletal protein binding and structural molecule activity. The differentially expressed genes were also annotated in several Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways, including focal adhesion (hsa04510), Regulation of actin cytoskeleton (hsa04810), and MAPK signaling pathway (hsa04010). ## 4. Differential Expression of Candidate Genes in CAFs before and after Chemotherapy We have picked up 6 genes from 35 differential genes and confirmed the different gene expression in CAFs before and after chemotherapy (treated with 20 ng/ml Taxotere) via Real-Time PCR, using triplicate samples. The candidate genes included up-regulated genes CXCL2, MMP1, IL8, as well as down-regulated genes RARRES1, FGF1, and CXCR7. It was found that there was significant difference between the expression of 6 candidate genes in CAFs before chemotherapy and after chemotherapy (*p*\<0.05). The pictures were represented in. # Discussion As is known, cancer is a systemic disease encompassing multiple components of both tumor cells themselves and tumor microenvironment. The notion is now widely accepted that the development and progression of cancer highly depends on the interactions between tumor cells and tumor microenvironment. Recently, many investigations have pointed to stromal cells as the major regulator in tumor initiation, progression, and metastasis of breast cancers. However, the origin of CAFs has been a debate. Based on different theories, CAFs might arise from activated resident fibroblasts, bone-marrow-derived mesenchymal stem cells, cancer cells that undergo epithelial-mesenchymal transition (EMT), or other undetermined mechanisms. Correspondingly, CAFs were reported to exhibit different expression of multiple biomarkers such α-SMA, FSP-1, FAP, platelet- derived growth factor-α receptor (PDGFR-α), platelet-derived growth factor-β receptor (PDGFR-β), vimentin, CAV-1, PTEN, p21, or TP53 mutation. According to our study, CAFs showed positive expression of α-SMA, vimentin, TE-7 (anti- fibroblast antibody) and negative expression of CK, CD34, as well as CD45, suggesting that these primary-cultured CAFs were more likely to arise from activated resident fibroblasts, rather than epithelial cells, endothelial cells or bone marrow. We are aware that the observation based on morphological characteristics is just weak evidence for explaining the origin hypothesis, which needs further and more fundamental studies. Moreover, our data showed that CAFs could promote the adhesion, invasion and proliferation of breast cancer cells, which was consistent with peer researches. Breast cancer is known as the leading cause of death in women worldwide. Neoadjuvant chemotherapy (NAC) has been considered as an effective way which could improve the outcomes especially in patients with advanced and inflammatory diseases. However, the resistance of tumor cells to a broad range of chemotherapeutic drugs and lack of useful predictive markers of response to NAC continue to be problems. Though the precise nature and molecular mechanism of chemotherapy resistance is still unclear, many current studies have focused on identifying novel predictors of chemotherapy efficiency. CAFs merit attention, in consideration of frequent association with chemotherapy resistance. Sonnenberg et al. reported highly variable response to cytotoxic chemotherapy in CAFs from lung and breast, which could explain some levels of resistance in stroma-positive tumors where stroma would not be sensitive to chemotherapy. Furthermore, an oral DNA vaccine targeting fibroblasts activation protein (FAP) was developed to suppress primary tumor cell growth and metastasis of multidrug- resistant murine breast carcinoma and allowed reversing resistance to chemotherapy, by increasing intratumoral drug uptake. Based on our data, the comparison of gene expression profiling between CAFs before and after chemotherapy indicated the solid gene changes and provided candidate markers that might participate in chemotherapy resistance. Considering the correlation of breast cancer treatment and tumor microenvironment (especially CAFs), we suppose that the genes about membrane protein and secreted factors will be likely to associate with chemotherapy resistance. Then we have looked through the relevant literatures, to evaluate the research status and prospect of these genes, and eventually chosen CXCL2, MMP1, IL8, RARRES1, FGF1, and CXCR7 as candidate genes. The differential expression of these genes in CAFs before and after chemotherapy was confirmed by RT-RCR (p\<0.05), suggesting potential predictors of response to treatment. CXCL2 is one member of a family of structurally related chemokines, which are also called ELR-positive subgroup of CXC-chemokines. It was reported that CXCL2 could enhance survival of primary chronic lymphocytic leukemia cells in vitro and differential expression of CXCL2 in colon cancer had impact on metastatic disease and survival. In addition, CXCL2 was found to show significantly different expression in 5-FU responder and nonresponder breast cancer cell lines, suggesting its relationship with chemotherapy response. Matrix metalloproteinase (MMP) 1 has been focused on, in view of the association between its five polymorphisms and lung cancer risk. Moreover, the study carried out by Li et al. established the relationship between TP and MMPs in cancer cell invasion. Recently, the up-regulation expression of MMP1 was observed during human triple negative breast cancer cell line progression to lymph node metastasis in a xenografted model in nude mice, indicating potential targets involved in the control of metastasis. Interleukin-8 (IL-8) is a pro- inflammatory cytokine which was indicated to correlate with the growth and progression of tumors. Some interesting observations were made with regard to the prognostic role of baseline plasma IL8 protein levels in breast cancer patients treated with weekly docetaxel. Besides, Snoussi et al. pointed out that the polymorphisms in IL-8 and its receptor CXCR2 are associated with increased breast cancer risk and disease progress, implying that IL-8 and CXCR2 might contribute to breast cancer pathogenesis and aggressiveness. Lee et al. found that increased expression of IL-8 in the tumor microenvironment enhanced colon cancer growth and metastasis, which is very inspiring for our research. All the above markers including CXCL2, MMP1 and IL-8 were up-regulated based on our study, while FGF1, RARRES1 and CXCR7 as follows were down-regulated. As is known, breast cancer cells overexpress fibroblast growth factor receptors. Fibroblast growth factor 1 (FGF1) was reported to be especially suitable as chemotherapeutic drug carrier in light of its biological activity. Additionally, FGF1-gold nanoparticle conjugates targeting FGFR could efficiently decrease breast cancer cell viability, suggesting the possibility for targeted therapy. According to these data, it could be assumed that decreased expression of FGF1 might be involved in the chemotherapy resistance. Retinoic acid receptor responder 1 (RARRES1) is a retinoid regulated gene, which is accounted as a tumor suppress gene and lost in many cancer cells. It has been demonstrated that the down-regulation of RARRES1 is related to tumor growth of colorectal cancer and nasopharyngeal carcinoma. The investigation on the role of RARRES1 in the chemotherapy resistance is still rare, therefore its decreased expression in CAFs after chemotherapy caused our attention. CXCR7, as well as CXCR4, have been known as the receptors of chemokine CXCL12. Liberman et al. considered that CXCR7 would elicit anti-tumorigenic functions, and may act as a regulator of CXCR4/CXCL12-mediated signaling in neuroblastoma. Recently Hernandez et al. found that CXCR7 impaired invasion of breast cancer, in contrast to CXCR4. We propose that CXCR7 would contribute to the chemotherapy resistance in the condition of treatment-induced damage to the tumor microenvironment. Moreover, in the abovementioned candidate genes, CXCL2, MMP1 and IL8 are recognized as secretory-type genes, which possibly could be tested in serum in the form of genes or proteins. Its relationship with prognosis will be studied. Overall, in this study we have primarily cultured CAFs, compared its gene expression profiling before and after chemotherapy, and picked up 6 candidate genes which are possibly associated with chemotherapy resistance in breast cancer. We hope that our study might supply the potential predictors for chemotherapy efficiency and possible targets for treatment, which could provide the patient with optimal therapeutic management and better prognosis. For further study, the molecular mechanism of these candidate markers will continue to be researched to elucidate their relationship with chemotherapy resistance. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: GR HK. Performed the experiments: GR. Analyzed the data: GR TH HS. Contributed reagents/materials/analysis tools: HK YW. Wrote the paper: GR.

# Introduction Age-related macular degeneration (AMD) is the leading cause of blindness in elderly residents of industrialized countries. The disease is classifiable into “wet” and “dry” forms based on distinct clinical features. In wet AMD, rapid visual loss is caused by macular choroidal neovascularization. Dry AMD entails a slower degeneration of the retinal pigment epithelium, choroid, and surrounding extracellular matrix in the macular area. Before diverging into these two distinct forms of the disease, both conditions are preceded by the accumulation of extracellular aggregates, termed drusen, between the retinal pigment epithelium and Bruch’s membrane. Unlike wet AMD, for which effective treatments exist, therapeutic options are lacking for dry AMD, partly because of a lack of appropriate animal models that recapitulate the complex clinical features of the condition. To date, mice have commonly been used to study isolated aspects of AMD because of their availability and suitability for genome manipulation. These efforts have greatly enhanced our understanding of the pathology of AMD, but insufficiently. Unfortunately, a fundamental limitation hampers the use of mice as animal models of macular degeneration: they lack a macula, a unique anatomic feature present only in humans and a subset of monkeys. Therefore, when studying macular disease in animals, monkeys with a macula are preferred over other species. Previous studies have reported that macular drusen are prevalent in various monkeys worldwide, including *Macaca mulatta* and *Macaca fascicularis*, which are frequently used in biomedical research. In *M*. *fascicularis*, a family with early onset drusen inherited in an autosomal-dominant manner has been reported, along with monkeys with drusen possibly inherited in a non-Mendelian manner within the same colony. During the past decade, several large-scale genetic studies targeting patients with AMD have identified disease-associated *genes* and single-nucleotide variants. Interestingly, many AMD-associated *genes* were found to encode members of complement pathways, including *CFH*, *C2/CFB*, *C3*, *CFI*, and *C9*. Complement pathways are ubiquitous inflammatory systems activated against pathogens and inflammatory stimuli in multiple organs throughout the body. Members of these pathways are often secreted into the bloodstream as soluble factors. Consequently, the alteration of complement pathways can affect systemic inflammatory biomarkers in the blood. For example, increased white blood cell count and C-reactive protein level are associated with AMD, which is consistent with the genetic findings. In monkeys, local ocular involvement of complement pathways has been detected by immunohistochemistry and proteome analysis using ocular samples from *M*. *fascicularis* with drusen. The same study group also conducted proteome analysis of plasma samples from *M*. *fascicularis* with and without drusen. They identified ApoE as a potential biomarker of the disease. However, each study examined only a few monkeys. Another study of *M*. *mulatta* implicated genetic risk *genes* shared between monkeys with drusen and human patients with AMD. In this study, we compared the results of standard blood tests in a large colony of *M*. *fascicularis* with and without drusen to identify systemic biomarkers of drusen and ascertain whether these markers overlap with those reported in humans. # Materials and Methods ## Animals We examined 1,174 *M*. *fascicularis* reared at Tsukuba Primate Research Center at the National Institutes of Biomedical Innovation, Health and Nutrition (NIBIOHN), Tsukuba, Japan. The monkeys ranged in age from 1–38 years. They were housed in an indoor environment where artificial lighting was used for 12 h each day. The animals were fed 70 g of commercial food (CMK-2; CLEA Japan, Inc., Tokyo, Japan) and 100 g of apples daily. Tap water was supplied *ad libitum*. Every morning their health status (e.g., viability, appetite, coat appearance) was monitored. The monkeys were provided with toys, branches, and music as a part of efforts to improve their enrichment. The maintenance of animals was conducted according to the rules for animal care of the Tsukuba Primate Research Center for the care and use of, and biohazard countermeasures related to, laboratory animals. All animal experiments were conducted in accordance with the guidelines for animal experiments of the NIBIOHN and with the *Guide for the Care and Use of Laboratory Animals* of the National Institutes of Health (Bethesda, MD, USA). The research protocol was approved by the ethics committee at the Tsukuba Primate Research Center. ## Fundus photography and blood test Approximately 20 min before examining the ocular fundi, a mixture of tropicamide and phenylephrine hydrochloride was instilled into both eyes of each animal to dilate the pupils. Then, the monkeys were anesthetized with an intramuscular injection of ketamine (10.0 mg/kg). Fundus photographs were taken with an ophthalmoscope camera (Kowa RC-2; Kowa Co. Ltd., Tokyo, Japan). A monkey was categorized as having drusen if one or more round yellowish spots with the characteristic appearance of drusen, regardless of their size or location, were identified in either eye or both eyes in a fundus photo of the posterior pole centered on the macula that encompassed \~23° vertically and \~19° horizontally. The quality of photos of nine monkeys was too poor to determine the presence or absence of drusen; thus, these monkeys were excluded from further analysis. All images were assessed by an experienced ophthalmologist and a veterinarian specializing in ophthalmology to ascertain the presence or absence of drusen. In most cases, the two assessors agreed on the interpretation of the photos (Cohen's kappa index value: 0.962). However, when there were disagreements, the fundus photos were reviewed together and decisions were made after a discussion. The body weight of each animal was measured. A blood sample was obtained from the femoral vein. A proportion of the blood was subjected to hematologic analysis. Serum was isolated from the remainder to perform biochemical analysis. The blood testing was performed as a part of a routine health-monitoring program unrelated to the current project by technical staff at the Tsukuba Primate Research Center under the direction of a veterinarian. ## Statistical analysis First, we evaluated the relationships among the parameters by calculating the Pearson correlation coefficient. For the pair of parameters that exhibited a high correlation coefficient (*r* \> 0.600), one was excluded from further analysis. Then, continuous variables were divided into quartiles with the first quartile as the reference group, to which two statistical analyses were applied before selecting variables to be analyzed by logistic regression analysis. Odds ratios (ORs) were calculated with and without adjustment for age and sex (either age or sex for subgroup analysis). To assess the ORs for the second, third, and fourth quartiles with the first quartile as reference for each variable, logistic regression analysis was applied. Then, the Cochran–Armitage test was used to objectively assess the trend for drusen frequency using the quartile data. Variables that showed an increasing or decreasing trend (P \< 0.050) with higher quartiles were further selected for logistic regression analysis to independently assess the effect of each of the selected variables. A multiple logistic regression model refined by stepwise procedures using the backward entry method was applied to estimate the risks of potential predictors, including age, sex, and selected blood parameters, for the development of macular drusen. The Akaike information criterion (AIC) was used to determine the variables to be added to or deleted from the model. Only the set of variables that minimized the AIC value was retained in the final model. We repeated the variable selection using the P-value (\< 0.150) to determine the best subset of variables for the model; the models also yielded similar results as that generated using the AIC. In addition, the collinearity of parameters retained in the final model was assessed using the variance inflation factor (VIF). All VIF values were less than 10.0, which meant that there was no collinearity in the model. Sub-population analysis was also carried out. The monkeys were divided into two groups, i.e., males (*n* = 232) and females (*n* = 713; subgroup analysis 1), and younger (1–6 years; *n* = 477) and older (older than 6 years; *n* = 468; subgroup analysis 2), and an identical analysis workflow as that applied to the monkeys as a single group was employed. R software (version 3.2.2; R Foundation for Statistical Computing, Vienna, Austria) was used for all calculations. # Results Fundus photographs were taken in 1,174 *M*. *fascicularis* between 2011 and 2013. The blood tests included a complete blood cell count and a standard biochemical analysis. Of the 1,174 monkeys, only those with a complete dataset, including biologic data and all basic blood data, and with discernable fundus photos were analyzed further. This analysis included 945 monkeys, comprising 317 with and 628 without drusen. The biologic distribution of each blood test parameter is presented in. Comparison of the biologic and basic blood test data between monkeys with and without drusen are presented in. Applying a Mann–Whitney *U*-test to each parameter revealed several statistically significantly different parameters between the two groups. To test if any of the variables were related, we calculated the Pearson correlation coefficient (*r*) between all the variables. Age and weight and red blood cell count and hemoglobin exhibited correlations (*r* \> 0.600). At this point, the data for weight and hemoglobin were excluded from the analysis. Then, quartile data for each variable were analyzed to detect significant associations with the frequency of drusen. When logistic regression analysis was applied to the quartile data, ORs were consistently increased for the three higher quartiles for age and white blood cell count. Then we assessed if the quartile data showed a significantly increasing or decreasing trend. We selected seven variables with an increasing or decreasing trend at P \< 0.050 and applied multivariate logistic regression analysis. As a result, only two factors remained. These were age (OR: 1.10 per year, 95% confidence interval \[CI\]: 1.03–1.18; P = 0.004) and white blood cell count (OR: 1.01 per 1 × 10³/μl, 95% CI: 1.00–1.02; P = 0.179). A sub-population analysis was conducted to further investigate the data. First, sex-related differences were assessed, because reports suggest that AMD- associated factors may differ between human males and females. The monkeys were divided into males (*n* = 232) and females (*n* = 713). Quartile analysis followed by a trend test for drusen frequency (Cochran–Armitage test;) was applied to each sex group following the protocol used to analyze all 945 monkeys. The quartile analysis showed an age-dependent increase in drusen development to be prominent in females but not in males. We then selected variables with an increasing or decreasing trend at P \< 0.050 and applied multivariate logistic regression analysis. When stepwise multiple logistic regression analysis was applied, three factors and one factor remained in the final model for males and females, respectively. In either sex, only age was statistically significant, with an OR of 1.07 per year (95% CI 1.02–1.13) for males and 1.1 per year (95% CI 1.07–1.13) for females. For males, red blood cell count and blood urea nitrogen remained in the final model, but they did not reach statistical significance. The VIF was ascertained to check for multicollinearity between the variables. All VIF values were less than 10.0, which meant that there was no collinearity in the model. Next, the monkeys were divided into two groups based on age, i.e., the younger (1–6 years; *n* = 477) and older (7 years or older; *n* = 468) groups. This was to test if the etiology of drusen present in younger animals differed from those present in older animals, because drusen are almost never observed in young humans. A quartile analysis followed by a trend test for drusen frequency (Cochran–Armitage test;) was applied to each of the two age groups. The quartile analysis showed the effect of the increase in white blood cell count on drusen frequency to be more prominent in young monkeys compared with older monkeys. We then selected variables with an increasing or decreasing trend at P \< 0.050 and applied multivariate logistic regression analysis. When stepwise multiple logistic regression analysis was applied, three factors and one factor remained in the final model for younger and older monkeys, respectively. In either group, age remained in the final formula. In both groups, age was statistically significant, with an OR of 1.3 (95% CI: 1.10–1.53) for younger monkeys and 1.08 (95% CI: 1.04–1.12) for older monkeys. For the younger monkeys, white blood cell count remained in the formula, with an OR of 1.01 (95% CI: 1.00–1.01), and showed statistical significance. In older monkeys, Albumin remained in the final model, but it did not reach statistical significance. The VIF was ascertained to check for multicollinearity between the variables. All VIF values were less than 10.0, which meant that there was no collinearity in the model. # Discussion In this study, we investigated non-ocular factors associated with drusen by analyzing biologic data and blood test results from 945 *M*. *fascicularis*. To the best of our knowledge, this is the largest study to screen the fundus for drusen and analyze blood samples to assess systemic involvement in monkeys with ocular diseases. As with humans, the development of drusen was associated most strongly with increasing age in these monkeys, consistent with previous reports. This was the case when all monkeys were assessed together or divided into male and female or young and old groups. Interestingly, age seemed to have the largest effect on young monkeys, with an OR (1.30) exceeding those of the other subgroups. In addition to age, an increased white blood cell count was also associated with drusen when all 945 monkeys were assessed together and when younger monkeys (6 years of age or younger) were selected for analysis. An increased white blood cell count is linked to an elevated incidence of early AMD in humans. It is particularly interesting that the association was evident in this study in younger monkeys and not in older monkeys, which implies that an age-related difference in the pathogenesis of drusen may exist at least in this species. In *M*. *fascicularis*, monkeys with early onset drusen inherited in an autosomal-dominant manner, and also monkeys with drusen inherited in a non- Mendelian manner, have been reported. Unfortunately, information related to their pedigrees is unavailable. Nevertheless, because drusen are quite common in monkeys of various species, and were found in approximately 35% of the monkeys examined in this study, with an age-dependent increase in its prevalence even in older monkeys, it is likely that drusen are not inherited as a discrete early onset autosomal-dominant disease in most monkeys. These findings, which are consistent with the idea that systemic inflammation also underlies the formation of drusen in monkeys, are intriguing. The OR for white blood cell count in the logistic regression model appeared low (1.01 per 1 × 10³/μl). However, the white blood cell count was highly variable between the monkeys (1.2–11.9 × 10³/μl, a range for mean ± two standard deviations). This implies the differential importance of the count as a potential risk factor for drusen in these monkeys. The reason for this increase in white blood cell count is uncertain. It is possible that it reflects chronic infection by pathogens, such as *Chlamydia pneumonia*, that might affect the clinical course of AMD. The main limitation of this study is the missing link between drusen formation and macular degeneration in monkeys. The phenotypic features of drusen in monkeys and humans are quite different. It seems that drusen can appear from a young age and are usually small, punctate, and concentrated around the fovea (see for typical macular drusen appearance) in monkeys, which differs from the typical profiles of drusen found in aged humans. In humans, an association between larger, “soft” drusen and the development of advanced AMD is established, whereas the pathogenicity of small, punctate, “hard” drusen is considered less significant. However, the molecular components of soft and hard drusen are not markedly different. Furthermore, evidence suggests that small, punctate, hard drusen that resemble those found in monkeys may also precede the development of dry AMD in humans. Therefore, taking into account all the discrepancies between drusen in humans and monkeys, there is little doubt that these monkeys are one of the best animal models of human macular drusen available. As the monkeys studied are from an inbred colony, some monkeys may be related to each other. This is another limitation of the study, because this indicates that the data from these monkeys are not likely to be independent. In conclusion, this study analyzed basic blood tests, including complete blood count and blood chemistry, in a large colony of *M*. *fascicularis* with and without drusen. Our results show associations of age and white blood cell count with drusen development. Systemic inflammation may underlie drusen formation in monkeys as it does in humans, which further highlights the relevance of monkeys with drusen as potential models of early AMD. # Supporting Information We thank Dr. Airi Takagi at the Clinical Research, Innovation and Education Center, Tohoku University Hospital for the constructive discussion and kindly providing advice for our statistical analysis. The manuscript was edited by a professional English editing service (Enago, Tokyo, Japan; [www.enago.jp](http://www.enago.jp/)). [^1]: The Department of Advanced Ophthalmic Medicine and Department of Retinal Disease Control are endowment departments within Tohoku University Graduate School of Medicine, supported with an unrestricted grant from SENJU Pharmaceutical Co., Ltd and Nidek Co., Ltd, respectively. This does not alter our adherence to PLOS ONE policies on sharing data and materials. [^2]: **Conceptualization:** KMN. **Data curation:** TF FO NS MT MS. **Formal analysis:** KMN YY YF KF YT RK. **Funding acquisition:** KMN TN. **Investigation:** KMN YF KF YT. **Methodology:** KMN. **Project administration:** KMN TN. **Resources:** TF FO NS MT MS. **Supervision:** KMN TN. **Writing – original draft:** KMN. **Writing – review & editing:** KMN TN.

# 1. Introduction Low back pain (LBP) is a very common condition, with a one-year period prevalence of approximately 50% in people from the Nordic populations. It is also the leading cause of years lived with disability worldwide. Identifying the etiology of LBP is challenging and consequently patients are often labelled as having non-specific LBP. In order to better understand non-specific LBP, groups of researchers have begun to test the hypothesis that LBP is not one condition, but more likely the predominant symptom of a number of, as yet, unidentified subgroups. In the search for nociceptive contributors to pain, magnetic resonance imaging (MRI) is increasingly used. Two recent systematic reviews have identified a number of lumbar MRI findings that are associated with LBP, and Modic Changes (MCs), i.e. endplate related signal changes in the vertebrae, have been proposed to constitute a diagnostic subgroup amongst patients with non- specific LBP. de Roos et al. were the first to describe endplate-related signal changes in the lumbar spine in 1987 and these were further examined by Modic et al., who classified them into three types: Modic changes type 1 (MCs1), Modic changes type 2 (MCs2), and Modic changes type 3 (MCs3), based on their appearance on T1-weighted and T2-weighted MRI. MCs1, seen as high signal on T2-weighted and low signal on T1-weighted magnetic resonance images, are considered to be the earliest stage of MCs, but also the most biologically active, and hypothesized to represent an inflammatory reaction in the bone marrow (edema type). MCs2, seen as high signal on T1 images and isointense or slightly hyperintense signal on T2 images, represent a fat infiltration of the bone marrow. MCs3, seen as low signal on both T1 and T2 images, represent a sclerotic change of the bone marrow. Histological samples of MCs1 and MCs2 have shown fissuring of the vertebral endplate and trabecular bone along with vascularized fibrous tissue (MCs1) and yellow fat (MCs2). The reported type and prevalence of MCs may depend on the field strength of the MRI scanner. In one study, more MCs overall and more MCs2 but fewer MCs1 were diagnosed in a 1.5 Tesla versus a 0.3 Tesla scanner. The appearance of MCs also depends on the MRI sequences used; e.g. on T2-weighted fat-suppression sequences, fat in MCs2 –but not edema in MCs1 –appears with a suppressed and lower signal. The association between MCs and non-specific LBP has been investigated in three systematic reviews: Jensen et al. 2008, Zhang et al. 2008 and Brinjikji et al. 2015. All three reviews found an association between MCs and LBP, but Brinjikji et al. only found it for MCs1. Although the title of the study by Zhang et al. indicates that this is a systematic review, it has the form of a narrative review. The study by Brinjiki et al. had strict inclusion criteria (studies in English only, with both symptomatic and asymptomatic participants between 15 and 50 years of age). Since 2008, when the last comprehensive review was published, many new studies have emerged and there is a need for an updated review. Furthermore, none of the previous reviews addressed the association between MCs and activity limitation, and none of them evaluated potential factors that could modify the associations (e.g. age, sex, MRI parameters, other degenerative findings such as disc degeneration, herniations and facet joint arthrosis). Albert et al. in 2013 found that MCs may have a bacterial etiology and can be treated with antibiotics. These findings created headlines worldwide and much debate among clinicians and researchers because of a potential risk of bias, conflicting results in studies investigating a bacterial etiology, and the prospect of treating a large group of LBP patients with long-term high-dose antibiotics. The heterogeneous results for associations between specific types of MCs and LBP in previous systematic reviews, our lack of knowledge about the association between MCs and activity limitation, controversies about MCs guiding antibiotic treatment and emerging new studies call for a comprehensive and updated systematic literature review to improve our understanding of the clinical relevance of MCs. Our objectives were to investigate 1) if the presence of MCs (including types and size) in the lumbar spine region is associated with non-specific LBP and/or activity limitation, and 2) if such associations are modified by other factors. # 2. Methods ## 2.1 Design A systematic, critical literature review with meta-analysis was performed. A research protocol was developed in advance and registered in the PROSPERO: International prospective register of systematic reviews (<http://www.crd.york.ac.uk/PROSPERO/display_record.asp?ID=CRD42015017350>). ## 2.2 Criteria for considering studies for this review ### 2.2.1 Types of studies Prospective or retrospective cross-sectional cohort studies and case-control studies were included. We chose to exclude studies with fewer than 26 individuals. This cut-off was chosen to minimize the risk of having cells in the 2x2 tables that included zero. ### 2.2.2 Participants People of all ages from general, working and clinical study populations were included. The following exclusion criteria were used: - Studies including participants diagnosed with specific LBP such as: spondylitis, discitis or spondylodiscitis, spondyloarthropathies (e.g. ankylosing spondylitis), fracture (including isthmic spondylolisthesis), spinal cord infarction, malignancy, hematological conditions and juvenile/idiopathic scoliosis. - Studies including participants treated with radiotherapy in the lumbar region. - Studies including participants treated with spinal surgery (although pre- intervention data were eligible for inclusion). ### 2.2.3 MRI findings and definitions (index test) We defined MCs as signal changes seen on MRI in the vertebral bone marrow, extending from the endplate. This definition included signal changes regardless of etiology and excluded signal changes only present in the bone marrow away from the endplate. We chose to include only studies evaluating MCs in the whole lumbar spine, disc levels L1-L2 to L5-S1 (except for studies using provocative discography), based on a previous report that the association between MCs and LBP is dependent upon disc level. We included studies using provocative discography, since this procedure intends to localize LBP to a specific disc level, allowing study of the association between MCs and LBP at that level (rather than between and LBP and MCs). ### 2.2.4 Target condition Non-specific LBP of all durations was included. ### 2.2.5 Outcomes (reference standards) The following outcomes were measured: 1. Presence and/or intensity of LBP measured by experimental tests (e.g. provocative discography or algometry) or patient-reported outcomes. 2. Presence and/or level of activity limitation, measured by the Oswestry Disability Index (ODI), the Roland Morris Disability Questionnaire (RMDQ) or similar tools. ## 2.3 Search methods for identification of studies ### 2.3.1 Electronic searches A systematic search of the literature was performed using a search strategy developed in collaboration with a research librarian. The three terms “lumbar spine”, “MRI” and “Modic changes” and their relevant synonyms were used as search terms, either as free text or as Medical Subject Headings. The MEDLINE, CINAHL and EMBASE databases were searched for relevant studies from first record to June 15^th 2016. No restrictions were used. The full electronic search strategy can be found in the. ### 2.3.2 Searching other resources Reference lists of all included studies were examined and all authors were asked to review the list of included studies for omissions. ## 2.4 Data collection and analysis ### 2.4.1 Selection of studies Two reviewers (CH, TSJ) independently screened the titles and abstracts to exclude clearly irrelevant papers. For each potentially eligible study, the full article was retrieved and independently assessed for inclusion (CH, TSJ). Any discrepancies were resolved by consensus. Where multiple publications used data from the same study sample, we chose the article with the most complete data related to the associations between MCs and LBP and/or activity limitation. In cases where association data were not presented in a format that we could use for data extraction, we contacted the authors to request additional data, as recommended by the Cochrane Handbook. We assessed the eligibility of non-English papers using Google Translate and, when this was impossible due to incomplete Optical Character Recognition obtained from scans of paper copies, with the help of a native speaker of the language in question. ### 2.4.2 Data extraction and management Data extraction and risk of bias assessment were completed by independent reviewers (CH, PK, AE, JStS, CLY, JK, JN, JSoS, KS, TSJ), allocated in pairs (except for non-English papers, where a single assessor was used), using spreadsheets (Tables –). All reviewers were pre-trained through pilot-testing of the process. Inconsistencies were resolved by consensus or, if needed, by including a third reviewer (CH or TSJ). Data regarding sample source, number of subjects, age, MRI parameters, observers, MCs (including types and size), clinical outcomes, and strength of associations between LBP and/or activity limitation and MCs were extracted from the papers. Data regarding possible modifiers or confounders of the associations between MCs and clinical outcomes were extracted and classified according to how the covariates were analysed: a) by matching on the covariate(s), b) by restricting participant selection so that all groups had the same covariate value, or c) by adjustment for covariates in the statistical analysis. For the purpose of this review, only analyses investigating single covariates were included. The reason for this was that adjustment by groups of covariates (e.g. age, sex, etc.) might change the estimate of the associations, but would not provide information as to which of the group covariates or combinations of covariates were modifying the associations. Because of the exploratory nature of this part of the review, we chose not to make a list of pre-defined candidate variables. ### 2.4.3 Risk of bias assessment We based our risk of bias assessment on the QUADAS 2 tool. This tool is used to evaluate the following four key domains: *study sample*, *index test*, *reference standard(s)*, *timing and data analysis* based on signaling questions and questions regarding applicability. For each domain, studies were classified as having ‘low risk of bias’, ‘high risk of bias’ or ‘unclear’ based on a number of signaling questions. Studies were classified as having an ‘overall low risk of bias’ if all four domains were scored as ‘low risk of bias’. We added additional signaling questions pertaining to items we found particularly important for the subject of this review. After pilot-testing, we modified some of the risk of bias questions and response options, making them more intuitive to answer. The questions regarding applicability were not used in this study. The result of our risk of bias assessment was not used as an inclusion criterion. ### 2.4.4 Statistical analysis and data synthesis Raw data for 2x2 tables or group differences were extracted where possible to calculate odds ratios (OR) with 95% confidence intervals (CI) for dichotomous outcomes or to perform t-tests for continuous outcomes. In cases where results were presented in the form of ORs or mean differences, without raw data, we present them as stated in the article, using data from the crude analysis, i.e. unadjusted. Data supplied by authors on request were treated in the same manner. ORs and 95% CIs were calculated for 2x2 tables. For tables containing 0 in one of the cells, we added 0.5 to all cells. Differences in means between groups were analysed using a t-test. Statistical analyses were performed using STATA (version 12.1, StataCorp, College Station, Texas, USA). Associations between subtypes and sizes of MCs and outcomes were determined with reference to participants with no MCs. A statistically significant association was defined as CIs not including 1.0 for dichotomous outcomes and a p-value below 0.05 for continuous outcomes. Studies were classified as having an association (‘positive study’ or ‘negative study’) if the association reported for one or more outcomes was statistically significant. If a single study reported both a statistically significant positive association and a statistically significant negative association, this would be classified as a contradictory association (‘contradictory study’). Results were pooled where it was deemed possible and appropriate (e.g. homogeneous in terms of study sample or outcome), and associations reported as ORs and 95% CIs. Due to the heterogeneity in terms of the prevalence estimates of MCs and study sampling, a random effect model was used. I² statistics were used to quantify inconsistency across studies. *Comprehensive Meta-Analysis* (version 3, Biostat, Englewood, USA) was used for meta-analysis. A single covariate was recorded to modify the associations between MCs and LBP/activity limitation if a) the estimates for unadjusted and adjusted/stratified analyses differed or b) an interaction term with MCs and the possible modifier was statistically significant, p\<0.05. Pre-determined sensitivity analyses were performed for publication bias and overall risk of bias. The classification of associations (positive, negative or contradictory) was then tested against mean age, year of publication, number of participants, and overall risk of bias using Fischer’s exact test. # 3. Results ## 3.1 Selection of studies In total, 5210 citations were identified, yielding 3834 records after removal of duplicates. After reviewing titles and abstracts, 3377 records were excluded, resulting in a total of 457 papers eligible for full text assessment. Another 420 studies were excluded in the full text assessment, resulting in 37 potential candidates. Two additional studies were found through manual search and by a person from the research team respectively, resulting in 39 potentially acceptable studies. We requested additional data needed for analysis from the authors \[, –\] for 11 of the 39 studies, but received these data from only three. We thus ended up including 31 studies. Almost all inconsistencies of data extraction and Risk of Bias assessment were solved by consensus among the pairs of assessors. A third reviewer was involved in reaching consensus for five data points. ## 3.2 Study characteristics Sixteen studies reported on clinical populations. These consisted of patients with or without leg pain, most of them being classified as having chronic LBP or referred for back surgery of some kind. A notable exception was the study by Nakamae et al., where the patients had lumbar degenerative scoliosis, and mixed LBP and leg pain were used as an exclusion criterion. Ten studies reported data from non-clinical populations, i.e. population-based cohorts, volunteers or working populations and five studies reported data from case-control studies. In relation to outcomes, 21 studies reported on self-reported LBP (13 studies on the presence of LBP and eight studies on the intensity of LBP). Nine studies reported on pain at provocative discography (no study reported on other experimental tests, e.g. algometry). Seven studies reported on activity limitation. The number of participants ranged from 36 to 2449, with a median of 200, with the proportion of women ranging from 0% to 96%. Mean age of the study samples ranged from 13 to 76. The majority of studies used MRI scanners with field strengths of 1.0–1.5 T. Four studies used a field strength below 1.0 T. Six studies used several scanners for their assessments. Five studies used only T2-weighted MRI sequences, making differentiation between different types of MCs impossible, whereas the remaining studies used both T1-weighted and T2-weighted sequences. One study used T2-weighted fat-suppression or post-contrast T1-weighted sequences. Three studies did not report on any MRI parameters at all. ### 3.2.1 Prevalence of MCs The prevalence of MCs, with all types taken into account, ranged from 3% to 80% in clinical samples (including cases from case-control studies), on a per individual basis, not including studies using provocative discography. For clinical studies using provocative discography, the range was 1% to 38%, reported per level assessed by discography. In non-clinical samples (including controls from case-control studies) the prevalence ranged from 0.5% to 88% on a per individual basis. Kovacs et al. reported the highest prevalence of MCs among non-LBP participants (88%), a number higher than even the highest prevalence (62%) found in the clinical studies. ### 3.2.2 Prevalence of LBP The prevalence estimates of LBP in non-clinical and case-control studies were measured with a variety of criteria, eg. ‘LPB last month’ and ‘lifetime LBP’, and thus varied considerably. ## 3.3 Risk of bias assessment The results from the risk of bias assessment can be seen in. Only one study was classified as having overall low risk of bias, i.e. with low risk of bias in all four key domains. Five studies had three domains with low risk of bias, three studies had two domains with low risk of bias, fourteen studies had one domain with low risk of bias and eight studies were classified as having no domains with low risk of bias. ## 3.4 Association between MCs and LBP ### 3.4.1 Association with LBP Across all included papers, 30 of 31 studies reported on the association between MCs (regardless of type) and LBP. Of these 30, 15 found statistically significant positive associations with ORs ranging from 1.53 (95% CI 1.02–2.29) to 83.10 (95% CI 4.85–1424.05), while only one found a statistically significant negative association with a mean difference between patients with and without MCs2 of -3.2 (-5.39 –-1.01) on a ‘back pain score’ ranging from 0 to 30. The remaining 14 studies reported statistically non-significant findings, of which eight reported negative (but non-significant) estimates on at least one of their outcome measures. No studies reported contradictory statistically significant associations. Across all included articles, 13 studies reported on the association between MCs1 and LBP. Six of these found statistically significant positive associations. Five reported ORs ranging from 2.06 (95% CI 1.12–3.79) to 51.67 (95% CI 11.43–233.51) for dichotomous outcomes whereas one study reported statistically significant positive associations using continuous outcomes. The remaining seven studies reported statistically non-significant findings for associations regarding MC1. Ten studies reported on the association between MCs2 and LBP. Four reported statistically significant positive associations with ORs ranging from 1.53 (95% CI 1.02–2.29) to 15.46 (95% CI 1.89–126.67). One study reported a statistically significant negative association with a mean difference between patients with and without MCs2 of -3.2 (-5.39 –-1.01). The remaining five studies reported statistically non-significant findings for associations regarding MCs2. The wide range of ORs and the broad and overlapping 95% CIs indicate that there is no significant difference between MCs1 and MCs2 in regard to their associations with LBP. Two studies reported on the association between MCs3 and LBP and one of these found a positive association with OR 2.51 (95% CI 1.05–5.97). The remaining study reported a statistically non-significant finding for association regarding MCs3. ### 3.4.2 Associations between different sizes of MCs and LBP Three studies reported statistically significant positive associations between extensive MCs and LBP; two reported ORs of 1.83 (95% CI 1.14–2.94) and 83.10 (95% CI 4.85–1424.05) and one reported on continuous outcomes, see. However, the estimates for extensive MCs were not different from those for MCs of any type, regardless of size. ### 3.4.3 Pooled results Due to the heterogeneity of the observational (non-discography) study samples (differences in outcome measures and in study sampling, see) no meta-analysis was performed for these studies. However, meta-analysis was performed for the nine studies using concordant pain with provocative discography as the outcome measure. Separate analyses were made for *MCs any*, *MCs1* and *MCs2* resulting in ORs (95% CI) of 4.01 (1.52–10.61), 6.14 (2.47–15.27), and 3.15 (1.00–9.93), respectively, indicating that there was no significant difference in the associations with LBP between the two types of MCs. Substantial heterogeneity was identified for all three analyses with I²-values of 84, 64 and 81, respectively. ### 3.4.4 LBP intensity in patients with and without MCs None of the six clinical studies investigating the difference in LBP intensity between patients with and without MCs found a significant difference between the two groups. ## 3.5 Association between MCs and activity limitation ### 3.5.1 Association with activity limitation One of seven studies (three clinical, three non-clinical and one case-control) reporting on activity limitation outcomes found a statistically significant association between MCs and activity limitation. Määttä et al. reported an association between activity limitation (ODI\>15%) and both any MCs and MCs2, OR 1.47 (95% CI 1.04–2.10) and 1.56 (95% CI 1.06–2.31), respectively, but not for MCs1. ### 3.5.2 Level of activity limitation in patients with and without MCs None of the four clinical studies investigating the difference in activity limitation levels between patients with and without MCs found a significant difference between the two groups. ## 3.6 Is the association between MCs and outcomes modified by other factors? In relation to identifying single modifiers of the association between MCs and clinical outcomes, five studies reported stratified analyses, four with stratification on disc levels and one on sex. Of the four studies that stratified by disc level, three identified statistically significant associations only for some levels. Two studies found the associations to be stronger at the two lower levels, while one found statistically significant positive associations at L1-L2, L4-L5 and L5-S1. In the one study reporting stratified analyses on sex, MCs were associated with LBP only for men when using the outcome measure ‘LBP month’, only for women when using the outcome measure ‘Seeking care’ and for both sexes when using the outcome measure ‘LBP year’. In two studies, the authors reported analyses where disc degeneration was included as a modifying factor. In one study, disc degeneration was included as an interaction term and in one study as a single covariate in two separate multivariable analyses. In both studies, disc degeneration was reported to reduce the estimates of association between MCs and LBP by 10–28%. With regard to possible modifiers of the association between MCs and activity limitation, two studies investigated this. In the study by Mok et al., the authors reported that disc level did not affect the association between MCs and activity limitation (as measured by the Oswestry Disability Index and Roland Morris Disability Questionnaire). In the study by Määttä et al., disc degeneration reduced the association between MCs and the Oswestry Disability Index. ## 3.7 Sensitivity analysis Of the 30 studies investigating LBP (one study did not), 15 studies reported statistically significant positive associations with MCs. The results of the sensitivity analysis are reported in. The publication of statistically significant positive associations between MCs and LBP were not related to year of publication (p\<0.79), classified as 1998–2004 (n = 5), 2005–2010 (n = 10) and 2011–2016 (n = 15), nor to the total number of participants (p\<0.14), divided into \<100 participants (n = 12), 100–500 participants (n = 11) and more than 500 participants (n = 7). As only one of seven studies evaluating activity limitation was classified as having a statistically significant positive association and the remaining studies showed non-significant associations, sensitivity analysis for this outcome was not meaningful. Only one study was classified as having ‘no overall risk of bias’ and performing a sensitivity analysis on the overall risk of bias assessment was therefore not meaningful. ### 3.7.1 Post hoc sensitivity analysis To further investigate the possible influence of bias and other factors in the reporting of a statistically significant association between MCs and LBP, we performed post hoc analyses of the classification of associations (statistically positive association, yes/no) for individual risk of bias domains and signaling questions, LBP outcomes, study design, and MRI field strength. There was a statistically significant difference (p\<0.01) in the distribution of studies using continuous or dichotomous outcomes. All 15 studies that reported significant positive associations used dichotomous outcomes, e.g. ‘LBP \< 6 weeks’, as compared to only half (53%) of the 15 studies that did not report significant positive association. No other statistically significant differences were identified between the two groups of studies. # 4. Discussion ## 4.1 Main findings In summary, the results show inconsistent associations between MCs and both LBP and activity limitation. Only half of the studies reported statistically significant positive associations between MCs and LBP. Both pooled and individual study data indicate that there is no difference in the strength of associations of MCs1 and MCs2 with LBP. Among patients with LBP, the *intensity* of LBP does not seem to differ between those with MCs and those without. Only one of seven studies found an association between MCs and activity limitation. Finally, our results indicate that disc level and disc degeneration modify the association between MCs and LBP. With respect to previous systematic reviews on this subject, these are new results and will be discussed in more detail below. ## 4.2 Discussion of findings ### 4.2.1 Inconsistent positive association between MCs and LBP The proportion of studies that showed a statistically significant positive association between MCs and LBP was lower in this review (50%) compared to the previous reviews, 88% and 70%. Amongst the studies classified as having a statistically significant positive association, almost a third also had estimates that were non-significant (with some of these being negative). This finding, along with eight of the 15 studies reporting non-significant associations (also including negative estimates on at least one outcome measure ) and one study that reported a statistically significant negative association, indicate that the association between MCs and LBP is more inconsistent than previously reported. It would be reasonable to assume that the heterogeneity of study quality, samples, sex, clinical outcomes, and the prevalence estimates of both MCs and LBP could explain the conflicting results. However, in an attempt to explain the differences between studies that did and studies that did not report significant positive associations between MCs and LBP, sensitivity analyses were performed for sample size, publication year, study design, type of LBP outcome, and MRI field strength. None of these explained the differences in the directions and strengths of associations. The prevalence of MCs is, of course, dependent on the definition of MCs used by the different authors, which could help explain the large variation seen in reported prevalence of MCs. There is large variation in the interpretation of when MCs are present in the included studies, e.g. “all signal changes in the vertebral bone, extending from the endplate, regardless of size” vs. “tiny spots of signal intensity change in the bone marrow adjacent to the vertebral corners, were not recorded.”. However, the lack of detailed reporting of definitions of MCs in the majority of studies made it impossible to analyse the impact of different phenotypes of MCs on the association with outcomes. Although the pooled results from the discography studies revealed statistically significant positive associations for all types of MCs with estimates ranging from OR 3.2 to OR 6.1, their 95% CIs are wide, and range from 1.00–15.27. Provocative discography carries inherent risks of bias when used as a diagnostic test in the presence of MRI findings. Patients subjected to discography are selected on the basis of clinical findings, including MRI. When the reference standard (LBP by provocative discography) is not blinded from the index test (MRI), there is a risk of circular reasoning that could confound the association. Furthermore, there are laboratory data showing increased intradiscal pressure at discs adjacent to the injected level in animal models, calling into question the validity of provocative discography. Because of the shortcomings mentioned, care must be taken when interpreting the results from the analyses of associations in this review. It is possible that future large scale high quality studies will affect the direction of the associations presented above. ### 4.2.2 Type and size of MCs do not seem to matter There was no significant difference in the strength of associations between MCs1 and LBP and MCs2 and LBP, either in the individual studies or according to the pooled results. Intuitively, one would believe that MCs1 would have a stronger association with pain than MCs2, due to the fact that MCs1 are supposed to occur in response to an inflammatory reaction, whereas MCs2 are considered a more biologically inactive entity. However, the lack of difference in strength of associations with pain could be attributed to the fact that (1) MCs1 and MCs2 can co-exist at the same disc level and/or within the same individual, and that (2) MCs2 often follow MCs1, making MCs2 a possible proxy for further degenerative changes (e.g. disc degeneration, protrusions/herniations) that are potentially painful. With regard to the size of MCs, we found that the estimates for the associations between ‘extensive’ MCs and LBP are not different from those between MCs of any size and LBP. One possible explanation for this is that considering a normal stimulus response curve for pain, the plateau for pain may be reached even for small discovertebral lesions. ### 4.2.3 No difference in LBP intensity between patients with and without MCs The results of the six clinical studies that investigated the LBP intensity in patients with and without MCs, indicate that patients with MCs may not experience more intense pain than those without MCs and thus, they may be difficult to identify solely based on pain intensity. The lack of difference in pain intensity between patients with and without MCs may be explained by the fact that all patients with LBP are in pain and that the pain experience is influenced by a multitude of factors other than nociception. Another explanation could be that MCs are only one finding among others in the degenerative chain of events, where disc degeneration, herniations and osteophyte formation each play their part and as such, MCs do not always stand out as the main contributor to LBP. ### 4.2.4 No support for association between MCs and activity limitation To the knowledge of the authors, this is the first systematic review to investigate the cross-sectional association between an MRI finding and pain- related activity limitation. Only one of the seven studies that reported on this association found a statistically significant positive association between activity limitation and MCs. In that study, by Määttä et al. the crude estimates for associations between MCs and activity limitation were mainly positive. Based on the results from the current review, there is no evidence to support that MCs are cross-sectionally associated with activity limitation. However, in support of a positive association, a recent longitudinal study by Järvinen et al. investigating patients with MCs and LBP, found that change in the extent of MCs1 was positively associated with 2-year changes in the Oswestry Disability Index, both unadjusted and adjusted for age, sex and size of MCs at baseline. ### 4.2.5 The association with LBP is likely modified by disc level and disc degeneration Although only based on three and two studies, respectively, disc level and disc degeneration were identified as potential modifiers of the association between MCs and LBP. A possible reason for MCs at the lower disc levels being more strongly associated with LBP is that this part of the lumbar spine is subjected to increased discovertebral load. Therefore, lower disc level is likely to be a proxy for other factors, e.g. increased physical load or injury to the discovertebral complex, that could lead to LBP. However, due to the low prevalence of MCs in the upper lumbar spine, the estimates for these levels are uncertain, and therefore more research would be needed to make it possible to more closely evaluate disc level as a possible modifying factor. The confounding of the association between MCs and LBP by disc degeneration may be explained by studies reporting that disc degeneration is an independent risk factor of LBP. ### 4.2.6. Overall risk of bias of included studies There was an overall risk of bias in all included studies but one. This risk was partly due to insufficient reporting and may not necessarily imply actual bias. Still, risk of bias needs to be taken into account when interpreting the current results and when performing new primary studies. The most common problems within each of the four bias domains were: 1) Lack of randomly or consecutively selected study participants, which could introduce a risk of selection bias, 2) Lack of reliability testing, raising concerns about misclassification which would influence prevalence rates of MCs, and thus also the strengths, directions and validity of associations as these are dependent on the prevalence, 3) Lack of blinding between assessment of outcome measure and MRI results, which was mainly an issue for discography studies where patients were referred for the procedure on the basis of the results of their MRI scan, which might have introduced beliefs that could affect their reporting of pain, and 4) Failure to report on the timing of the MRI and clinical outcome assessments, with longer periods increasing the risk of change in either MRI appearance or LBP/activity limitation status. ## 4.3 Limitations and strengths ### 4.3.1 Limitations While our sensitivity analysis did not show that study design influenced the result, we did include case-control studies, although they are less suited for our purpose due to the fact that the groups are from different samples, thus introducing a potential bias, as described in the Cochrane Handbook. By only including studies that had evaluated MCs at all lumbar levels, it is possible that we excluded high quality studies that could have informed on the association between MCs and LBP. However, that decision was made to allow us to investigate the modifying effect of disc level on the association between MCs and LBP, which was novel analysis. ### 4.3.2 Strengths We did a broad search without language restriction in three major databases, supplemented with a hand search and query for additional studies from experts. The search was not restricted to terms for “low back pain” and “activity limitation”/“back-related disability”, since we originally wanted to investigate the prevalence of MCs as well. This strategy reduced the risk of missing important studies. That the relatively large group of reviewers, all with a special interest in MCs, only found one additional article, likely indicates that our search strategy was comprehensive. The results from this review are based on three times as many studies than previous reviews. The increased number of included studies helps estimate the direction and strength of the associations even though the heterogeneity of the studies made it inappropriate to perform meta-analysis on the relationships between MCs and self-reported LBP and pain-related activity limitation respectively. We based our risk of bias on the QUADAS 2, a recommended and validated tool for the task, but modified it after pilot-testing without further validation. A pilot study was performed on both the data extraction and risk of bias assessment, which familiarized the assessors with the process, and highlighted problematic areas, which were changed before initiating the study. For the extraction of data and risk of bias assessment, assessors were blinded to the assessments of their fellow co-assessors up until consensus. Our review was performed by a large number of reviewers from different research groups, all of whom were familiar with the subject of MCs. ## 4.4 Recommendations for further research The widely different prevalence rates reported for MCs in similar populations may indicate inconsistent phenotyping of MCs. Agreement on the characterization of MCs across studies is needed, including criteria for size and for differentiation from other signal changes (e.g. fat or edema in osteophytes, inhomogeneous bone marrow, hemangiomas abutting the endplate), or at least a concise reporting of the methods used to evaluate these findings (including all relevant MRI parameters), in order to be able to compare results between studies. In light of the results of our risk of bias assessment, we urge researchers to improve their reporting of the methods used. In particular, we found weaknesses related to the selection of study samples, reliability testing on MRI assessments, blinding and study logistics (timing of assessments). Researchers might also assess whether other characteristics of MCs (e.g. location, extent, their signal after fat suppression) may be more relevant to pain than are the type of MCs based on conventional T1- and T2-weighted MRI. To be able to further our understanding of the details of the association between MCs and LBP, we need large population-based cohort studies with low risk of bias that allow for stratified or multivariable analyses including known and suspected modifiers. ## 4.5 Clinical implications of our findings The lack of difference in pain intensity between patients with MCs and patients without MCs, along with the sparse knowledge around other distinguishing clinical characteristics, makes identification of patients with MCs difficult, without the use of MRI. However, this may be without clinical relevance, as our finding of a more inconsistent association between LBP and MCs than previously shown should call for caution when using ‘Modic changes’ as a diagnosis, explanation for LBP, and indication for specific treatment in patients with non-specific LBP. ## 4.6 Conclusion The results from this systematic review show that the associations between Modic changes and both outcomes of low back pain and activity limitation are inconsistent. Heterogeneity in terms of study samples, classification of Modic changes, clinical outcomes and prevalence of Modic changes and low back pain may explain the inconsistent associations. Also, no difference in low back pain intensity or level of activity limitation was found between patients with and without Modic changes. These results question the conclusions from previously published reviews that Modic changes may constitute a specific clinically relevant subgroup among people with low back pain. Disc level and disc degeneration were identified as factors potential modifying the association between Modic changes and low back pain. New studies with low risk of bias are likely to affect the direction and strength of these associations. # Supporting information We would like to thank Youting Bentzen for the assistance with translation of papers in Chinese and Suzanne Capell for assistance with copy-editing. [^1]: These authors have declared that no competing interests exist: CH, PK, AE, JStS, CLY, JN, JSoS, KS, TSJ. JK: Paid lectures by MSD and Pfizer, Stocks of Orion Pharma Ltd, Membership of Scientific Advisory Board of Axsome Therapeutics Inc.

# Introduction *Plasmodium falciparum* is responsible for the most severe forms of human malaria that include cerebral malaria, pregnancy-associated (placental) malaria, and acute anemia. A major aspect of the virulence of *P. falciparum* derives from the ability of parasitized erythrocytes to adhere to different endothelial cell types in the deep vasculature of the body, thus resulting in a sequestration of the parasites away from splenic clearance. The specificity of malaria parasite cytoadherence is mediated by variants of the parasite-specific adhesin, ***P****.* ***f****alciparum* **e**rythrocyte **m**embrane **p**rotein 1 (PfEMP1) that are exported and assembled on the infected host cell surface where they interact with diverse cellular receptors in the microvasculature. Whereas parasite sequestration in the peripheral microvasculature is associated with parasitized erythrocytes that bind to CD36, ICAM-1, VCAM or E-selectin receptors, sequestration in the placenta mainly involves chondroitin sulfate A (CSA) that is abundantly expressed by placental syncytiotrophoblasts. Furthermore, evidence from targeted gene disruption studies have established that the PfEMP1 variant var2CSA is the main ligand mediating the cytoadherence process against placental CSA receptors. In support of these observations, var2CSA-dependent binding of parasitized erythrocytes to placental BeWo cells can be efficiently inhibited by pretreatment with soluble CSA proteins. This suggests that a disruption of var2CSA protein export and assembly on the erythrocyte surface, or inhibition through chemotherapy of its interaction with placental CSA receptors could limit the disease severity reversing the pathophysiology of placental malaria. However, despite the availability of several drug compounds targeting the asexual development and growth of malaria parasites *in vitro* and *in vivo*, none of the current antimalarial drugs and those in clinical development is able to protect against placental malaria. For example, post-mortem studies of severe malaria in previously treated patients often show high levels of infected erythrocytes bound to the microvasculature in spite of clearance of the peripheral parasitaemia. These observations strongly underscore the urgent need for new antimalarial drugs targeting the cytoadherence process of malaria parasites, particularly in high-risk pregnancy cases. Here, we describe the development of a novel image-based assay for the high- throughput screening and identification of small molecule inhibitors of parasitized erythrocyte cytoadherence in the placenta during pregnancy. # Materials and Methods ## *P. falciparum* and BeWo cell cultures *Plasmodium falciparum* FCR3 strain was maintained in RPMI 1640 media supplemented with *L*-glutamine, 25 mM HEPES, sodium bicarbonate, 0.5% Albumax, 0.1 mM hypoxanthine and 16 µM thymidine in human O⁺ erythrocytes. To enrich for the PfEMP1^var2CSA expressing phenotypes, parasitized erythrocytes were panned once every three weeks, and also before use in our binding assays against a monolayer of BeWo cells in culture dishes. Additionally, parasites were synchronized by two sequential treatments with 5% sorbitol at 10 hours interval, and cultivated for at least one complete cycle prior to the drug assays with early ring stage parasites (∼6 hpi). The human choriocarcinoma placental cell line (BeWo) was obtained from the American Type Culture Collection (ATCC CCL-98) and grown in Ham's F12 medium supplemented with *L*-glutamine (Invitrogen) and 10% fetal bovine serum (Gibco). For the cytoadherence assays, adherent BeWo cells were detached using trypsin- EDTA (Gibco) and washed with assay-complete medium (ACM) comprising Ham's F12 medium supplemented with 4.7% human serum (Sigma), 0.23% Albumax I (Gibco) and 0.7% fetal bovine serum. The cells were then seeded at a density of 2000 BeWo cells per 50 µl culture in 384-well plates (Greiner), and grown for 4 days at 37°C to achieve approximately 80% confluency prior to the cytoadherence assays. ## Cytoadherence assay and image acquisition Panned *P. falciparum* FCR3-infected erythrocytes were diluted in complete culture media (50 µl per 384-well) to a final parasitaemia of 6% and hematocrit 2%, and then cultivated with or without drugs for 24 h at 37°C. After mixing by vortexing at 1700 rpm for 45 seconds (Mixmate), 5 µl of the samples were transferred onto a monolayer of BeWo cells in a second microtiter plate and incubated for 1 h at 37°C to allow for binding of the infected erythrocytes. Unbound erythrocytes were washed three times with assay complete media using an EL406 combination washer (Biotek), and the attached cells fixed with 4% paraformaldehyde at RT for 15 minutes. This was followed by nucleic acid staining with Syto60 (Molecular Probes) diluted in PBS (1∶4000) and erythrocyte membrane labeling with anti-glycophorin A FITC-conjugated antibody (Caltag Laboratories) at a 1∶1000 dilution in PBS. The plates were washed again and imaged using an ImageXpress Ultra automated-confocal microscope (Molecular Devices). Four images (2000 pixel×2000 pixel each) were acquired from each test well using a 20×-magnifying lens, and analyzed using customized algorithms that were developed in-house. ## Image mining algorithms and data analysis To quantitatively determine the effect of small molecule inhibitors of *P. falciparum* cytoadherence to BeWo cells, we developed specific algorithms capable of measuring the proportion of overlapping BeWo cell area with bound parasitized erythrocytes. We assumed that all parasitized erythrocytes are of the same sizes and that the proportion of BeWo cell area occupied by the bound erythrocytes directly correlates with the number of adhering erythrocytes. We confirmed such correlations by measuring the proportion of overlapping infected red blood cell area per BeWo cell area with increasing amounts (parasitaemia) of panned erythrocytes. For both the attached RBCs (green fluorescence channel) and BeWo cells (red fluorescence channel), a Gaussian low-pass filter, was used for noise filtering whereas adaptive thresholding was used for the cell segmentations. This adaptive threshold was based on a k-means clustering algorithm that separates image pixels into either foreground (BeWo or iRBC) or background. The above-described algorithm was then implemented as a plugin (programming language C-Sharp) to Institut Pasteur Korea's High Content Screening platform that is currently accessible only to authorized users (cf Moon and Genovesio, 2008). ## Drug effects on parasite cytoadherence and viability To validate the assay protocol, we investigated the effects of the cytoadherence competitive inhibitor chondroitin sulfate A (CSA), the protein transport inhibitor brefeldin A (BFA), and the antimalarial compound artemisinin (ART) on cytoadherence to the BeWo cells and parasite growth *in vitro*. Dose-response experiments were designed with final concentration ranges from 1 mg/ml to 50 µg/ml for CSA (3-fold dilution), 160 µM to 313 nM for BFA (2-fold dilution), and 400–0.78 nM for ART (2-fold dilution) in a 384-well plate. Panned parasites at early ring stages were then added and grown for 24 h to allow for the expression and assembly of cytoadherence factors on the infected RBCs. Five microliters (5 µl) of each culture was then transferred onto a previously prepared plate with BeWo cells and analyzed for cytoadherence as described above. To assess the effects of each compound on the parasite growth, plates with the remaining culture (45 µl per well) were incubated for a further 24 h to allow for schizont egress and invasion. The parasite growth relative to non-treated controls (wells containing DMSO at 0.5%) was then determined using the pLDH assay, and their EC₅₀ values determined using GraphPad Prism 5.0. All assays were done in triplicates and the calculated standard deviations used to assess the assay reproducibility. Additionally, z-factors which inform on the reliability of each test procedure were determined for both the cytoadherence and viability assays using DMSO as positive control and CSA (1 mg/ml) or ART (400 nM), respectively, as negative controls. To confirm the involvement of CSA receptor binding in the assays, CSA and CD36-panned erythrocytes, were analyzed at equal parasitaemias (6% parasitaemia) for binding to the attached BeWo cells. # Results To facilitate the discovery of new antimalarial drugs targeting the cytoadherence process of *Plasmodium*-infected erythrocytes during pregnancy, we have developed an image-based drug susceptibility assay suitable for high- throughput screening of diverse chemical libraries. This assay was designed for identifying drug compounds capable of inhibiting either the export (and assembly) of malaria parasite cytoadherence molecules on the infected erythrocyte surface, or their interactions with cellular receptors in the placenta microvasculature. Parasitized erythrocytes at the late developmental stages were layered on a monolayer of placental BeWo cells and allowed to bind for 1 h at 37°C. The attached erythrocytes were then labeled with glycophorin A antibody conjugates whereas the underlying BeWo cells were stained with the nucleic acid dye Syto60 for fluorescence microscopy imaging. For an unbiased quantification of binding ratios, specific image mining algorithms were developed and validated for its performance in detecting the effects of two known cytoadherence inhibitors, CSA and BFA, and the antimalarial drug artemisinin. As shown in, a positive correlation of 0.79 was obtained between the binding ratios (proportion of placental BeWo cells with attached erythrocytes) as determined by the developed algorithms and the culture parasitaemias, thus indicating a high performance of the assay in detecting changes in erythrocyte binding capabilities. To confirm the involvement of host CSA receptors in binding of the infected erythrocyte to placental BeWo cells, CSA and CD36-panned erythrocytes were allowed to bind on the BeWo cells and their binding ratios were compared. As shown in, the CSA-panned erythrocytes exhibited over 4-fold increase in binding to the placental cells when compared to the CD36-panned cells (p-value of 0.004). Likewise, binding of the BeWo- panned erythrocytes to the BeWo containing plates was highly sensitive to the presence of soluble CSA in the binding assay, thus indicating a high specificity of the binding interactions in our developed system. Together, these findings strongly indicate that the assay is highly suitable and relevant for drug studies specifically targeting the cytoadherence of *P. falciparum*-infected erythrocytes in the placenta. To validate the assay for the quantitative assessment of diverse drug effects *in vitro*, dose-response experiments were done using the cytoadherence inhibitor CSA, as well as the protein transport inhibitor BFA, and the schizonticide artemisinin. In parallel, parasite viability assays were done with all three agents as described in the “” section and in order to distinguish between potentially cytotoxic agents from the cytoadherence-specific inhibitors. As shown in, binding of the BeWo-panned infected erythrocytes to the plated BeWo cells was highly sensitive to increasing concentrations of all three agents with half maximum effective concentration (EC₅₀) values of 12 µg/ml for CSA, 8 µM for BFA and 16 nM for artemisinin. A comparison of parasite viability in the presence of the different drug concentrations following an extended period of additional 24 hours indicates that only artemisinin and BFA were toxic to the parasite growth with EC₅₀ values of 72 nM and 12 µM, respectively. These data indicate that the observed artemisinin or BFA inhibition of parasitized erythrocyte binding to the BeWo cells presumably occurred due to the drug effects on parasite growth, and that soluble CSA is a specific inhibitor of *Plasmodium* cytoadherence. Taken together, the data suggest that our combined assay is capable of distinguishing between cytoadherence-specific agents and cytotoxic compounds that might influence the hit-selection process. To further validate our developed assay system in terms of its reproducibility and/or hit detection accuracy, replicate experiments (n = 192 wells/plate×3 plates) were done using untreated parasites as positive controls or CSA (1 mg/ml)-treated cultures as negative controls. Using the calculated mean binding ratios and standard deviations from both controls, a z′ value of 0.4 was then determined for the new assay (data not shown). These data suggest that the developed assay is reliable for the use in high-throughput screening of diverse compound libraries. # Discussion To facilitate the rapid discovery and development of new antimalarials targeting the cytoadherence of *Plasmodium*-infected erythrocytes in the placenta microvasculature, we have developed a robust and technically simple image-based phenotypic assay for malaria based on the use of placental BeWo cells in a 384-well plate format. Specific image-mining algorithms were developed for automated quantification of binding ratios, and for accurate assessment of various drug effects. As evidenced by i) the high correlation coefficient (R²) of 0.79 between the determined binding ratios and parasitized erythrocyte ratios (parasitaemias), ii) the high sensitivity of our assay (detection limit \<0.38% parasitaemia), and the calculated z′ value of 0.4, our developed assay is suitable for high-throughput drug studies targeting placental malaria. Additionally, our assay represents a more relevant drug susceptibility system for placental malaria when compared to assays with immobilized CSA since the herein used assay utilizes whole cells that were derived from human placental tissues. BeWo cells are heterogenous cells expressing several diverse parasite cytoadhesion receptors including CSA in abundant amounts, and to a lesser extent, intercellular adhesion molecule-1 (ICAM-1) and neonatal Fc receptors, all of which are capable of interacting with *P. falciparum*-parasitized erythrocytes,. Unlike peripheral cell-types such as in the brain microvasculature, placental BeWo cells do not express other known cytoadhesion receptors including CD36, CD31, E-selectin and vascular cell adhesion molecule-1 (VCAM-1). In support of the absence of CD36 receptors on BeWo cells, parasitized erythrocytes that were panned against immobilized CD36 proteins did not significantly bind to the adherent BeWo cells when compared to the CSA-panned cells. These observations suggest that the developed assay is highly specific and suitable for drug studies targeting CSA-dependent binding in the placenta. To further validate the developed assay for use in anti-cytoadhesion drug studies, we investigated its performance in quantifying the inhibitory effects of soluble CSA, the protein transport inhibitor BFA, and the fast-acting antimalarial drug artemisinin. Consistent with a predominant involvement of the CSA receptor in binding to the BeWo cells , soluble CSA efficiently inhibited the cytoadherence process with an EC₅₀ value of ∼12 µg/ml. This EC₅₀ value is consistent with the activity range of CSA using FCR3-parasitized erythrocytes and BeWo cells. However, the obtained EC₅₀ value is significantly high when compared to published activity data using immobilized CSA (∼0.2 µg/ml), or using other parasite strains (approximately 70% inhibition of CS2 binding with 10 µg/ml CSA). These observations suggest that BeWo cells presumably bind the parasitized erythrocytes with much higher avidity than would be obtained with purified receptors, and that the associated binding interactions are strain-dependent. In contrast to the effects of artemisinin and BFA that inhibited the parasite growth, CSA was non-cytotoxic suggesting that CSA is a specific inhibitor of placental cytoadherence processes. Meanwhile, the observed binding inhibition by artemisinin and BFA in this study presumably resulted from the drug effects on the parasite growth as previously reported. Taken together, our findings strongly indicate that the developed assay, when combined with a growth inhibition assay, can reliably discriminate between cytoadherence-specific effects and parasite cytotoxicity effects. Interestingly, both BFA (cytoadherence protein export inhibitor) and CSA (cytoadherence inhibitor) inhibited the cytoadherence process in our assay. This suggests that compounds with anti-adherent activities may function either in the secretory pathway of cytoadherence molecules, or by direct interference with the cytoadherence process as observed with CSA. This new assay represents a significant advancement in antimalarial drug discovery that will facilitate the development of novel anti-adhesive therapies against the disease. A unique attribute of the assay is in the customized algorithms used in quantifying the parasite binding ratios. These algorithms are easy to develop and can be readily replicated in any screening facility that is equipped with a medium-size computer laboratory and a plate microscope for imaging. The authors would like to thank all members of CND3 for their support. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MK FMD LFJ. Performed the experiments: MK FMD. Analyzed the data: MK FMD MAEH AG LFJ. Contributed reagents/materials/analysis tools: MAEH AG LFJ. Wrote the paper: MK FMD MAEH LA LFJ. [^3]: Current address: Broad Institute of MIT and Harvard, Cambridge, Massachusetts, United States of America

# Introduction Rugby union and rugby league are intensely physical contact sports that place players at risk for frequent traumatic injury. This risk has increased in parallel with the introduction of professionalism to the sport, given that a greater advantage can be gained from larger, stronger players. The changes in physiology and anthropometrics of professional rugby players, have led to a greater physicality in the game and an increase in the incidence of traumatic musculoskeletal injury of almost 20% since the 2009/2010 season. Both codes of rugby involve significant contact and physicality, and the use of little or no body padding. The multiple physical collisions, tackles and high impact hit-ups repeatedly expose players to very high magnitudes of direct force, with reports of ‘g’ forces as high as 7–10 g during professional games of rugby. Fractures occur when loads exceed the capacity of a bone to withstand them, and constitute up to 6% of all rugby injuries, although not frequently reported at the spine. In the general population, vertebral fractures (VFs) are under recognised and therefore under-reported, with only around 1 in 4 receiving clinical attention. This lack of recognition is due both to the absence of notable symptoms, given that VFs commonly occur without pain of a sufficient magnitude to arouse concern, and there is often difficulty in determining the cause of symptoms. This may be particularly the case for rugby players who, with a higher than average pain threshold and being accustomed to injury, may be less likely to report minor pain or symptoms. There is a high risk for symptomatic spinal injuries (including vertebral nerve root and disc injuries) in professional rugby union players; with incidence quantified at 10.9 per 1000 match hours. However, the specific incidence or prevalence of spine fracture (asymptomatic or symptomatic) in professional players is not known. Although VFs are infrequently reported, all are associated with morbidity. In professional rugby players, back injuries have the highest rate of recurrence of all injuries, and there is an increased risk of further fracture following an initial VF. Vertebral fracture may also be a catalyst for degenerative disease, discopathy, kyphosis, back pain and neuroplaxia. Dual energy X-ray absorptiometry (DXA) has previously been used to assess and monitor body composition in professional rugby players. Recent models of DXA now include a morphologic Vertebral Fracture Assessment (VFA) facility, which provides an accurate, valid and reliable method of identifying and grading VF, to a standard comparable to conventional radiography and accepted by the International Society for Clinical Densitometry (ISCD). The densitometer acquires a high resolution radiographic image of almost the entire spine at a radiation dose of less than 1% of a comparable radiograph. VFA is used clinically in patients with a high risk of fracture, and has been used in published research studies of VF in adolescents, and men and women aged 18 to 87 years. Despite the known increased risk for back injuries in contact sports, no study to date has investigated the prevalence of vertebral fracture in a professional rugby or sports cohort. The objective of this study was to investigate the prevalence of VFs in professional male rugby union and rugby league players, through whole squad screening using DXA-VFA. # Methods ## Ethics statement Ethical approval for the study was granted by the Carnegie Faculty Research Ethics Committee (Leeds Metropolitan University, UK) in accordance with the Declaration of Helsinki, and all participants provided their signed, informed consent prior to taking part in the study.An observational study was conducted and all tests were performed at the same research centre, between November 2010 and May 2012. Ninety five UK-based male professional rugby league (n = 53) and rugby union (n = 42) players provided signed informed consent to participate in the study. There were no players who declined the invitation to participate in the study. Rugby league players were recruited from one Super League and one Championship team (all first team members), and rugby union players from one English Premiership club (all first team members). The minimum age for participation in the study was 20 years. All participants had been signed on professional rugby contracts for at least the preceding three years, and all were currently signed to contracts. ## Measurements Participants wore light-weight clothing and removed shoes and any jewellery for all physical measurements. Standing height was measured using a stadiometer (SECA, Birmingham, UK) and recorded to the nearest millimetre. Body mass was measured with calibrated electronic scales (SECA, Birmingham, UK) and recorded in kilograms (kg) to the nearest 0.1 kg. Vertebral fracture assessment (VFA) was performed in all participants, using a GE Lunar iDXA densitometer with the VFA software installed (Lunar iDXA fan beam densitometer with enCORE software version 13.5, GE Medical Systems, UK). The participant lay in a left lateral decubitus position according to the manufacturer's instructions for patient positioning, to ensure the spine was parallel to the table. The arm of the machine moved to the lateral position and then a lateral fan-beam high resolution X-ray image of the whole spine was obtained. Whilst spinal radiographs are generally considered to be the gold standard for the diagnosis of vertebral fractures, the VFA method is recognised as advantageous for being easy to use, precise and using low radiation,. Furthermore, the iDXA VFA provides minimal interference from soft tissue artefacts around vertebral bodies, especially in the thoracic regions, whereas conventional spine radiographs require adjustment for soft tissue. The images acquired were of high quality and all vertebrae between L4 and T6 could be visualised clearly. The automated quantitative morphometric analysis labelled vertebral deformations using a 6 point measurement of the anterior, posterior and mid points of the vertebras. In accord with the ISCD guidelines, each vertebra on all 95 scan images, were also visually inspected by a certified and VFA trained (ISCD) densitometrist, in order to minimise false positive or false negative results. Particular care was given to distinguish normal anatomic variants and non fracture abnormalities, notably Sheuermann's, degenerative disease, osteophytes, diffuse idiopathic skeletal hyperstosis, Pagets, Schmorl's nodes, cupid's bow defect, and rib and scapula shadows. Visualisation of vertebra was optimised using the ClearView facility which enhances images by adjusting contrast and brightness for bone edge enhancement. Once markers were agreed on, the software calculated the degree and type of vertebral shape anomalies using the semi- quantitative Genant classification, which is considered the most appropriate and standardised method. Anomalies were classified as either wedge (when the anterior height was the lowest), biconcave (when the middle height was the lowest) or crush (when the posterior height was the lowest). A relative height reduction (with reference to posterior-mid-anterior heights) between 20–25% was graded ‘mild’, 26–40% ‘moderate’ and \>40% ‘severe’. Precision for VFA by DXA has been reported at 0.84% CV for the average height of vertebrae in men aged 38 to 87 years. Lumbar spine (L1-L4) areal bone mineral density (BMD) was evaluated using standard DXA procedures. Age- and sex-specific reference data was used to calculate BMD Z-scores. Body composition was measured in all participants also using DXA of the total body, and variables included fat percentage and lean mass. Local precision values for our Centre are 0.4% for lumbar spine BMD and 0.5–0.9% for body composition (in healthy subjects, aged 34.6 years). The observed in-vitro coefficient of variation was low at less than 0.5% for the regular quality control scans of the Lunar (soft tissue and bone) calibration phantom. In this study, repeat scanning to perform a precision analysis of the vertebral deformity grading by VFA DXA was not feasible. However, published data demonstrates excellent agreement between DXA VFA and conventional radiography, with very good sensitivities and specificities, particularly for moderate (grade 2) and severe (grade 3) fractures. ## Statistical analysis All statistical evaluations were performed using SPSS version 18.0 (LEAD Technologies Inc). Descriptive statistics (mean, standard deviation of the mean, minimum and maximum values) were used to characterise the sample. Data were normally distributed, therefore comparisons of descriptive results between groups (with VF *v* without VF; code: rugby union *v* rugby league; position: backs *v* forwards) were made using independent t-tests. Pearson's correlation analyses were computed to investigate relationships between dependent and independent variables. The level of significance for all tests was set at p\<0.05. # Results Participant descriptive results are summarised in. There were no differences in age, BMI, body fat or BMD between rugby union and rugby league players (p\>0.05). Rugby union players were taller (186.7 (6.9) *v* 182.3 (6.1) cm, p = 0.001), heavier (105.1 (12.9) *v* 96.1 (10.1) kg, p\<0.001), and had greater lean mass (81.6 (7.6) *v* 75.4 (7.6) kg, p\<0.001) than rugby league players. When participants were grouped by playing position (backs or forwards), there were 49 backs and 46 forwards. Forwards were taller, heavier, with a higher percentage body fat, lean mass, total body and lumbar spine BMD than backs (p\<0.05). In total, VFA identified 120 morphometric VFs in 51 players. Examples of VFA images and morphometry from our sample are presented in to. Seventy four were graded as mild (grade 1) (62%), 40 as moderate (grade 2) (33%) and 6 as severe (grade 3) (5%). VFs by DXA were observed in 24 (57%) rugby union players and 27 (51%) rugby league players. Multiple fractures (\>2) were identified in 37 players (39%). Moderate to severe VFs were identified in 31 players (33%). Three players (3%) were each found to have five VFs and each of these players had a deformity identified as ‘severe’. Five players (5%) had four VFs each (all with ‘moderate’, and two with ‘severe’). There were no differences in average number of VFs per player between rugby union and rugby league (n = 1.2 (1.4) *v* n = 1.4 (1.6); p = 0.731) players. There were no differences in the average number of VFs per player between forwards and backs (n = 1.2 (1.4) *v* n = 1.4 (1.6); p = 0.477). Furthermore, there were no differences in participant descriptive results between players with, and without VF. The majority of the identified VFs were located in T8 (n = 23), T9 (n = 18) and T10 (n = 21). Fractures classified as severe (grade 3) were prevalent in T7, T8, T12, L3 and L4. The most common type identified was wedge (n = 50), followed by biconcave (n = 44) then compression (n = 26). All 6 severe VFs were wedge type. There were also no differences in the type or grade of fracture between rugby union and rugby league players. Bivariate correlation analyses revealed positive associations between lean mass and BMD at the total body (r = 0.462, p\<0.001) and lumbar spine (r = 0.366, p\<0.001). There were no associations between height and BMD. There were no associations between any independent variable and number, type or grade of VF. # Discussion This is the first study to investigate the prevalence of VF using DXA Vertebral Fracture Assessment imaging in professional rugby players and to examine differences in prevalence between rugby union and rugby league codes. The principal finding was that of a high prevalence of VF in a group of professional rugby league and rugby union players. The majority of VFs were found in the thoracic spine region, between T8 and T10. The rates of over 50% for all grades of fracture, and 33% for grades 2–3 (moderate - severe), are greater than prevalence rates cited for the general population. There are few published reports of thoracic or lumbar spine fractures in rugby players. Most of the existing data are related to trauma to the cervical spine; likely due to the potentially devastating clinical consequences of injuries at this region. The first report of thoracic spine fractures in a rugby league player was in 1997. Investigations were conducted after the player complained of persistent dorsal mid-thoracic pain, and revealed fracture to T6 and T7 without neurological complications. Elsewhere, symptomatic spinal injuries have been reported in rugby union players, including nerve root injury. Investigations in the current study were not based on symptoms, except for one case. One rugby league player (full-back) visited us with back pain the week following a significant collision with another player. He had received chiropractic treatment for his pain, which worsened and was assumed to be muscular. VFA revealed multiple clinical (with pain) fractures between T6 and T10 (severe at T8 and moderate at T9 and T10). The player rested but returned to play within one month. Our findings suggest a higher risk for vertebral injury in players engaged in professional rugby of both codes, than previously recognised. There were no player characteristics that were predictive of number, type or grade of VF, and there were no differences in characteristics between players with and without VFs. It is common knowledge that a weak bone is less capable of withstanding force than a strong bone, and in the general population, fractures to the spine are most frequently associated with bone fragility. Athletes from contact sports generally have superior BMD compared to non-athletes, supporting Frost's Mechanostat theory, which proposes that when all else is equal, individuals who load their skeletons to a high degree should have stronger bones than their less active peers. The mean total body and lumbar spine BMD of our sample was well above age-matched normative values, and the lack of relationship between BMD and vertebral deformity suggests an underlying aetiology unrelated to bone fragility. Fracture risk is also dependent on the force of impact and the structure and geometric properties of bone. The architecture of the vertebrae optimises strength and flexibility while minimising weight. It is not known if rugby player vertebral bone structure (such as bone size, cortical thickness, trabecular number, thickness, and connectivity) differs from the non-rugby playing population. Nor have the forces to which the spine is exposed during rugby playing been directly quantified. As a consequence, the magnitude of force to the spine in rugby that will exceed the threshold of vertebral bone tolerance is unclear. The prevalence of the observed VFs in rugby players however, may indicate that peak strain magnitudes in the spine generated during tackling and in collisions, or significant hyperextension, can exceed the vertebral bone strain threshold. It has been estimated that more than two thirds of a ton in force is shared across the front row of players during the engagement of the scrum in rugby union and that this is associated with a greater risk for spinal cord injury, prompting calls for a ban of the contested scrum. Our finding of a similar prevalence of VFs in rugby league and rugby union players, despite the absence of the contested scrum in rugby league, suggests additional causes of VFs may also be at play. Indeed it has been suggested that collisions have the highest propensity to cause traumatic injury during rugby. Collisions are not only limited to game situations - an average of 77 collisions per player have been recorded during National Rugby League team training sessions. The demands on professional players to maintain training and game time may increase vertebral micro-damage and muscle fatigue thereby exposing the spine to even greater mechanical stress during impacts. Micro-damage to bone tissue that occurs in the course of regular mechanical loading is repaired through a process of bone remodelling so that overall bone strength is maintained. When the rate of micro- damage accumulation is greater than the rate of repair, bone is more susceptible to structural failure. Medical management of pain in professional rugby players, including the use of steroid injections, local anaesthetic and chiropractor therapy may exacerbate vertebral bone injury. These interventions may also return players to the game before micro-damage can be repaired, thereby exposing them to increased risk of fatigue fracture at the spine. Research is required into both the direct and indirect localised effects of analgesic interventions on bone in elite sports populations. The long term consequences of VF in professional athletes from contact sports, is unknown. A recent study reported an 11% increased risk of mortality among Olympic athletes from contact sports compared with other athletes. The authors suggest their finding reflects the impact of repeated collisions over time and hypothesised that the risk may be underestimated for the current generation. Of further clinical relevance, peak bone mass can continue to accrue in the third decade of life, therefore the skeletons of many of the current study participants may not have stopped growing (mean age 25.9 years). The effects of repetitive high trauma to the growing spine are unclear. Our sample was restricted to professional rugby players participating at the highest standard, and at this level, the intensity of collisions and other physical contact is likely to be considerably greater than at the amateur level. Our results are therefore not to be generalised to all rugby players. Although we did not test a control group, epidemiological data elsewhere has reported 1.0 to 4.0 incident VF cases per 10,000 population/year in men, 20–35 years of age. It was also not feasible to conduct spinal radiographs, however, the VFA method is widely recognised as advantageous for being accurate and precise, mostly for grade 2 and grade 3 vertebral fractures. There are no sensitivity data specific to young adult male athletes, however precision of 0.84% in men aged 38–87 years for VFA has been reported, and VFA has also been used in younger age groups with success. Due to the observational design of this study, we cannot assume cause and effect, or rule out other potential causes of VF linked to general risky behaviour. However, an increased risk for musculoskeletal trauma is inherent to contact sports. The health consequences of repetitive physical trauma, during muscular fatigue and over time, in professional rugby players remain unclear. Chronic back pain is one of the most common complaints of retired rugby players, although back pain is not always attributable to VF. Further studies should be prospective in design, and would benefit from the inclusion of an assessment of back pain scores in players. In conclusion, this study is the first to investigate and reveal a high number of vertebral fractures in professional male rugby union and rugby league players. This identified risk warrants exploration in prospective studies in order to quantify seasonal incidence, relative risk, and investigate possible aetiologies. We recommend that pre and post season vertebral fracture screening protocols for all professional players, are considered by rugby league and rugby union governing bodies and clubs. We also recommend the development of sport- specific vertebral fracture safe-management guidelines. Finally, the effects of contact sports on the development of the growing spine, and the short and long term consequences of VFs in professional rugby players, are unknown and represent additional avenues for timely, future research. [^1]: The authors have declared that no competing interests exist. [^2]: Analyzed the data: KH. Wrote the paper: KH BB FB.

# Introduction The scientific interest in the information-content hidden in the frequency statistics of words and letters in a text goes at least back to Islamic scholars in the ninth century. The first practical application of these early endeavors seems to have been the use of frequency statistics of letters to decipher cryptic messages. The more specific question of what *linguistic* information is hidden in the *shape* of the word-frequency distribution stems from the first part of the twentieth century when it was discovered that the words in a text typically have a broad “fat-tailed” shape, which often can be well approximated with a power law over a large range. This led to the empirical concept of Zipf’s law which states that the probability that a word occurs *k*-times in a text, *P*(*k*), is proportional to 1/*k*²\[–\]. The question is then what principle or property of a language causes this power law distribution of word- frequencies and this is still an ongoing research. In the middle of the twentieth century Simon in instead suggested that since quite a few completely different systems also seemed to follow Zipf’s law in their corresponding frequency distributions, the explanation of the law must be more general and stochastic in nature and hence independent of any specific information of the language itself. Instead he proposed a random stochastic growth model for a book written one word at a time from beginning to end. This became a very influential model and has served as a starting point for much later works. However, it was recently pointed out that the Simon-model has a fundamental flaw: the rare words in the text are more often to be found in the later part of the text, whereas a real text is to very good approximation translational invariant: the first half of a *real* text has, provided it is written by the same author, the same word- frequency distribution as the second. So, although the Simon-model is very general and contains a stochastic element, it is still history dependent and, in this sense, it leads to a less random frequency distribution than a real text. An extreme random model was proposed in the middle of the twentieth century by Miller in: the resulting text can be described as being produced by a monkey randomly typing away on a typewriter. The monkey book is definitely translational invariant, but its properties are quite unrealistic and different from a real text. The RGF (random group formation)-model, which is the basis for the present analysis, can be seen as a next step along Simon’s suggestion of system- independence. Instead of introducing randomness from a stochastic growth model, RGF introduces randomness directly from the maximum entropy principle. An important point of the RGF-theory is that it is predictive: if the only knowledge of the text is *M* (total number of words), *N* (number of distinct words), and *k*_*max* (number of repetitions of the most common word), then RGF provides a complete prediction of the probability distribution *P*(*k*). This prediction includes the functional form, which embraces Gaussian- like, exponential-like and power-law-like shapes; the form is determined by the sole knowledge of (*M*, *N*, *k*_*max*). A crucial point is that, if the maximum entropy principle, through RGF, gives a very good description of the data, then this implies that the values (*M*, *N*, *k*_*max*) incorporate all information contained in the distribution *P*(*k*), which makes the prediction neutral and void of more specific characteristic features. More specific text information is, from this view-point, associated with systematic deviations from the RGF-prediction. Texts sometimes deviate significantly from the empirical Zipf’s law and a substantial part of work has been devoted to explain such deviations. These explanations usually involve text- and language specific features. However, from the RGF point of view, such explanations appear rather redundant and arbitrary, whenever the RGF-prediction agrees with the data. This point of view has been further elucidated in for the case of species divided into taxa in biology. In a recent paper by L. Lü *et al.* it was pointed out that the character frequency-distribution for a text written in Chinese characters differs significantly from Zipf’s law, as had also been noticed earlier. Chinese characters carry specific meanings. For example, ‘huí’ and ‘$\text{ji}\bar{\text{a}}$’ are two Chinese characters carrying the elementary meanings of “return” and “home”, respectively. In general a Chinese character can also carry multiple meanings, where the relevant meaning has to be deduced from the context. A Chinese word corresponds to one, two or more characters, *e.g.* the two characters ‘huí’ and ‘$\text{ji}\bar{\text{a}}$’ can be combined into the Chinese word ‘huí,$\text{ji}\bar{\text{a}}$’ denoting the concept of “returning home”. Thus both Chinese characters and Chinese words carry meanings which can be single or multiple. Roughly a word in Chinese corresponds to about 1.5 characters on the average and typically more than 90% of the words in a novel are written with one or two characters, where about 50% of the words are written by one character and 40% with two. The remaining ones are made up of more than two Chinese characters. The Chinese character frequency distribution is illustrated in. The straight line in the figure is the Zipf’s law expectation. From a Zipf’s law perspective one might then be tempted to conclude that the deviations between the data and Zipf’s law have something to do specifically with the Chinese language or the representation in terms of Chinese characters, or perhaps a bit of both. However, the dashed curve in the figure is the RGF-prediction. This prediction is very close to the data, which suggests that beyond the three characteristic numbers (*M*, *N*, *k*_*max*) \[total number of Chinese characters, distinct characters, and the number of repetitions of the most common character\] there is *no* specifically Chinese feature, which can be extracted from the data. A crucial point for reaching our conclusions in the present paper is the distinction between a predictive model like RGF and conventional curve-fitting. This can be illustrated by : if your aim is to fit the lowest *k*-data points in (*e.g.* *k* = 1 to 10) with an *ad hoc* two parameter curve you can obviously do slightly better than the dashed curve in the. However, the dashed curve is a *prediction* solely based on the knowledge of the right-most point in (*k*_*max* = 747) and the average number of times a character is used (*M*/*N* = 11.5). RGF predicts where the data points in the interval *k* = 1−10 in should fall *without* any explicit a priori knowledge of their whereabouts and with very little knowledge of anything else. This is the crucial difference between a prediction from a model and a fitting procedure and this difference carries over into the different conclusions which can be drawn from the two procedures. Another illustration is the fact that although the data in cannot be described by a Zipf’s-line with slope -1, such a line can be fitted to the data over a narrow range somewhere in the middle. Such an *ad hoc* fitting has no predictive value. Specific information about the system may be reflected in deviations from the RGF-prediction. One such possible deviation is discussed. It is also suggested that the cause of this deviation is multiple meanings of Chinese characters. A statistical information based argument for this conclusion is presented together with an extended RGF-model. The method section starts with a brief recapitulation of the RGF-theory, as well as the Random Book Transformation, which allows for the analysis of sub-parts of the novels. Both these methods are used as starting points when analyzing the frequency distribution for two Chinese novels. The Chinese character-frequency distributions are compared to the corresponding word-frequency distributions for both novels, as well as for parts of the novels. The results from these comparisons lead to an information theory which makes it possible to approximately include the multiple meanings of Chinese characters. It is pointed out that the existence of words with multiple meanings isn’t a characteristic specific to Chinese, but a general feature of languages. The frequency distribution of the elementary entities of a written language (words or characters) is therefore influenced by the distribution of meanings over these entities, in a characteristic way. Conclusions are discussed in a last section. # Methods ## Random Group Formation The random group formation describes the general situation in which *M* objects are randomly grouped together into *N* groups. The simplest case is when the objects are denumerable. Then if you know *M* and *N* the most likely distribution of group sizes, *N*(*k*) (number of sizes with *k* objects), can be obtained by minimizing the information average *I*\[*N*(*k*)\] = *N*⁻¹∑*N*(*k*)ln(*kN*(*k*)) with respect to the functional form of *N*(*k*), subject to the two constraints that *N*⁻¹∑*N*(*k*)*k* = \< *k* \> = *M*/*N* and ∑*N*(*k*) = *N*. Note that the information to localize an object in one of the groups of size *k* is log₂(*kN*(*k*)) in bits and ln(*kN*(*k*)) in nats. Minimizing the average information *I*\[*N*(*k*)\] is equivalent to maximizing the entropy. Thus RGF is a way to apply the maximum entropy principle to this particular class of problems. The result of the simplest case is the prediction *N*(*k*) = *A*exp(−*bk*)/*k*. However, in more general cases there might be many additional constraints and in addition all the objects might not lend themselves to a simple denumerization. The point is that in many applications you *do* know that there must be additional constraints relative to the simplest case *but* you have no idea what they might be. The RGF-idea is then based on the observation that any deviation from the simplest case will be reflected in a change of the entropy *S*\[*N*(*k*)\] = −∑_*k* *N*(*k*)/*N*ln(*kN*(*k*)/*N*). This can then be taken into account by incorporating the actual value of the entropy *S* as an additional constraint in the minimizing of *I*\[*N*(*k*)\]. The resulting more general prediction then becomes *N*(*k*) = *A*exp(−*bk*)/*k*^*γ*. Thus RGF transforms the three values (*M*, *N*, *S*) into a complete prediction of the group-size distribution. This also means that the form of the distribution is determined by the values (*M*, *N*, *S*) and includes a Gaussian limit (when *γ* = (*M*/*N*)*b* and (*M*/*N*)²/*γ* is small), exponential (when *γ* = 0), power-law (when *b* = 0) and anything in between. In comparison with earlier work, one may note that the functional form *P*(*k*) = *A*exp(−*bk*)/*k*^*γ* has been used before when parameterizating distributions as described *e.g.* by Clauset *et al* and that such a functional form can obtained from a maximum entropy as described *e.g.* by Visser. The difference with our approach is the connection to minimal information which opens up the predictive part of the RGF. As emphasized in the Introduction, it is this predictive aspect which is crucial in the present approach and which lends itself to the generalization of including multiple meanings of characters. The RGF-distribution was in shown to apply to a variety of systems like words in texts, population in counties, family names, distribution of richness, distribution of species into taxa, node sizes in metabolic networks, etc. In case of words, *N* is the number of different words, *M* is the total number of words, and *N*(*k*) is the number of different words which appears *k* times in the text. In English the largest group consists of the word “the” and its occurrence in a text written by an author is a statistically very well defined: it is typically about 4% of the total number of words. As a consequence one may replace the three values (*M*, *N*, *S*) by the three values (*M*, *N*, *k*_*max*). Both choices completely determine the parameters (*A*, *b*, *γ*) in the RGF-prediction. However, the latter choice has the practical advantage that *k*_*max*, the number of repetitions of the most common word, is more directly accessible and statistically very well-defined. For example, if *k*_*max* is close to the average \< *k* \> = *M*/*N*, such that (*k*_*max*− \< *k* \>)/ \< *k* \> \< \< 1 then the RGF- prediction approaches a Gaussian, which comes as no surprise because a Gaussian is just the outcome of the maximum entropy principle for such a narrow distribution. ## Random Book Transformation In general, the distribution for a system, which falls into the RGF-class, has a distribution with a shape which depends on *M*. Since *M* for a text is the total number of words, this means that the frequency distribution is text-length dependent. The reason is that if you start from a text characterized by (*M*, *N*, *k*_*max*), then the corresponding value for a half of the text is characterized by (*M*_1/2, *N*_1/2, *k*_{*max*_1/2}). Here *M*_1/2 = *M*/2 by definition, *k*_{*max*_1/2} = \[*k*_*max*\]/2 because the most common word is to good approximation equally distributed within the text, but *N*_1/2 is non trivial. In the present investigation we need a method to separate between changes in the frequency distribution due to multiple meanings and due to the size of the text. For this purpose we use the Random Book Transformation (RBT) discussed in, where it was shown that the text- length dependence of the average *N*, when taking a part of a given text, is to good approximation a neutral feature: it is to good approximation the same as when you randomly delete the corresponding amount of words from the text. The process of changing the length of a text by randomly deleting words is a simple statistical process which transforms the probability distribution *P*_*M*(*k*) = *N*(*k*)/*N* for the full text into *P*_*M*/*n*(*k*) for the n^*th* part of the text by $$\begin{array}{r} {\mathbf{P}_{M/n}(k) = B\sum\limits_{k^{\prime} = k}^{M}\mathbf{A}_{kk^{\prime}}\mathbf{P}_{M}\left( k^{\prime} \right),} \\ \end{array}$$ where **P**_*M*/*n* and **P**_*M* are column matrices corresponding to *P*_*M*/*n* and *P*_*M*. The transformation matrix **A**_*k*′*k* is given by $$\begin{array}{r} {\mathbf{A}_{kk^{\prime}} = (n - 1)^{k^{\prime} - k}n^{k^{\prime}}C_{k}^{k^{\prime}},} \\ \end{array}$$ where $C_{k}^{k^{\prime}}$ is binomial coefficient. *B* is given by the normalization condition $$\begin{array}{r} {B^{- 1} = \sum\limits_{k}^{M}\sum\limits_{k^{\prime} = k}^{M}\mathbf{A}_{k^{\prime}k}\mathbf{P}_{M}\left( k^{\prime} \right).} \\ \end{array}$$ As shown in the next section, this simple random book transformation also to good approximation applies to text written in Chinese characters. # Results ## RGF and size transformation for Chinese texts shows that the data for the novel *A Q Zheng Zhuan* is well described by the neutral-model prediction provided by RGF. This implies that the frequency distribution of both words and characters is to large extent directly determined by the “state”-variable triple (*M*, *N*, *k*_*max*). At first sight this might appear surprising because the development of a spoken language and its written counterpart is a long and intricate process. However, in statistical physics this type of emergent simple properties from a complex system is well established. A well-known example is the ideal gas law *P* = *NT*/*V* which predicts the pressure, *P*, that an ideal gas inside a closed container exerts on the walls from the three “state”-variables (*N*, *V*, *T*), where *N* the number of gas particles, *V* is the volume of the container and *T* is the absolute temperature of the gas. Yet each gas particle follows its own deterministic trajectory including collisions with other particles and the walls. Since the number of particles is enormous it is in practice impossible to predict the outcome by deterministically following what happens in time to all the particles. The emergence of the simple ideal gas law stems from the fact that, with an enormous number of possibilities, the actual one is very likely to be close to the most likely outcome, assuming that all possibilities are equally likely. The basis for the maximum entropy principle in the present context is precisely the assumption that all distinct possibilities are equally likely. A crucial point is that, provided RGF does give a good description of the data, this means that it is the deviations between the data and the RGF-prediction which may carry interesting system-specific information. From this perspective Zipf’s law is just an approximation of the RGF *i.e.* the straight line in should be regarded as an approximation of the dashed curve. It follows that the deviation between Zipf’s law and the data does not reflect any characteristic property of the underlying system. Following this line of argument, it is essential to establish just how well the RGF does describe the data. gives such a quality test: if all that matters is the “state”-variables (*M*, *N*, *k*_*max*), then one could equally well translate the same novel from Chinese characters to words. As seen in, the word-frequency distribution for the novel *A Q Zheng Zhuan* is completely different from the character-frequency and also the “state”-variables are totally different (see for “state”-variables and RGF prediction values). Yet according to RGF the change in shape only depends on the value of the “state”-variables and not if they relate to characters or words. As seen from, RGF does indeed give a very good description in both cases. The translation of *A Q Zheng Zhuan* from characters to words is in itself an example of a deterministic process. Yet, as illustrated in, it is a complicated process in the sense that the resulting word-frequency distribution, through RGF, can be obtained to very good approximation without having any knowledge about the actual deterministic translation-process! This can again be viewed as a case when complexity results in simplicity. gives a second example for a longer novel, *Ping Fan De Shi Jie* by Lu Yao (about 40 times as many characters as *A Q Zheng Zhuan*, see). In this case the word-frequency is very well accounted for by RGF. Note that in this particular case the Zipf’s law prediction agrees very well with both the RGF-prediction and the data (Zipf’s law is a straight line with slope -2 in Fig). RGF also provides a reasonable approximation of the character-frequency, whereas Zipf’s law fails completely for this case. This is consistent with the interpretation that Zipf’s law is just an approximation of RGF; an approximation which sometimes works and sometimes does not. However, as will be argued below, the discernible deviation between RGF and the data may reflect some specific linguistic feature. As shown above, the shape of the frequency curve for a given text changes when translating between characters and words and this change is well accounted for by the RGF and the corresponding change in “state”-variables. This is quite similar to the change of shape when more generally translating a novel to different languages. This analogy is demonstrated on the basis of the Russian short story *The Man in a Case* by A. Chekhov and its translations into English words and Chinese characters. As shown in, the respective RGF-predictions match the corresponding frequency distributions very well. The same is true for the English novel *The Old Man and the Sea* by E. Hemmingway (compare). These findings confirm that the information contained in the triple (*M*, *N*, *k*_*max*) is sufficient to describe the frequency distribution of the fundamental entities of a written language, independent if those are words or characters in Chinese and irrespective of the underlying language. In order to gain further insight into what causes the difference in word- frequency and character-frequency of a text written in Chinese one can compare text-parts of different lengths from a given novel. As described in, text-parts of different length of a novel have different frequency distributions. For example if you start from *A Q Zheng Zhuan* and take an 10^*th*-part, then the shape changes, as shown in. According to RGF this new shape should now to good approximation be directly predicted from the new “state” $(M/10,N^{\prime},k_{max}^{\prime})$ (see for the precise values) As seen in this is to good approximation the case. As explained in **Methods** and can be verified from, $k_{max}^{\prime} \approx k_{max}/10$. One may then ask if the transformation from *N* to *N*′ involves some system specific feature. In order to check this one can compare the process of taking an *n*^*th*-part of a text with the process of randomly deleting characters until only a *n*^*th*-part of them remains. This latter process is a trivial statistical transformation described in **Methods** under the name RBT (Random- Book-Transformation). also shows the predicted frequency distribution obtained from the “state”-variable triple $(M^{\prime},N^{\prime},k_{max}^{\prime})$ *derived* from RBT and used as input in RGF. (The actual RBT-derived value for *N*′ is given in). The close agreements signal that the change of shape due to a reduction in text length, to large extent, is a general totally system- independent feature. shows the change of the frequency-distribution, when taking parts of the longer novel *Ping Fan De Shi Jie* written in characters and compares the parts with the RGF-prediction, as well as with the combined RGF+RBT-prediction. The conclusion is that the change of shape carries very little system specific information. By comparing Fig and, one notices that whereas RGF gives a very good account of the shorter novel *A Q Zheng Zhuan*, there appears to be some deviation for the longer novel *Ping Fan De Shi Jie*. In we compare a 40^*th* part of *Ping Fan De Shi Jie* with the full length of *A Q Zheng Zhuan*. As seen from the two texts have very closely the same character-frequency distribution. From the point of view of RGF, it would mean that the “state”-variables (*M*, *N*, *k*_*max*) are closely the same. This is indeed the case, as seen in and from the direct comparison with RGF in. *Ping Fan De Shi Jie* and its partitioning suggest a possible specific additional feature for written texts: a deviation from RGF for longer texts, which becomes negligible for shorter. In the following section we suggest what type of feature this might be. ## Systematic deviations, information loss and multiple meanings of words As suggested in the previous section, the clearly discernible deviation in between the character-frequency distribution for the data and the RGF-prediction in case of *Ping Fan De Shi Jie* could be a systematic difference. The cause of this deviation should then be such that it becomes almost undetectable for a 40^*th*-part of the same text, as seen in. We here propose that this deviation is caused by the specific linguistic feature that a written word can have more than one meaning. Let us start from an English alphabetic text. A word is then defined as a collection of letters partitioned by blanks (or other partitioning signs). Such a written word could then *within* the text have more than one meaning. Multiple meanings here means that a word in a dictionary is listed to have several meanings *i.e.* a written word may consists of a group of words with different meanings. We will call the members of these under-groups primary words. So in order to pick a distinct primary word, you first have to pick a written word and then one of its meanings within the text. It follows that the longer the text is, the larger the chance that several meanings of a written word appear in the text. Our explanation is based on an earlier proposed specific linguistic feature that a more frequently written word occurring in the text, has a tendency of having more meanings. This means that a written word which occurs *k* times in the text on the average consists of a larger number of primary words than a written word which occurs fewer times. Thus if the text consists of *N*(*k*) written words which occur *k* times in the text, then the average number of primary words is *N*_*P*(*k*) = *N*(*k*)*f*(*k*) where *f*(*k*) describes how the number of multiple meanings depend on the frequency of the written word. In the case of texts written with Chinese characters, it is, as explained the introduction, the characters are the elementary entities carrying individual meanings and hence play the role of words. It is possible to incorporate the concept of multiple meanings into a RGF-type formulation. The point to note is that the distributed entities are really the primary words/Chinese-characters and the information needed to localize a primary word/Chinese-character belonging to a written word/Chinese-character which occurs *k* times in the text is log₂(*kN*_*P*(*k*)) = log₂(*kN*(*k*)*f*(*k*)). We want to determine the distribution *N*(*k*) taking into account that the information lost, −log₂(*f*(*k*)), caused by the number of multiple multiple meanings (on the average) of a word which occurs *k* times in the text. It follows the information which then needs to be minimized in order to obtain the maximum entropy solution is the average of log₂(*kN*(*k*))−log₂(*f*(*k*)) or equivalently $$\begin{array}{r} {I\left\lbrack N(k) \right\rbrack = N^{- 1}\sum\limits_{k}N(k)\ln\left( kN(k)f^{- 1}(k) \right)} \\ \end{array}$$ and following the same steps as in **Methods** and this predicts the functional form $$\begin{array}{r} {P(k) = A^{{}^{\prime}}\frac{\exp( - bk)}{\left( kf^{- 1}(k) \right)^{\gamma^{\prime}}}.} \\ \end{array}$$ Basically the specific linguistic character is that *f*(*k*) is an increasing function and that *f*(*k* = 1) = 1, because a word which only occurs a single time in the text can only have one meaning within the text. The simplest approximation is then just a linear increase. gives some support for this supposition: the average frequency, $\bar{k}(f_{D})$, of Chinese characters in *Ping Fan De Shi Jie*, which have *f*_*D* dictionary meanings,is plotted against *f*_*D*. The plot shows that the $\bar{k}(f_{D})$ to fair approximation has a linear increase of the form $\bar{k} = f_{D}/c^{\prime} - 1/c^{\prime} + 1$ or equivalently $f_{D} = c^{\prime}\bar{k} + 1 - c^{\prime}$. corresponds to the full text and to a 40^*th* part. Note that the slope *c*′ changes with text size. This is easily understood: shortening the text is, as explained in the previous section, basically the same as randomly removing characters. This means that a character with a smaller *k* has a larger chance to be completely removed from the text than one with higher. But since the characters with higher frequency on average have a larger number of multiple meanings, this means that the resulting characters with low *k* will on average have more multiple meanings. Also note that the *dictionary* meanings and the meanings *within* a text is not the same; the former is larger than the latter, but the longer the text the more equal they become. However, it is reasonable to assume that also the number of meanings *within* a text follows a similar linear relationship. Next we make the further simplification by replacing the average $\bar{k}$ with just *k* *i.e.* we are ignoring the spread in frequency of characters having a specific number of meanings within the text. However, this approximation still catches the increase in meanings with frequency. We will take this linear increase as our ansatz and include a cut-off *k*_*c* for large *k*, since the most frequent Chinese characters has few multiple meanings. This is a general linguistic feature, the most frequent English words, “the”, has only one meaning. Thus we use the approximate ansatz *f*(*k*) ∝ *k*/(1+*k*/*k*_*c*). This approximation reduces the RGF functional form to $$\begin{array}{r} {P(k) = A^{{}^{\prime}}\frac{\exp( - bk)}{k^{\gamma}\left( 1 + \frac{1}{dk} \right)^{\gamma}},} \\ \end{array}$$ where *d* = 1/*k*_*c*. In addition to the “state”-variable triple (*N*, *M*, *k*_*max*) we should specify an a priori knowledge of *f*(*k*). The knowledge of this linguistic constraint is limited and enters through its *approximate* form *f*(*k*) ∝ *k*/(1+*kd*). This enables us to determine the value *d* = 1/*k*_*c* from the RFG-method by including the value of the entropy *S* as an additional constraint. Thus we use RGF in the form of together with the “state”-variable quadruple (*N*, *M*, *k*_*max*, *S*). This follows since the four constants (*A*′, *b*, *γ*, *d*) in, through RGF-formulation completely determine the quadruple (*N*, *M*, *k*_*max*, *S*) and vice versa. In this form of extended RGF is tested on data from three novels written in Chinese characters. The corresponding “state”-quadruples (*N*, *M*, *k*_*max*, *S*) are given in together with the corresponding predicted output-quadruple (*γ*, *b*, *k*_*max*, *d*). The agreement with the data is in all cases excellent (dashed curves in the). The dotted curves are the usual RGF-prediction based on the “state”-triples (*M*, *N*, *k*_*max*). Note that for a 100^*th*-part of *Ping Fan De Shi Jie*, the usual RGF and the extended RGF agrees equally well with the data. This means that any effect of multiple meanings is in this case already taken care of by the usual RGF. However as the text size is increased to 40^*th*-, 10^*th* part and full novel, the extended RGF agrees equally well, whereas the usual RGF-start to deviate. It is this systematic difference, which suggest that there is specific effect beyond the neutral-model prediction given by the usual RGF. Is the multiple meaning explanation sensible? To investigate this we estimate the average number of multiple meanings \< *f*(*k*) \> using the ansatz form for *f* including the condition that a single character can only have a single meaning in the text *f*(*k* = 1) = 1 *i.e.* *f*(*k*) = (1+*d*)*k*/(1+*kd*) together with the obtained values of *d* $$\begin{array}{r} {\sum\limits_{k = 1}N(k)f(k)N = \sum\limits_{k = 1}^{k_{c}}P(k)\frac{(1 + d)k}{1 + dk}.} \\ \end{array}$$ These estimated values for \< *f* \> are given in. shows that \< *f* \> increases with the text length. This is consistent with the fact that the number of uses of a character increases and hence the chance that more of its multiples meanings appears in the text. For the same reason \< *f* \> increases with the average number of uses of a character \< *k* \> as shown in. In addition the chance for a larger number of dictionary meanings is larger for a more frequent character. Thus it appears that the connection between \< *f* \> and multible meanings makes sense. Multiple meaning is of course not a unique feature of Chinese, it is a common feature of many languages. Therefore, it is unsurprising that we can also observe systematic deviations from the RGF-prediction in other languages, such as English and Russian. However, the average meaning of English words are much less than that of Chinese character: in modern Chinese there are only about 3,500 commonly used characters and even for a novel including more than one million of characters, the number of distinct characters involved is less than 4,000; but for the same novel written in English, the number of distinct words is more than 20,000. Therefore, the systematic deviation caused by multiple meaning can be neglected for short English text, as shown in. Even for a rather long text, the deviation is still very slight and, as shown in, the usual RGF gives a good prediction (RGF with multiple meaning constraint incorporates more *a priori* information and may consequently be expected to give a better prediction but the difference is very small). Taken together, Chinese uses a small amount of characters to describe the primary word, resulting in a high degree of multiple meanings, further leading to that the head of the character- frequency distribution (or tail of the frequency-rank distribution) deviates somewhat from the RGF-prediction. But such deviations are not special to Chinese, as we have demonstrated in, it is just more pronounced in Chinese than in some other languages. # Discussion The view taken in the present paper is somewhat different and heretical compared to a large body of earlier work. First of all we argue that Zipf’s law is not a good starting point, when trying to extract information from word/character frequency distributions. Our starting point is instead a neutral-model containing a minimal *a priori* information about the system. From this minimal information, the frequency distribution is predicted through a maximum entropy principle. The minimal information consists of the “state”-variable triple (*M*, *N*, *k*_*max*) corresponding to the *(total number of-, number of different-, maximum occurrence of most frequent-)* word/character, respectively. The shape of the distribution is entirely determined by the triple (*M*, *N*, *k*_*max*). Within this RGF-approach, Zipf’s-law (or any other power law with an exponent different from the Zipf’s law exponent) distribution only results for seemingly accidental triples of (*M*, *N*, *k*_*max*). The first question is then if these Zipf’s law triples are really accidental or if they carry some additional information about the system. According to our findings there is nothing special about these power-law cases. First of all in the examples discussed here, Zipf’s law is in most cases not a good approximation of the data, whereas the RGF-prediction in general gives a very good account of all the data *including* the rare cases when the distribution is close to a Zipf’s law. Second, translating a novel between languages, or between words and Chinese characters, or taking parts of the novel, all changes the triple (*M*, *N*, *k*_*max*). This means that the shape of the distribution changes, such that if it happened to be close to a Zipf’s law before the change, it deviates after. Furthermore, in the case of taking parts of a novel, the change in the triple (*M*, *N*, *k*_*max*) is to large extent trivial, which means that there is no subtle constraint for preferring special values of (*M*, *N*, *k*_*max*). All what this leads up to is that the distributions you find in word/character frequencies are very general and apply to any system which can be similarly described in terms of the triple (*M*, *N*, *k*_*max*) as discussed in. From this point of view the word/character frequency carries little specific information about languages. In a wider context, this generality and lack of system-dependence was also expressed in as: *…we can safely exclude the possibility that the processes that led to the distribution of avian species over families also wrote the United States’ declaration of independence, yet both are described by RGF*, and earlier and more drastically by Herbert Simon in: *No one supposes that there is any connection between horse-kicks suffered by soldiers in the German army and blood cells on a microscopic slide other than that the same urn scheme provides a satisfactory abstract model for both phenomena*. The urn scheme used in the present paper is the maximum entropy principle in the form of RGF. Herbert Simon’s own urn model is called the Simon model. The problem with the Simon model in the context of written text is that it does presume a specific relation between the parameters of the “state”-triple (*M*, *N*, *k*_*max*): for a text with a given *M* and *N*, the Simon model *predicts* a *k*_*max*. This value of *k*_*max* is quite different from the ones describing the real data analyzed here. For example in case of the “state” triple for *A Q Zheng Zhuan* in Chinese characters the values of *M* and *N* are 17,915 and 1,552, respectively and the Simon model predicts *k*_*max* = 9,256 and *P*(*k*) in the form of a power law given by ∝ 1/*k*^2.1. Thus the most common character accounts for about 50% of the total text, which does not correspond to any realistic language. compares this Simon model result with the real data, as well as with the corresponding RGF-predictions. You could perhaps imagine that you in each case could modify the Simon model so as to produce the correct “state”-triple. However, even so a modified Simon models will anyway have a serious problem, as discussed in: if you take a novel written by the Simon stochastic model and divide it into two equally sized parts, then the first part has a quite different triple (*M*/2, *N*_1/2, *k*_*max*/2) than the second. Yet both parts of a real book are described by the same “state”-variable triple. This means that the change in shape of the distribution by partitioning cannot be correctly described within any stochastic Simon-type model. From the point of view of the present approach, the fact that the data is very well described by the RGF-model gives a tentative handle to get one step further: since RGF is a neutral-model prediction, the implication is that any systematic deviations between the data and the RGF-prediction might carry additional specific information about the system. Such a deviation was shown to become more discernable the longer the text written in Chinese characters is. The multiple meaning of Chinese characters was suggested as an explanatory factor of this phenomenon. This is based on the notion that characters/words used with larger frequency have a tendency to have more multiple meanings within a text. Some supports for this was gained be comparing to the dictionary meanings of a Chinese character. It was also argued that this tendency of more multiple meanings could be entered as an additional constraint within the RGF- formulation. Comparison with data suggested that this is indeed a sensible contender for an explanation. Our view is that the neutral-model provided by RGF provides a useful starting point for extracting information from word/character distributions in texts. It has the advantage, compared to most other approaches, in that it actually predicts the real data from a very limited amount of *a priori* information. It also has the advantage of being a general approach which can be applied to a great variety of different systems. # Supporting Information Economic support from IceLab is gratefully acknowledged. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: PM XY. Performed the experiments: XY. Analyzed the data: XY. Wrote the paper: PM.

# Introduction Soccer is a high-intensity intermittent sport characterized by high variability within and between games. Various contextual factors such as match score, tactics, fixture time and location, environmental conditions, League ranking, championship or cup competition, travel demands proximal to competition, amongst others, can influence match-play characteristics and post-game psycho- physiological responses. These factors conspire to elicit psychological and physiological perturbations to homeostasis relative but not limited to, muscular, endocrine, and immune systems. The adaptations of autonomic nervous system (sympathetic and parasympathetic systems) and other body systems during and response to various stressors have been extensively explored. Adaptive changes to stressors can be classed as behavioural or physical, which can interact along with the brain and its peripheral components to simulate the hypothalamic-pituitary-adrenal axis (HPA) and the autonomic sympathetic system. These systems therefore have interrelated psychological (e.g., anxiety and mood state) and physiological (e.g., cortisol and testosterone responses) components (e.g., responses are psycho-physiological in nature). Consequently, monitoring the psycho-physiological response to training and competition, including the HPA axis, hypothalamic—pituitary—gonadal (HPG) axis and hypothalamic—pituitary—thyroid (HPT) axis has been seen within soccer. Such HPT/HPA data has been used to determine player preparedness for subsequent training and competition, and inform individualized recovery strategies. Circulating testosterone and cortisol levels fluctuate during stressors or exercises. As such, they have been used in an attempt to quantitatively determine the psycho-physiological stress/effort imposed by a soccer competition often alongside appropriate psychometric tools and/or inventories that enable researchers/practitioners to subjectively quantify the psychological status, such as the mood states \[the profile of mood states (POMS)\], the perception of exertion \[rating of perceived exertion (RPE:)\] and the anxiety states (anxiety rating). However, large intra- and inter-individual variability in testosterone and cortisol responses to soccer competition are seen, hence conflicting results within the literature. This variability has been attributed to a host of physical/physiological and/or cognitive factors, including but not limited to, match-play activity profile, match-outcome \[win or loss, type of contest (*i*.*e*., competitive *versus* non-competitive fixtures), competitive level, player coping style (psychological apparatus to deal with personal and public pressure), training status, player support network, and gender. Given the variability present, it appears logical to compile and subsequently systematically review the available evidence, to determine which factors are indeed moderator/mediator variables relative to the psycho-physiological responses to a soccer match-play. If psycho-physiological relationships are apparent (e.g., increased anxiety is related to match loss), it may proffer the opportunity for targeted intervention(s) by practitioners to favorably influence performance and/or recovery agendas. Therefore, the aim of the present systematic review and meta-analysis was to determine the effects of soccer match-play on steroid hormones (i.e., testosterone and cortisol) and psychosocial responses (mood state, competitive anxiety, psychological stress, social connectedness), identifying key moderator/mediator variables like match-outcome, gender, type of contest and competitive level. It is hypothesized that (1) match outcome, type of contest and competitive level may moderate the soccer match-hormonal responses relationship, and (2) cortisol and testosterone changes would be associated with changes in anxiety and mood state respectively. # Methods ## Search strategy The present systematic review and meta-analysis was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) guidelines. The systematic search was conducted using different databases, as recommended by the Cochrane Association, namely PubMed/MEDLINE, Scopus (Elsevier), SciVerse ScienceDirect (Elsevier), Institute for Scientific Information (ISI)/Web of Science (WoS), SPORTDiscus, ProQuest, Chemical Abstracts Service (CAS), the Directory of Open Access Journals (DOAJ), the Cochrane Database of Systematic Reviews (CDSR) of the Cochrane Library, the Cumulative Index to Nursing and Allied Health Literature (CINAHL), the Scientific Electronic Library Online (SciELO), and Google Scholar with dates ranging from the earliest record to April 2016. All study designs were included. The search terms included the following keywords: “soccer match”, “hormonal response”, “testosterone”, “cortisol”, and “stress”, connected using proper Boolean connectors and using Medical Subject Headings (MeSH) and wild-card options, when appropriate. Target journals have been hand-searched for capturing all potentially relevant studies. ## Inclusion and exclusion criteria Studies were included in the review if they met all the following Population/Intervention /Comparison/Outcome(s) (PICO) criteria: 1. *Population*: studies recruiting male and female novice and/or high- level soccer players as participants. 2. *Intervention or Exposure*: investigations studying the psychological changes over the duration of a soccer match and focusing on the hormonal responses using any hormonal measurements methods and collecting blood, urine or saliva samples. Salivary and serum derived hormone values were deemed equally valid, exhibiting strong positive correlations with one another. 3. *Comparison*: steroid hormones and psychological responses to a soccer match relative to match-outcome, gender, type of contest, and competitive level. 4. *Outcome(s)*: pre- to post-steroid hormones (i.e., testosterone and cortisol) changes to a soccer match-play and the correlation between the steroid hormones response and the psychosocial status. 5. *Design*: original investigations published in scholarly, peer-reviewed journals. 6. *Time filter*: from inception to April 2016. 7. *Language filter*: English. Studies were excluded according to the following criteria: 1. Reviews, comments, opinions and commentaries, interviews, letter to editor, editorial, posters, conference abstracts, book chapters, and books were excluded; available reviews have been anyways scanned for increasing the chance of including potentially relevant articles. 2. Focusing on methodological issues (for example, comparing endocrinological assays). 3. Comparing hormonal responses between footballers and lay people and not over soccer matches. 4. Assessing hormonal change after nutritional supplements or other kinds of intervention (interventional studies). 5. Lacking quantitative information and details. ## Screening strategy The studies have been independently screened by two authors (MS, NLB) looking at study titles and abstracts for potential eligibility. Screening questions have been *ad hoc* developed and pilot-tested with a subset of records before implementation. Disagreement has been assessed using κ statistics and has been resolved through discussion until consensus was reached; a third reviewer (JSB) and a forth reviewer (LT) have been involved when necessary. ## Review questions The review questions of the present review are the following: 1. What hormonal changes (i.e., testosterone and cortisol) can be induced by a soccer match-play? 2. Is there an association between steroid hormones responses and psychosocial status (e.g., mood state, anxiety, psychological stress and social connectedness) during a soccer match? 3. Which are the possible moderator and mediator variables? The research literature suggests many potential moderators/mediators of the hormonal changes-soccer match relationship such as: gender, match outcome, competitive levels and the nature of the competition itself, among others. Are there differences in steroid hormones responses between genders? Is there a difference between winners and losers in testosterone or cortisol responses to a soccer match? Is there a difference between novice and high- level players in testosterone or cortisol response to a soccer match? Is there a difference between competitiveand non-competitive fixtures matches in testosterone or cortisol response to a soccer match? Has the change of cortisol or testosterone levels to a soccer match been driven by the changes in psychosocial status? ## Statistical analysis ### Effect sizes Effect sizes (ES) were calculated with its 95% confidence interval (CI) according to Cohen and represent the difference between pre- to post- match only, means hormonal values divided by the baseline standard deviation. This method permits the determination of the magnitude of the differences or the changes between pre- to post- match for each study that provided absolute mean data and standard deviations. ES was interpreted with the following rule of thumb: ES \<0.2 was defined as trivial; 0.2–0.6 was defined as small; 0.6–1.2 was defined as moderate; 1.2–2.0 was defined as large; \>2.0 was defined as very large; and \>4.0 was defined as extremely large. A significance level of p \<0.05 was considered for all the analysis. ### Correlation Correlated t-tests (two-tailed), Pearson correlation coefficient, Spearman rank correlation coefficients and Pearson Product-Moment Correlation were used to determine the association (r and/or p values) between testosterone or cortisol levels and psychosocial variables in the included studies. Correlations were evaluated as follows: small (0.1–0.29), moderate (0.30–0.49), large (0.50–0.69), very large (0.70–0.89), nearly perfect (0.90–0.99), and perfect (1.0). ### Heterogeneity Statistical heterogeneity in our systematic review and meta-analysis was assessed using the Q and I² statistics. If the I² was \>50%, this was regarded as substantial heterogeneity. To identify sources of variation, further stratification was performed relative to the main characteristics of included studies, carrying out meta-regression analyses in order to quantitatively investigate the relationship between one or more covariates (moderators) at the study level and a dependent variable (that is to say, the effect size). In addition, for the sensitivity analyses, the stability of the pooled estimate with respect to each study was investigated by excluding individual studies from the analysis. # Results ## Study selection The search strategies yielded a preliminary pool of 921 possible papers. The full text of 39 articles were retrieved and assessed for eligibility against the inclusion criteria. After a careful review of their full texts, 22 articles were excluded with reason and the remaining 17 articles were eligible for inclusion in the current review. More specifically, 8 qualitative studies and 9 quantitative studies were noted. More specifically, qualitative study generates rich, detailed and valid process data that contribute to the in-depth understanding of a context. Quantitative study, on the other hand, generates reliable population-based and generalizable data that is suited to establishing cause-and-effect relationships. The main differences between quantitative and qualitative researches concern data sample, data collection, data analysis, and outcomes. From the quantitative studies, three interventions studied the testosterone and cortisol responses to a soccer match; two trials investigated the cortisol response; and one study assessed the testosterone response. ## Characteristics of included studies In total, 17 articles were identified and retained in the present research. The characteristics of the study population included novice (i.e., healthy participants) and high-level (i.e., elite, sub-elite, semiprofessional, professional, national) participants (Tables). In addition, 16 studies included high-level players as sample participants and one studies used novice players as sample subjects. The total number of participants included in this review was 333 (177 males, 130 females and 26 participants’ gender was not specified). Sample size ranged between 7 and 42, with age ranging from 8 to 31 years. All studies were characterized by a cross-sectional design and performed a pre- and post-match hormonal analysis (Tables). Participant’s characteristics, such as gender, competitive level and age were extracted and tabulated for each selected study (Tables). Testosterone and cortisol percentage change values (Δ%) were derived from the nmol/L post-match value relative to the pre-match value. ## Pooled effect-sizes The forest plot of cortisol \[nmol/L\] changes induced by soccer match-play is shown in. Fixed-Effects Model Pooled ES for cortisol \[nmol/L\] was 1.01 (\[95% CI -1.66/-0.36\], p = 0.002) when including Coelho et al. study’ and 0.67 (\[95% CI -1.01/-0.33\], p = 0.001) when removing Coelho et al. study’. There was heterogeneity (Q = 22.82, I² = 73.70, p = 0.001) when including Coelho et al. study’, however, there was no heterogeneity (Q = 4.74, I² = 0.00, p = 0.448) when removing Coelho et al. study’. The forest plot of testosterone \[nmol/L\] changes induced by soccer match-play is shown in. Fixed-effects model pooled ES for testosterone \[nmol/L\] was 0.46 (\[95% CI -1.43/0.51\], p = 0.35) when including Coelho et al. study’ and 0.36 (\[95% CI -1.73/1.02\], p = 0.60) when removing Coelho et al. study’. There was heterogeneity when including Coelho et al. study’ (Q = 18.42, I² = 83.72, p\<0.001;) and when removing Coelho et al. study’ (Q = 18.28, I² = 89.06, p\<0.001;). ## Potential moderator and mediator variables Male novice soccer match contest caused a large increase in cortisol levels compared to pre-match levels (Δ% = 44.36; ES = 3.73; p\<0.001), the magnitude of response was significantly higher (Q = 18.08, p\<0.001) than that seen in response to a high-level soccer match (Δ% = 32.16; ES = 0.57). Male players reported a lower percentage increase in cortisol levels (Δ% = 34.60; ES = 1.20) compared to females (Δ% = 162.7; ES = 0.98), even though not statistically significant. Additionally, the meta-regression analyses showed that the type of contest moderate a soccer match-cortisol response relationship (p \<0.001) (Tables). Only within competitive matches testosterone, regardless of gender and type of contest, demonstrated a small increase pre-to-post match (Δ% = 20.38; ES = 0.45), without statistical significance. Regardless of match contest, male players reported a lower pre- to post-match percentage increase in testosterone levels (Δ% = 6.26; ES = 0.28) than females (Δ% = 49.16; ES = 1.00), without statistical significance. A statistically significant moderator variable relationship was seen with respect to experience level of the players (p \<0.001, Tables). Finally, mood state and competitive anxiety resulted to be mediator variables of hormonal change in response to competition in soccer players. Cortisol changes were driven by changes in cognitive anxiety (very large correlation) for starter female soccer players, while testosterone changes were driven by changes in mood state (moderate correlation) for females and social connectedness (large correlation) for male soccer players. # Discussion The aim of the current systematic review and meta-analysis was to determine the hormonal (i.e., testosterone and cortisol) responses to a soccer match according to match outcome (i.e., win, loss), gender, type of contest (*i*.*e*., competitive vs. non-competitive fixtures) and competitive level. Testosterone response was found to vary according to the game outcome, with a larger response in winners compared to losers whereas cortisol concentrations did not vary with regard to match outcome. Competitive level may have moderated the cortisol response-soccer match relationship, with greater levels of cortisol reactivity in male novice compared to high-level soccer players. Thus, competitive soccer matches increased cortisol levels to a greater magnitude compared to non- competitive fixtures (i.e., collegiate tournament). Additionally, regardless of gender differences, higher testosterone reactivity in high-level compared to novice players was shown. When psycho-physiological stress was evaluated post- match, cortisol changes appeared to be driven by changes in cognitive anxiety, while testosterone changes were driven by changes in mood state and social connectedness. ## Moderator variables ### Match outcome and type of contest Match outcome moderated the testosterone level in response to a soccer match (higher in winners) within the present data, in agreement with earlier meta- analyses. This differential testosterone response between winners and losers proffers the opportunity for practitioners to employ a precompetitive or half- time cognitive intervention to enhance positive psychological states, such as the mood state and motivation to win, given they are related to game outcome. In contrast, cortisol increase appeared to follow general physical activity with no difference between winners and losers, yet competitive compared to non- competitive matches demonstrated greater cortisol reactivity. Competitive level differences on cortisol are likely due to the higher psycho-physiological effort during competitive matches than non-competitive fixtures, specifically mental toughness, self-confidence, aerobic and anaerobic capacities. Practitioners should therefore seek to make training sessions/matches as externally valid to competition as possible, to encourage players to develop robust coping apparatus with and familiarity to highly-competitive externally valid competition scenarios. ### Competitive levels It has been previously demonstrated that the competitive level is a moderator variable for the testosterone response to a soccer match, likely due to differences in players psycho-physiological competences and traits. Increasing expertise is associated with greater physical capacities and more robust psychological apparatus in soccer players. Indeed, high-level soccer players (high-level vs. ‘other’ standards of athlete) are well acknowledged as psycho- physiologically unique. Thus, it is worth noting that high-level players possess a high capability to cope with the game demands and stress as shown by the fast recovery pattern and lower performance impairments. On the basis of the available scientific investigations, it appears that distinguishing competitive levels in regards to the hormonal responses to a soccer match-play may be useful for trainers and coaches in the development processes. More specifically, practitioners should plan specific match/training session soliciting the endocrine system in novice soccer players, in order to simultaneously decrease the psycho-physiological stress and improve the recovery pattern and performances as high-level players, and adopt the best strategy to control the stress-recovery balance. ### Gender It seems that gender did not moderate the hormonal response to a soccer match- play. However, this might be attributable to the small sample size and number of included studies (two studies recruiting female players;) and so further research is warranted to determine effect of gender on hormonal responses during a soccer match. In contrast, a previous review showed that the stress response was different between males and females. The differences were attributed to (a) the higher hormone concentrations for men at rest (particularly testosterone levels) (b) higher psychological stress in men immediately before the test which represents a challenge and (c) the differences in body composition (e.g., higher fat mass in women) and/or in sexual hormonal status. ## Hormonal changes during the recovery period after a soccer match-play Regarding the hormonal responses during the recovery period, the competitive soccer match did not altered the testosterone level during the recovery period; however, plasma cortisol concentration significantly increased at 24 h and 48 h as compared to baseline, which returned to baseline after 72 h of recovery. Furthermore, cortisol and testosterone concentrations were found to be diminished after 24 (cortisol \~ -36%, testosterone \~ -25%) and 48 h post- simulated match (cortisol \~ -32%, testosterone \~ -30%). However, in line with the findings of Silva et al., the authors did not observe alterations in plasma testosterone (free testosterone) concentration throughout the 144 h of the recovery period. In summary, high-level non-competitive and competitive soccer matches led to an increased catabolic hormonal environment until 48 h into the recovery period. This finding suggests that an interval of at least 72–144 h between competitions should be considered by coaches working with high-level soccer players. ## Psychological mediators Alterations in the hormonal and psychological variables over a soccer game are often observed, suggesting that combined psychological and hormonal changes during competition could be useful to monitor the stress in relation to soccer match performance. The assessment of psycho-physiological stress after a soccer match showed contradictory results throughout different studies. The correlation between cortisol changes and psychological states over a soccer match were investigated in five studies. Some studies showed no significant correlations between changes in cortisol over the game and changes in mood state, anxiety state in female soccer players, whereas only one study demonstrated a significant correlation between cortisol changes and cognitive anxiety in starter female soccer players. Concerning testosterone changes after a soccer match, some studies showed significant correlations between testosterone changes and social connectedness in male soccer players and mood state changes in female players. In contrast, there was no significant correlation between testosterone changes, anxiety state and social connectedness changes in women players. For these reasons, coaches and scientists should be attentive to the role of social and environmental stress, personality, and other psychological metrics as measurable experimental variables that can influence data outcomes during a soccer match. These data also proffers the opportunity for practitioners to employ a precompetitive, or half-time, cognitive intervention that improves the participants’ mood state and consequently the participant’s chances of winning, as well as, manage the cognitive anxiety and psychosocial stress and consequently help cope both the physical and psychological demands as elite players. ## Limitations A number of limitations affecting both the primary data and the current systematic review and meta-analysis should be properly acknowledged. First, there was a considerable amount of small numbers of included studies, particularly in female and novice players. While this review could identify important moderators of soccer match-hormonal changes, it is possible that other factors that were not assessed could also explain the observed heterogeneity. For example, training programs and time of the day when competition was played can affect the hormonal stress affecting the chronobiological system. Casanova et al. observed a decrease in testosterone and cortisol levels (pre-to-post- matches), which might be explained by the circadian effect, the time of collecting samples, rather than the effect of the match *per se*. Furthermore, because of the cross-sectional design of included studies, the observed correlation between hormonal and psychological changes should not implicit as a causal relationship. # Conclusions This systematic review and meta-analysis provides readers with the first rigorous analytical synthesis of data concerning psychological and hormonal changes induced by soccer matches-play. In fact, the present review showed significant difference in the testosterone response to soccer games between winners and losers, with positive and negative changes in winners and losers, respectively. Furthermore, cortisol concentrations did not vary with regard to the contest outcome. Thus, it has been shown that testosterone reactivity was higher in high-level compared to novice players. Male novice soccer match contests increased cortisol levels to a greater magnitude compared to high-level soccer match. When psycho-physiological stress was evaluated after soccer matches, cortisol changes were found to be driven by changes in cognitive anxiety, while testosterone changes were driven by changes in mood state and social connectedness. The current review highlights that match outcome and competitive levels should be considered as the key moderator variables of the soccer match-hormonal changes relationship. A psycho-physiological assessment of soccer players could give sports coaches and managers the opportunity to (1) understand the processes involved in the stress response, (2) identify how an athlete copes with stress induced by a competition, (3) reduce and mitigate ‘stress’ response of players pre, post or between games, (4) modulating the training/play load according to the specific hormonal response and (5) to design and implement various *ad hoc* mental/recovery/coping strategies for performance enhancement and optimization. Particularly, positive reevaluation and active recovery should be recommended. # Supporting information The authors would like to declare that no sources of funding were used in the preparation of this review. They would also like to affirm that they have no conflict of interest that is directly or indirectly relevant to the content of the present review. [^1]: The authors have declared that no competing interests exist.

# Introduction Published cytogenetic comparisons clearly show size differences among the Y chromosomes of our nearest relatives, the chimpanzee (*Pan troglodytes*) and the bonobo (*Pan paniscus*). The chimpanzee Y chromosome is the smallest of the complement and is almost metacentric in morphology, while the bonobo Y is submetacentric, and similar in size to the G-group chromosomes of this species. The difference is attributable to a large early replicating euchromatic segment present in the proximal long arm of the bonobo Y, that is absent in the chimpanzee. Additionally, both the chimpanzee and the bonobo Y chromosomes exhibit a C-band positive heterochromatic segment at the tip of their short arms (Yp) – a characteristic feature of the distal long arm of the human Y chromosome (Yq). The pseudoautosomal region (PAR), together with the sex-determining region on the Y (*SRY*) located at the tip of human Yp, is located at the tip of Yq in chimpanzee and bonobo. Single-copy genes (X-degenerated genes), that map to human Yp and proximal Yq, and which survive as relicts from ancient autosomes from which the X and Y evolved, are shown to be conserved and arranged as single-copy genes along the distal half of the chimpanzee and bonobo Yq. The non-recombining ampliconic fertility genes *TSPY* and *RBMY* are shown to be highly amplified on the bonobo Y when compared to the chimpanzee Y, while these are significantly rearranged on the human Y,. The male-specific regions of the Y chromosome (MSY) of the human and the chimpanzee (*P. troglodytes*) are fully sequenced now. Comparison of the MSYs of the two species has shown dramatic rearrangements especially in the non-recombining parts that harbour ampliconic and repeated fertility genes. Comparable data from the MSY of the bonobo (*P. paniscus*) the common chimpanzee's closest relative are still missing. However, such data from bonobo as an outgroup species are essential to draw conclusions on the evolutionary fate of the Y chromosome in different primate species. Interestingly, a recent study showed intra-species variation in the copy number of the Y-specific ampliconic fertility gene *DAZ* within chimpanzee, but not bonobo. The restriction of *DAZ*-variation to chimpanzee prompted us to scrutinize the variability of the non-recombining part of the Y chromosome in a number of chimpanzees and bonobos more closely. Our results disclose a high intra-species variation in number and arrangement of ampliconic fertility genes among chimpanzee Y chromosomes while no variation was evident among bonobo Y chromosomes. # Results and Discussion Our focus was initially directed at the structural arrangement of the ampliconic fertility genes *DAZ* and *CDY*, both of which are expressed exclusively in the testis of human and chimpanzee. As a consequence, we mapped human-derived DNA probes specific for *DAZ* and *CDY* to metaphase Y chromosomes of 17 chimpanzees derived from 11 wild-born males, and 16 bonobos representing seven wild-born males by fluorescence *in situ* hybridization (FISH). Our results revealed highly diverse signal copy numbers and Y-chromosomal locations for *DAZ* and *CDY* genes for the chimpanzees. We detected ten Y-chromosomal variants among the 11 male chimpanzee lineages represented in our investigation. Only “Bobby” and “Tommy”, both wild-born chimpanzees, presented the same Y chromosome. Furthermore, two wild-born chimpanzees, “Max(1)” and “Moritz”, exhibited an identical pattern for *DAZ* and *CDY* but, importantly, the “Moritz” Y chromosome differed by a pericentromeric inversion as well as by the addition of a DAPI-positive segment on its long arm telomere distal to the PAR (see online ); these features are responsible for the submetacentric appearance of the Y chromosome in this specimen. A morphologically even more conspicuous Y chromosome variant was detected in “Max(2)” which exhibited strong FISH-signals for both *DAZ* and *CDY* in pericentromeric positions, and the presence of an additional *CDY*-signal on the subtelomeric short arm. Compared to the size of a “normal” chimpanzee Y chromosome, that of “Max(2)” showed a considerable increase in total length. To further cytogenetically dissect the subchromosomal structure of this chromosome we employed FISH analysis using DNA probes specific for ampliconic and X-degenerate genes on the Y chromosome. Our FISH results clearly showed that the Y chromosome of “Max(2)” exhibits a drastic increase of signals for *DUXY* sequences representing segmental duplications mapping in human Yq11.1/Yq11.21. In addition, a signal increase for ampliconic *TSPY* and *RBMY* genes was visible in the Y chromosome long arm of “Max(2)”. With the exception of *USP9Y* that maps close to the centromere, the X-degenerate genes maintain their expected location in the distal half of Yq proximal to the PAR. Although *USP9Y* is thought to be required for spermatogenesis in human males, it seems that *USP9Y* is dispensable in chimpanzees, as the two chimpanzee Y chromosomes sequenced both carry inactive forms of USP9Y,. Thus, the translocation of *USP9Y* close to the centromere on the Y chromosome of “Max(2)” may be of no relevance. The finished MSY sequence of the index specimen “Clint”, enabled us to schematically map the *DAZ* and *CDY* loci on “Clint's” Y chromosome providing yet another variant to our chimpanzee sample. In marked contrast to the situation in *P. troglodytes*, no variation in copy number or location was detected for either *DAZ* or *CDY* on the Y chromosomes of 16 male bonobo (*P. paniscus*) specimens, representing seven wild-born bonobos. In all animals investigated, *DAZ* and *CDY* map to Yp11.2 with mostly overlapping single FISH signals for both genes. The detection of a single signal for *DAZ* is consistent with the number of *DAZ* genes reported in 10 male bonobos. Additionally, no variation was found in the X-degenerate genes (proximal to distal) *USP9Y*, *DDX3Y* (formerly *DBY*), *UTY*, *KAL*, *AMELY*, *PRKY*, and most distally *SHOX* (the PAR-gene) that are arranged on distal Yq of the bonobos. These results are in agreement with the linear stability of X-degenerate genes for both bonobo and chimpanzee. By comparing the Y chromosome arrangement of MSY genes between 11 chimpanzee and seven bonobo male lineages, we conclude: (i) There is high variation for ampliconic fertility genes *DAZ* and *CDY* among the chimpanzees. Of 11 chimpanzee Y-chromosomal lines, 10 variants were detected. In contrast, the fertility genes on a morphologically stable submetacentric Y chromosome were invariant among seven bonobo Y-chromosomal lines. (ii) There is minimal variation for ampliconic genes *RBMY* and *TSPY* among both chimpanzees and bonobos. That said, all bonobo Y chromosomes share an amplification of *RBMY* sequences in their proximal long arm. (iii) X-degenerate single copy genes (with exception of *USP9Y*) show stable positions on all bonobo and chimpanzee Y chromosomes investigated. Although sample size is limited, especially in our invariant bonobo sample, it should be noted that the global captive bonobo population is derived from only 35 specimens –18 males and 17 females. Of these 18 males, two potential founders have still to reproduce. All founders and potential founders in this population are considered to be unrelated to each other since they originated from at least four distinct populations located east to west across the bonobos' range in the Democratic Republic of the Congo (DRC). Also, wild-caught bonobos do not appear to have come from severely inbred populations with high levels of individual homozygosity. The natural distribution of bonobos is confined to a single area - Cuvette Centrale in the centre of the Congo Basin in the DRC. This remote region of moist evergreen forests is encircled by the Congo River in the North, and the Kasai and Sankuru rivers in the South. In contrast, three to four subspecies, of chimpanzee are recognized, distributed across semideciduous forests in Central Africa. Given these data one could argue that the invariant bonobo Y chromosome in our sample is the result of a founder effect, while the Y-variants in chimpanzee simply reflect subspecific variation. The latter is considered unlikely given that the number of chimpanzee Y chromosome variants detected exceeds the recognition of geographic variants and both “Hans” and “Moritz” (wild-caught in West-Africa and are attributable to *P. t. verus*), present markedly different Y-variants. In other parts of the genome there is also considerable genetic variability within chimpanzee main subpopulations while, in contrast, bonobos are less polymorphic than each of the chimpanzee subpopulations. Thus, loss of genetic variation by founder effect or genetics drift must further be considered as a possible explanation in the bonobo. We posit rather that the geographic isolation by the Congo River probably permitted the establishment of different social systems in *P. troglodytes* and *P. paniscus* promoted, in part, by ecological and behavioural adaption. Male chimpanzees remain in their natal communities, and establish dominance hierarchies with a clear alpha male. Sexually receptive females show conspicuous periovulatory swellings and mate promiscuously –. Hormonal patterns indicate that they ovulate when they are maximally tumescent, and males can therefore monitor female receptivity. Under such conditions the opportunistic mating strategy in chimpanzee communities offers an opportunity for sperm competition in this species –. In contrast, the high social status of females in bonobo communities is unique to this species of *Pan*, which, coupled to concealed ovulation, could allow greater female choice, and this may act as an evolutionary counterstrategy that diminishes sexual selection via sperm competition. This view is further supported by the observation that relative to chimpanzees, adult bonobos show reduced sex dimorphism in both body size and the canine teeth. In addition, adult testosterone levels of male bonobos are much lower than those of adult chimpanzees,. As a consequence spermatogenic gene variation might be low in modern male bonobos. We conclude that, although chimpanzee and bonobo both show polyandrous mating behaviour with potentially high levels of sperm competition,, the contrasting patterns of Y-chromosomal variation in these closely related species might have an explanation in the context of their markedly different social structures. In chimpanzees, multiple males copulate with a receptive female during a short period of visible anogenital swelling, and this may place significant selection on fertility genes. In bonobos, however, female mate choice may make sperm competition redundant (leading to monomorphism of fertility genes), since ovulation in this species is concealed by the prolonged anogenital swelling, and because female bonobos can occupy high-ranking positions in the group and are thus able to determine mate choice more freely. We may speculate that the evolutionary history of a primate species Y chromosome is not simply encrypted in its DNA sequences but is shaped by social and behavioural circumstances. It is interesting that FISH studies in gorillas and orangutans similarly failed to detect intra-species variation in spermatogenesis genes (Greve et al., in preparation). It is apparent that monoandrous mating behaviour in gorillas, as well as the preference of female mate choice in orangutans , similarly diminishes sperm competition thus mirroring the situation in bonobo. # Material and Methods ## Blood samples Peripheral blood samples from all chimpanzee and bonobo specimens used in our studies were provided by zoo physicians. Details about the origin and status of the chimpanzee and bonobo specimens used in our study are presented in and. Pedigrees tracing the genealogy of the chimpanzee and bonobo specimens to the wild-born founders are shown in and. ## Chromosome preparations Chromosome preparations of all chimpanzee and bonobo specimens were made directly from peripheral blood lymphocytes – only in the case of the chimpanzee “Sascha” from a lymphoblastoid cell line established from peripheral blood in our lab – according to standard methods with minor modifications. Slides carrying interphase cells and metaphase spreads were dehydrated in a series of ice-cold ethanol (70%, 90% and 100% each for 3 min) then air dried and stored at −80°C. Before using for *in situ* hybridization, the slides were dehydrated again (70%, 90% and 100% each for 3 min) and then air dried. ## FISH analysis All FISH-assays were performed on metaphase and prometaphase spreads following Schempp et al. Prior to FISH, the slides were treated with RNase followed by pepsin digestion as described. Chromosome *in situ* suppression (CISS) was applied to gene clones listed in. For two-color detection, double-hybridization experiments were performed with biotinylated and digoxigenin (DIG)-labeled probes. Biotinylated probes were detected with FITC-conjugated avidin and DIG- labeled probes with anti-DIG-mouse antibodies (Sigma) followed by TRITC- conjugated goat anti-mouse antibodies (Sigma). After FISH the slides were counterstained with DAPI (4′,6-doamidino-2-phenolindole; 0.14 μg/ml) and mounted in Vectashield (Vector Laboratories). Preparations were evaluated using a Zeiss Axiophot epifluorescence microscope equipped with single-bandpass filters for excitation of red, green, and blue (Chroma Technologies, Brattleboro, VT). During exposures, only excitation filters were changed allowing for pixel-shift- free image recording. Images of high magnification and resolution were obtained using a black-and-white CCD camera (Photometrics Kodak KAF 1400; Kodak, Tucson, AZ) connected to the Axiophot. Camera control and digital image acquisition involved the use of an Apple Macintosh Quadra 950 computer. # Supporting Information We thank T. Robinson for comments on the manuscript and stylistic revisions, and I. Malheiro and S. Kirsch for discussions in the initial phase of the project. We also thank several zoo physicians for providing blood samples from chimpanzee and bonobo specimens over the years. [^1]: Conceived and designed the experiments: WS. Performed the experiments: FS AMF CH CM JJP. Analyzed the data: FS AMF CH CM JJP WS. Contributed reagents/materials/analysis tools: WR. Wrote the paper: WS. Supervised chimpanzee and bonobo paternity ascertainment: WR. [^2]: Current address: Developmental Biology, Institute of Biology 1, University Freiburg, Freiburg, Germany [^3]: Current address: Institute of Pathology, University Freiburg, Freiburg, Germany [^4]: The authors have declared that no competing interests exist.

# Introduction Terrestrial systems have generally shown an increase in species numbers from the poles to the tropics. Likewise, early marine studies confirmed this trend, and discussed its ecological implications for the marine environment. Since these early studies, others have examined various aspects of latitudinal biodiversity gradients in marine systems, although with varying results, suggesting that while such trends may be general they are not ubiquitous. Similar to these general latitudinal studies, some studies have focused on macroalgal biodiversity patterns along latitudinal gradients. Early studies on macroalgae suggested that there is no evidence of a latitudinal trend of increasing species numbers towards the tropics. In fact, areas of both low and high species richness have been identified at sites throughout temperate and tropical waters. Studies since this early work have reported varying results such as increased species richness at mid latitudes and also towards the equator or decreasing species richness towards the equator. A recent literature review covering 387 sites throughout the Atlantic, Indo-Pacific, and Southern Oceans, which spanned 140° of latitude found that in general, temperate oceans tended to have the highest numbers (350–450) of macroalgal genera, particularly between 110° and 160°E longitude. Interestingly, Kerswell also generally found that the number of algal genera had distinct hotspots, namely around Japan and southern Australia. Other studies have identified hotspots in the Mediterranean, the Philippines, the Pacific coast of North America, the Atlantic European coast, and the Caribbean. The current belief is that while lower species richness occurs at the poles, macroalgae generally exhibit variable species richness patterns in different areas. Sometimes these patterns show an increase with latitude, sometimes they decrease, or sometimes they peak at mid-latitudes. Previous studies on macroalgae have focused on alpha diversity, which examines macroalgal richness within the full extent of a single community, typically homogenizing the various depth and/or intertidal height strata. However, when examining nearshore latitudinal gradients, it is important to consider the intertidal height or water depth from where the samples are taken. This is particularly important for point diversity studies, which focus on a predetermined subset of species from the total site. Since macroalgal species typically occupy particular locations (strata) along a latitudinal gradient, point diversity studies must standardize the strata and the sampling design from which the samples are taken. For example, south-western Iceland, southern Alaska, and the Magellanic region all have recognizable species depth distribution patterns with species diversity increasing seaward in the intertidal. Hence, point diversity samples taken from the high intertidal stratum in one region are not comparable to samples from the low stratum of another region or even of the same region. Similarly, across the Gulf of Alaska, both species richness and abundance/biomass displayed depth strata-related patterns depending on the taxon group being examined. Because of this, it is important to keep intertidal heights and water depths consistent during latitudinal gradient analyses for point diversity. Most of the previous work on spatial patterns of macroalgal diversity mentioned above was based on non-structured meta-analyses of the existing literature and available species lists. One of the main problems with these previous types of studies is that methodological problems may obscure or artificially impose spatial trends. A potentially more powerful analysis to examine latitudinal gradients would be based on standardized sampling protocols to avoid any biases introduced by varying collection methods. This also would improve diversity- biomass comparisons if data were obtained from the same samples. Using a standardized protocol, however, does introduce its own issues. This is because sites are not similar as far as size of the dominant species or the overall dispersion of the various species. Hence, it would be difficult if not impossible to obtain a true measure of alpha diversity for a number of sites using a standardized protocol. However, point diversity lends itself very well to the use of a standardized protocol because it is only examining a subset of the overall richness within each site. Another aspect of macroalgal community organization that is sometimes explored is the relationship between different diversity-related attributes. Specifically, the importance of biomass in predicting species richness has been examined. While the Engelhardt and Ritchie study found higher algal and macrophyte biomass in mesocosms associated with a greater macrophyte species richness, Gough et al. showed that environmental variables explained much more of the variation in potential species richness than biomass. However, when sites exposed to extreme environmental conditions were eliminated from the analysis, biomass became the primary predictor of realized richness. In Portugal, macroalgal species richness was found to be significantly correlated with total biomass on intertidal boulders. An inverse relationship was found in South Africa where high algal biomass and low species richness along the cool and warm temperate region of the coastline was linked to upwelling activity and wave action indices. This influence of upwelling on macroalgal biomass has been described elsewhere. Conversely, low algal biomass and high species richness has been attributed to warmer immersion and emersion temperatures along the sub- tropical region of the coastline. While these latter studies have provided some regional knowledge, there have been few larger scale studies on macroalgal community organization and diversity-related attributes to make any general statements regarding diversity/biomass relationships. In the current paper, we explore the relationships between macroalgal taxon numbers and their respective biomass with latitude by depth strata using a standardized sampling design. We also determine if correlations exist between the number of macroalgal taxa and biomass with all depths pooled together. We hypothesize that similar to studies on alpha diversity, species density, as a proxy for point diversity, and macroalgal biomass will show latitudinal trends with higher numbers in mid latitudes. We also hypothesize that using a standardized protocol where species richness data are taken from the same samples as biomass data, we will find that, similar to others, macroalgal species richness will be correlated with total biomass. # Methods Macroalgal communities were sampled at 69 rocky substrate sites from approximately 10°N to 60°N latitude (Supplementary). Sites were primarily sampled between 2005 and 2009, except in Alaska, USA, where some sites were sampled in 2003. Although a balanced distribution was attempted, not all regions were sampled equally and in many regions sites were spatially clumped. This was an artifact of the location where researchers involved in this program were based. Several important regions, such as Asia were not adequately sampled, while others, such as Alaska were heavily sampled. Species richness is defined in this paper as point diversity or species density, where richness describes a subset of the community. The use of a standardized protocol is an adequate tool for point diversity comparisons but does not collect absolute site species richness (alpha diversity). For the purposes of this study, we wanted sample numbers and sizes to be equal for our comparisons. All sites were sampled when diversity was thought to be highest for that site (i.e. when annual species were present). Most sites had similar structure with a canopy and understory cover accompanied by algal turf. The standardized protocol used in this study was developed during a workshop for the Natural Geography In Shore Areas (NaGISA) program within the Census of Marine Life initiative. The NaGISA protocol uses a stratified random sampling design at each site in which five replicate random samples are taken along a 30–50 m horizontal transect at the high, mid, and low intertidal strata and 1 m, 5 m, and 10 m below MLLW. Five samples were deemed the best compromise between sufficient replication and practicality of sampling multiple depth strata at each site, especially when the focus of the comparison is point diversity and not alpha diversity. Intertidal heights were determined based on prevailing biobands for that region, such as barnacles, red algae, and brown algae that often typify the high, mid, and low zone, respectively. Not all strata were sampled at all sites because some sites did not have all strata. For example, only the 5 m depth stratum was sampled in the Arctic Beaufort Sea as this is the only depth with hard substrate for macroalgal growth. At each stratum at every site, all macroalgae were removed from within five 50×50 cm quadrats along a horizontal transect line following the stratum. Algae were sorted to the lowest taxonomic level (usually species) and their wet weights determined by taxon using an analytical scale with 1g precision. Taxonomic affinities were verified using the AlgaeBase web site ([www.algaebase.org](http://www.algaebase.org)). All encrusting algae were excluded from this study because they could not be completely cleared from the substrate. Data for the five replicate quadrats per stratum were averaged at each site. Macroalgal assemblages were graphically presented with all strata combined to illustrate general latitudinal trends. Pearson correlations were completed on species numbers and biomass by latitude for the northern hemisphere using StatView (v5.0.1, SAS Institute Inc.). # Results A total of 629 macroalgal species, or higher taxonomic affiliations, were identified during this study. When all sites were combined for each stratum, generally the greatest numbers of taxa were found at the 1 m subtidal depth, with taxon richness decreasing farther into the intertidal and deeper subtidal. In the intertidal, fewer taxa were found in the high than in the low stratum. In the subtidal, there were no noticeable differences between the 5 and 10 m water depths. Similar to taxon richness, the greatest macroalgal biomass was found at the 1 m intertidal height with biomass decreasing into the intertidal and subtidal strata. However, unlike taxon richness, biomass differences were not observed among intertidal heights or subtidal depths, although a slight trend of decreasing biomass with increasing depth was observed in the subtidal. In general, biomass was generally greater in the subtidal than it was in the intertidal. When all strata per site were pooled for a single analysis, significant correlations were not found between latitude and either average taxon numbers or biomass per quadrat (r = 0.27, p = 0.32, n = 176 and r = 0.32, p = 0.19, n = 176 for taxon numbers and biomass, respectively;). However, there was a slight trend for both taxa number and biomass to increase at mid latitudes, particularly between 45 to 60 N°. When strata per site were analyzed separately, the highest taxon numbers were typically found at higher latitudes for most strata, specifically around 60°N except at 5 and 10 m where some high values also were seen at around 25°N. Significant positive correlations in latitudinal trends were found for all three intertidal strata and at 1 m, but not for other subtidal strata. Some of the highest r values were found in this analysis, with 0.79 and 0.70 in the mid and 1 m strata, respectively. Overall, highest macroalgal biomass were found at some sites in the high, 1 m, 5 m, and 10 m strata with upwards of 5400 g/0.25 m² at 5 m depth. These high biomass sites were generally around 57°N and occasionally around 45° N. Similar to the number of taxa, biomass in the mid and low strata had significant positive correlations with latitude. Overall, r values were relatively low, with 0.46 and 0.52 in the mid and low strata, respectively. Taxon numbers were not correlated with macroalgal biomass (Pearson correlation r = 0.34, p = 0.13, n = 176;). Interestingly, the site (in Alaska, USA) with the greatest average biomass (5345 g/0.25 m²) was found with an average of 5.8 taxa/0.25 m², while the site (in Portugal) with the most taxa (an average of 29 species/0.25 m²) averaged only 76 g/0.25 m² of biomass. Overall the sites with the greatest biomass all had less than ten taxa. # Discussion It is difficult to make generalizations about biodiversity in natural systems because of their inherent spatial and temporal variation. However, if generalizations can be proposed, a better understanding of processes and underlying causes may result. This study presents some generalizations regarding macroalgal taxon numbers and biomass along various depth and latitudinal gradients. This paper differs from others in that it examines species density as a proxy for point diversity using a standardized protocol rather than the more typical alpha or beta diversity. It also scrutinized depth strata separately rather than just concurrently examining species richness in the entire nearshore zone at a given site. One important generalization found in this study was that mean taxon numbers and mean biomass were greatest at the 1 m depth stratum, with lower numbers in the intertidal and deeper subtidal. Similar trends have been seen in eastern Canada, where macroalgal species numbers were negatively correlated with elevation, with fewer species in the higher zones. In the Gulf of Alaska, the macroalgal taxon number also was generally higher at 1 m depth and decreased towards shallower and deeper depths. Although this appears to be a common trend, variation does exist. For example, while macroalgal taxon numbers were greatest in the low intertidal at Kodiak Island (Alaska USA), they were highest at the 5 m stratum in neighboring Prince William Sound just 500 km away. Other local or regional studies examining macroalgal biomass with depth have found similar results to this larger scale study. Macroalgal biomass in the Gulf of Alaska was generally more abundant at the 1 m stratum and decreased with increasing intertidal height and subtidal depth, although study site variation was evident. In Iceland, macroalgal biomass increased seawards from the high intertidal, and in California, macroalgal biomass decreased with increasing subtidal depth. Explanations for the high richness and biomass at 1m depth may be related to the special conditions at the interface between the intertidal and the subtidal. On the one hand, since the 1 m stratum is typically only exposed at extreme low tides, it does not experience the harsh conditions that the shallower intertidal strata are subjected to, e.g. desiccation, freezing, and heat, which may lead to lower species richness and biomass in the intertidal. On the other hand, the 1 m stratum experiences higher light conditions than are common at deeper depths and may be less structured by herbivores than the subtidal. This likely optimizes the overall conditions at the 1 m depth stratum for macroalgae, with variations to this pattern based on locally different conditions. Another general finding was that the number of taxa and average biomass per site decreased in the northern hemisphere from higher to lower latitudes. Peaks were found in the mid latitudes around 45–60°N, with a sharp drop at 70°N in the Arctic (only at the 5 m depth stratum). The only depth stratum that we were able to sample in the high Arctic was 5 m. The drop in taxon numbers at this depth confirms the general observation that macroalgal species richness decreases at the poles. Our observations also support our first hypothesis, that similar to studies on alpha diversity, point diversity measured as species density, and macroalgal biomass show latitudinal trends with higher numbers in mid latitudes. This contrasts to a study completed thirty years ago that found large peaks in macroalgal species numbers at 20°N and a smaller peak at 48°N, but no real trend going from north to south. More recently, Kerswell's study on macroalgae found no difference in genus numbers along a global latitudinal gradient but did find hotspots at various latitudes. Other more regional studies have been completed and resulted in various types of trends, including increased species numbers with latitude on the west coast of South Africa and the temperate regions of the Pacific South America, decreased species numbers on the east coast of South Africa and the Atlantic coast of Europe, and mid latitude peaks in the North and Central Americas. Some of the differences among these studies may be attributable to site selection, analyses (alpha versus point diversity), or methodology (literature searches and existing databases in past studies versus a standardized protocol in the present study). Some known northern hemisphere algal diversity hotspots, such as Japan, the Mediterranean, the Philippines, the Atlantic European coast, and the Caribbean, were not extensively sampled in the current study. While these latter regions may be actual hotspots attributed to drivers such as sea surface temperature, upwellings, disturbance, spatial heterogeneity, or species interactions it could also be that they are simply regions that are better studied. The use of a standardized protocol allowed us to examine species density separately for each intertidal and subtidal stratum. This analysis showed that there was an increase in taxon number and biomass with latitude in the mid and low zones. There was also an increase in taxon number alone in the high and 1 m strata. The lack of any trends in the subtidal compared to the intertidal strata may be due to the more benign physical conditions in the subtidal than the intertidal. Temperature extremes are greater in the intertidal, as are the problems associated with desiccation. Abiotically generated stressors such as temperature and desiccation typically occur in more unfavorable habitats such as the upper intertidal. It may be that disturbances and other harsh conditions that cause sudden mortality also increase species diversity, although in some circumstances, they may also reduce species diversity. The more benign conditions found in the subtidal may also play a role in reducing taxon number and biomass variation with depth. It is known that in general, abiotically generated stressors decrease in more favorable environments, such as increasing water cover. However, while abiotic stressors decrease, biotic stresses increase in these more abiotically favorable environments, resulting in competitive exclusion. In this study, more differences may have been found between the mid and low intertidal zones than between 5 and 10 m because of the associated environmental stressors. In terrestrial communities, species richness is related to community biomass in a “hump-shaped” fashion, suggesting two different drivers. At very low biomass, richness is probably limited by abiotic factors causing low survivorship. At very high biomass, it is thought that competitive exclusion may reduce species richness. Some marine studies have found that macroalgal biomass is positively correlated with species richness, while others have had conflicting results. The present study found that taxon numbers and biomass were not positively correlated. This does not support our second hypothesis, that similar to other studies, macroalgal species richness is correlated with total biomass. In this study, many sites with very high biomass had very low taxon numbers while other sites with very low biomass had very high taxon numbers and equally, many sites had both very low biomass and very low taxon numbers. The drivers of these relationships are unknown and need to be further investigated. While the data in this study have limitations, primarily related to small sample sizes and unevenly distributed sites, they have demonstrated that there is a common trend of more taxa and more biomass at mid latitudes in the northern hemisphere, particularly for the intertidal strata. The use of the standardized protocols probably eliminated some of the biases associated with sampling sites using different methodologies making this study powerful in regards to equal effort. The use of species density as a proxy for point diversity was helpful in that it allowed for the comparison of depth strata, rather than looking at the typical all site alpha diversity. # Supporting Information We thank Sandra Lindstrom (University of British Columbia) for supplying data for this project and providing comments on an earlier draft. We also thank Angela Mead (University of Cape Town), Bernabé Santelices (Pontifica Universidad Catolica de Chile), and one anonymous reviewer for their insightful comments on a previous draft of the manuscript. The authors acknowledge the Census of Marine Life and Jesse Ausubel for their leadership and Ron O'Dor for his help and support of NaGISA over the years. [^1]: Conceived and designed the experiments: BK KI JJCM LBC GP PM EK YS. Performed the experiments: BK KI JJCM LBC ALK GP PM ME TT EK RRR MW SJ AS ISP YS. Analyzed the data: BK. Contributed reagents/materials/analysis tools: BK KI JJCM LBC ALK GP PM ME TT EK RRR MW SJ AS ISP YS. Wrote the paper: BK KI JJCM LBC ALK GP PM ME TT EK RRR MW SJ AS ISP YS. [^2]: The authors have declared that no competing interests exist.

# Introduction Gingivo buccal squamous cell carcinoma (GBSCC) is one of the most prevalent (∼60%) cancers in oral cavity, especially among the tobacco users in India. Entire set of head and neck squamous cell carcinoma stands as the fifth most common malignancy worldwide. But head and neck cancer comprises ∼24% of total cancers in India as recorded at a tertiary hospital, Tata Memorial Centre, Mumbai and about ∼13.5% of them are from the oral cavity. Five years survival rate (∼50%) of patients suffering from head and neck squamous cell carcinoma has not improved much even after intense research during last 15 years, so, early detection is still a key issue for better survival. Since 1993, miRNA has emerged to be one of the most prominent biological regulators, which play important roles in controlling and fine tuning its target's (mRNA) expression. In recent years, it has been demonstrated that microRNAs (miRNAs) are also involved in human tumorigenesis and could act either as tumorigenic/oncogenic or anti-tumorigenic molecules. Thus, it is making a new layer in the molecular events in human cancer. Gene expression studies revealed that many miRNAs are deregulated in different cancer types and functional studies clarified that miRNAs are involved in several molecular and biological processes that drive tumorigenesis. Thus, in addition to different scales of variability in terms of new mutations in genome or genetic background of the patients (i.e. germ line mutation), variability in expression of different miRNAs is also a major aspect for investigation. With this belief, we studied expression deregulation of 762 miRNAs in GBSCC and also checked whether similar kind of expression deregulation could be observed in precancerous leukoplakia and lichen planus tissues from oral cavity. Efforts have been made to understand how these miRNAs may be involved in cancer using pathway analysis and information from lit*e*ratures and databases. # Materials and Methods ## Samples This study was approved by “Review committee for protection of research risk to humans, Indian Statistical Institute”. All individuals in this study have provided written informed consents to publish case details. All participants signed a questionnaire containing demography, tobacco habit and a statement describing that he/she has no objection for use of blood and tissue samples in this study and is participating in this study voluntarily. Unrelated patients suffering from cancer (n = 18), lichen planus (n = 12) and leukoplakia (n = 18) were considered for this study from Guru Nanak Institute of Dental Science and Research, Panihati, Kolkata. One tissue punch from cancer/precancer site and another punch from adjacent “clinically” normal site (at least 1.5 inch away from the border of the lesion) were biopsied by oral pathologist. One portion of biopsy tissues was used for histopathological confirmation and remaining part of tissues were kept separately in “RNA Later” at −80°C and used within 2 months. ## RNA isolation and TLDA assay RNA isolation was done using mirVana kit (Life Technologies, USA). RNA yield quantification was done by Nano Drop and integrity was checked with Agilient's Bioanalyzer (Agilent RNA6000 Nano Kit), respectively. RIN number cut-off was chosen as ≥6.9. Parallel expression assay of 762 miRNAs was performed using TLDA-A (V2) and TLDA-B (V3) card in 7900HT FAST Real Time PCR system (Applied Biosystems, USA) using TLDA flat block. Primary analysis was done using SDS and Data Assist (Life Technologies, USA) software packages to get expression in terms of Ct, ΔCt and ΔΔCt values where Ct =  Cycles at which the PCR product quantity reaches a defined threshold, ΔCt =  Ct _{of a miRNA in cancer tissue} - Ct _{of geometric mean of expression of 3 most stable endogenous control miRNAs in that tissue} and ΔΔCt =  ΔCt _{of a miRNA in cancer tissue} - ΔCt _{of that miRNA in control tissue}. Out of the assayed 762 miRNAs, five were candidates for endogenous controls. However on the basis of their expression stability across all samples, RNU-48, RNU-44 and U6/mmu-6 were selected as endogenous controls. To get ΔCt as the normalized measure of expression of a miRNA in both cancer/precancer and control tissues, expression of all miRNAs in tissues was normalized independently using geometric mean of expression of selected 3 endogenous control miRNAs in the same tissue. Ct value less than 40 were only considered for further analysis. ## Analysis ### Data Pruning If expression of any particular miRNA was observed to be present in at-least 9 out of 18 cancer-normal paired tissues, then only, it has been considered for further downstream analysis. Annotation of TLDA assay V2 (A-card) and V3 (B-card) follows miRBase release-14. Now, expression data of those miRNAs were considered for further analysis whose annotation is still valid in miRBase release-19. In this way expression of 531 miRNAs was considered for subsequent down-stream analysis. ### Statistical Analysis Initially, one sample Kolmogorov-Smirnov (KS) test for expression data of all 531 genes was performed independently so as to check for normality of the differential expression values across all samples \[(Zi =  {(Ci-Ni)-(mean C- mean N)}, so Zp =  ∑Zi/SD Z, KS test was done on Zp; Ci: i^th Cancer sample's ΔCt value, Ni: i^th Normal sample's ΔCt value\] and eventually expression was found to be normally distributed. It is to be noted that ΔCt values are already log transformed, so, no further log transformation was done with this set of expression data. Followed by the normality test, one tailed paired t-test was performed \[(Ci-Ni)\>1/\<1\]. Null hypothesis of one tailed paired t-test was “expression of a particular miRNA is not greater than 2 fold up/down regulated”. So naturally, alternative hypothesis was “expression of a particular miRNA is \>2 fold up/down regulated. Test for up-regulation (lower tail test) was performed if the median of a particular miRNA's ΔΔCt is less than zero. Similarly, test for down-regulation was performed if the median of a particular miRNA's ΔΔCt is more than zero. Multiple testing corrections using Benjamini-Hochberg method was performed at 5% level of significance and corrected cut off p-value was 0.00065. ### Cluster Analysis of miRNAs expressed in cancer tissues K-median clustering method was used for the entire pruned data set to see whether genome wide miRNA expression variations between individuals were large enough to reliably divide the samples into a number of sub clusters. For this clustering, chosen distance metric was Euclidean. Since, expression of many miRNAs was absent in our data set, it was creating “Absent data” situation. This situation is known to bias the clustering if the most popular clustering method- K-means was adopted. So, the choice of clustering was K-median. The number of reliable clusters, the data would form, was determined using “the elbow” method. ### TaqMan Assay Expression of significantly de-regulated miRNAs, detected in cancer tissues by TLDA method, were also validated in same cancer tissues and examined in leukoplakia and lichen planus tissues by TaqMan assay (7900HT Fast Real Time PCR system, Applied Biosystems, USA). Probes and primers were supplied by Invitrogen India Ltd and data were retrieved as “fold change” compared to adjacent controls. Normalization of expression of each gene in each sample was done using geometric mean of the same 3 endogenous controls, viz. *RNU-44, RNU-48* and *mmu-6*, to get ΔΔCt value of that miRNA. # Results ## Expression profile in cancer tissues by TLDA Expression data of 762 miRNAs were assayed by this method but after data pruning, expression data of 531 miRNAs (including 5 endogenous controls) was used for other down-stream analysis. During statistical test, at least 2-fold expression change in cancer tissue compared to its paired-control tissue was considered to be the bench-mark of expression deregulation. Thus, expression of 7 miRNAs was found to be significantly deregulated after multiple test correction and all of these seven miRNAs had \>4-fold average expression deregulation (i.e. either up- or down-regulation). Validation of expression by miRNA specific TaqMan assay also reconfirmed expression deregulation in cancer tissues although fold-changes were diminished compared to TLDA outcome. The difference in sensitivity of these two RT-PCR based (TLDA and miR-specific TaqMan Assay) experimental methods, coupled with the fact that these two sets of experiments were performed at two different time points, might be the influencing factors for difference in degrees of expression deregulation of miRNAs. Relative locations of these 7 miRNAs along with other miRNA genes across different chromosomes revealed that *hsa-miR-133a* and *hsa*-*mir-7* are located on two (Chr 18 and Chr 20) and three chromosomes (Chr 9, Chr 15 and Chr 19), respectively. Here, assays were performed for only mature miRNAs, thus, expression of *hsa-miR-133a* and *hsa*-*mir-7* might be cumulative sum of all mature forms of respective miRNAs. Among these 7 miRNAs, expression of 4 miRNAs viz. *hsa-miR-1293, hsa-miR-31, hsa-miR-31\** and *hsa-miR-7* were significantly up-regulated and those of 3 miRNAs viz. *hsa-miR-206, hsa-miR-204* and *hsa- miR-133a* were significantly down regulated in cancer samples. Heat map was constructed according to two ways unsupervised hierarchical clustering. So, all down-regulated miRNAs clustered together at the upper part of the plot whereas all up-regulated miRNAs clustered at the bottom. It showed values of expression (i.e. ΔΔCt) compared to adjacent control as well as number of samples providing expression data. Expression of *hsa-mir-1293* was obtained from 9 cancer-control paired samples but 6 of them had ΔΔCt values ≤−2 and remaining 3 samples had ΔΔCt values between 0 and −2. So, for *hsa-mir-1293*, all these 9 samples had ΔΔCt values with -ve sign, meaning; whenever expression was obtained, it is always up-regulated but 6 of them showed more than 4-fold expression change. Similarly, expression data for *hsa-mir-31\** were obtained from 16 cancer- control paired samples. Out of these 16 samples, one sample had ΔΔCt between 0 & +2, two samples had ΔΔCt values between 0 & −2 and remaining 13 samples had ΔΔCt values ≤−2. Normalized expression (ΔCt) of these 7 miRNAs in cancer and control tissues were mostly non-overlapping and ΔCt values of different miRNAs across the control tissues also showed quite a wide range of variation. So to get correct relative expression, it is important to compare expression of miRNAs in cancer tissue with those of control tissue from the same individual. In this figure, miRNAs were placed according to increasing p-values (from top to bottom) vertically. The ΔCt values of 8 miRNAs in different tumor and control samples had been plotted on the horizontal axis to show distribution of ΔCt values across all the samples. It is evident that more the overlap between ranges of expression of miRNAs in cancer and control tissues, less is the level of significance. Here, *hsa-mir1293* with lowest p value 0.000028 had been positioned on the top and *hsa-mir-1* with p value of 0.00094 (just below the corrected level of significance) was placed at the bottom on the vertical axis. ## Expression of miRNAs in precancerous leukoplakia and lichen planus samples Among 7 miRNAs which was observed to be significantly deregulated in cancer samples, expression of only *miR-31* was significantly up-regulated in leukoplakia tissues where as expression of none of these 7 miRNAs was deregulated significantly in lichen planus tissues. Expression of *miR-31\** and *miR-204* was also up- (4.75 folds) and down-regulated (1.99 folds), respectively, in leukoplakia tissues but not significantly different. ## Database mining and bioinformatics analyses Published reports on these 7 miRNAs had also shown similar direction of expression deregulation in some cancers including head and neck,. A total of 561 unique targets were identified when these 7 miRNAs were used to search targets using miRWalk and further cross-validated from Pubmed (<http://www.ncbi.nlm.nih.gov/pubmed>). The *hsa-miR-31\** has validated target, *RhoA*, which is reportedly implicated in mouth neoplasm. The *hsa-miR-1293* till now is known to target *GCN1L1* and *hsa-miR-1293* mediated down regulation of this tumor antigen gene (i.e. *GCN1L1*) could contribute to poor prognosis of the tumor. IPA tool was used for “disease term search” (Ingenuity® Systems, [www.ingenuity.com](http://www.ingenuity.com)) and most significant “disease term” in cancer category of IPA was “head and neck” cancer. IPA could not provide any hit by *hsa-miR-1293* and considered *hsa-miR-31* and *hsa-miR-31\** as synonymous so the output was from 5 miRNAs. It was also noticed that IPA considered *hsa-miR-206* synonymous to *hsa-miR-1* since they have identical seed sequence. It was also observed that expression deregulation status of *hsa- miR-1* and *hsa-miR-206* across the samples were very similar to each other (r = 0.94). Inclusion of *hsa-miR-1* in target search list increased the number of targets to 702. In KEGG pathway mapping portal – validated targets of these 8 miRNAs have been used. Top most pathways in the mapped list were “microRNAs in cancer” followed by “proteoglycans in cancer” and “global cancer pathway” (Table S1). Other top relevant pathway was *PI3K-AKT* which is one of the most common pathways implicated in cancer. Other most significant changes that could have happened due to these miRNAs are disruption of actin cytoskeleton maintenance and focal adhesion. Other probable implicated signaling pathways were *RAS*, *RAP1, MAPK, HIF-1*, *FOXO, TNF, ErBB*, apoptosis etc (Table S1). “GO” term enrichment search was performed using input of 702 validated targets of these eight miRNAs (data not shown) and it was observed that most prominent disrupted biological processes are primarily related to cell migration, loss of apoptosis, cell proliferation etc. But DAVID based “GO” term enrichment for biological process showed most prominent biological process is “Apoptosis” (Table S2). All these observations support that these 8 miRNAs may play important functional roles in gingivo buccal cancer. Our RNA-Seq data shows that expression of *FN1, MSN* and *MMP9*, which belongs to “Proteoglycans in cancer” gene list and are targets of down-regulated miRNAs, was up-regulated in 10 of 13 tissue samples (on average, 6.83, 2.43 and 21.89 folds, respectively, compared to adjacent normal). Similarly predicted up-regulation of *LAMC* and down regulation of *PAI*, which belong to *PI3K-AKT* pathway, was also validated in RNA-Seq data in a sub set of these tissue samples (unpublished data). These observations also support our predictive pathway analysis regarding involved biological process/pathways that could be targeted by this set of 8 miRNAs. Selected samples had little variation in terms of its differentiation status or clinical staging. Cluster analysis was performed to check whether genome wide miRNA profile could explain such characteristic variation. It was performed using expression deregulation data (ΔΔCt) of 531 miRNAs from 18 cancer samples using K-median method. Optimally, only two distinct clusters were obtained; one consisted of 13 paired tumor-normal samples and other consisted of 5 paired tumor-normal samples. It was evident that these two clusters of samples were not formed on the basis of their differentiation status or clinical stages. Expression of 30 miRNAs was significantly deregulated in the cluster of 13 samples (p-value cut off 0.00298 after Benjamini-Hochberg Correction,) but none of the miRNAs were significantly deregulated in the cluster of 5 samples (data not shown). Fold expression (up- or down-regulated) of a miRNA in a tumor tissue compared to its adjacent control and number of samples providing expression data of a miRNA could be observed in Heat map diagrams of cluster of 13 samples. It showed that expression of all miRNAs was not available from all samples. Expression of some miRNAs was up-regulated in most of the samples (e.g. expression of *miR-31\*, miR-7, miR-21* shown at the bottom of the figure) and expression of some miRNAs was down-regulated in most of the samples (e.g. *miR-206, miR-1, miR-133a* shown at the middle of the figure). Out of these 30 miRNAs, expression of 28 miRNAs showed similar expression deregulation pattern as it was already reported in earlier studies on cancer,. Of the remaining two miRNAs, one miRNA *(hsa-miR-770-5p*) has not been associated yet with any cancer. Expression of remaining one miRNA (*hsa- miR-211*) has been deregulated in opposite direction in this study compared to previous reports on colorectal cancer. Actually, 7 miRNAs, whose expression was significantly deregulated in the analysis of 18 samples, are a subset of these 30 miRNAs. However, expression of *hsa-miR-411\** was observed to be significantly deregulated in this study, but no previous report had shown its involvement with cancer. Again, it is known that *hsa-miR-411* and *miR-411\** are originated from the same precursor miRNA and *hsa-miR-411* has a strong association with cancer. When we checked expression of mature *hsa-miR-411* and *hsa-miR-411\**, a strong positive correlation was observed (r = 0.95) although expression of *hsa-miR-411* was not found to be statistically significant. Similarly, we have observed significant expression deregulation of *hsa- miR-135b\** and *hsa*-miR-99a\* and reports of association of cancer with *hsa- miR-135b* and *hsa*-*miR-99a*. Searching in similar way, 1207 unique target genes were obtained for these 30 miRNAs. IPA and KEGG mapping were performed using these 30 miRNAs and their known 1207 target genes. In IPA analysis, three cancers, which were found to be associated with expression deregulation of a major subset of these miRNAs, were hypo pharyngeal, esophageal and head and neck cancer (p-values 1.71×10⁻⁰⁷, 2.20×10⁻⁰⁷ and 1.02×10⁻⁰⁷ respectively). Interestingly, same biological signaling pathways (relevant to cancer) were reported to be involved with these set of 30 miRNAs as it was observed with 8 miRNAs earlier. But the difference lies in the repertoire and number of targeted nodes for all these pathways (Table S1 in, Table S2 in, Table S3 in and Table S4). # Discussion and Conclusions Expressions of 7 miRNAs were significantly deregulated in 18 cancer samples and significant up-regulation of *hsa-miR-1293* is being reported for the first time in gingivo buccal cancer. As of now, functional role of *hsa-miR-1293* in cancer is very limited. According to miRWalk and StarBase, one of the predicted targets of *hsa-miR-1293* is *MAPK14* (p38) which has been shown to be associated with tumor's sensitivity to *cis*-platinum treatment. If this predicted relationship could be validated, then expression of this miRNA may be useful in prediction of patient's sensitivity to *cis*-platinum treatment. Two miRNAs, *hsa-miR-206* and *hsa-miR-1*, are known to play anti-tumorigenic role – and *hsa-miR-*206 may indirectly activate apoptosis, inhibition of cell migration and focus formation. Down regulation of expression of *hsa-miR-1* and *hsa-miR-133a*, which has been observed in this study, has already been reported in an earlier study with oral squamous cell carcinoma. In fact, *hsa-miR-133a* targets several oncogenes and is reported to be commonly down regulated in a number of other oral cancer studies. Expression of both *hsa-miR-31* and *hsa-miR-31\** was reported in some earlier study on cancer. Study on OSCC cell line showed that exogenous delivery of *pre-mir-31*, which boosts up quantity of mature *hsa-miR-31* and *hsa- miR-31\**, enhanced OSCC oncogenicity. Hence, our observation of up-regulation of *hsa-miR-31* and *hsa-miR-31\**, corroborates with existing reports on oral and other cancers. Similar to a previous report on head and neck cancer, here also, expression of *hsa-miR-204* was also found to be significantly down regulated. Expression of *hsa-miR-7*, a known OncomiR, is reportedly up- regulated in OSCC. It targets primarily tumor suppressor transcripts from *RECK*. In this study, expression of *hsa-miR-7* was also significantly up- regulated and thus corroborates with previous findings related to oral cancer. Efforts have been made to understand possible biological implication by mining different databases. Pathway analysis revealed that most disrupted biological processes would be cell migration, apoptosis and proliferation (Table S2 in). Most disrupted pathways were predicted to be “proteoglycans in cancer” and *PI3K-AKT* (Table S1) which are also supported by our RNA-Seq data on expression deregulation of some proteoglycan genes and *LAMC* and *PAI* which belong to *PI3K-AKT* pathway (unpublished data). These observations also support our predictive pathway analysis regarding involved biological process/pathways that could be targeted by this set of 8 miRNAs. Cluster analysis of 18 samples had shown that expression of 30 miRNAs was found to be significantly deregulated in the cluster of 13 samples. Literature search and database mining with these 30 miRNAs showed relevance of these genes with oral and other cancers. Although, pathway analysis using 8 miRNAs (identified from expression data of 18 samples) and 30 miRNAs (identified from expression data of cluster of 13 samples) narrowed down to similar signaling pathways but the number and repertoire of nodes targeted by these set of miRNAs are quite different (data not shown). This cluster analysis is actually revealing existence of molecular heterogeneity that may chew up resolution of finding finer molecular marker. Interestingly, observed molecular heterogeneity pattern in no way related to differentiation subtype that exists within the tissues or clinical stages of samples. Among two pre-cancers, leukoplakia is known to be associated with tobacco habit, whereas lichen planus is an auto immune disease. It is reported that all oral cancers are preceded by precancerous lesions. So, we checked whether, similar to cancer samples, expression deregulation of miRNAs could be observed in these two precancerous lesions. Here, TaqMan data showed that expression of only *hsa- mir-31* was significantly up-regulated (4.55 fold more than adjacent control tissue) in leukoplakia samples but not in lichen planus tissue. Though, expression of another miRNA, *hsa-mir-31\**, was also up-regulated by 4.75 folds but significantly not different due to greater standard deviation among samples. So, expression of these two miRNAs needs to be checked in more leukoplakia samples. This once again reiterates about molecular proximity of leukoplakia with GBSCC than other pre-cancer, lichen planus. So, expression of *hsa-mir-31 and hsa-mir-31\** in leukoplakia tissues could be potential risk-markers for progression of precancer to GBSCC. Small number and mixed stage of tumor samples and expression assay of only 762 miRNAs by TLDA (available at the time of our experiment) limited this study to infer that expression of only 7 miRNAs were significantly deregulated in GBSCC tissues compared to adjacent control tissues. There is high chance that number of deregulated miRNAs will increase if we could assay expression of ∼2578 miRNAs presently known in human tissue as mentioned in miRBase-20. As a result, expression of more miRNAs could have been checked in cancer and leukoplakia tissues to infer more about molecular markers. More importantly, role of these miRNAs in carcinogenesis is to be validated by functional study. # Supporting Information We heartily thank Dr. Analabha Basu and Dr. Saurabh Ghosh for assessing the aptness of statistical methods used in this study. We thank Dr. Raghunath Chatterjee who guided us in understanding some of the pathway analysis outputs. We thank Ms. Richa Singh whose valuable suggestion has made the shape of the manuscript better. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: NDS BR. Performed the experiments: NDS RR SG. Analyzed the data: NDS JKM AC IM BR. Contributed reagents/materials/analysis tools: RRP SG BR. Contributed to the writing of the manuscript: NDS BR. [^3]: Current address: Department of Dentistry, College of Medicine and Sagore Dutta Hospital, Kolkata, India.

# Introduction According to physical activity recommendations healthy adults need moderate intensity aerobic physical activity (PA) for a minimum of 150 or vigorous intensity PA for a minimum of 75 min on three days each week accumulating from PA bouts of at least 10 minutes. Moderate intensity activities have energy expenditure between 3.0 to 5.9 metabolic equivalents (MET) and vigorous intensity activities higher than 6.0 METs. One MET is defined as the resting metabolic rate for quietly sitting and is about 3.5 ml · kg^-1 · min^-1, when expressed as oxygen consumption (VO₂) rate. Then 3.0 MET and 6.0 MET physical activities correspond 10.5 ml · kg^-1 · min^-1 and 21.0 ml · kg^-1 · min^-1 VO₂, respectively. Accelerometry provides a useful and feasible method to characterize PA during free living conditions. It permits objective measurements of the intensity, duration and frequency of daily PA and exercise including assessment of short activity bouts that cannot be captured with questionnaires, interviews or diaries. However, the main challenges in accelerometry pertain to the use of proprietary algorithms, and lack of valid physiologically determined cut-points for intensity of PA, which both compromise the cross-study comparisons. Storing and processing the raw acceleration data have been proposed as the means to improve the comparability between different devices and studies. We recently developed a novel method for universal analysis of PA from raw tri- axial accelerometer data. In that study, raw acceleration data were collected during various sedentary and ambulatory activities and analysed with several classifiers in both time and frequency domain. Of these the mean amplitude deviation (MAD) of the resultant acceleration signal consistently provided the best performance in separating different PA intensity levels from each other. Most importantly the MAD enabled a direct comparison between the results of different accelerometer brands despite clearly different technical specifications of these devices. The MAD value describes the mean value of the dynamic acceleration component. It is calculated from the resultant value of the measured tri-axial acceleration, which comprises both tri-axial dynamic components due to velocity changes and static component due to gravity. The static component is removed from the analysed time period (epoch) and the remaining dynamic component is rectified. The MAD value is the mean of the rectified signal within the epoch and independent of static component. There is a strong correlation between incident MAD values and heart rate both among adults and adolescents. We therefore hypothesized that the incident MAD value also correlates strongly with actual VO₂ at individual level. Previously similar acceleration signal-derived traits have been evaluated during treadmill walking and running and found to have strong correlation with VO₂ and be better predictor of VO₂ than heart rate. However treadmill derived relationship may overestimate the intensity of PA \[, \], thus, accurate comparison is best achieved during actual locomotion. The present study was carried out to develop a MAD-based model for predicting VO₂ during locomotion within a wide range of speeds and to determine its accuracy. The main objective was to determine the MAD-based universal cut- points for light, moderate and vigorous PA corresponding to commonly accepted MET values for these intensity levels. # Methods ## Participants The study group consisted of 29 healthy volunteers, 15 males and 14 females. Prior to testing body height, weight and waist circumference were measured with standard methods. Participants were informed of the experimental test protocol and they gave their written informed consent. This study conformed to the code of Ethics of the World Medical Association (Declaration of Helsinki) and it was approved by the Ethics Committee of Pirkanmaa Hospital District (R13040). ## Test procedure Participants performed a pace-conducted non-stop test on a 200 m long oval indoor track with slightly banked bends. Pace was verified by a so called “light rabbit” system, which comprises light sources alongside of the track. The light sources were installed at every 2 meters and they automatically lited-up one at a time according to the required pace. Every light source was also visible to tester. So participant’s location on the track could be determined with the accuracy of 2 meters. Initial speed was 0.6 m/s and it was increased by 0.4 m/s at every 2.5 minutes. The direction of travel was counter-clockwise. The participants had to keep up with the light sources and they could freely decide whether they preferred walking or running. The selected gait type was recorded for each stage. The test was continued until volitional exhaustion when the participant could not keep up with the lights anymore. End point was visually inspected by the tester and time at the end point was recorded. ## Measurements A tri-axial accelerometer (Hookie AM20, Traxmeet Ltd, Espoo, Finland) was attached to the hip-mounted elastic belt at the level of the iliac crest. The accelerometer had ± 16 *g* measurement range and acceleration was measured at 100 Hz sampling frequency. Because the indoor track has banked turns only to left, it might have effect on the accelerometer output. Thus the accelerometer was placed either at the right (r-hip) or left side (l-hip) of the hip and the side was randomly selected. The left side (i.e. inner side of curve) was assigned to 13 participants and the right side (i.e. outer side of curve) to 16 participants. In addition, all participants carried one accelerometer in the middle of the back (mid) to emulate a situation where misplacement was the longest in terms of the preferred lateral location. During the test procedure, VO₂ was continuously measured with a portable breath-by-breath mobile metabolic analyser (Oxycon, Carefusion, Yorba Linda, CA, USA) and the data were recorded with a telemetry system. The metabolic cart was calibrated before each test according the manufacturer’s instructions. ## Data analysis The analysis of the acceleration signal was based on the MAD value described recently. Tri-axial acceleration was measured in raw mode from all three orthogonal measurement axes in actual g-units and stored for further analysis. In short, each measurement point (*i*) consisted of samples *x*_*i*, *y*_*i* and *z*_*i*. The resultant acceleration (*r*_*i*), which defines the magnitude of the acceleration vector and contains both dynamic and static component of acceleration, was calculated for each (i) time point as $$\text{r}_{\text{i}} = \sqrt{\text{x}_{\text{i}}^{2} + \text{y}_{\text{i}}^{2} + \text{z}_{\text{i}}^{2}}$$ The epoch length used in the analysis was 6 s, and for each analysed epoch the mean resultant value (*R*_*ave*) describing the static component of acceleration was calculated as $$R_{ave} = ~\frac{1}{\text{N}}{\sum_{\text{i} = \text{j}}^{\text{j+N-1}}\text{r}_{\text{i}}}$$ The MAD value of the given epoch was calculated as $$\text{MAD} = \frac{1}{\text{N}}{\sum_{\text{i} = \text{j}}^{\text{j+N-1}}\left| {\text{r}_{\text{i}} - \text{R}_{\text{ave}}} \right|}$$ where *N* is the number of samples in the epoch (ie, 600) and *j* is the start point of the epoch. The unit of the MAD is milligravity (m*g*); ie, the Earth’s gravity 1*g* is equal to 1000 m*g*. To allow the participants to find the steady rhythm the acceleration was analysed for the final 2 minutes of each stage. As a measure of steady-stage VO₂ for given speed mean VO₂ of the final minute of the corresponding stage was used. VO_2peak value was the highest measured VO₂ for one minute period during the test. Measured end time made it possible to calculate the mean speed during the final two and half minutes of the test (v_max). The v_max value depends on the maximum speed achieved during the test and time it has been maintained. The oxygen cost of movement (ml · kg^-1 · km^-1) was calculated for each stage as the ratio of measured VO₂ (ml · kg^-1 · min^-1) to known speed (km/min). ## Statistical methods Data were analysed with SPSS 21.0 (SPSS Science, Chicago, USA) software. First, independent two-sample t-test was performed to investigate whether the MAD values were different between right and left side for each speed. In addition, Pearson correlation between VO₂ and the MAD was determined for each participant. Mean correlation coefficient was calculated by first z-transforming the individual correlation coefficients, taking the arithmetic mean of transformed coefficients, and then by back-transforming the mean. The generalized linear model was used to estimate VO₂. Since the data were not normally distributed, gamma regression model was used. Incident VO₂ was the dependent variable and the incident MAD value, physical characteristics (age, weight, height, and waist circumference) and performance values (VO_2peak and v_max) served as the independent predictor variables. Furthermore, because the data during walking were normally distributed, the estimation of the VO₂ during walking from the MAD value was based on linear regression model. Stages with respiratory exchange ratio over 1.0 or not fully completed were excluded from the analysis. To find optimal intensity based cut-points for the MAD values the receiver operator characteristics (ROC) analysis was used. The measured VO₂ values were used as a golden standard. VO₂ values were converted to MET values by using the standard conversion factor (1 MET = 3.5 ml · kg^-1 · min^-1) and performances were classified to light (\< 3.0 MET), moderate (3.0–5.9 MET) and vigorous (\> 6.0 MET) activity. For both 3 MET and 6 MET limits a new dichotomous variable was created to define the outcome of the test. If a measured MET value was less than limit, the outcome was negative and the variable value was set to 0. Otherwise the outcome was positive and the variable was set to 1. After that ROC curve analysis were conducted to determine sensitivity and specificity values. Sensitivity and specificity defines correctly identified positive and negative values. The MAD value which maximized the sum of the specificity and sensitivity was selected as the optimal cut-point. Also the area under curve (AUC) value was determined. AUC value of 1 indicates a perfect classifier whereas AUC value of 0.5 denotes no discriminatory value. # Results The illustrates one test performance, where the participant walked the first four stages and changed to running at stage five (2.2 m/s). In the beginning of the test there was some trouble in achieving steady pace, which can be seen from MAD and VO₂ curves. The required pace was maintained almost 27 minutes. The shows the number of fully completed stages and the preferred gait. Typically the participants changed the gait type (from walking to running) at the beginning of the stage. At speed 2.2 m/s (7.9 km/h) four participants changed the gait type in the middle of the stage. The mean VO_2peak was 56.0 ± 7.1 ml · kg^-1 · min^-1 (range 45–69 ml · kg^-1 · min^-1). The range of the v_max was 3.1–5.1 m/s (11.1 km/h– 18.4 km/h). At the group level the oxygen cost of locomotion reached the minimum at the speed 1.4 m/s. During running, after initial increase the energy cost remained quite steady despite increasing speed. At the individual level the curve of the oxygen cost was u-shaped both for walking and running. For walking, 26 participants have the lowest oxygen cost at speed 1.4 m/ (5.0 km/h), while for running the minimum value varied between speed from 2.2 m/s (7.9 km/h) to 3.8 m/s (13.7 km/h). VO₂ showed a curvilinear increase during walking and linear during running with increasing speed. The highest MAD and VO₂ value for the stage containing barely walking was 651 m*g* and 30.2 ml · kg^-1 · min^-1, whereas the lowest value for stage containing barely running was 581 m*g* and 26.1 ml · kg^-1 · min^-1, respectively. ## Sensor placement and MAD Sensor placement conferred a slight effect on measured MAD values, the effect being largest during running. With slow speed walking the mid position values and with running the right-side position values were slightly lower. The MAD values increased with increasing gait speed. ## Individual correlations Within individuals, the correlations between both MAD and VO₂, and MAD and speed were very high. In all participants, the MAD value increased with increasing VO₂ or speed for both walking and running. Mean correlation values were highest for walking, and somewhat lower for data containing both walking and running, or running alone. This is due to different regression slopes between MAD and VO₂, and MAD and speed during walking and running, and larger variation in data during running. ## Prediction models between MAD and VO₂ The direct relationship between the incident MAD and VO₂ values were estimated with following equation: VO₂ (ml/kg/min) = 10.015 · e^{0.0017 · MAD (mg)} (r = 0.958, standard error of the estimate (SEE) = 6.05 ml/kg/min), where *mg* denotes milligravity. For walking only, the prediction model based on linear regression was: VO₂ = 7.920 + 0.0331 · MAD (*mg*) (r = 0.943, SEE = 1.66 ml/kg/min). By using all measured values the estimation equation was following: VO₂ (ml/kg/min) = 2.351 · e^{(0.00177 · MAD (mg)- 0.282 · vmax (m/s) + 0.0183 · VO2peak (ml · kg-1 · min-1) + 0.0117 · height (cm)– 0.0142 · weight (kg) + 0.00693 · waist circumference (cm)– 0.00211 · age (years))} (r = 0.975, SEE = 4.46 ml/kg/min). Parameters in the equation are in the order of significance and for each parameter p-value was less than 0.05. ## Optimal cut-points According to the ROC curve analysis the optimal MAD cut-point for intensity of 3.0 MET was 91 *mg* and for 6.0 MET 414 *mg*. Sensitivity and specificity values were 100% and 96% for the 3.0 MET cut-point, and 96% and 95% for 6.0 the MET cut-point. The AUC and 95% confidence interval for both limits is shown in the. # Discussion The present study demonstrated that the MAD is a highly valid method to estimate the intensity of PA within a wide range of locomotion from slow walking to fast running. The study also produced valid cut-points for accurate determination of light, moderate and vigorous intensity levels of PA that are much needed in epidemiological population studies. Because the calculation of the MAD is based on the raw acceleration data and has been shown to be device-independent, the MAD approach offers a possibility to obtain directly comparable and accurate results on the intensity of PA with all accelerometers which provide tri-axial raw data within a sufficient dynamic range. The accuracy of the MAD method to predict VO₂ during walking and running compares well with other commonly used methods. As is the case with other methods, the accuracy was better for walking than running. However the MAD value is not compromised by the common ceiling effect, where the accelerometer output values reach a peak at a certain speed and do not increase in response to further speed increments. In the present study increasing speed produced increasing MAD values in a dose-response manner for each participant. The oxygen cost of the locomotion was at typical level for all except the first stage of the test (i.e. slow walking), when the cost was higher than expected. Apparently the emotional excitement at the beginning of the test contributed to relatively high measured VO₂ values. Also some participants had difficulties in achieving a steady pace at the beginning of the first stage. The measured oxygen cost was 303 ± 39 ml/kg/km at the 0.6 m/s speed, while in the previous study the oxygen cost was below 250 ml/kg/km at the 0.67 m/s speed. The most economical movement speed in the present study was 1.4 m/s with the oxygen cost of 188 ± 17 ml/kg/km which is line with other studies. Preferred gait transition speed from walking to running is slightly lower than energetically optimal transition speed and it seems to depend on various metabolic and biomechanical factors related to transition. In the present study preferred speed for gait transition was in line with literature. With running the oxygen cost was slightly lower for the higher speeds. This is probably consequence of the fact that only participants with good running economy reached the higher speeds and could maintain the speed for several minutes. The association between the MAD and VO₂ was very strong both at individual level and at group level. By adding individual anthropometric and performance data into the prediction model only a slight improvement in the estimation of VO₂ could be attained, especially during running. However, when the MAD value was excluded, the most significant individual predictors were VO₂peak and v_max values, which might be difficult to obtain in practise. The contribution of these predictors is apparently related to running economy and the bouncing characteristics of running. The mechanics and energetics of running depend on the kinetic and potential energy of the whole body and the body segments, besides storage and release of mechanical energy by the contracting muscles and tendons. The accelerometer cannot separate whether the energy for the speed change is produced actively by the muscle or passively by the tendon. The simpler mechanics and energetics characteristics of the walking seem to account for more straightforward estimation of VO₂ with accelerometer. Nevertheless, the prediction of the incident VO₂ was excellent whether or not the additional predictors were known, and importantly, the estimation was sufficient enough for accurate classification of PA in terms of light, moderate and vigorous intensity. Previously it has been shown that sensor placement on either hip or waist area can have effect on accelerometer output. With the MAD values the effect of sensor placement was marginal in relation to the wide range of MAD values in different speeds. However, some observations are worth discussing. First, at low walking speeds the mid position showed slightly lower MAD values than the other positions. This is explained by the sensor movement which is apparently higher on the side due to pelvis tilting. Second, during running the right side MAD values were slightly lower. This is attributable to the characteristics of the indoor track used in the present study. It had banked curves and participants moved only to anti-clockwise direction. The inner (left) and outer legs (right) apparently experienced somewhat asymmetrical loading in the curves. These observations not only show how sensitive method the accelerometry can be in detecting subtle differences in movements, but also underline the importance of keeping placement of the sensor as constant as possible. Some analysis methods of accelerometry data can produce inaccurate results, if the sensor orientation in relation to gravity is not controlled for. With the MAD method this is not a concern, as illustrated by the following examples. Assuming that the orientation of the sensor x-axis is perpendicular to ground (ie, parallel to the gravity vector) while both y- and z-axes are parallel to ground, then the measurement vector M = (x, y, z) is M = (1.000, 0.000, 0.000) in *g*-units. The resultant acceleration R is in this static situation equal to Earth’s gravity 1.000 *g*. If the sensor moved downwards with 0.5 g acceleration and thereafter upwards with 0.5 *g* acceleration then the sensor readings would be M = (0.500, 0.000, 0.000) and M = (1.500, 0.000, 0.000) and corresponding R values 0.500 *g* and 1.500 *g*. Both values deviate 0.500 *g* i.e. 500 *mg* from the static value. Assuming next that the sensor is rotated 45° around the z-axis resulting in M = (0.707, 0.707, 0) and R = 1.000 g in the static situation. For the above described dynamic 0.5 g downwards and upwards accelerations the sensor readings would be M = (0.354, 0.354, 0.000) and M = (1.061, 1.061, 0). Again, the corresponding R values are 0.500 *g* and 1.500 *g* and deviation from the static value is 0.500 *g* despite the different position of the sensor. Another problem with accelerometers is their offset, which means the difference between the measured value of the accelerometer and the true acceleration. Although the sensors are calibrated during manufacturing process, the offset cannot be avoided, because it is sensitive to external conditions, like temperature. Using the conditions of the previous example and assuming that all axes have 0.05 g offset the calculations are as follows: When the x-axis is perpendicular to ground, sensor readings for static condition would be M = (1.050, 0.050, 0.050) and R value 1.052 *g*. For the 0.5 *g* dynamic upward and downward accelerations the corresponding values are M = (0.550, 0.050, 0.050) and M = (1.550, 0.050, 0.050), and the R values 0.555 *g* and 1.552 *g* resulting in 0.499 *g* mean deviation from the static value. In the 45° rotated case the sensor readings for the static condition would be M = (0.757, 0.757, 0.050) and for dynamic conditions M = (0.404, 0.404, 0.050) and M = (1.111, 1.111, 0.050). The respective R values would be 1.072 *g*, 0.573 *g* and 1.572 *g* resulting in 0.499 *g* mean deviation from the static value. Inferred from the above described examples, the offset can have minor effect on the results but there is no need to use excessive methods to calibrate the sensor when the MAD approach is used. Participant’s aerobic fitness in the present study was high as can be judged from their high VO_2max values. All but three participants reached the highest class in seven level fitness classifications and the remaining three belonged to classes 5–6. This can be considered a strength of the study because a wide speed range of locomotion and oxygen consumption was attained from very slow walking to high speed running. Other strengths were the relatively large sample size and the direct measurement of the expiratory gases during actual bipedal movement on the track instead of a treadmill. The main weakness of the study was that no other activities than walking and running on the track were studied. On the other hand, locomotion is the most common type of physical activity among ordinary people. # Conclusion In conclusion, the MAD is a valid method to estimate the intensity of PA within a wide range of bipedal human locomotion. The MAD values highly reflect the incident oxygen consumption within a wide range of walking and running speeds. The proposed cut-points offer a valid base for assessing the health effects of PA. Sensor positioning does not compromise the results. This study further underscores the utility of the simple and universal MAD approach as the means to overcome the challenges for comparisons between studies and different accelerometers. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: HV-Y AM ToV TiV HS. Performed the experiments: AM TiV. Analyzed the data: HV-Y. Wrote the paper: HV-Y JS PH. [^3]: ‡ These authors also contributed equally to this work.

# Introduction The CD4⁺CD25⁺FOXP3⁺ regulatory T (Treg) cells are required for proper maintenance of immunological self-tolerance and immune homeostasis. Treg cells develop in the thymus as an independent CD4⁺ T cell lineage. It is believed that epigenetic modifications serve as an important regulatory mechanism that mediates cell fate choice between the conventional T (Tconv) cells and Tregs, but there is a paucity of information related to the epigenetic changes that occur during Treg differentiation. Epigenetic modifications, such as methylation, acetylation and phosphorylation of histones, are implicated in regulating gene expression by controlling chromatin structure and facilitating DNA accessibility. In T cells, histone modifications and nucleosome positioning have been correlated with gene expression or repression. The functional program of Treg cells has been demonstrated to be, at least partially, controlled by miRNA epigenetic modulation –. Moreover, constitutive gene expression of the Treg lineage- directing transcription factor (TF) forkhead box P3 (FOXP3) has been found to be dependent upon the DNA methylation status of its cell-specific enhancer. More than 100 differentially methylated regions (DMRs) have been identified in Treg or Tconv cell type-specific or highly expressed genes such as FOXP3, interleukin 2 receptor alpha (IL2RA), CTL-associated molecule-4 (CTLA4), CD40 ligand (CD40LG) and interferon gamma (IFNG). Unfortunately, very little information has been gleaned about the regulatory role of histone methylation during Treg lineage commitment, differentiation or cell type-specific gene regulation. Determining the global methylation profile in the distinct T cell lineages, as related to gene expression status and regulatory regions, such as promoters and enhancers, will provide significant insight into differentiation and lineage commitment processes and Treg-specific function. General studies on histone acetylation have revealed that this particular modification is associated with the euchromatin form of DNA and active gene transcription. On the other hand, histone methylation has exhibited a more complex relationship with chromatin states. The monomethylations of one of the four core histones, H3, at lysines 27, 9 (H3K27, H3K9), H4K20, and H2BK5 are all linked to gene activation, whereas trimethylations of H3K27 and H3K9 are linked to repression. As for H3K4, both monomethylation and trimethylation are linked to gene activation. Acetylation has been found to be enriched in the promoter regions and at the 5′-ends of coding regions. Within the promoters, the two nucleosomes that flank the transcription start sites (TSSs) are hypoacetylated at certain lysines and are enriched in the histone H2A variant Htz1 in yeast. In yeast genome, the TSSs themselves are devoid of nucleosomes. However, nucleosome occupancy in promoter regions (and at the TSS) is dependent on Pol II occupancy in the human genome. Three forms of histone methylation, monomethylated histone (H3K4me1), the dimethylated form (H3K4me2) and the trimethylated form (H3K4me3), have been characterized as strongly enriched around the TSSs, whereas H3K36me3 peaks near the 3′-ends of genes. The chromatin immunoprecipitation-sequencing (ChIP-Seq) technique developed in recent years combines the use of modification-specific antibodies for ChIP with next-generation high-throughput sequencing-by-synthesis, and has revolutionized our ability to monitor the global incidence of histone modifications. ChIP-Seq profiles for protein–DNA association have been successfully used to identify distal and proximal regulatory elements with high spatial resolution. In this study, we aimed to take advantage of the fine resolution afforded by the ChIP- Seq assay to generate, for the first time, genome-wide distribution profiles of H3K4me1 and H3K4me3 in human Treg cells and activated (a)Tconv cells. Previous ChIP analysis followed by microarray sequencing-by-hybridization of the 1% of the human genome represented by the ENCODE regions indicated that H3K4me1, but not H3K4me3, was enriched around distal *cis*-elements for the E1A binding protein p300 (EP300), while both modifications were enriched at promoters. Furthermore, the chromatin state at promoters was found to be largely invariant across diverse cell types. In contrast, the enhancers identified in different cell types appeared to have cell type-dependent chromatin modification patterns, and the cell type-specific presence of chromatin marks at enhancers, such as of H3K4me1, was closely correlated with cell type-specific expression of the putative gene targets of these enhancers. Thus, enhancers may be more dynamically regulated in different cell types and are likely principal mediators of cell type-specific gene expression. Using the global profile of methylation distribution in Tregs and aTconv cells, we also aimed to discover novel enhancer regions that mediated differential gene expression. Here we present the comprehensive genome-wide dataset of lineage-specific H3K4me1 and H3K4me3 patterns in Treg and aTconv cells. The majority of the H3K4me1 regions found to differ between Treg and aTconv cells were located at promoter-distal sites. I*n vitro* reporter gene assays were used to evaluate and identify novel enhancer activity. These global methylation profiles represent a crucial foundation from which future studies will elucidate the genetic mechanisms that regulate differentiation decisions, lineage commitment and gene regulation in Tregs. # Materials and Methods ## Cell purification and culture Mononuclear cells (MNCs) were isolated from leukapheresis products of healthy volunteers by density gradient centrifugation over Ficoll-Hypaque solution (Biochrome AG, Germany). CD4⁺CD25⁺ T cells were enriched using the human CD4⁺CD25⁺ Regulatory T Cell Isolation Kit and the Midi-MACS separation system (both by Miltenyi Biotec, Germany). The isolated CD4⁺CD25⁺ T cells were then stained with CD4-FITC and CD25-PE (both from BD Biosciences, USA), and their purity was detected with a FACS-Aria high-speed cell sorter (BD Biosciences). The purity of cells after sortings was determined to reach above 98%. MACS-purified CD4⁺CD25⁺ regulatory T cell populations were monoclonally expanded *in vitro* over a period of eight to nine weeks using the Dynabeads® Human Treg Expander (Invitrogen, USA). Briefly, isolated cells were stimulated with magnetic polysterene beads coated with a mixture of monoclonal antibodies against CD3 and CD28 in the presence of high-dose recombinant human IL-2 (rhIL-2, 300 U/mL: Proleukin; Chiron, USA), as described in the manufacturer's instructions. The expanded cells were then stained with cell surface anti-CD4-FITC, anti-CD25-PE antibodies and intracellular anti-FOXP3-APC antibody (eBioscience, USA), and the fixed cells were separated by fluorescence activated cell sorting (FACS) into batches of CD4⁺CD25⁺FOXP3⁻ activated Tconv (aTconv) and CD4⁺CD25⁺FOXP3⁺ regulatory T cells. Jurkat cells (human T cell leukemia) were grown in 90% 1640-RPMI (PAN Biotech GmbH, Germany) plus 10% fetal bovine serum (FBS) supplemented with 2 mM L-glutamine (Biochrome, Germany), MEM non-essential amino acids, sodium pyruvate, MEM vitamins, 50 U/mL penicillin/streptomycin, and 50 nM beta- mercaptoethanol (all from Gibco, USA) in a humidified atmosphere at 37°C and 5% CO₂. Written, informed consent was obtained from all subjects prior to participation, and this study was approved by the ethics committee of the Third Military Medical University, Chongqing, China. ## Suppression assay CD4⁺CD25^- T cells selected from PBMCs with anti-CD4 MACS- beads were labeled with 2 µM of the intracellular fluorescent dye 5-(and -6)-carboxyfluorescein diacetate succinimidyl ester (CFSE; Invitrogen) for 10 minutes at 37°C, and washed twice with PBS. Aliquots of 2×10⁴ sorted CD4⁺CD25⁻ T cells were seeded in wells on a 96-well U-bottom plate pre-coated with anti-CD3 (2 µg/mL; BD Biosciences) and co- stimulated with either soluble anti-CD28 (1 µg/ml; BD Biosciences) alone or in the presence of expanded CD4⁺CD25⁺ T cells at different ratios, as indicated. Co-cultures were harvested after four to five days of incubation and analyzed on a FACS Calibur flow cytometer. ## ChIP and ChIP-Seq The procedure of ChIP-Seq was carried out as previously described. The Treg and aTconv cells were used for ChIP analysis. To map enzyme target sites, 2×10⁶ cells were crosslinked with formaldehyde and sonicated to obtain chromatin fragments of 200 to 300 bp. Sonicated chromatin was pre-cleared and incubated with 2 µg of anti-H3K4me1 (Abcam, United Kingdom), anti-H3K4me3 (Abcam) or anti-rabbit IgG (Upstate, USA) overnight at 4°C. The crosslinks were reversed, and DNA was treated sequentially with Proteinase K and RNase A, and purified using the Qiaquick PCR-purification kit (Qiagen, Germany). ChIP samples were tested for enrichment by qPCR. For ChIP-Seq, the precipitated DNA was repaired using PNK and Klenow enzyme, and ligated to adapters according to manufacturer's instructions. Subsequently, PCR-amplified fragments of approximately 220 bp were sequenced using the Solexa 1 G Genome Analyzer following manufacturer's protocols ([www.illumina.com](http://www.illumina.com)). The ChIP-Seq data have been now accessible in NCBI's Gene Expression Omnibus (GSE26427). ## ChIP-seq reads mapping to genomic regions ChIP-seq reads of ∼35 bp were mapped to the University of California, Santa Cruz (UCSC) human genome (hg18) by SOAP, which allowed a uniquely aligned read to have up to two mismatching bases. The output of the SOAP analysis data was converted to browser-extensible data (BED) files in order to view the data in the UCSC Genome Browser. ## Identification of H3K4me1 and H3K4me3 peaks The uniquely aligned reads by SOAP were considered in peaks calling. To eliminate noise and account for unequal total numbers, we used a defined analysis model (Model-based Analysis of ChIP-Seq, MACS) with default parameters to find peaks, which were called “peaks” of H3K4me1 and H3K4me3. The results include peak location, peak sequence, etc. ## Distribution of H3K4me1 and H3K4me3 peaks The overall profile of the H3K4me1 and H3K4me3 distribution was generated by dividing the human genome into four regions : proximal promoters (1 kb upstream and downstream of the transcription start site (txStart), based on annotated “known genes” from the UCSC Genome Browser); exons; introns; and intergenic sequences. ## Identification of common and lineage-specific H3K4me1 and H3K4me3 peaks We compared the location of each peak in Treg and aTconv cells for H3K4me1 and H3K4me3. For identification of common peaks, the location of peaks has to be overlapped in both lineages with a minimal distance of 1 bp. Furthermore, the lineage-specific peaks were defined as peaks in one lineage that did not overlap with any other peaks in the other lineage. ## Identification of common and lineage-specific proximal promoters or genes Proximal promoters enriched by H3K4me3 in the two lineages were compared each other to determine the common and cell-type specific proximal promoters. As a single proximal promoter is usually associated with one or more genes, we compared the gene(s) associated with each of the proximal promoters in each cell type to determine the common and cell-type specific genes. ## Profiles of the tag density of modifications For each gene, uniquely mapped tags (reads) were summed in 125bp windows (40 windows per region) for the regions ranging from 5 kb upstream of txStart to the txStart itself and from the transcription end site (txEnd) to 5 kb downstream of the txEnd, respectively. Within the gene body, every gene was splitted into 40 windows. All window tag counts were normalized by the total number of bases encompassed within the windows and the total read number from sequencing of the given library. ## Quantitative real-time PCR Total RNA was extracted from the expanded and sorted Treg and aTconv cells by Trizol Reagent (Invitrogen, USA). The quantity of total RNA was measured by a NanoDrop spectrophotometer (Agilent Technologies, USA), and 500 ng was used to synthesize cDNA with a Reverse Transcription Kit (TaKaRa, Japan). GAPDH was used as the endogenous control. PCR was carried out in a 25 µl reaction with 0.5 mM gene-specific primers and using a SYBR Green Kit (TaKaRa) for 40 cycles in a Rotor-Gene 6000 (Gene Company Limited, Australia). The 2^−ΔΔ CT method was used to calculate expression relative to the GAPDH housekeeping control. ## Reporter assays The selected H3K4me1 and H3K4me3 enriched regions (500–1000 bp) were PCR- amplified from human genomic DNA and cloned directly into the pGL3-promoter vector (Promega, USA). Primer sequences are listed in. All inserts were verified by sequencing. One-million Jurkat cells were co-transfected using DEAE-dextran, with 1.0 mg of each reporter plasmid and 0.15 mg of a Renilla control vector (Promega). After transfection, cells were left untreated or stimulated with 20 ng/mL PMA and 1 mM ionomycin (1 mg/mL). Triplicate transfections were harvested after 24 h of incubation. Cell lysates were assayed for firefly and Renilla luciferase activities using the Dual-Luciferase Reporter Assay System (Promega) on a Lumat LB9501 luminometer (Berthold, Germany). Firefly luciferase activity of individual transfections was normalized against the Renilla luciferase activity. ## Statistical analysis The two-tailed Student's *t-*test was used in the analysis of mRNA expression and of luciferase activity. Significance was defined by a *P*-value \<0.05. # Results ## Expansion and purification of human CD4⁺CD25⁺FOXP3⁺ Treg cells Due to the low frequency of Treg cells present in peripheral blood, we expanded the Treg cell population *in vitro* to obtain enough cells for analysis. First, we purified CD4⁺CD25⁺ T cells from PBMCs of healthy human volunteer subjects by using the Magnetic-activated cell-sorting method (MACS). The purity of products was determined to be \>93%. Then, the CD4⁺CD25⁺ T cells were exposed to artificial antigen- presenting cells for repeated stimulation via CD3 and CD28 in the presence of high-dose IL-2, which resulted in profound monoclonal proliferation of up to 1000-fold expansion within an eight to nine week span . We then tested the suppressive activity of the expanded CD4⁺CD25⁺ T cells by evaluating their ability to inhibit the proliferation of autologous CD4⁺CD25⁻ T cells after allogeneic stimulation. Results from the mixed lymphocyte reaction (MLR) assay showed that the proliferation of CD4⁺CD25⁻ responder T cells was inhibited, in a dose-dependent manner, by the expanded CD4⁺CD25⁺ T cells. Because CD25 is known to be expressed on activated T cells derived from CD4⁺CD25⁻FOXP3⁻ T cells, the expanded CD4⁺CD25⁺ T cells were expected to include FOXP3⁻ T cells. To obtain high-purity of CD4⁺CD25⁺FOXP3⁺ Treg cells, which represent the intrinsic Treg cells, we purified the triple-positive T cells from the expanded CD4⁺CD25⁺ T cells by the FACS method, and the purity of CD4⁺CD25⁺FOXP3⁺ Treg and activated conventional CD4⁺CD25⁺FOXP3⁻ T cells (aTconv) reached 99.0% and 99.4%, respectively . ## Direct sequencing analysis of ChIP DNA samples We used the high-throughput ChIP-Seq approach to generate genome-wide H3K4me1 and H3K4me3 maps of human CD4⁺CD25⁺FOXP3⁺ Treg cells and CD4⁺CD25⁺FOXP3⁻ aTconv cells. The sequencing procedure required a one-step adaptor ligation and limited PCR amplification (18 cycles) of ChIP DNA molecules, followed by cluster generation and sequencing-by-synthesis. The read/peak numbers for each library in each cell type were shown in. Prior to and post sequencing on the Solexa 1 G Genome Analyzer, the ChIP samples were confirmed for the target sites in both cell types by regular ChIP-qPCR with the indicated primers. The qPCR results were highly consistent with ChIP-Seq data as expected. ## Genome-wide maps of H3K4me3 modifications in human CD4⁺CD25⁺ T cell lineages with or without FOXP3 expression To obtain an overall picture of the H3K4me3 distribution, we divided the entire human genome into four distinct regions, according to the annotated “known genes” from the UCSC Genome Browser, : proximal promoters (1 kb upstream and downstream of the TSSs), exons, introns, and intergenic sequences. Results showed that about 35% and 49% of H3K4me3 islands were located in proximal promoter regions for Treg cells and aTconv cells, respectively. Examination of those H3K4me3 tags (reads) located within gene bodies, and their 5′- and 3′-end 5 kb extended regions, also revealed enrichment of H3K4me3 islands near TSSs. These results are consistent with recent observations from others that have indicated that H3K4me3 associates extensively with proximal promoters of active genes in human T cells, as well as in human and murine embryonic stem cells,. We compared the H3K4me3 enriched regions between Treg and aTconv cells, and found that the coefficient correlation was 0.92 and that there were 20784 H3K4me3 islands that overlapped in the two cell types. Furthermore, about 75% of these overlapping islands were located in proximal promoters. We then compared the H3K4me3 enriched proximal promoters of Treg and aTconv cells, and determined that the coefficient correlation was 0.83 and that there were 15508 overlapping H3K4me3 enriched proximal promoters. In addition to these overlapping islands we also found that nearly 30000 H3K4me3 islands were Treg cell-type specific, and about 7% of those were associated with proximal promoters. We also analyzed the particular genes related to the H3K4me3 enriched proximal promoters, and found that most of the genes were common between the two cell types; only 1220 related genes were Treg cell-type specific. These results suggested that most of the genes were expressed in both Treg cells and aTconv cells, and the distinct properties of development and function of Treg cells might, in fact, be due to the unique H3K4me3 modification of Treg cell type-specific genes like FOXP3. ## Disparate H3K4me3 modification of signature genes between Treg and aTconv cells Because the T cell subsets represented distinct and stable cell lineages, we inferred that the signature genes corresponding to their respective phenotypes would harbor unique H3K4me3 marks in their proximal promoters, consistent with the corresponding gene expressions in that particular lineage. We first examined the H3K4me3 pattern for IL2RA, CTLA4, TNFRSF18 and FOXP3 genes, each of which encodes the defining lineage markers for Treg cells. Results showed that IL2RA, CTLA4 and TNFRSF18 genes were marked in their promoters by H3K4me3 in both Treg and aTconv cells; this finding was consistent with their respective expression levels detected in activated T cells derived from CD4⁺CD25⁻ T cells. In contrast, FOXP3, a gene that is required for Treg cell development and functions, was marked in its proximal promoter by H3K4me3 in Treg cells, but not in aTconv cells. We detected a 50-fold increase in the expression level of FOXP3 mRNA in Treg cells, as compared to aTconv cells; comparable expression levels of IL2RA, CTLA4 and TNFRSF18 mRNA were observed between the two cell types. We also examined the mRNA expression levels of other genes that were found to be marked in their proximal promoters by H3K4me3 in Treg and/or aTconv cells, such as the STATs and CCR7. The mRNA expression levels of these genes were consistent with the H3K4me3 status observed for their proximal promoter. For example, STAT family TFs are crucial for proper T cell differentiation; however, their expression is not sufficient to drive lineage commitment. Consistent with their ubiquitous expression patterns, we found that most STATs were marked in their promoter regions by H3K4me3, in both the Treg and aTconv cells (-I). Real-time PCR assays showed that the mRNA expression levels of all the STATs were comparable among the two lineages. In contrast, the promoter for the CCR7 gene was marked by H3K4me3 only in Treg cells, and the real-time PCR assay showed an approximate 20-fold increase in its expression as compared to that in aTconv cells. Based on the above results, we predict that Treg differentiation and lineage commitment are associated with specific H3K4me3 events in the 1220 cell- type specific genes that were marked in their proximal promoters by H3K4me3 only in Treg cells and not in aTconv cells. Apart from the H3K4me3 islands in promoters, there were about 60% H3K4me3 islands located in non-promoter regions, a finding which may be indicative of enhancers. Two regions of particular interest were the intragenic H3K4me3 islands located about 6 kb (ChrX:49001600-49002200) and 4 kb (ChrX:49004100-49005100) downstream of the FOXP3 promoter in Treg cells. By using online tool “TFSEARCH: Searching Transcription Factor Binding Sites (ver 1.3)”, we found both islands contained multiple TF target sites, including those for p300, AML1 and STATs. As such, this region may serve as an enhancer to regulate the transcription of the FOXP3 gene in Treg cells. ## Genome-wide maps of H3K4me1 modifications in human CD4⁺CD25⁺ T cell lineages with or without FOXP3 expression Previous studies have suggested that H3K4me1 at promoter-distal sites is often associated with the presence of an enhancer. We, thus, generated genome-wide H3K4me1 maps in human Treg and aTconv cells to compare the predicted enhancers in both cell types. Results showed that more than half of the total identified H3K4me1 islands were located in introns in both aTconv and Treg cells. Interestingly, examination of those H3K4me1 tags found within gene bodies and their 5′- and 3′-end 5 kb extended regions also revealed that the H3K4me1 enrichment status of proximal promoters was higher than those in other regions. When comparing the H3K4me1 enriched regions in Treg and aTconv cells, we found that the coefficient correlation was only 0.48 and there were only 8897 overlapping H3K4me1 islands present among the 115391 total regions between both cell types. These results indicated that most of the H3K4me1 islands were cell- type specific. More importantly, they suggested that enhancers represent the most variable class of transcriptional regulatory element between Treg and aTconv cells, and were probably primary mediators of Treg cell-type specific patterns of gene expression. **H3K4me1 modifications of cell signature genes and verification of enhancer activity** Among the 18081 total genes that were H3K4me1 modified in Treg cells, we selected a subset of the cell signature genes to further examine the H3K4me1 patterns and verify the activities of enhancers predicted to be related to these genes. The signature gene subset included IL2RA, CTLA4, TNFRSF18 and FOXP3 genes, which are known to be highly or specifically expressed in Treg cells. We identified some Treg cell-specific H3K4me1 regions, including: a region in intron 1 of FOXP3 that was also enriched by H3K4me3 \[ChrX:49001620–49002192\]; a region in the last exon of FOXP3 \[ChrX:48994400–48995097\]; a region in intron 2 of the CTLA4 \[Chr2:204444600–204445077\]; a region in intron 1 of IL2RA \[Chr10:6136100–6136695\] and a region upstream of IL2RA \[Chr10:6148000–6148784\]. We were unable to identify any Treg cell-specific H3K4me1 regions for the TNFRSF18 gene. Previous studies have suggested that H3K4me1 at promoter-distal sites are often associated with enhancer function. A general property of such enhancers is the ability to increase transcriptional activity in a heterologous context. As this type of function can be readily studied using traditional reporter gene assays, we selected the five Treg cell-specific H3K4me1 regions described above to evaluate their heterologous enhancer activities. As shown in, only two of the five regions examined showed enhancer activity. Interestingly, the majority of regions that did not show enhancer activity in Jurkat cells corresponded to Treg cell-specific H3K4me1 enriched regions. In line with this finding, a H3K4me1 region in intron 1 of the IL2RA gene \[Chr10:6131603–6132187\] and a H3K4me1 region upstream of the TNFRSF18 gene \[Chr1:1133645–1134389\], which were enriched in both Treg and aTconv cells, did exhibit enhancer activity in Jurkat cells. Since Jurkat T cells represent a leukemic counterpart of conventional T cells, it is very possible that they lack Treg cell-specific TFs that are necessary for enhancer functions of these particular regions. However, some Treg cell-specific H3K4me1 regions did function even in Jurkat cells, suggesting that the relevant TFs required for enhancer activity at these sites were, at least, available. ## Comparison of H3K4me3 and H3K4me1 enriched regions in Treg or aTconv cells We also compared the H3K4me1 and H3K4me3 enriched regions in the same sample, and determined that the coefficient correlation was only 0.16 in Treg cells and 0.19 in aTconv cells. Furthermore, there were only 5030 overlapping H3K4me1 regions and 7063 overlapping H3K4me3 regions. These results indicated that H3K4me1 modified regions with potential regulatory function were seldom overlapped with H3K4me3 modified regions in the whole genome of human Treg and aTconv cells. # Discussion In this study, we obtained high-purity of human CD4⁺CD25⁺FOXP3⁺ Treg cells and aTconv cells by combining *in vitro* expansion, MACS- and FACS- sorting methods. As reported by other researchers employing this technique, these cells maintained all phenotypic, functional and epigenetic Treg cell characteristics, even after extensive *in vitro* expansion. We utilized these cells for ChIP-Seq analysis to generate high-resolution maps of the genome-wide distribution of H3K4me1 and H3K4me3 in both cell subtypes. Ultimately, we identified a number of cell type- specific H3K4me1 regions and H3K4me3 marked proximal promoters in Treg cells. The majority of the differential H3K4me1 regions were found to be located in promoter-distal sites, and we selected some for verification of their enhancer activity by using reporter gene assays. CD4-positivity and CD25-positivity have long been considered as the cell- specific indicators of Treg cells. However, CD4⁺CD25^- T cells were demonstrated to be able to up-regulate their CD25 expression upon activation by antigen, indicating that CD4 and CD25 double-positive T cells actually represent a heterogeneous cell population and these surface markers are not sufficient identifiers of Treg cells. Thus, we used CD4⁺CD25⁺FOXP3⁺ triple expression to define Treg cells since FOXP3 gene expression is essential for Treg cell function. We carried out comparative analysis of the genome wide epigenetic methylation status for H3K4 in CD4⁺CD25⁺FOXP3⁺ (Treg) and CD4⁺CD25⁺FOXP3^- T cells (aTconv). The low frequency of Treg cells in normal human peripheral blood has thus far limited the detailed characterization and potential clinical application of human Treg cells. In many previous studies, Treg expansion has been carried out to obtain enough cell material to perform analysis. Unfortunately, CD4 and CD25 were usually used to identify the Treg cells. Here, we found that although CD4⁺CD25⁺ T cells were expanded up to 1000-fold, most of the expanded cells were FOXP3-negative. Thus, we performed FACS-sorting to obtain high-purity Treg cells with CD4, CD25 and FOXP3 expressions immediately prior to our ChIP-Seq assay. Previously, Heintzman determined the chromatin modification states at high resolution along 30 Mb of the human genome, and found that active promoters were marked by H3K4me3 and enhancers by H3K4me1. We also found that most proximal promoters enriched by H3K4me3 were common between the Treg and aTconv cells, suggesting that the related genes of the proximal promoters were co-expressed in the two lineages. Although some genes are widely used as markers for Treg cells, such as IL2RA, CTLA4 and TNFRSF18, accumulating evidence has unfortunately suggested that these markers are not strictly Treg-specific. Upon activation, all T cells express CD25, the alpha-chain of the IL-2 receptor, and its combination with IL-2 is essential for T cell clonal expansion. CTLA-4, which is the receptor for APC-B7, negatively regulates the IL-2 production of the newly activated T cell and inhibits further T cell proliferation upon binding of B7 and is up-regulated on all CD4⁺ and CD8⁺ T cells, two to three days following activation. Similarly, the expression of TNFRSF18, which is a possible target molecule in cell contact-dependent suppression, is induced in T cells upon activation. This could explain why we observed H3K4me3 in the proximal promoters of these genes in aTconv cells. STAT family TFs are critical for T cell differentiation; however, their expression is not sufficient to drive lineage commitment. Consistent with the ubiquitous expression patterns of STAT family TFs, we found that most STATs were marked in their promoter regions by H3K4me3, in both Treg and aTconv cells. Based on these results, we predicted that the common 19927 genes between Treg and aTconv cells may be expressed in both lineages. We found that, apart from the common H3K4me3 promoters, there were also some Treg cell-type specific proximal promoters marked by H3K4me3, such as FOXP3. It may be these types of specific proximal promoters, especially the FOXP3, that are responsible for the differences between Treg cell and aTconv cells. The proximal promoter of FOXP3, which is believed to serve as a master regulator of Treg cells, was found to be enriched by H3K4me3 in Treg cells. Moreover, a 50-fold higher mRNA expression level was observed in Treg cells, as compared to aTconv cells. This nearly exclusive expression of FOXP3 in Treg cells was in accordance with the current concept that FOXP3 represents the critical TF of Treg cells. In addition, we also found that the CCR7 gene was marked by H3K4me3 in its proximal promoter only in Treg cells, and exhibited a nearly 20-fold increase of mRNA expression in Treg, as compared to aTconv cells. H3K4me3 is usually associated with promoters, and its occurrence at enhancers remains a topic of debate. Whereas Heintzman, *et al*. found little or no H3K4me3 at p300-associated enhancers, Barski, *et al*. identified all three methylation states at the related functional enhancers. It is, therefore, unclear whether the promoter-distal H3K4me3 sites identified in this study are associated with uncharacterized functional transcription units, or whether they are able to act as enhancer regions themselves. For example, we found there was a region located about 6 kb downstream of the FOXP3 promoter (ChrX:49001620–49002192), which showed enhancer activity in transient transfection assays; the existence of this region suggests that there may be some non-promoter H3K4me3 regions associated with enhancers. Another region located about 4 kb downstream of the FOXP3 promoter was specifically enriched by H3K4me3 in Treg cells (ChrX:49004128–49005080), and also exhibited enhancer activity; interestingly, previous studies have shown that this region was enriched for STAT5 consensus sites. Treg cell survival critically depends on interaction with IL-2. The TF STAT5 is activated through the IL-2 receptor, has an essential role in Treg cell homeostasis, and is known to regulate the lineage-specific TF FOXP3 through an intronic, methylation-sensitive enhancer. Together, all the data indicate that certain promoter-distal H3K4me3 modified regions may have enhancer activity. Moreover, it is likely that some of the 27000 Treg cell-type specific H3K4me3 non-promoter regions that were identified in this study might be important for Treg cell-type specific patterns of gene expression. Although the promoter region represents a primary element of gene expression, it is controlled by distal regulatory elements like enhancers and silencers. Previous studies have shown that H3K4me1 at promoter-distal sites was often associated with enhancer function. Our results indicated that most of the H3K4me1 islands were cell-type specific, suggesting that enhancers are the most variable class of transcriptional regulatory element between Treg and aTconv cells, and are probably of primary importance in driving Treg cell-type specific patterns of gene expression. Our study identified a number of putative regulatory elements for genes that are highly important for Treg cell functions. For instance, we found that there was a region located in intron 1 of the FOXP3 gene (ChrX:49001620–49002192) enriched by both H3K4me1 and H3K4me3 in Treg cells and which showed enhancer activity in transfected Jurkat cells. A region located upstream of IL2RA (Chr10:6148000–6148784) also showed enhancer activity. Since cultured and expanded conventional T cells express high levels of CD25 as a consequence of TCR activation, it is possible that this region may contribute to regulating constitutive (rather than activation-induced) CD25 expression in Treg cells. In addition, we found that most H3K4me1 enriched regions were not enriched by H3K4me3, suggesting that most potential regulatory elements were only enriched by H3K4me1 but lacked H3K4me3 in the whole genome of human Treg and aTconv cells. This finding is consistent with the observations of p300-associated enhancers that were found to have little or no H3K4me3. However, there were also some regions simultaneously enriched by H3K4me1 and H3K4me3, such as the region located in intron 1 of the FOXP3 gene, which did show enhancer activity. Whether or not the regions enriched by the two types of histone methylations may harbor more potential to act as enhancers remains unknown. In conclusion, we identified genome-wide H3K4me1 and H3K4me3 modification regions in Treg and aTconv cells. The H3K4me3 modifications located in proximal- promoter regions were nearly identical in both Treg and aTconv cells, with the exception of a few promoters of genes, such as FOXP3 and CCR7, which are expressed uniquely in Treg cells. In contrast to the H3K4me3 modification, H3K4me1 exhibited cell-type specific locations, indicating that enhancers are the most variable class of transcriptional regulatory elements between Treg and aTconv cells. Furthermore, enhancers are likely to be of primary importance in driving Treg cell-type specific patterns of gene expression. The Treg- and aTconv-specific H3K4me1 and H3K43 patterns may function as significant mediators of differentiation events, lineage commitment and cell type-specific gene expression. It is likely that this basic principle is not confined to these two closely related T cell populations, but may apply generally to somatic cell lineages in adult organisms. # Supporting Information [^1]: Conceived and designed the experiments: YZW BN JCV YW. Performed the experiments: YZ ZJ ZH JT Y. Zheng Y. Tian QW ZT DY Y. Zang XF SL Y. Tang JS YW. Analyzed the data: JW BN. Wrote the paper: JCV Y. Tian ZJ JW YW BN. [^2]: The authors have declared that no competing interests exist.

# Introduction Keratins are filament forming proteins of epithelial cells and are essential for normal tissue structure and function. In contrast to actin filaments and microtubules, keratins are encoded by a large family of genes clustered at two divergent chromosomal sites: 17q21.2 (type I keratins, except K18) and 12q13.13 (type II keratins, including K18). These are also expressed in tissue and differentiation state-specific manner and play an important role in protecting epithelial cells from mechanical and non-mechanical stress and injury. Epithelial tumors continue to express keratins that are characteristic of their site of origin and therefore keratins are extensively used as immunohistochemical markers in diagnostic tumor pathology. Accumulating evidence points to the importance of keratins as prognostic markers and, more interestingly, as active regulators of epithelial tumorigenesis and treatment responsiveness. Previous studies have reported alterations in keratin expression during oral carcinogenesis,. Further, many keratins are recognized as independent markers of prognosis in OSCC. Within the oral cavity there is a complex pattern of keratin expression, reflecting both the type of epithelium and stage of differentiation specific expression. The basal proliferative layer of all oral epithelia expresses K5/K14 and K19. The suprabasal, differentiating layers of keratinized (cornified) epithelia express K1 and K10, while the differentiating layers of non- keratinized epithelia such as buccal mucosa and esophagus synthesize predominantly K4 and K13. Suprabasal epithelial cells of the hard palate and gingiva express K6, K16, and K76. Previous studies have reported altered terminal differentiation and keratin expression patterns in oral tumors, such as downregulation of K4, K5, K13 and K19. Conversely, increased expression of K8/K18, K17 and K14 is reported in oral tumor tissues compared to the normal counterparts. Various studies using *in-vitro* system have elucidated mechanistic role of keratins (K8/18, K19) in tumor invasion and metastasis. However, *in-vitro* data may not fully reflect the *in-vivo* condition. Interestingly, alterations of keratin expression pattern marks the common signature in human oral cancers and experimental oral tumors developed in animal models. Hence, we selected *in-vivo* model systems: the hamster model to demonstrate K76 downregulation during sequential progression of oral cancer, and the KO mice model to evaluate the effect of *KRT76* loss. Gene expression analysis from our laboratory has revealed downregulation of *KRT76* in tumors of the oral cavity. *KRT76*, a type II epithelial keratin (previously designated as K2p), is specifically expressed in the suprabasal cell layers of oral masticatory epithelium (the slightly orthokeratinized stratified squamous epithelium lining the gingiva and the hard palate). We now present data indicating that *KRT76* is downregulated prior to tumor development and its potential association with hyperproliferation in the formation of preneoplastic lesions. # Materials and Methods ## Human Tissue Specimen Collection The Institutional Review Board and the Local Ethics Committee of Tata Memorial Hospital (TMH) and Nair Hospital Dental College, approved the study. Written informed consents were obtained from all the study participants. Treatment naive neoprimary frozen tissues (n = 57) and paraffin embedded tissue blocks (n = 102) of different cohort of patients with gingivobuccal cancer (GBC) were obtained from the ICMR National Tumor Tissue Repository and Department of Pathology TMH, Mumbai respectively. Precancerous lesions (incident leukoplakia cases which are histopathologically hyperplastic lesions with focal mild to moderate dysplasia n = 61), independent normal tissues (n = 35), and inflamed tissues not associated with oral malignancy or pre-malignant conditions (n = 7) were collected from the Department of Oral Pathology, Nair Hospital Dental College, Mumbai; all these tissues were from gingivobuccal region. Tumor tissues with more than 70% tumor content were subjected to RNA extraction. ## Animal Models The study on hamsters was conducted after approval from the Institutional Animal Ethics Committee (IAEC) of ACTREC, endorsed by the Committee for the Purpose of Control and Supervision of Experiments on Animals (CPCSEA), Government of India guidelines. Inbred male Syrian hamsters (6–8 weeks old; Animal house, ACTREC, India) were randomized (10 animals per group) and maintained under standard conditions: 22±2°C, 45% ±10% relative humidity, and 12-h light/dark cycle (7∶00 to 19∶00 light; 19∶00 to 7∶00 dark). The animals received an autoclaved standard pellet diet and plain drinking water *ad libitum.* Hamsters (3–5) were housed in the polypropylene cages provided with autoclaved rice husk bedding material available locally. The hamsters were topically treated with 7,12-dimethylbenz\[α\]anthracene (DMBA) (0.5%) in corn oil using a Gilson pipette (80 µl ≈ 0.4 mg) on their right buccal pouch, thrice a week for 16 weeks. The ‘corn oil’ was used for the treatment in vehicle control group. Animals in all groups were observed for apparent signs of toxicity such as weight loss or mortality during the entire study period. Following 1, 2, 4, 6, 8, 10, 12 and 16 weeks of DMBA applications, hamsters were euthanized (by CO₂ chamber) 24 h after the last DMBA dose. Their buccal pouches were excised and fixed in 10% buffered formalin. The animal research ethical review committees of the Cancer Research UK Cambridge Research Institute and Cambridge University approved all the studies involving mice. *KRT76*-KO mice were obtained from the Wellcome Trust Sanger Institute (<http://www.sanger.ac.uk/mouseportal/search?query=KRT76>), and were maintained under the terms of a UK Government Home Office license 80/2378 (license holder Fiona M. Watt). ## RNA Isolation from Tissues RNA was isolated from human tumor and normal tissues using the RNeasy mini kit (Qiagen, Germany) according to the manufacturer’s protocol. Briefly, 15–20 mg tissue was pulverized by grinding with liquid nitrogen, followed by addition of RLT buffer with β-mercaptoethanol (Sigma-Aldrich, USA). The homogenate was processed for column purification and isolation of RNA. DNA contamination was avoided by treating the column with RNase free DNase I (Ambion, USA). The quantity and quality of RNA was determined using Nanodrop ND-1000 (NanoDrop Technologies, Wilmington, DE, USA) and RNA 6000 Nano LabChip Kit on an Agilent 2100 Bioanalyzer (Agilent Technologies, CA) respectively. ## Quantitative Reverse Transcriptase-Polymerase Chain Reaction (qRT-PCR) For complementary DNA (cDNA) synthesis, 1.5 µg of total RNA was reverse- transcribed with the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems, USA) following the manufacturer’s protocol. Twenty ng of cDNA were used for TaqMan qRT-PCR analysis and experiments were performed in duplicate (*KRT76* Assay Id: Hs00210581_m1, 18S RNA Assay Id: Hs99999901). Results were analyzed using SDS 2.3 and RQ manager software (Applied Biosystems). The relative expression of *KRT76* messenger RNA (mRNA) was determined using 18S ribosomal RNA as an endogenous control. These were compared between GBC cancers and unrelated normal tissues from the same site. The expression of *KRT76* in each sample was analyzed using the comparative CT method (also known as the 2^−ΔΔCT method) where ΔΔCT = \[CT gene of interest − CT internal control (18S)\] of test sample – \[CT gene of interest − CT internal control (18S)\] of reference sample. Fold change values for qRT-PCR data were calculated as 2^−ΔΔCT. ## Immunostaining of K76 in Human Oral Tissues Formalin-fixed, paraffin-embedded GBC tissues (n = 102), OPLs (n = 61) and normal oral tissues (n = 21) were used for immunohistochemical (IHC) analysis. Five micron tissue sections were deparaffinized with xylene, rehydrated with sequential ethanol washes (100%, 90% and 70%). To quench the endogenous peroxidase activity, sections were incubated with 3% hydrogen peroxide in methanol for 30 min in dark. After heat based antigen retrieval with sodium citrate buffer (pH = 5.8), sections were incubated with normal horse serum. The sections were incubated overnight with rabbit polyclonal anti-human K76 antibody (1∶225, HPA019696, Sigma-Aldrich) at 4°C. For negative or isotype control, the primary antibody was replaced with rabbit serum used at respective antibody concentration. Sections were then incubated with biotinylated universal secondary antibody solution for 30 min followed by incubation with VectastainVR elite ABC reagent for the same time. The immunoreaction in tissue sections was visualized using 3,3′–diaminobenzidine tetrahydrochloridehydrate (Sigma- Aldrich). The slides were finally counterstained with hematoxylin and examined under microscope. For immunofluorescence, deparaffinization and antigen retrieval steps were similar to those for IHC. Tissues were fixed in cold methanol for 10 min followed by blocking with 5% normal goat serum, 0.3% (v/v) Triton X-100 in PBS for 1 hr at room temperature. Tissues were next incubated with K76 antibody at a dilution of 1∶250 overnight at 4°C, followed by incubation with an Alexa Fluor 488 anti-rabbit antibody (Life technologies, USA) at 1∶200 dilution, for 1 hr at room temperature. Cells were counterstained with DAPI and viewed under a fluorescence microscope (Ziess; LSM-510 Meta Germany). ## Immunostaining of K76 in Animal Models Formalin fixed hamster buccal pouch tissues were used from the following experimental groups for IHC analysis: 1) Control group: 1^st, 2^nd, 4^th, 6^th, 8^th, 10^th, 12^th and 16^th week hamsters buccal pouch topically treated with vehicle (no DMBA); 2) DMBA treated group: 1^st, 2^nd, 4^th, 6^th, 8^th, 10^th, 12^th and 16^th week hamsters buccal pouch topically treated with DMBA. Formalin fixed tissues from *KRT76*-Wild type (WT) and *KRT76*-KO mice were used for immumostaining and histopathological analysis. For experimental models, the IHC staining procedure was similar to that described earlier with minor changes in blocking, which was performed with 3% BSA and 2% goat serum; while secondary antibody was biotin conjugated anti- rabbit secondary raised in goat (Santa Cruz Biotechnology, USA). ## Immunohistochemical Assessment and Scoring For assessment of K76 protein expression, the cytoplasmic staining intensity was categorized as 0 (absence of staining in any cell), +1 (weak staining in less than 10% of cells), +2 (moderate staining and/or 10 to 50% of positive cells), or +3 (strong staining in more than 50% cells) by pathologist (AP). For further statistical analysis, the stained tissues were categorized in two groups: 0 and +1 as mild to no expression, while +2 and +3 as moderate to strong expression. ## Statistical Analysis All statistical analyses were performed using IBM SPSS version 21. The Mann Whitney test was performed to analyze the difference between ΔCT values of tumor and normal samples obtained by qRT-PCR. The Chi-square test was used to determine the correlation between expression levels of K76 protein and tissue type, as well as clinicopathological characteristics. Polytomous logistic regression was used to evaluate the relationship of protein expression scores to the risk of OPL and OSCC development, with normal tissue as a reference; odds ratio (OR) were computed by adjusting for age and gender. Disease-specific survival (DSS) was calculated as the time from surgical diagnosis to the date of death due to cancer or to the last clinical follow-up prior to death. DSS was examined visually with Kaplan-Meier curves and analyzed by log rank tests. All p-values \<0.05 were considered statistically significant. # Results ## Patient Characteristics The clinicopathological and demographic characteristics of all OPLs and tumor samples are summarized in. The patients in this study cohort were predominantly male tobacco habitués and tobacco chewing was the most prevalent habit. Most of the tumor samples were of moderate or poor grade, and mainly of pTNM stages III or IV. Approximately 50% of the cases showed lymph node invasion. Majority of OPLs had mild to severe hyperplasia and few showed presence of focal mild to moderate dysplasia. ## Validation of Microarray Results by qRT-PCR Microarray analysis of 27 GBC cases showed a significant downregulation of *KRT76*, as reported previously. We observed downregulation of many genes associated with structural molecule activity Gene Ontology: 0005198 of which *KRT76* showed the highest fold change. The Oncomine data source illustrated two more studies reporting consistent downregulation of *KRT76* in OSCC. To confirm the findings of the microarray analysis, we performed qRT-PCR using primers specific for *KRT76* in 57 OSCC and 14 normal tissues. qRT-PCR analysis revealed significant downregulation of *KRT76* RNA in tumor samples compared to normal samples. ## Sequential Downregulation of K76 in Oral Carcinogenesis K76 expression was analyzed in 184 oral tissues by immunohistochemistry. Normal gingivobuccal tissues expressed higher levels of K76 protein compared to OPL and invasive OSCC. Distribution of K76 expression was confirmed by immunofluorescence as illustrated in. Normal oral epithelium showed K76 expression confined to the suprabasal, differentiating cell layers while, there was a gradual overall loss of K76 expression in OPLs and tumors. The frequency of K76 positive staining significantly decreased across the transition from normal tissue (100% positive) to OPL (44%) to oral tumor (35%). To examine whether *KRT76* downregulation was associated with benign epithelial hyperproliferation (injured normal tissue without any association with oral preinvasive and invasive lesions), we performed IHC on inflamed buccal mucosa (n = 7). Even though these epithelia histologically appeared hyperproliferative, K76 staining was consistent with that seen in normal buccal epithelium. These results indicates that downregulation of *KRT76* expression is not associated with injury related proliferation and acute inflammation. ## Correlation of K76 Expression with Clinicopathological Parameters Statistical analysis to determine the association of K76 expression and different clinical parameters, such as node, stage, grade, habit profile and outcome (recurrence and survival) was performed. Reduced expression of K76 showed a very weak association with survival (p = 0.096), whereas other parameters analyzed did not show any association. Polytomous Logistic regression with normal as the reference group showed a significant correlation of K76 downregulation with risk of developing OPL (p = 0.002) and OSCC (p≤0.0001). ## Loss of K76 Expression in an Experimental Model of Oral Carcinogenesis K76 expression was analyzed by IHC in the buccal epithelium of DMBA treated hamsters (group details described in methods). Interestingly gradual decrease in staining intensity was observed with disease progression in hamster buccal epithelium. Irrespective of duration of treatment, control group showed higher levels of K76, while reduced expression was observed in premalignant lesions and oral tumors, which was similar to that seen in human hyperplastic lesions and OSCC. ## Mice Lacking *KRT76* Develop Hyperplastic Oral Lesions To determine whether loss of *KRT76* is sufficient to induce premalignant lesions in the oral cavity, we examined the oral epithelia of *KRT76*-KO and *KRT76*-WT mice. Immunohistochemical analysis showed specific K76 staining in buccal epithelium of WT mice, whereas no staining was observed in *KRT76*-KO buccal epithelium, confirming specificity of K76 antibody ( A, B). Histological examination of the buccal mucosa of *KRT76*-KO mice showed development of hyperplastic lesions along with increased keratinization across the epithelium, which was not observed in *KRT76*-WT mice ( C, D). In contrast, the epithelium of the dorsal tongue, which is normally *KRT76*-negative, exhibited normal homeostasis in *KRT76-*KO mice indicating that *KRT76* loss associated abnormalities are highly sub-site specific in oral cavity. However, none of the *KRT76*-KO mice in the entire life span developed spontaneous oral tumors. # Discussion Deregulated keratin expression is associated with impaired epithelial differentiation and organization during OSCC progression. Our microarray based gene expression profile of 27 advanced stage gingivobuccal cancers previously revealed deregulation of several keratins, namely *KRT4, KRT13, KRT19, KRT76*, which are normally expressed in the oral cavity. *KRT76* was found to be the topmost downregulated gene amongst all differentially expressed genes. Gene expression profiles of oral cancer obtained by other groups have also shown consistent downregulation of *KRT76*. We now report, for the first time, differential expression of *KRT76* in human and hamster oral precancerous and cancerous lesions, and show that loss of *KRT76* is sufficient to cause hyperplasia in the oral cavity of the mice. We validated our previous microarray findings in an independent patient cohort by qRT-PCR and IHC; both these techniques showed reduced expression of *KRT76*. While previous reports have demonstrated changes in keratin gene expression associated with severe dysplasia and poorly differentiated SCC, reflecting gross changes in epithelial differentiation and maturation, our studies are the first to indicate that loss of a specific keratin is sufficient to initiate preneoplastic changes. We did not find association of K76 downregulation with clinicopatholgical parameters such as node, grade, clinical outcome; nor with benign inflammation-associated hyperproliferation. Although, the fact that K76 downregulation is observed in leukoplakia, a preinvasive oral lesion and is sustained during the development of frank malignancy, indicates its association with the early stages of oral carcinogenesis. Interestingly, we observed gradual decrease in K76 expression during the sequential process of tumor development in DMBA treated buccal epithelium of hamster. The K76 downregulation was consistent with human OPL and OSCC. Although hamster cheek pouch model has several areas of uniqueness, it also lacks lymphatic drainage as observed in humans, mice, or rats, which makes it immunoprotected. However none of the existing animal models in studies on oral cancer are fully satisfactory and simulate tobacco chewing. Hamster is one of the extensively used models, as the oral epithelium has similar histological and genetic events involved in the development of premalignant lesions and tumors as in humans. In order to investigate the effect of *KRT76* loss, we used *KRT76*-KO mice. The transgenic and knockout mouse models provide unique advantage of genetic manipulation of specific target gene/s, it also has similar intracellular signaling pathways as of humans. In-vivo systems over comes the weakness of in- vitro experiments which fails to replicate the complex cellular and tissue interaction in an organism; hence, better suited for observing the overall effects of a target gene in a living system. *KRT76*-KO mice displayed hyperplastic changes in buccal epithelium, however they do not spontaneously develop tumors similar to previous reports on other keratin knockout mice models. Our current findings suggest that the loss of *KRT76* may not be a sole molecular event leading to oral cancer development. However, the hyperplastic changes observed in *KRT76*-KO mice points to an indirect role of *KRT76* in regulating proliferation of the basal layers of buccal mucosa similar to previous findings of *KRT10* loss. Overall, our data implies the fact that carcinogenesis being multifactorial and multistep process, potential role of *KRT76* as one of the factor, which alone is not sufficient for cell transformation; however, its contribution in oral carcinogenesis cannot be ruled out. We envision a number of possible ways in which *KRT76* loss contributes to cancer development. One is that it contributes to a barrier defect in the epithelium, which may render the tissue more susceptible to penetration by carcinogens. Another is that *KRT76* loss may lead to a disturbed inflammatory infiltrate; which is observed in human and mouse epidermis on loss of structural proteins. We did not see loss of *KRT76* in benign hyperproliferative oral epithelium, with associated inflammation, nevertheless, altered immune infiltrates are a hallmark of OSCC. Future investigations are needed to assess the impact of *KRT76* loss in predicting high-risk precancerous lesions of oral cavity. We observed *KRT76* downregulation in patients with gingivobuccal cancers – a sub site of oral cancer, which is etiologically associated with peculiar tobacco and betel quid chewing habit common in India. These results have to be generalized with caution to other etiologies associated with development of oral tumors. Although, *KRT76* loss is characteristic of gingivobuccal tumors it is not associated with cell transformation, our results warrant future studies to understand other key players driving the process of oral carcinogenesis. # Supporting Information The authors thank all participants of the study. Mrs. Sadhana Kannan is acknowledged for her help in statistical analysis. ICMR National Tumor Tissue Repository, Tata Memorial Centre, Mumbai is acknowledged for providing tumor tissues. The authors sincerely acknowledge Dr. Miriam Rosin and Dr. Hector Hernandez-Vargas for their critical suggestions in improving the manuscript. [^1]: All the authors have approved the submission and have declared no conflict of interest. [^2]: Conceived and designed the experiments: SA PGB MM. Performed the experiments: SA PGB MP. Analyzed the data: SA PGB AP. Contributed reagents/materials/analysis tools: EH GK SK GBM RSD FMW MM. Wrote the paper: SA PGB GBM MM. [^3]: Current address: Epigenetics Group, Section of Mechanisms of Carcinogenesis, International Agency for Research on Cancer, Lyon, France

# Introduction Understanding mosquito ecology has recently been prioritized as a prerequisite for malaria eradication. Ferguson *et al.* stressed that our knowledge of mosquito ecology is minimal compared to that of other agricultural pests and model organisms, and suggested the reasons for this are institutional compartmentalization and cultural effects, research having focused on medical issues, largely overlooking the mechanisms and ecology of vector transmission. As mosquito vectors are embedded within ecological communities as predators, prey and competitors, an understanding of their ecology is essential to avoid any interventions triggering cascades of ecological effects that could lead to enhanced malaria transmission. With over thirty different primary vectors dominating transmission, an understanding of the competitive interactions and species specific niche adaptations is critical for effective vector management. Although many studies have shown that the growth rates of larval mosquito vectors are negatively correlated with their population size, resulting in smaller, more robust and fecund populations, the mechanisms underlying this plasticity are largely unexplored. Body size has also been shown to have important fitness implications, however individual body size frequency distributions within a population remain under-investigated in insects in general. One of the key factors controlling population dynamics and body size is larval nutrition, and previous studies have shown that nitrogen (N) and phosphorus (P) availabilities are important ecological determinants in other insects. However, it is extremely difficult to study nutritional impacts on such small insects and generally methods of analysis are laborious and complex, often limiting the scope of the studies conducted. Here we present some rapid techniques that may overcome some of these constraints opening up opportunities for more holistic ecosystem based research. Advances in elemental analysis and pyrolysis techniques to measure fatty acid concentrations, mean that it is now possible to investigate nutritional impacts on mosquito larvae development and survival on an individual basis. This allows us to explore mosquito larval development within the larger ecological framework and relate it to current paradigms in ecological thinking, such as ecological stoichiometry. Ecological stoichiometry has been heralded as the unifying theory of ecology. It is based on simple laws of physics such as mass balance and energy dissipation meshed with the biological principles of energy tradeoffs at biochemical and individual levels. These principles have been cleverly honed to explain the dynamics of individuals, populations, communities and ecosystems. At the very base of ecological stoichiometry theory is the concept that at the organism level there is a unique balance of multiple chemical substances, mainly ratios of carbon:nitrogen:phosphorus (C:N:P) and the consequence of this homeostasis is that nutrient cycles and processes at higher scales in the ecosystem are driven. Fundamentally the theory suggests that living organisms are constrained and different from their environment, and in almost all circumstances will be limited by one element; usually but not exclusively, nitrogen or phosphorous. Although this is a universal phenomenon, little is known of the extent to which stoichiometry drives population dynamics and its consequences for general mosquito biology. Stoichiometric theory contrasts to the current theory that mosquito larval nutrition is a complex combination of dietary requirements. In this study we set out to test whether these theories hold up for *Anopheles arabiensis* mosquitoes and whether they might explain some observed phenomenon of population plasticity. Mosquito larval nutrition has been extensively studied; it is known that proteins (or amino acids), sugar (glucose or sucrose), polyunsaturated fatty acids (PUFAs), sterols, vitamins and nucleotides are all essential for mosquito development. It has been shown that at least fourteen amino acids are essential for larval growth and survival. Additionally, minimal concentrations of essential vitamins are required to ensure optimal growth of several mosquito species. Dadd & Kleinjan also showed that less that 5% of *Cx. pipiens* larvae reached adult stage in diets lacking a combination of three nucleotides, demonstrating their role in nutrition. It is well documented that mosquito larval diet quality and quantity influences both adult quality and in turn sexual competitiveness. Efficient and economic mass rearing of any insect requires an in-depth understanding of the dietary components which influence insect quality. A broad understanding of dietary requirements and influences can also yield an interesting insight into the natural ecology and biology of the insect. For example, laboratory experiments have shown that supplementary protein feeding of fruit flies led to more successful mating behavior, a critical issue in both insect ecology and sterile insect technique programmes. The stoichiometric paradigm suggests that many of the reductionist investigations that determine the specific chemical requirements for the successful nutrition of an organism, overlook the ubiquitous presence of the individual components in the ecosystem as a whole. It suggests in fact that systems are generally constrained by specific macro-nutritional requirements which have the individual components embedded within them, and that primary producers and thus secondary and higher level consumers are ultimately constrained by biogeochemistry. It is well documented that primary producers respond positively to inputs of nitrogen and phosphorous, as these are the limiting elements in most natural systems. Extending from this, evolutionary theory states that the fittest individuals will use the available resources for reproduction most efficiently and therefore their genes will dominate. Evolutionary logic would suggest that generalist secondary consumers would thus adapt to utilize the most commonly present components in primary producers and that these would be used in the most energy efficient manner. Indeed current trophic interaction research suggests that it is energetically more efficient to incorporate dietary fatty acids (FAs) directly into the consumer’s tissue without degradation or modification, a process termed dietary routing. Fatty acid profiles provide a large amount of information on the development, reproduction, health and feeding ecology of organisms. Previous studies have stressed that mosquitoes are unable to elongate the 18C poly unsaturated FAs and thus C18, C20 and C22 polyunsaturated fatty acids are essential for larval development, adult survival and flight. However in natural aquatic environments these fatty acids may be present and are possibly not limiting constraints of larval nutrition *per se*. In most ecological systems energy transfer and energy flux is thought to be the primary constraint on total ecosystem productivity. Here we present a framework with which we test the hypothesis that dietary quality, or nutrient content, i.e. stoichiometry, plays a significant role in regulating ecosystem energy and nutrient transfers in primary consumers and can be used as a predictive tool of population response. Studying larval nutrition within the context of developing diets for mass reared anopheles mosquitoes gave us the opportunity to test a number of hypotheses that would support or discredit the stoichiometry theory. The pyrolysis GCMS system offered us the chance to engage in a simple comprehensive analysis of the fatty acids in the diet and their effects on individual mosquito FA profiles. Elemental analysis and isotope ratio mass spectrometry of whole mosquitoes meant that we could compare whole body carbon, nitrogen and phosphorous and macro nutrient levels, in addition to the fatty acid profile of the diet, giving us insight into the important features of diet composition. Finally, the isotope analysis enabled us to investigate some of the current assumptions in isotope ecology and test those assumptions in controlled environments, allowing us to confidently apply these techniques in future field studies. Central null hypotheses we set out to explore were: 1. Fatty acid composition of mosquitoes is fixed and not influenced by the diet. 2. Mosquito stoichiometry is fixed and not influenced by the larval diet. # Materials and Methods ## Mosquito Stocks and Rearing Methods All experiments were conducted at the Insect Pest Control Laboratory (Joint FAO/IAEA Division) in Seibersdorf, Austria, in climate-controlled rooms maintained at 27°C ±1°C and 60% RH ±10%, with LD 12∶12 h photoperiod, including dusk (1 h) and dawn (1 h). A stock strain of *Anopheles arabiensis* MRA-856 (available from MR4, MRA-856), was used in all the experiments. Having originated from Dongola, Northern Sudan (2005) the strain has since been maintained on a Koi Floating Blend® diet for approximately 105 generations. Eggs were hatched at a low density. ## Experimental Designs ### Experiment 1. Initial investigations of TBN, TBC, C:N ratios and wing length This experiment set out to investigate the relationship between total body carbon (TBC) and nitrogen (TBN) and wing length, thus a range of diet concentrations were fed to the mosquitoes to get a range of mosquito size classes, but basically fed on the same diet. Less than 4 hours after eclosion, 32 larvae (1^st instar) were transferred into six 9 cm diameter Petri dishes containing 32 ml of deionised water and these were fed daily with 2 ml of a 1, 1,5, or 2% solution of a KD1 diet which was a mixture of ground wheat, corn, bean, chick pea, rice, bovine liver powder (BLP) and Vita mix in the following ratio: 2∶2:2∶2:2∶2:2.6. Six replicate dishes were set up for each treatment. All pupae were collected daily and live un-fed teneral adults collected within 12 hours. Wing length was determined for each specimen: briefly, a wing was clipped and mounted on a slide, and a digital image taken using a camera mounted on a stereo microscope (CC-12 camera, Olympus Soft Imaging Solutions). Wing length was measured from the alula notch to the wing tip; measurements were performed with AnalySIS FIVE software (Olympus Soft Imaging Solutions). Wings were re-united with their bodies in the tin cups used for analysis and thus whole adult mosquitoes could be analysed for total body carbon and nitrogen and their isotopic ratios as described below. This meant TBC and TBN could be compared against wing length at the level of individual mosquitoes. ### Experiment 2. Feeding experiments for In-depth Dietary analysis In this experiment we set out to determine which factors of nutritional quality could influence mosquito size and whether fatty acids were directly routed from diet to the consuming larva and consequently preserved in the adult mosquito. The sixteen diets tested in this experiments were AP100 (Zeilgler USA, a commercially available shrimp larval diet), bean powder, bovine liver powder (BLP), brewer’s yeast, carrot, chick pea, corn, rice, soy hydrolysate, spirulina, squid liver powder (SLP), tuna, vitamin-mix, wheat, wheat bran and yeast hydrolysate. Five hundred larvae (L1 instar) were counted into a tray (30×40 cm) containing 1.5 L of de-ionized water. Due to the sensitivities of anopheline species to overfeeding and larval habitat fouling, the mosquitoes were fed on demand (approx. 0.25 mg/larva/day). This diet was added in a ground form in quantities aiming to attain maximum adult survival based on the colour and state of the larval water and the previous experience of the technicians. Newly formed pupae were transferred to emergence tubes. Upon adult emergence, ten males and ten females were transferred to Eppendorf tubes and frozen. Care was taken to sample the first ten males and first ten females that emerged from each treatment to overcome any emergence date bias. These were randomly divided into three batches and triplicate whole mosquito samples of each sex were analysed for fatty acids (Py-GCMS), TBN and TBC and their respective ¹⁵N and ¹³C values (Elemental-IRMS), and TBP (Total body phosphorous). ### Experiment 3. Determining the influence of dietary N and P on mosquito survival and production This experiment was set up to determine the influence of dietary N and P concentration on adult and pupal survival. Triplicate sets of 16 1^st instar larvae were loaded into 35 mm diameter petri dishes containing 16 ml of water. Each dish daily received 1 ml of a 1% solution of one the sixteen different larval foods listed above. It has been shown in previous experiments that a 1% concentration was the concentration where sufficient food was available but was least likely to produce water fouling and associated effect on the population size. Pupation date was noted and adults collected as described above. ## Sample Analysis ### Pyrolysis GCMS for fatty acid analysis Typically 100 µg of diet or a complete mosquito specimen were put into a quartz tube and 4 µl of a diluted, aqueous solution of tetramethylammonium hydroxide (TMAH) was added. The samples were subsequently pyrolyzed at 450°C for 10 s with a CDS 5250 pyrolysis autosampler attached to a Thermo Trace GC Ultra/MD 800 gas chromatography/mass spectrometry system. Volatile products were separated on a Supelco SP 2330 column (30 m, ID 0.32 mm, 0.2 µm film thickness) with helium 4.6 as carrier gas (2 ml.min⁻¹) and identified by interpretation of their EI mass spectra and comparison to NIST 2002, Wiley, and NBS electronic libraries. The pyrolysis interface was kept at 300°C, the GC/MS interface at 280°C; the GC was programmed from 100°C (1 min) to 230°C (5 min) at a rate of 10°C min⁻¹. The mass spectrometer was operated in EI mode (70 eV) at a source temperature of 200°C. The method was optimised based on the standard linseed oil, a triglyceride based oil, with several unsaturated fatty acids. An optimal thermally assisted hydrolysis and methylation method was developed to avoid the known problems of isomerisation. Py-GCMS requires no sample preparation apart from the addition of TMAH. Analysis typically takes 20 minutes per sample and 100 µg C is the ideal sample size. In contrast fatty acids are conventionally measured by gas-liquid chromatography (GLC) using a flame ionisation detector, following a complex procedure of lipid extraction, purification, transesterification and methylation of approximately 50–70 mg of sample. The sample preparation procedure typically takes 2–3 days. ### Elemental and stable isotope analysis Whole single mosquito samples and ground diet samples were dried at 60°C for 24 h, placed into 8 by 5 mm tin cups and analyzed at SILVER, Vienna University for total N, C,¹⁵N and ¹³C, using an isotope ratio mass spectrometer (Delta PLUS, Thermo Finnigan, Germany) interfaced with an elemental analyzer (Flash EA, CE Instruments, UK). Samples were combusted in an atmosphere of oxygen at 1,020°C and passed over chromium oxide and silvered cobalt oxide for complete oxidation, and subsequently over hot copper (640°C) to reduce oxides of nitrogen to elemental nitrogen (N₂). The resultant gas was carried in a stream of helium through a scrubber to remove residual water and was then passed over a gas chromatographic column to separate N₂ and CO_2. Peaks were bled into the mass spectrometer to determine the isotopic ratios. A full complement of internal and external standards was run with the samples to calculate isotopic ratios, % N and % C values. The isotope ratios were expressed as parts per thousand per mille (‰) or δ deviation from the internationally recognized standards, Vienna Pee Dee Belemnite (VPDB) and atmospheric nitrogen. ¹⁵N makes up 0.3663% of all N at atoms natural abundance levels in air, the delta notation is basically deviation from this value multiplied by a thousand. Similarly the deviation from VPDB multiplied by a thousand gives the delta notation for ¹³C. ## Phosphorous Analysis Phosphorous analysis of whole mosquitoes or ground diet samples was conducted based on a wet digestion. Dried single whole mosquitoes or 3–5 mg of diet (noted to nearest µg) were placed into individual 10 ml test tubes and 1.0 ml of 98% H₂SO₄ added. The tubes were heated in an aluminium heating block until they reached a temperature of 150°C, at which point 0.75 ml of 30% H₂O₂ was added drop by drop until the solid disappeared, excess H₂O₂ was boiled off at 150°C, samples cooled and their volume noted. Aliquots (0.5 ml) of sample were diluted 1∶10 with deionised distilled H₂O and brought to pH 5 with NaOH, determined using a phenolphthalein indicator, and made up to a known volume. Samples were analysed using the microtitre plate Malachite green method. 200 µl of sample or potassium di-hydrogen phosphate (KH₂PO₄) standard was mixed with 40 µL of Reagent 1 (14.2 mmol L⁻¹ ammonium molybdate tetrahydrate in 3.1 *M* H₂SO₄) and shaken for 10 minutes. Following shaking 40 µL of Reagent 2 was added and the plate again shaken for a further 20 minutes before being read at 630 nm on a Tecan, micro-titre plate reader. Reagent 2 was prepared by adding 3.5 g L⁻¹ aqueous polyvinyl alcohol (PVA) reagent (molecular weight between 31 000 and 50 000) to 500 ml of 80°C deionised distilled water and stirring it, after cooling to room temperature, 0.35 g of Malachite Green oxalate (Merck, Art. No. 1398) was added and made up to 1 litre using distilled deionised water. All values were compared to standards and calculations done to give µg P per mosquito or % P of the diet. ## Statistical Analysis Statistical analyses were performed using Microsoft Excel, Statgraphics Plus, Centurian USA and Primer 6 version 6.18. software. In all cases, the significant alpha level was taken as *P*\<0.05. # Results ## Experiment 1. Initial Exploratory Experiment to Investigate the Relationship between Total Body Carbon and Nitrogen and Mosquito Wing Length Simple regression and multiple regression analyses were used to determine the interactions, well aware of the possible interdependence of some of the variables. There was a weak but significant correlation between total body carbon (TBC) and wing length (r² = 0.47 p\<0.0001). However, there was a stronger correlation for total body nitrogen (TBN) and wing length (r² = 0.61 p\<0.0001), and this interaction was stronger in females than males. In addition, deviation from average wing length appeared to be nitrogen (N) dependent. There was a strong correlation between TBN and TBC (r² = 0.84 p\<0.0001). In addition mosquito C:N ratio correlated weakly but significantly with wing length (r² = 0.01 p\<0.036) and more strongly with the independent variable δ¹³C (r² = 0.49 p\<0.0000); there was no correlation with δ ¹⁵N. This latter point could be explained by the fact that mosquitoes with greater N and subsequently greater C accumulation underwent enhanced lipogenesis which has been shown to be initiated in fruit flies fed N-rich diets. This theory is supported to some extent by the non-independent relationship between C:N ratio and TBC, which was best described by a polynomial function (r² = 0.48 p\<0.000) rather than a direct linear function, which would suggest lipogenesis above a threshold N value. Lipids are known to be depleted in ¹³C in relation to bulk tissues due to isotopic discrimination by key enzymes. In short this indicates that beyond a certain N threshold mosquitoes just got fatter rather than bigger. Although TBN and TBC increased linearly with emergence date, the differences between sampling dates, as determined by ANOVA, were small but significant (p = 0.0009 for TBC and P = 0.0235 for TBN). Approximately 1% of the mosquitoes emerged on the initial emergence day followed by 44% on day two, with 38, 6 and 3% emerging on the following consecutive days. Thus the slight increase in TBN and TBC could be explained by the lack of competition for resources. ## Experiment 2. Feeding Experiments for In-depth Dietary Analysis The pyrolysis method gave similar patterns of relative fatty acid composition of the three diets BLP, SLP and Tuna compared to the conventional method, revealed using simple regression analysis (r² = 0.816, 0.726, 0.778 for the three diets, respectively) as has been previously demonstrated. It was impossible to compare measurements of single mosquitoes using the conventional methods due to the constraints described in the methods section, so based on these results, previous conclusions, and the appropriate analysis of standards it was assumed that the method was suitable for the analysis of FAs in single mosquitoes. The wheat spectra were unusual in that only one fatty acid (palmitic acid 16∶0) appeared to be present, compared to published data in which palmitic 16∶0, stearic (18∶0), behenic (22∶0) oleic (18∶1) and linoleic (18∶2) acids were measured in significant concentrations in wheat flour , therefore this sample was excluded from the fatty acid analysis. On some diets, notably the carrot, yeast hydrolysate, bean, vitamin mix and soy hydrolysate, mosquitoes failed to grow to adulthood and so these were also excluded from the analysis (.). The most common fatty acids present in all the diets were the palmitic (16∶0, 100% occurrence, 12–40% relative fatty acid composition (RFAC)), palmitoleic (16∶1 n7, 100% occurrence, 3–16% RFAC), stearic (18∶0, 100% occurrence, 0.5–18% RFAC), oleic (18∶1 n9, 100% occurrence, 8.2–37% RFAC) and linoleic (18∶2 n6, 100%, 2–51% RFAC) acids. In addition, spirulina contained the characteristic gamma linolenic acid (18∶3 n6) and all leguminous samples contained the characteristic (18∶3 n3) rumelenic acid. In a matrix of larval diet versus relative fatty acid composition of diet, cluster analysis clustered the cereals closely and the fish products closely, as would be expected (insert). Simply plotting RFAC of diet, against the RFAC of the resultant mosquitoes, yielded highly significant relationships; in most cases accounting for over 50% of the variability in the data, suggesting that the majority of fats are taken up and preserved indiscriminately, as was evident from the patterns observed in. On closer analysis it became clear that the significant relationships were a reflection of the dominant fatty acid profiles of the diets skewing the data. Average relative fatty acid composition of the most common acids in the diets were palmitic (16∶0, 31.1% s.d. 22.9), palmitoleic (16∶1n7 4.5% s.d.4.6), stearic (18∶0, 7.5% s.d. 6.1), oleic (18∶1 n9, 16.5% s.d.10.2) and linoleic (18∶2 n6, 18.5% s.d. 17.7). In the mosquitoes the average dominant acids were palmitic (16∶0, 30.3% s.d. 6.1), palmitoleic (16∶1n7 14.4% s.d. 8.4), stearic (18∶0, 3.2% s.d. 2.1), oleic (18,1 n9, 20.0% s.d.5.6), linoleic (18∶2 n6, 16.4% s.d. 9.9). The diet data showed broad similarities in FA profiles, somehow reflecting the uniformity of the building blocks required for the essential structures of life; however the lower standard deviation values in the mosquitoes suggested that as consumers they also have some influence on their specific fatty acid profiles. These data also suggest there is preferential accumulation from the diet of palmitoleic and oleic FAs by the mosquitoes. To establish whether “you are what you eat”, a matrix of mosquito and diet versus relative fatty acid composition was constructed. At the first level, mosquitoes of the same sex which were fed the same diet showed the highest degree of similarity, with Euclidean distances of less than 10. At the next level the mosquitoes fed the same diet showed the greatest degree of similarity, at the next level it appeared that there was some clustering based on whether the mosquitoes were fed a cereal, fish or legume diet. Notably, diet had a stronger influence than gender on the fatty acid profile of the individual mosquitoes, suggesting that there is a high degree of nutritional plasticity and providing strong evidence for dietary routing (.). This led to the rejection of hypothesis 1 that fatty acid composition of mosquitoes is fixed and not influenced by the diet. In an attempt to generate comparable information from the large data set and reveal the flow and synthesis of individual fatty acids up the food chain, graphically evident from, in a simple mathematical manner, both the food and the mosquitoes were scored depending on the presence (1) or absence (0) of a particular fatty acid. This allowed us to compute and graphically present direct uptake, *de-Novo* synthesis and frequency of occurrence of each fatty acid in all the mosquitoes fed on the range of diets presented. Presence or absence classifications based on binary systems have been widely used in medical studies and ecological modelling. In this system the cut off threshold for absence was 0% RFAC and presence was deemed anything above 0% RFAC. The sum of the binary values for the mosquitoes over the sum of the binary values for the diet factorised (i.e. multiplied by the number of mosquitoes (typically n = 3) obtained and measured from that diet) was computed. If the value for mosquitoes was greater than the value of diet (factorised), they were scored with a 1 and designated *de-novo* synthesis. The sum of these values was computed for each fatty acid and calculated as a percentage of the number of diets used (to successfully rear mosquitoes) in the overall analysis (11 diets male, 8 diets female), this allowed male and female values to be compared. To determine the occurrence of direct uptake, if the sum of the binary value of the mosquitoes and the binary value of the factorised diet was greater than number of mosquitoes used in the analysis it was assumed that there was direct uptake and a value of 1 was assigned and again calculated as a percentage of the number of diets used for both males and females. This has a simple logic, as if there was no fatty acid present in the diet the factorised value of the diet would be zero, and thus only when the fatty acid was present in both the diet and the mosquito would the sum of the values be greater than the number of mosquitoes used, and thus the number 1 can be assigned to indicate direct uptake. The fatty acids present in all the mosquitoes were the palmitic (16∶0, 100% occurrence, 12–40% RFAC), palmitoleic (16∶1 n7 100% occurrence, 3–16% RFAC), stearic (18∶0, 100% occurrence, 0.5–18% RFAC), oleic (18∶1 n9, 100% occurrence, 8.2–37% RFAC) and linoleic (18∶2 n6, 100%, 2.0–51% RFAC) acids; these are all common fatty acids found in a range of food stuffs. The analysis suggested that these were all directly taken up from the diet, this fits the hypothesis of dietary routing, which suggests that organisms will take up and use FAs in their original form to avoid energy loss or the cost associated with modification (.). To further examine the level of dietary routing a cluster analysis was performed on the “raw” data of the 28 possible fatty acids, by producing a matrix of each average diet RFAC versus average mosquito RFAC (n = 11 male, n = 8 female). This allowed us to estimate, on an individual fatty acid basis, the apparent transfer and conservation of RFAC profile up the food chain from a range of diets. For this analysis resemblance matrices were constructed based on Bray Curtis similarity. In female mosquitoes the highest diet to mosquito similarities were in the linoleic t (98%) linoleic c (80%), oleic (80%) palmitic (80%) and rumelenic (78%), gamma linolenic (64%) fatty acids. It could be argued that these are also the fatty acids most commonly found in both diets and the mosquitoes analysed, with only the rumelenic (25% *de-novo* synthesis, 50% direct uptake) and gamma linolenic (13% *de-novo* synthesis, 13% direct uptake) not present in all of the mosquitoes and diets analysed. In the males the highest diet to mosquito similarities were in palmitic (82%) rumelenic (78%), gamma linolenic (70%), argaric (60%) acids. The percentage occurrence of *de- novo* synthesis was similar for rumelenic (20% *de-novo*, 40% direct uptake) and gamma linolenic (10% *De-novo* 10% direct uptake) acids in both males and females. However argaric acid was exclusively produced by *de-novo* synthesis in only 30% of the females and 54% of males, and obtained by direct uptake in about 5% of males from the squid liver powder (SLP) treatment. Although females of the SLP treatment were viable and of average size based on total body carbon data, they were not successfully analysed for fatty acids due to technical problems, possibly leading to this discrepancy. There was a weak but highly significant correlation between the independent variables TBN and % N of the diet the mosquitoes were fed (r² = 0.17, F_1,86 = 14.22, p = 0.0003), but not with % C or % P of the diet. There were no simple correlations between TBC or TBP and elemental dietary composition. There were strong correlations and highly significant relationships between TBC and TBN for both males (R² 0.794, F_1,34 = 188.40 p = \<0.0001) and females (R² 0.85, F_1,32 = 188.40 p = \<0.0001) in experiment 2. It could be argued that these are not independent variables, however there were no significant relationships between the TBC and TBP either, for either males or females. Multiple regression analysis revealed that 66% of the variation in the TBC of male mosquitoes could be explained by the fatty acids and stoichiometry of the diet; this was a significant interaction (p = \<0.000004). For TBN, 93% of the variation could be explained by the diet (p = \<0.00000) and for total body P (TBP) an astonishing 99.7% of the variation could be explained for by the dietary composition (p = \<0.00000). In the females only 19% of the variation in TBC (p = \<0.05), 80% of TBN variation (p = \<0.00000) and a similarly high 99.8% of TBP variation (p = \<0.00000) could be explained by the dietary composition. Regression analysis of δ¹⁵N diet against δ¹⁵N of the subsequent mosquito yielded an equation of y = 0.877×+2.495 with an r² of 0.931, the overall shift being around 2.5 ‰. Diet to mosquito shifts in carbon isotope signatures were within the expected range, of around 1 ‰; regression analysis of δ¹³C diet against δ¹³C mosquito yielded the equation y = 1,017×+0.732 with an r² = 0.991. ## Experiment 3. Determining the Influence of Dietary N and P on Mosquito Survival and Production In experiment three percentage P in the diet appeared to have a greater impact on adult “production” per dish than dietary % N, both showing weak but highly significant correlations (p\<0.00000, r² = 0.40, F_1,46 = 22.60 and p\<0.0006, r² = 0.23, F_1,46 = 13.52 for P and N respectively). # Discussion These analyses taken together suggest that larval dietary quality and quantity have a substantial impact on adult population size, wing length, survival and possibly fecundity in *An. arabiensis* mosquitoes. We have shown that fatty acids such as arachidonic acid which have previously been shown to be essential for Culex species were only present in around 80% of male and 60% of female mosquitoes, with *de-novo* synthesis evident in at least 50% of the mosquitoes sampled. The highest levels of similarity between mosquito arachidonic RFAC and diet were in steric acid in the males (40%) and eicosaptaenoic acid in the females (35%), suggesting that these acids may play a role in the synthesis of arachidonic acid in mosquitoes. In addition there were high levels of *de-novo* synthesis of arachidonic acid (\>5% RFAC) in the wheat bran fed mosquitoes; wheat bran had high levels of linoleic acid (50% RFAC) which is an established precursor of arachidonic acid. Arachidonic acid is rare in the plant kingdom, but it can be found in some fungi, mosses and ferns and is a major component of several microalgae, where it reaches up to 47% of the triglyceride pool. This could explain the link with mosquito larval nutrition, macro-benthic algae being omnipresent in most natural larval ecosystems. Previous work has suggested that mosquitoes have to obtain the C18-22 fatty acids from their diet as they are unable to elongate the 18C acids. They have shown that eicosapentaenoic acid (EPA), docosahexaenoic acid (DHA) and arachidonic acid (AA), are all essential acids; however in these experiments there was substantial evidence of *de-novo* synthesis in the 18C plus group, in both males and females. Whether this is the result of elongation or shortening, is unclear from our results, but by using individually stable isotope labelled fatty acids and a pyrolysis system linked to an isotope ratio mass spectrometer it should be possible to further elucidate these pathways. Only decohexaenoic acid (DHA) in both males and females and gadoleic acid in males, were exclusively directly routed from the diet but were not present in all the mosquitoes, suggesting they are not essential, with the caveat that these experiments did not study the full life cycle of these *An. arabiensis* mosquitoes. These fatty acids previously identified as essential for mosquitoes may be necessary to complete the full life cycle. Whether the *de-novo* synthesis of fatty acids takes place within the mosquito or within the microbial or macro-benthic biofilm in the larval trays is unclear from these experiments, but a number of bacteria have been shown to be able to synthesise these FAs, and mosquitoes are known to graze on bacterial cells in the water column. An attempt was made to detect this intermediate trophic level by using the isotopic data from both the food and diet. It is commonly quoted in isotopic circles that “you are what you eat plus a few per mille (‰)”, since there is a ubiquitous and characteristic shift in the δ¹⁵N signal as you move up the trophic ladder, due to the discrimination against heavier ¹⁵N atoms by the enzymatic and kinetic reactions. One step up the trophic ladder usually results in delta shift of 2–3 ‰, thus we hypothesised that if larval grazing of bacteria was a dominant source in the larval diet we should be able to see a characteristic shift of around 4–6% from the diet to the mosquito as it would reflect the two trophic levels. Regression analysis of δ¹⁵N of food against δ¹⁵N of subsequent mosquitoes was around 2.5 ‰, suggesting that direct food uptake rather than bacterial grazing was the dominant process. Unfortunately, published data for isotopic shifts from diet to bacteria is scarce to non-existent, despite extensive data mining, thus this result suggests either that bacterial grazing does not contribute significantly to larval nutrition or that bacterial grazing does contribute significantly, but there is no characteristic isotopic shift from substrate to product during bacterial growth. There was a weak but significant interaction (R² = 0.241, F_1,68, p = 0.0149) between ratio of the diet and delta shift from diet to mosquito which could hint at the role of bacterial processing of diet in the higher C:N treatments, or could be a reflection of starvation which has been shown to increase diet to organism delta shifts. In retrospect it would have been astute to do both isotopic and fatty acid analysis of the biofilms which are commonly present in detectable quantities on the bottom of the larval trays. An attempt was made to determine the influence of dietary fatty acids profiles and dietary stoichiometry on mosquito stoichiometry and some measure of mosquito fitness or competitiveness (fitness in the context of sterile mosquitoes could be construed as a misnomer). Wing length as a measure of mosquito size is a well-accepted determinant of mosquito competiveness. Given that we found significant correlations between wing length and both TBC and TBN in the initial exploratory experiment and that it is logistically easier to analyse for TBN and TBC than wing length when running isotope analysis, we used TBC and TBN as a measures of mosquito competitiveness. It is important to remember that these mosquitoes were raised at a low larval density and non-limiting conditions and were sampled as tenerals and not fed as adults. In experiment 2, regression analyses of % N, C and % P in the diet versus mosquito TBN and TBC showed that only % N of diet was significant otherwise no significant interactions were observed; this suggests that % N of the diet has an influence on the TBN and thus wing length, and competiveness of the mosquito. Additionally in experiment 3, we showed that both % N and % P of the diet had a significant impact on population size. In experiment 2 multivariate analysis suggested there was a greater predictability of mosquito TBP from the dietary composition, which was clearly a result of the degree of its bioavailability and stability. Phosphorous, most probably being phospholipid derived, was therefore the link with the fatty acid profile and not lost from the system, in contrast to the nitrogen which can be complexed within lignin type substances and not readily nutritionally available. In addition, excess N can be lost from the system as gas through the processes of denitrification. Oleic and linoleic acid c appeared to have most consistent influence on total body stoichiometry, with % N and % P in the diet significantly contributing to the model of µg N and µg P in the female mosquitoes (.). Notably, although there were strong correlations and highly significant relationships between TBC and TBN for males and females, respectively, in experiment 2, there were no significant relationships between the TBC and TBP for either males or females. The average C:N:P ratio of all the diets was 106∶9:1, ranging from 243∶10:1 in bean to 9∶4:1 in tuna meal, the average value for all the female mosquitoes was 28∶7:1, and for all the males was 48∶10:1. These results reveal divergent male and female stoichiometry, females having a much higher P requirement with an average of 5 µg P per female and 2.9 µg P per male (s.d 1.0 in both cases). In addition the C:N ratio of the mosquitoes was more tightly bound than their C:P ratios with %SD of the C:P ratio five times greater than the %SD of the C:N ratio. Previous research has shown that approximately 50% of the total body carbon and nitrogen of teneral mosquitoes is structural and does not turnover within the lifetime of the mosquito, and that these mosquitoes can accumulate up to three times their initial body carbon from sucrose solutions. Results from experiment 1 and 2 therefore suggest that mosquito size is primarily determined by nitrogen availability. Experiment 3 was set up to determine the impact of diet quality or stoichiometry on production and survival akin to “mosquito production per unit food”. This was achieved by keeping dietary carbon concentrations constant across treatments and initial larval density to a minimum to overcome any negative feedbacks of overfeeding. Although % P in the diet had a greater impact on adult “production” compared to dietary % N, the stronger influence of dietary P on survival may once again have been the result of greater dietary P bio-availability. Nitrogen and phosphorous interactions are notoriously difficult to untwine and it is apparent that dietary quality has a significant influence on the production and survival of the adult mosquito, as seen in. When we combine the evidence from all the experiments presented, it appears that mosquito size and consequent competitiveness is controlled by the nitrogen content and maybe more importantly nutritional bio-availability of that nitrogen from the larval food source. Conversely, concentration of P is often quoted as the degree of overall productivity of the aquatic system. The general stoichiometric mismatch between diet and mosquitoes would suggest that nitrogen is indeed limiting both in the laboratory and the natural environment. In the tuna meal, squid liver powder, and brewer’s yeast treatments this was not the case, indicating that maybe these larvae were carbon rather than nitrogen limited. However, it is unlikely that these dietary configurations would be observed in nature. We hypothesise that in *An. arabiensis* mosquitoes nitrogen content and thus mosquito size is controlled by upper and lower limits of nitrogen cycling, the lower limit being determined by the nitrogen availability of the diet and the upper limit being posed by the fouling of the larval water due to build up of toxic ammonium products either as the result of excretion or mineralisation, the breakdown of organic nitrogen to inorganic nitrogen by the microbial communities in the larval water. Larval overfeeding often leads to high larval mortality and *An. arabiensis* is known as a clean water species. Indeed additional simple chemical analysis showed that larval water ammonium concentration in the healthy *arabiensis* trays was around 2 ppm NH⁺₄ compared to 100 ppm NH⁺₄ in the larval trays of *albopictus* species. We hypothesise that total larval nitrogen availability linearly determines the overall size of the mosquito which ranged from 20 µg N/80 µg C to 58 µg N/261 µg C, almost a threefold difference in TBC or mosquito size, and that P is not only present as a structural component linked to specific phospholipids, but is also a more flexible storage component, evident from the data shown in. This hypothesis would explain why there is little correlation between total body C and P values but a strong correlation between fatty acid profiles and total body P, contrary to stoichiometric theory. Therefore in essence we reject Hypothesis 2 and replace it with “Teneral individual C:N is fixed and not influenced by the larval diet”, as it appears that teneral C:N ratios are fairly fixed with values of 4.2∶1 and 4.5∶1 and % SD of less than 7 and 11% for males and females, respectively. It could be argued that stoichiometric theory is more applicable to aquatic environments but these mosquitoes were sampled as non-fed adults and as such were not subjected to a terrestrial dietary environment. Some explanation could be offered by the terrestrial feeding ecology of the mosquito: male mosquitoes notoriously only feed on sugar sources in the adult stage, which are presumably low in phosphorus. Dietary restriction, or even having a flexible P store, could possibly improve longevity, as most aquatic systems are either P or N limited. This mechanism of N determining size, and P determining longevity and mosquito abundance would result in an evolutionarily successful flexible trade off strategy between size and longevity. Small mosquitoes live longer increasing probability of finding a mate, whereas larger individuals play hard, win the mate and die young, as has been shown in other insect species such as crickets. Body size is often strongly correlated with fighting ability, or resource- holding potential (RHP), such that the larger of two competing males usually wins the contest. Larger male mosquitoes have been reported to be more successful in mating than smaller ones. Intriguingly female body size has also an advantage in mate selection, larger females of *An. gambiae* s.s. being preferentially selected for mating. Field studies have noted a positive correlation between female body size, which is presumably influenced by larval nutrition and competition, and parity status. On the other hand, dietary restriction in insects has generally been shown to increase longevity, and longevity can in turn lead to an increased chance of mating. Indeed this flexible hypothesis for mosquitoes is backed up by a very recent, as yet unpublished, study of *Anopheles gambiae* s.l. in which mean adult male body size significantly influenced adult survival (F-value = 51.847; P\<0.01) and correlated with larval nutrition (r = 0.946; P\<0.01). Males that consumed the greatest amounts of food had the lowest survival (F-value = 4.491, P = 0.012) with a mean survival of 11 days. This data suggests that the smallest ones had the highest levels of longevity. The lack of influence of phosphorus on overall insect size contrasts with the findings of Visanuvimol and Bertram who found that P availability in the diet influenced cricket weight and size (although intriguingly not total carbon); however they also found that dietary P had little influence on cricket life span. These contradictory findings could reflect the extremely different life cycles of cricket and mosquitoes. Indeed what the authors did stress was that there was a significant relationship between total body carbon and nitrogen, but they also found that neither total body carbon nor nitrogen were correlated with total body phosphorous, thus also failing to demonstrate strict stoichiometric interactions. In line with our hypothesis they did find that older insects were more depleted in P, suggesting that P stores are used up as insects get older and are not replenished. Woods *et al*., also suggested that P content may be only weakly related to body mass. They suggested that several taxa exhibit inverse dependence of P content on body size (e.g. plants).; Nielsen *et al.* again supporting our nascent hypothesis. ## Conclusions In conclusion, pyrolysis GCMS allowed a comprehensive analysis of fatty acid profiles of single mosquitoes and their diets to be undertaken which revealed the common occurrence of *de-novo* synthesis of a number of important fatty acids. It also suggested that fatty acids play an important role in the P nutrition of mosquitoes. The analysis revealed that diet has a greater influence on fatty acid profiles than gender, suggesting that dietary routing is an important mechanism in mosquitoes. *An. arabiensis* mosquitoes appear to exhibit a highly plastic feeding strategy characteristic of generalist feeders and are able to feed on a range of fatty acids and diet qualities, an ability that allows them to exploit a range of micro habitats dominated by different primary producer species. The stoichiometric-centric analysis suggested that *An. arabiensis* individual adult size is determined by the upper and lower limits of nitrogen availability and that population-size is determined by the total phosphorus availability of the system with the consequence that phosphorous is a flexible storage product of the adult mosquito. These findings are in line with new paradigms about quality/quantity issues in ecology, which shift away from a biomass density variable to include a two state paradigm, which represents populations or groups in a food web in terms of both their quality and quantity. Given the simplicity and rapidity of the sample analysis described herein, we suggest that these methods could be useful to further test the models presented by Getz and Owen on a logistically feasible scale. These results lay out an experimental foundation on which to conduct future research both in the field and laboratory and may explain why increased malaria incidences have been observed and reported in areas with higher inorganic fertiliser usage. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: RHN JG CS. Performed the experiments: RHN BS CS SS OM MA MW JG. Analyzed the data: RHN BS CS SS OM MA MW JG. Contributed reagents/materials/analysis tools: RHN BS CS SS OM MA MW JG. Wrote the paper: RHN CS.

# Introduction In 2014, the U.S. harvested over 33 million ha of soybean with a total value of over \$40 million dollars. Soybean (*Glycine max*) is primarily grown for oil content (industrial and household) and meal production (animal feed). Its nitrogen-fixing capability makes it very important in cropping systems and crop rotations, particularly with corn (*Zea mays*). Many diseases cause production issues for soybean throughout the U.S., including several viruses. Viruses such as *Soybean mosaic virus* (SMV), *Bean pod mottle virus* (BPMV), and *Alfalfa mosaic virus* (AMV) commonly occur in soybean. Recently, a new soybean virus which causes vein necrosis was identified and named *Soybean vein necrosis virus* (SVNV). SVNV is a species within the genus *Tospovirus*. In the U.S. there have been very few instances where a *Tospovirus* was found to infect soybean. Nischwitz et al. demonstrated that soybeans in Georgia were infected with *Tomato spotted wilt virus* (TSWV) and were asymptomatic. In other countries *Tospoviruses* have been more readily identified on soybean and include TSWV, *Tomato yellow ring virus*, *Groundnut ringspot virus*, *and Groundnut bud necrosis virus*. SVNV is like other viruses in this genus in that its genome consists of a large negative-sense RNA component (L) and two smaller ambisense RNA components (M and S) that encode proteins in both the positive and negative-sense. The sizes of the M and S components are similar to other tospoviruses. The L component is the largest of the genus at 9,010 nucleotides. The encoded proteins \[nucleocapsid (N); nonstructural protein (NSs); glycoprotein (G_N/G_C); nonstructural protein (NSm); RNA-dependent RNA polymerase (RdRp)\] are typical of the genus in general. However, SVNV does not conform to the typical sub- groups within the *Tospovirus* genus in that the homology of these proteins to other members of the genus is quite distinct. Tospovirus species within this genus are typically split between two distinct genetic clades called the ‘New World’ viruses and the ‘Old World’ viruses. All viruses within the *Tospovirus* genus fall in these two sub-groups with the exception of SVNV and another closely related virus species, *Bean necrotic mosaic virus* (BeNMV). SVNV was not identified as a pathogen of soybean until 2008 when it was documented in Tennessee, Arkansas, and several other southern states. More recently it has been found in north central U.S. states such as Wisconsin and Iowa. Considering the uniqueness of SVNV and its relatively recent emergence as a soybean pathogen, it has been hypothesized that this virus has only recently adapted to soybean. Several of the plant-infecting tospoviruses have very broad host-plant ranges, including SVNV. Zhou and Tzanetakis identified several weed hosts capable of being alternative hosts for SVNV. Tospoviruses are transmitted exclusively by thrips, in a persistent propagative fashion (i.e., they replicate inside their thrips vector and are transstadially passed from molt-to-molt). SVNV can be transmitted by soybean thrips (*Neohydatothrips variabilis)*, but it is not known if other thrips species occurring on, or reproducing upon soybean can transmit the virus, or if there are other means of transmission. Some soybean infecting-viruses like *Tobacco ringspot virus* (TRSV) can be transmitted via seed. While the rate of seed transmission can be low, the impact on yield of this nepovirus can be high. Yield can be reduced as much as 100% through reduced pod set and seed formation. Seed from plants infected with TRSV often have higher total protein and lower total oil content than seed from non- infected plants. It is generally accepted that members of the *Tospovirus* genus are not seed transmitted. However, tospoviruses have been detected in pods of other legumes. Pappu et al. demonstrated that TSWV, type species of the genus *Tospovirus*, localized to the peanut pod and testa using enzyme-linked immunosorbent assay (ELISA) and polymerase chain reaction (PCR) on plants that were symptomatic for TSWV infection. Sequences specific for TSWV were detected occasionally in cotyledons using PCR. Seedlings grown from seed harvested from both symptomatic and asymptomatic plants were tested for TSWV infection using ELISA, but no positively infected plants were identified. These data suggest that in peanut, accumulation of the TSWV is localized to the shell and testa and is not passed to the progeny. In 2013 a seed lot of commercial soybean seed from Nevada, Iowa that had a high level of discolored, deformed and undersized seed (‘damaged’ seed lot) was obtained. Virus infection was suspected, including SVNV. A separate seed lot from the same field that was normal in appearance (‘normal’ seed lot) was also acquired. In this paper, we provide evidence that i) the ‘damaged’ seed lot contained seed that was infected/infested with SVNV; ii) the SVNV-positive seed passed the virus to the emerged seedlings when planted. This is the first time that a Tospovirus has been found to be seed transmitted. # Materials and Methods ## Soybean seed used in this study The ‘damaged’ soybean seed (*Glycine max* (L.) Merr.) used in this study was hand harvested from a commercial soybean field (variety: Asgrow AG2433, Monsanto Company, St. Louis, MO) with symptoms indicative of virus infection in Nevada, Iowa in 2013. The geographical coordinates of the field where the sample was collected were 42.041262, -93.472741. The collection of seed was performed on private land with the consent of the farmer. In 2013, SVNV was widespread in Iowa and confirmed in soybean fields in every county by the Iowa State University Disease Diagnostic Clinic. The seed was maintained at room temperature in the University of Wisconsin-Madison Field Crops Pathology seed storage collection until use. An additional seed lot not visibly damaged and considered ‘normal’ in appearance, was also collected from the same field. This seed lot was used in the following experiments for comparison with the ‘damaged’ seed. ## Assessing Seed Quality Variables Seed germination, 100-seed weight, protein content, oil content and fiber content were assessed for each of the two seed lots described above. Germination was determined by placing 10 seeds on filter paper saturated with de-ionized water and placed in a Petri plate. Each Petri plate was considered a technical replicate and five technical replicates were completed for each repetition. Two repetitions were conducted for each seed lot for a total of 100 seeds evaluated for each seed lot. One hundred-seed weight (indicator of seed size) was determined by counting out five technical replicates of 100 seed for each seed lot. Two repetitions (2 separate days) were conducted for a total of 10 observations for each seed lot. A grain analyzer (Foss 1241; Eden Prairie, MN) was used to determine total protein content, total oil content, and total fiber content of five technical replicates for each seed lot. The grain analyzer was programmed using a standard curve and five sub-samples were completed for each replicate. All seed quality variables were subjected to a mixed model analysis of variance (ANOVA) with seed lot as the dependent variable and seed quality variables as the independent variables. Replicate and repetition were considered random effects. F-tests were performed at α = 0.05. All analyses were performed using SAS v. 9.4 (SAS Institute, Cary, NC). Subsamples of seed from the ‘damaged’ and ‘normal’ lots were surface disinfested (1 min in 95% ethanol and 1 min in 10% Clorox) and plated onto potato dextrose agar (PDA) media amended with 25 μg/ml ampicillin, 10 μg/ml rifampicin, and 25 μg/ml streptomycin in Petri plates and incubated for 3–5 days at room temperature. Colonies of three different fungi were identified from seed sampled from the ‘damaged’ lot. Fungi were isolated and purified. Mycelia were subsequently collected from the culture plates and DNA extracted using the FastDNA spin kit (MP Biomedicals, Santa Ana, CA) according to the manufacturer’s instructions. DNA was subjected to PCR using ITS 4 and ITS 5 primers. Resulting PCR products were purified for sequencing using the Wizard SV Gel and PCR Clean- Up System (Promega Corporation, Madison, WI). Sequencing was performed using ABI 3730xl DNA Analyzers (Applied Biosystems, Foster City, CA) at the University of Wisconsin-Madison Biotechnology Center. Sequences were subjected to BLASTn on GenBank for positive identification of fungal organisms. ## Initial virus detection assays In April 2014, a random sample of soybean seed from the ‘damaged’ seed lot indicated above was planted in a growth room (14 hrs light/10 hrs dark, 22°C) at the University of Wisconsin-Madison. The growth room is located in the basement of Russell Laboratories and is lit by a combination of fluorescent and ultraviolet bulbs suitable for growing plants, and was isolated from any source of SVNV inoculum or thrips infestation. Soybean seed was planted into 32-cell propagation trays containing Sunshine Redi-earth Potting Mix (Sun Gro Horticulture, Agawam, MA). Germination of the seed was greater than 90% and the resulting seedlings appeared normal. Four composite samples of six plants each were tested using reverse transcription polymerase chain reaction (RT-PCR; described below) by selecting leaves (unifoliate or trifoliate) from each of six plants and combining them to create a composite sample. Subsequently, twelve individual plants (a single leaflet from each plant) were further tested after initial composite tests indicated the presence of SVNV. ## RNA extraction and RT-PCR The harvested leaves were placed into individual plastic bags, flash-frozen in liquid nitrogen and stored at -80°C prior to RNA extraction. Approximately 200 mg of leaf or cotyledon tissue were ground in liquid nitrogen and RNA was extracted using the TRIzol® Plus RNA Purification Kit (Life Technologies, Carlsbad, CA) according to the manufacturer’s instructions. RNA quantity was estimated on a NanoDrop UV-Vis Spectrophotometer (Thermo Fisher Scientific, Inc., Waltham, MA). Approximately 1 μg of total RNA was used to generate first- strand cDNA with random or specific primers using the iScript™ Reverse Transcription Supermix for RT-PCR (Bio-Rad Laboratories, Hercules, CA) or Superscript III one step RT-PCR kit with platinum Taq DNA polymerase (Invitrogen/ Life Technologies, Carlsbad, CA) according to the manufacturer’s instructions. cDNAs were diluted ten- to twenty five-fold in nuclease free water for use in subsequent PCR. Specific primers were used to amplify portions of all three genomic segments of SVNV. Primers designed to amplify the gene encoding the entire nucleocapsid protein were used in nested PCR reactions using SVNaV-f1/SVNaV-r1 and SVNaV-f2/SVNaV-r2 primer sets. Published primers were used to amplify regions of the SVNV L segment. To amplify a portion of the M segment, primers were designed using PrimerQuest Software Package (IDT DNA Technologies, Coralville IA;). PCR were performed using GoTaq Green Master Mix (Promega Corporation, Madison, WI) containing: GoTaq DNA Polymerase in 1X Green GoTaq Reaction Buffer (pH 8.5), 200 uM dNTPS, and 1.5 mM MgCl₂; 1 uM of each primer; and 5–7.5 ul cDNA in the total volume of 25 ul per reaction. PCR was performed under the following conditions in an Eppendorf MasterCycler Pro S programmable thermal cycler (Eppendorf AG, Hamburg, Germany): denaturing at 95°C for 2 min followed by 32–40 cycles of 30 sec denaturation at 94°C, 30 sec annealing at 52°C, and 1 min elongation at 72°C, followed by a final extension step at 72°C for five to ten min. Nested PCR (S segment) was performed as described above using specific internal primers and 5 μl of the first round PCR product as the template. Reaction conditions for all PCR was similar to that described above. Proper negative (water blanks) and positive (cDNA from a SVNV positive field isolate) controls were used in all PCR experiments. PCR fragments were visualized by electrophoresis in 1X TAE on a 1.2% agarose gel containing SYBR Safe DNA gel stain (Life Technologies, Carlsbad, CA), using PCR Marker (Promega Corporation, Madison, WI) or 1 kb DNA Ladder (Promega Corporation, Madison, WI) to estimate the size of the fragments. Select PCR fragments were purified for sequencing using the Wizard SV Gel and PCR Clean-Up System. Sequencing was performed using ABI 3730xl DNA Analyzers at the University of Wisconsin-Madison Biotechnology Center. Template RNA was also checked for the presence of endogenous DNA by using the above PCR primers and conditions. RNA from the initial extraction was used as template instead of cDNA as described above. Genomic DNA extracted from the same plant material was used as template and subjected to PCR using the primers described above and soybean genomic primers to rule out the presence of sequences homologous to SVNV that could be transcribed and result in false positives in the PCR experiments. ## Controlled environment studies A thrips-proof cage large enough to contain three-32 cell propagation trays was used to cultivate and maintain plants in a controlled environment. The thrips proof cage was housed in an environmentally controlled growth room in the basement of Russell Laboratories, and was maintained at 22°C, under a 14 hr day/10 hr night photoperiod and was lit by a combination of fluorescent and ultraviolet bulbs suitable for growing plants. Seeds from the Nevada, Iowa seed lot were randomly planted in the trays, with no bias as to the physical appearance of the seed. The seed was planted into Sunshine Redi-earth Potting Mix, as described above. Once the plants emerged, and the first trifoliate began to open (usually 10–11 days post planting) one cotyledon from each sample plant was removed. Eight randomly selected plants (random number generator was used) from each tray were sampled. For each experimental run: a single tray (32 cells) represented a replication, thus, there were 3 replications per experimental repetition. Sample size was 8 plants per tray (sub-sample from the population) and 3 replications per experimental repetition, for a total of 24 plants per experimental repetition. Two repetitions were conducted. RNA was extracted from each plant and cDNA was synthesized and subjected to RT-PCR, as described above, in order to determine the presence of SVNV in all plants sampled for each repetition. ## Total RNA sequencing (RNA-seq) Total RNA was extracted and quantified as described previously. The total RNA extracts from a composite sample (multiple soybean leaflets), and also from a SVNV-symptomatic leaflet collected from a Wisconsin field, both of which were PCR-positive using SVNV primers were sent to ProteinCT Biotechnologies LLC (Madison, WI) where they were treated with DNase and the total RNA was used to construct a micro RNA library using the SeqMatic Micro RNA Sample Preparation kit (SeqMatic LLC, Fremont, CA). The total RNA library was subjected to high throughput sequencing (1x50bp) using the Illumina platform (Illumina, Inc., San Diego, CA). RNA reads were quality-filtered to remove low quality and very short reads and to remove adaptors. The clean reads were then aligned to the published soybean genome (the v2.0 assembly, *Glycine max Wm82*.*a2*.*v1* <http://phytozome.jgi.doe.gov/pz/portal.html#!info?alias=Org_Gmax>) and to the SVNV genome (GenBank accession numbers, HQ728385.1, HQ728386.1, and HQ728387.1) using Bowtie short read aligner. # Results and Discussion ## Seed Quality Assessment Germination, 100-seed weight, total protein content and total fiber content were not significantly different between ‘normal’ and ‘damaged’ seed lots. Total oil content was significantly lower in the ‘damaged’ seed lot compared to the ‘normal’ seed lot. These results suggest that virus infection, including infection by SVNV of parent plants may influence the chemical composition of soybeans. Decreased oil content is of concern as soybean is an oilseed crop, grown in part for industrial oil needs. The reduction in oil content and increase in protein content is consistent with previous work on soybean. Demski et al. and Demski and Jellum found that *Tobacco ringspot virus* alone, or in mixed infection with other viruses, typically reduced oil content of soybeans and increased total protein content. Furthermore, virus infection was found to change the fatty acid composition of soybean oil. While we did not test the fatty acid content of the damaged seed examined here, trends in reduced oil content and increased protein in soybean seed infected with a systemic virus are similar to previous reports. Fungi isolated from the ‘damaged’ seed lot included *Diaporthe phaseolorum*, *Alternaria alternata*, and *Fusarium equiseti* all with 98–99% homology to sequences in GenBank (*data not shown*). No pathogenic fungal organisms were isolated from the ‘normal’ seed lot. All three organisms isolated from the ‘damaged’ seed lot have been reported as pathogens of soybean. *Diaporthe phaseolorum*, and *Alternaria alternata* have been implicated in causing pod blights and *Fusarium equiseti* may also be found on soybean plant parts including pods. These findings suggest that there might be correlation between virus infection and increased presence of fungal pathogens of seed, which further reduce overall seed quality. However, it is difficult to determine if virus infection increases susceptibility of seed to fungal infection or vice versa. ## SVNV positives in initial RT-PCR In April 2014, a random sample of soybean seed from the Nevada, Iowa ‘damaged’ seed lot was planted in the growth room to determine whether or not the resulting plants would be symptomatic for viral infection. Germination of the seed was greater than 90% and the plants looked normal. Leaves were selected from the flats of plants to test with RT-PCR in ‘composite’ sample (six plants were represented in each composite). RT-PCR products were obtained from two of the four composite samples tested in initial experiments using the nested primers for the SVNV S segment. illustrates a composite sample extraction and RT-PCR amplification using the three primer sets (Lanes 3, 9, and 15). In subsequent experiments with individual plants from this same group of plants, two of 12 individual plants tested, yielded RT-PCR products with the nested primers (SVNaV-f1/SVNaV-r1 followed by PCR using SVNaV-f2/SVNaV-r2 primers). illustrates one of the single plant extractions and RT-PCR amplifications with the three primer sets, which is also compared to PCR amplification from RNA extracted from a symptomatic field isolate from Wisconsin in 2013 (Lanes 5, 11, and 17). The PCR products from the composite samples and the individual samples were sequenced and the resulting sequences were at least 98% identical to published SVNV S segment sequences including HQ728387. These results demonstrate that SVNV was present in leaves of the plants tested, and most likely moved from infested seed to emerging seedlings. Further testing of plant material was also conducted using the enzyme-linked immunosorbent assay (ELISA) described by Khatabi et al.. No positive confirmations of SVNV seed-transmission were identified using this approach, including confirmation from plant material previously found to be positive using RT-PCR. This could be due to extremely low virus titer in plants containing the seed-transmitted isolate of SVNV identified here. In previous work we found that RT-PCR using the N gene nested primers in was a highly sensitive method for detecting SVNV in low copy number. PCR protocols were consistently more reliable than ELISA for detection purposes for SVNV and were used for all subsequent experiments. PCR reactions using RNA as template instead of cDNA resulted in no products after gel electrophoresis for all but three samples. In those three samples, the bands (\~150bp in size) did not correspond to the expected sizes for each primer pair. These small bands were excised and sequenced and checked using BLASTn. No matches were identified (data not shown). Therefore, no DNA contamination was identified in the RNA preparations used in these studies. Additionally, PCR reactions using DNA extracted from the same plant material resulted in no products after gel electrophoresis, indicating that transcribed viral DNA was not integrated in the plant genome. ## SVNV detection in controlled environment studies After the initial detection of SVNV in plants grown in the growth room, additional measures were taken to further control the growing environment by restricting access to any possible thrips infestation. Subsequent controlled environment studies were performed in a thrips-proof cage to eliminate any possibility of SVNV infection due to thrips feeding. Yellow sticky traps used for catching insects were placed both inside and around the thrips-proof cage to monitor for any insect activity. Thrips cards were periodically examined for presence of thrips. Thrips were never trapped inside the thrips-proof cage. Occasionally eastern flower thrips (*Frankliniella tritici*) were trapped outside the thrips cage. Soybean thrips have been identified as vectors of SVNV. It has not been demonstrated that eastern flower thrips can transmit SVNV. A total of 48 plants were tested (8 plants x 3 replicates x two repetitions). Three plants from the 48 sampled were determined by RT-PCR (using primers as indicated) to be infected with SVNV. These results confirm the presence of SVNV in the AG2433 seed lot examined. The mean proportion of plants infected with SVNV was therefore 0.06. ## RNA-seq results From the RNA-seq analysis, 1,790,826 total reads were obtained from the composite sample of plants from seed grow-outs, of which 9,376 reads (0.60%) mapped to the SVNV genome. For the symptomatic field sample, 6,457,522 total reads were obtained with 2,025,891 reads (35.10%) mapped to the SVNV genome. The total RNA extract from the composite sample that was PCR positive using SVNV primers, and the symptomatic field sample, were identified to have numerous reads of small RNA segments with sizes ranging from 18–23 nt, which would be appropriate for RNA viruses, and with hotspots in the 21–22 nt range. Further analyses revealed that both sense and antisense reads for specific genes for SVNV were identified for both the composite sample and the symptomatic field sample and were spread throughout the entire three genomic segments of SVNV (Figs and). Our intention in conducting the RNA sequencing experiment was not to do a thorough bioinformatics analysis of the response of soybean to SVNV infection but to use the data to establish the presence of the virus in the plant material sample subjected to analysis, as has been previously done. The RNA used for library construction and sequencing for the seed-transmitted strain was from a composite sample so the number of infected leaves was likely a low number obviating the large number of viral reads needed for a more in depth analysis. However, the fact that the size distribution of the small RNAs mapping to the SVNV genome is consistent with that expected for siRNAs, and that the location of reads and pattern is similar to that of the symptomatic field isolate, indicates that the SVNV virus is present in the RNA extract of the composite sample taken from seed grown in isolation. In this study the size distribution of small RNAs clustered between 21 nt and 22 nt similar to that from TSWV-infected *Nicotiana benthamiana* and tomato. The distribution of matches along the genomic and complimentary strands of all three viral RNAs, and their mapping to all viral open reading frames provides independent, corroborating evidence for the RT-PCR data indicating the presence of SVNV in plant material grown from seed. Furthermore, these results are consistent with the occurrence of virus replication in the emerging seedlings and systemic transport throughout the plant. Zhou and Tzanetakis reported that soybean was a local-lesion host for SVNV and that the virus did not systemically infect this host. However, the isolate of SVNV present in the seed lot examined here, appears to be systemic without causing typical symptoms on soybean. Follow-up testing of newly developing leaves of plants previously found to be positive for SVNV seed-transmission have also yielded positive PCR products. In addition, roots from plants grown in isolation have been tested for SVNV and positive PCR products have been identified using the S genomic segment primers. Sequencing of the PCR product identified a 98% match to the known SVNV segment in GenBank. These results further substantiate the fact that the SVNV isolate found in the seed lot tested here is systemically transported in soybean plants. In a recent paper by Hajimorad et al., they identified soybean plants with SVNV grown in a greenhouse. The SVNV-positive plants were grown from seed collected from the field. The authors postulate that the presence of SVNV in greenhouse grown plants was likely a result of feeding by overwintering, viruliferous adult thrips. However, three soybean varieties were grown in the same greenhouse and only two of the varieties were infected with SVNV. If viruliferous thrips were the source of SVNV inoculum, it would be presumed that all three varieties would have been infected, as genetic resistance toward SVNV is not known. In addition, distribution of viral symptoms should have been fairly uniform in this environment, if thrips were the source of inoculum. Based on the results of our study, we hypothesize an alternative explanation that the SVNV-infected greenhouse plants described by Hajimorad et.al. likely could have arisen from seed-borne SVNV transmission. ## Conclusions To our knowledge this is the first report of seed-transmission of a *Tospovirus*. We have shown that SVNV can be transmitted in seed and systemically transmitted to the emerging seedlings at a rate of approximately 6%. Previously, it was reported that this virus was not systemically transmitted in soybean. The seed-transmitted isolate of SVNV identified in this work causes no foliar symptoms on soybean. However, the seed lot containing the seed- transmitted isolate of SVNV was found to have reduced total oil and slightly elevated protein content. In addition, three fungal organisms were also found parasitizing the same seed. It is not clear if SVNV or mixed virus infections predispose seed to increased infection by fungal pathogens. These findings are especially important for soybean seed farmers who seek high quality, pathogen- free seed and who desire high oil output from a soybean crop. SVNV might also interact with other viruses in the same soybean plant to cause synergistic reactions that could result in increased damage to soybean. Therefore, identifying SVNV seed transmission and removing contaminated seed from a seed lot could help reduce damage caused by plant pathogenic virus synergism. Finally, it is possible that the asymptomatic, seed-transmitted isolate identified here could recombine with symptomatic isolates present in the field resulting in a more aggressive and destructive seed-transmissible strain that could cause significant damage to soybeans in the field. Therefore, it is important to identify seed lots with SVNV and remove contaminated seed before it makes it to breeding nurseries or farmer fields. Experiments are in progress to determine the additional agronomic importance and to determine the mechanism by which seed transmission of SVNV is occurring. The authors would like to thank Hyun Jin Hwang for assistance in conducting experiments and Anna Whitfield and Medhi Kabbage for thorough review of this manuscript. Funding for this work was provided in part by the Wisconsin Soybean Marketing Board and the USDA-ARS Floriculture and Nursery Research Initiative. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: CG TG RD DM DLS. Performed the experiments: CG TG RD. Analyzed the data: CG TG RD DM DLS. Contributed reagents/materials/analysis tools: CG TG RD DM DLS. Wrote the paper: CG TG RD DM DLS. Made initial field sample collection: DM. Conducted initial experiments to identify SVNV positive seed samples: CG.

# Introduction X-linked adrenoleukodystrophy (X-ALD) is a peroxisomal disorder characterized by impaired β-oxidation of very long-chain fatty acids (VLCFA) and accumulation of these VLCFA in tissues. It is caused by mutations in the *ABCD1* gene ([www.x-ald.nl](http://www.x-ald.nl)). The disease is highly variable in clinical expression, however, in adulthood it most frequently manifests as a gradually progressive myelopathy and peripheral neuropathy (adrenomyeloneuropathy phenotype or AMN). Treatment for AMN is purely symptomatic and currently there is no proven intervention that can halt progression of the disease. We identified ELOVL1 as the enzyme responsible for the synthesis of VLCFA, and demonstrated that siRNA-mediated knockdown of ELOVL1 lowers VLCFA levels in X-ALD fibroblasts. Next, we showed that bezafibrate (BF) reduces VLCFA levels in X-ALD fibroblasts by directly inhibiting ELOVL1. BF is a drug of the fibrate class for the treatment of dyslipidaemia and has a proven safety profile for (long-term) use in humans. We therefore designed a proof of principal clinical trial to test whether BF can reduce VLCFA levels in the plasma and lymphocytes of patients with X-ALD. # Methods The protocol for this trial and supporting CONSORT checklist are available as supporting information; see Checklist S1 and Protocol S1. The BEZA trial study protocol was approved by the Institutional Review Board (Medisch Ethische Toetsings Commissie) of the Academic Medical Center. The trial is registered at clinicaltrials.gov (NCT01165060). Adult men with biochemically and genetically proven X-ALD without contra-indications for the use of BF were eligible for inclusion. All participating patients were evaluated at baseline for eligibility and received trial medication after written informed consent was obtained. They were evaluated at intervals of 4 weeks until the end of the trial at 24 weeks. The initial dose of BF was 400 mg per day, which was subsequently increased to 800 mg per day at week 12. At each visit side effects were monitored, a general physical examination including weight was performed and blood samples taken. Blood samples were taken in the morning after an overnight fast before the first medication dose. Blood samples were analyzed at the laboratory for clinical chemistry for routine laboratory tests. VLCFA and BF levels were analyzed as previously described. Lysophosphatidylcholine-C26∶0 (C26∶0 lysoPC) was analyzed in bloodspots. Data were analyzed with PASW statistics, version 18 (IBM). Statistical significance was evaluated with a paired t-test. # Results Ten males with AMN participated in the trial. No side effects that necessitated discontinuation of the trial medication occurred. Body weight was unchanged. There was a clear reduction in plasma triglycerides (1.34 mmol/L to 0.70 mmol/L at BF 400 mg and 0.71 mmol/L at BF 800 mg), and to a lesser extent a decrease in total cholesterol and LDL-cholesterol. There was also an increase in HDL- cholesterol. These are known effects of BF and confirm patient adherence. There was no consistent reduction in C26∶0 in plasma or lymphocytes, neither at 400 nor at 800 mg BF per day. We observed an increase in plasma C22∶0 and C24∶0 at a dose of 800 mg BF per day. The amount of C26∶0 lysoPC was unchanged in blood spots after 24 weeks of treatment with BF. The plasma level of BF did not exceed 25 µmol/L at the highest dose of 800 mg BF per day. # Discussion The pathophysiology of X-ALD is not well understood, although it seems likely that accumulation of VLCFA is toxic and related to neurodegeneration. Therefore drugs that reduce the level of VLCFA might be effective in halting or slowing progression of the disease. Recently, we showed that it is possible to reduce VLCFA in fibroblasts from X-ALD patients by inhibiting the synthesis of VLCFA by the enzyme ELOVL1. We later showed that this can also be accomplished by incubating fibroblasts from X-ALD patients with BF. BF is a drug that has been in use for decades for the treatment of hypertriglyceridaemia and has an excellent safety profile. Therefore we decided to initiate this small scale proof of principle clinical trial to investigate whether BF reduces VLCFA in plasma and lymphocytes of X-ALD patients. In a previous clinical trial with lovastatin we demonstrated that reduction of plasma VLCFA can be an artifact of LDL reduction and does not reflect a reduction in blood cells. Unfortunately, we could not show a reduction on plasma or lymphocyte VLCFA levels. Conversely, there was an unexpected increase in C22∶0 and C24∶0 levels in plasma. We did not observe this in blood cells or bloodspots. Our results show that there is no rationale for a large follow-up trial with clinical endpoints utilizing this compound. The concept of treating X-ALD patients with an inhibitor of VLCFA synthesis remains a feasible option. It seems that BF is simply not efficacious enough. Our previous work suggests that BF is a competitive inhibitor of ELOVL1. In our cell culture experiments a high concentration of BF of 400 µmol/L was required to achieve a maximal effect on the level of VLCFA. At this concentration the *de novo* VLCFA synthesis was reduced to the level in control cells. It is likely that even with the high dose of 800 mg BF per day, the intracellular levels of BF remained inadequate. Indeed, at the highest BF dosage plasma levels did not exceed 25 µmol/L with an average of 10 µmol/L. These levels are not peak levels, but rather residual plasma levels. It is unlikely that concentrations even approaching 400 µmol/L were reached. This may explain the lack of *in vivo* efficacy of BF on our outcome parameters. To achieve the effect of VLCFA reduction, significantly higher BF concentrations are necessary as compared to concentrations indicated for reduction of TG. Future research will be focused on the identification of specific inhibitors of ELOVL1 that act at much lower concentrations than BF and are well-tolerated. In conclusion, BF appears to have no therapeutic utility in X-ALD. # Supporting Information [^1]: Conceived and designed the experiments: ME RW BTP SK. Performed the experiments: ME LT RO JB AM ID. Analyzed the data: ME RO SK. Wrote the paper: ME BTP RW SK. [^2]: The authors have declared that no competing interests exist.

# Introduction Alcoholism is a most devastating disease affecting broad parts of the western society. In the US alone, 14 million people meet standard criteria for alcohol abuse or alcoholism, triggering massive expenses in health care (approximately \$180 billion dollars per year). The devastating disease process often causes enormous harm to the patients, their family members and their social environment. Alcohol abuse is a prototypic complex disease which is determined jointly by multiple genes and environmental influences. Numerous studies have found that stress increases alcohol consumption in animals and humans. Individual differences are based on both environmental influences and genetic factors. Studies in genetically modified animals or with distinct pharmacological interventions described that individual alcohol consumption is influenced by endogenous peptidergic systems. Our own experiments focusing on the renin-angiotensin system (RAS) confirm this concept. Interestingly, numerous peptides involved in these motivational procedures (e.g. opioids, CRH, angiotensin) are enzymatically degradable by a small group of peptidases, including neprilysin (NEP; neutral endopeptidase, EC.3.4.24.11). NEP is a widely distributed transmembranal metallopeptidase with a broad spectrum of endogenous substrates and being associated with human diseases like Alzheimer disease and obesity. Therefore, we wanted to investigate in detail the effect of the lack of NEP activity on voluntary alcohol consumption. # Results and Discussion The catabolic action of NEP on several peptides involved in individual alcohol intake (e.g. opioids, angiotensin II, substance P) led us to investigate the impact of its deficiency on alcohol consumption. While we primarily identified such NEP-deficient animals as mice with significantly more alcohol intake, a repetition of this approach in new animal facilities could only demonstrate a tendency for an increased alcohol intake in such mice having free access to alcohol for 4 weeks in comparison to their age-and gender-matched wild-type controls. Taking the slight increase in total fluid consumption (second left panel) into consideration, such trend in alcohol intake has been almost blunted when the alcohol/total fluid ratio was calculated (second right panel). Importantly, we could also show in an independent set of animals that the kinetics of alcohol degradation was not different between NEP-deficient mice and their wild-type controls. The only difference between both animal facilities were possible stress components within the first round of experiments due to livestock keeping of animals in a frequently used open-plan room and an experimental setting using groups of 4–5 animals. In the second setting however, the experiments were performed in a separate experimental room and constructed by 2 animals per cage separated by a perforated plexiglas wall, avoiding the development of social stress. Since NEP also degrades peptides being involved in stress-motivated drinking (e.g. CRH;), the authors therefore hypothesized that NEP is a strong candidate for the elusive genetic factor in the relationship between stress and alcoholism. Therefore, we carefully arranged two-bottle-free-choice experiments under stress-free conditions and under social stress conditions in male NEP-deficient mice (n = 22) and corresponding wild-type controls (n = 16). Four-month-old animals were observed for 21 days and as before no difference in alcohol intake (10% alcohol) between both genotypes could be identified under such new conditions avoiding the development of stress. On day 22, animals were stressed by intruder as described in detail by Sillaber and colleagues. As shown in, NEP- deficient males drank significantly more alcohol after stress, while the wild- types did not show increase in alcohol consumption (day 26–48). Notably, such increase occurred very fast since the second measurement after stressing the mice showed already the significantly increased alcohol consumption in the NEP knockouts. However, approx. 3 weeks after stress, NEP deficient animals returned back to normal. After such normalization phase (day 49–70), the animals were re- stressed with the same stress method. As described above for the first intruder- mediated stress, NEP–deficient mice drank significantly more alcohol (day 75–97). However, in contrast to the first stress, there was no fading and NEP- deficient mice remained drinking significantly more alcohol until the end of the experiment (day 158). Although the number of candidate peptides affected by NEP alteration is significant, we started already first investigations showing e.g. lower enkephalin degradation in a variety of brain regions of NEP-deficient mice. Such higher enkephalin levels, however, did not influence the alcohol consumption under basal (unstressed conditions) (data not shown). Thus, still lacking a mechanistic explanation for such finding, the alcohol data might illustrate the decisive role of NEP for alcohol preference in mice and probably also for alcoholism in men. To evaluate whether the well-documented relation between stress sensitivity and increased alcohol intake in men could also be the basis for our findings in NEP- deficient mice, we tested the hypothesis that NEP-deficient animals might have a different stress response and therefore a higher alcohol preference. Both genetic lines were tested in two independent behavioral paradigms. In the EPM, total number of arm entries and number of closed-arm entries are usually considered as measures of general activity. Its reduction and a longer time in the closed arm are interpreted as anxious behavior. However, the NEP- knockout males did not show any alteration for such parameters implicating no increase in anxiety in mice with NEP deficiency. On the other hand, an overwhelming urge to escape from an environment or situation can be an important feature of panic disorder. In line with this, Blanchard and colleagues have proposed that this pattern of behavior shown by mice in threatening contexts may be analogous to the behavioral symptoms of panic. The EPM test is basing on a conflict between exploration and aversion to elevated open places. As shown in, NEP-deficient animals crossed significantly more often the area between the arms, implicating that these animals are more responsive in unfamiliar situations and thus show a significant stronger stress response. To further evaluate the stress response of the NEP-deficient mice, a locomotor activity test was performed to measure further components of emotional behavior. The test was performed under different illumination levels (30 lux vs. 450 lux). It is well-known that the conditions of illumination can dramatically alter the response to a novel environment , whereby intense light conditions are considered to stimulate aversion in rodents. When tested under low-light conditions (30 lux), there were no differences between NEP-deficient mice and their wild-type controls in all parameters measured in this test, indicating no differences in the emotional state under less aversive conditions (data not shown). In contrast, when the Moti-test was performed under intense light conditions (450 lux), locomotor activity (in terms of distance and time travelled) and percent of time spent in the centre of the test box (which reflects components of emotional behavior) were significantly decreased in NEP- knockout mice. This identifies such mice to be more responsive in aversive situations, confirming the EPM data that NEP deficiency leads to a stronger behavioral response under stress conditions. To test whether the stress-mediated increase in alcohol consumption requires life-long NEP deficiency as in the genetically deficient NEP-knockout mice, we used a pharmacological approach in an independent experimental set of wild-type mice. Four-month-old animals were fed either with standard diet or food pellets with 200 mg/kg/day candoxatril (each group n = 8), a NEP-specific inhibitor, as a food additive. Under such conditions, candoxatril exclusively reduced the peripheral NEP activity. Notably, other peptidases like ACE and ECE were not altered in their activity under candoxatril treatment (data not shown). Fourteen days after treatment start, animals received 10% alcohol or water in a two- bottle-free-choice experiment. In contrast to the approach with NEP knockouts (2 animals per cage separated by a perforated plexiglas wall), social stress was not avoided (no separation) in the candoxatril experiment. As shown in, also the pharmacological NEP inhibition resulted in a highly significant increase in alcohol consumption. Notably, the increase in the preference ratio related only to more alcohol intake while the general fluid consumption was not influenced by candoxatril. Thus, taken together our data shows that the deficiency in NEP and the pharmacological inhibition of this peptidase lead to a stress-induced increase in alcohol intake in a 2-bottle-free-choice paradigm. Nowadays, a variety of pharmaceutical companies developed NEP inhibitors for the treatment of cardiovascular diseases. However, our data implicate that their use might have a significant side effect for people under social stress or with financial problems, since the NEP inhibition in such individuals might stimulate an alcohol abuse. Our findings are all the more important, since we used candoxatril, a NEP inhibitor unable to pass the blood-brain barrier, illustrating the crucial role of peripheral NEP for alcohol preference. Recently, a variety of authors have suggested the existence of a gut-brain axis, e.g. demonstrating the impact of signaling initiated by stimuli in the gut on exteroceptively generated emotions. Therefore, there are substantial mechanistic explanations for our findings that peripheral accumulation of endogenous NEP substrates is regulating alcohol consumption. Although it is still speculative which peptides, NEP is decisive for their generation or degradation, are responsible for the observed effects in mice lacking NEP-activity, we could already demonstrate in previous studies that NEP deficiency modifies the concentration of endogenous peptides like GLP-1, NPY and galanin, which has been demonstrated to influence behavior. Therefore, a better understanding of the molecular mechanisms involved in such gut-brain cross-talk for the motivation to voluntarily drink alcohol is an important prerequisite for the development of new therapies. Although a link between stress and alcohol is well accepted, there are only a very few papers showing a strict genetic determinant that is involved in that interrelation. Since the absence of NEP activity itself does not lead to signs of alcohol preference in mice but requires an environmental stimulus, our findings build a bridge between stress components and genetic factors in the development of alcoholism. Therefore, stimulating NEP activity might be a very attractive approach for the treatment of alcohol abuse. # Materials and Methods ## Animals We used male NEP-knockout mice that were originally generated by Lu and maintained in the breeding stocks of T.W. at the Charité, Campus Benjamin Franklin (CBF), Berlin, Germany. Homozygous NEP-knockout mice and wild-type controls were bred from parents, which were F2 after hemizygous mating and being on a C57Bl/6N background. Animals were housed in litters separated according to sex at 22±1°C in a 12 h/12 h light/dark cycle with unrestricted access to food and water. Recordings of physiological parameters were performed between 9:00 a.m. and 1:00 p.m. All experiments were performed under the regulation and permission of the Animal Care Committee of the Erasmus MC, Rotterdam, The Netherlands and the local authorities in Germany (Landesamt für Gesundheit und Soziales des Landes Berlin). The investigation conforms to the *Guide for the Care and Use of Laboratory Animals* published by the US National Institutes of Health (NIH Publication No. 85–23, revised 1996). ## Preference test in two-bottle-free-choice paradigm Experimental animals were kept in groups of two animals per cage. Each cage was divided into two equal compartments by a Plexiglas divider as we have described previously. The mice were held in a free-choice paradigm with a bottle of tap water and a bottle containing a 10% (v/v) ethanol solution. Food and beverage were available *ad libitum*. Water and alcohol consumption were recorded every third day. ## Two-bottle-free-choice paradigm under social stress Social stress experiments were performed as previously described. In brief, single experimental animals have been set into a cage with three unfamiliar age- matched male mice, which have been together in their cage for at least 1 week. The new animal has been immediately attacked by the home-cage mice (less than a minute). Directly after such physical contact, the experimental mouse has been separated from the group within the cage by introducing a Plexiglas wall between the 3 home-cage mice and the single mouse. Such visual contact has been kept for 15 min. Then, the experimental mouse has been transferred back to its own home cage and the two-bottle-free choice experiment has been continued. ## Treatment with candoxatril Standard food (“ssniff SM/R/N-H (10 mm)”; ssniff Spezialdiäten GmbH, Soest, Germany) was cold milled, supplemented with the frequently used, well tolerated, orally active and tasteless NEP-inhibitor candoxatril (Pfizer Deutschland GmbH, Karlsruhe, Germany), mixed and grouted under pressure (2 g/kg); this corresponds to a daily consumption of approx. 200 mg/kg/day, a dose that has been identified in dose-finding experiments. The manufacturing of candoxatril-containing food was performed as usually by ssniff Spezialdiäten GmbH. The used control food was prepared at the same day by the same management. ## Kinetics of alcohol degradation The measurement of blood alcohol concentration after oral application of 10% alcohol was performed using the microphotometer LP20, according to the manufacturer's manual (LZC32, 2000) of Dr. Lange GmbH & Co D-14163 Berlin (Germany). ## Measurement of NEP activity NEP activity was measured in kidney and forebrain homogenates on the basis of a general method described by Winkler. \[D-Ala², Leu⁵\] enkephalin (DALEK, 200 μM) in a 50 mM Tris-buffer (pH 7.4) was used as substrate. To prevent the degradation of DALEK by ACE and by aminopeptidases, lisinopril (10⁻⁶ M) and bestatin (10⁻⁴ M) were added. The reaction was performed in a thermo-shaker at 37°C, and stopped by addition of 0.35 M perchloric acid (HClO₄), centrifuged (5.000 g), and the supernatants stored till HPLC-analysis at 4°C. The resulting degradation product Tyr–D-Ala–Gly (TAG), as well as the remaining DALEK concentration, were measured by high performance liquid chromatography (HPLC) using a RP-C18 column and an isocratic fluid phase consisting of an acid perchlorate-phosphate-buffer containing 6% acetonitrile. The peptide peaks as well as standard concentrations of TAG and DALEK were detected at 216 nm (UV). The NEP-specificity of the reaction was characterized in parallel assays using the NEP inhibitor 10⁻⁵ M candoxatrilat (UK73,967; Pfizer, Karlsruhe, Germany). ## ACE activity ACE activity was measured with a fluorimetric method using Hip-His-Leu as substrate and His-Leu as standard reagent as described previously. Fluorescence arising from His-Leu after reaction with o-phthalaldehyde was measured at 365 nm (excitation) and 500 nm (emission). The activity is expressed as nmol His- Leu/min/mg protein. ## ECE activity ECE-1 activity was measured as described in detail by Johnson and Ahn using the DNP-quenched fluorogenic substrate Mca-Arg-Pro-Pro-Gly-Phe-Ser-Ala-Phe-Lys-DNP. ## \[Leu⁵\]enkephalin degradation \[Leu⁵\]enkephalin degradation by membrane preparations was investigated as previously described. In brief, 100 µM \[Leu⁵\]enkephalin in 50 mM Tris-buffer (pH 7.4) was incubated with membrane preparations of a final protein concentration of 0.5 mg/ml. The reaction was stopped by addition of 0.35 M HClO₄. The remaining \[Leu⁵\]enkephalin concentration was quantified by HPLC (Shimadzu, isocratic gradient of acetonitril \[25%\] and 75% NaClO₄ \[0.15 M\]/NaH₂PO₄ \[0.01 M\] buffer \[pH 2.2\], nucleosil column \[100 C18\]) at 216 nm. ## Elevated plus maze Fear was measured in an elevated plus maze (EPM). The maze was made of black polyvinyl chloride and had two open and two closed arms (50×10×40 cm) mounted 50 cm above the floor. The floor of the arms was smooth. Light level was 30 lux. A mouse was placed on the central platform of the apparatus facing a closed arm. A camera on the ceiling of the test room was used to score and tape the animal's behavior from an adjacent room for a period of 7 min. Number of entries into open arms and time spent on open and closed arms were recorded. Arm entry was defined as all four feet in the arm. Animals were tested in a random order, and the maze was cleaned after each trial. Parameters were taken from 16 wild-type and 29 NEP-deficient mice. The person scoring the behavior was blind to the genetic constitution of the animals. ## Locomotor activity Locomotor activity was monitored over a period of 15 min at 30 lx or 450 lx in a computerized activity meter (46×46 cm) with 15 photocells in each dimension (MOTITEST, TSE Bad Homburg, Germany). Travel distance and mean total activity (horizontal plus vertical) and time on the central field (25×25 cm) of 16 wild- type and 29 NEP-deficient mice were recorded. The person performing the test was unaware of the genetic constitution of the animals. ## Statistical analysis Results are represented as mean±SEM. For determination of intergroup differences, Student's *t* test was used (InStat 2002, GraphPad, San Diego, USA). Significance was considered from a value of *P*\<0.05. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: WES TW. Performed the experiments: BM MB FG. Analyzed the data: AB HPS WES TW. Contributed reagents/materials/analysis tools: AB HPS. Wrote the paper: FG WES TW.

# Introduction Rheumatoid arthritis (RA) is a common chronic inflammatory joint disease. RA patients suffer from chronic fatigue. Pain, joint stiffness and functional disability are most prominent in the morning. These symptoms reflect abnormal circadian rhythms of circulating inflammatory cytokines TNFα and IL-6 as well as serum cortisol in RA. Many physiological and pathological processes are under circadian regulation. A central circadian pacemaker is located in the suprachiasmatic nucleus (SCN) of the hypothalamus. Because circadian rhythm of the SCN is not exactly 24h in humans, light adjusts the rhythm of the central pacemaker. The central circadian pacemaker synchronizes the peripheral molecular pacemakers located in all other tissues. The function of the molecular clock is based on rhythmic oscillation of transcription and translation of reciprocal clock genes. Brain and muscle Arnt- like protein-1 (BMAL1 or ARNTL) and Circadian Locomotor Output Cycles Kaput (CLOCK) form a heterodimer which upregulates clock controlled genes by binding to an E-box element in the promoter of the clock controlled genes. Some of the upregulated genes, Periods and Cryptochromes, form the components of the best known negative feedback loop. The clock molecules, DEC1 (BHLHE40) and DEC2 (BHLHE41), form another less known negative feedback loop because they compete with BMAL1/CLOCK for E-box binding. Clock genes are needed for proper immune cell function. Notably, autoimmune diseases develop in aging DEC1 deficient mice which display increased production of IL-4 and IL-10 without affecting IFN-γ. In contrast, its paralogue DEC2 is selectively expressed in Th2 cells and enhances their development in mice leading to improper Th2 responses in asthma and parasite response models. In addition, a connection between circadian clock and arthritis has been described. Arthritis is exacerbated in Cry1 and Cry2 deficient mice and dysfunctional clock is present in RA patients. TNFα affects the clock and in human cells, in contrast to that of mice, the upregulated genes are ARNTL2 and NPAS2, functional paralogues of BMAL1 and CLOCK, respectively. Paradoxically, however, TNFα suppresses clock controlled genes DBP and PER3. Thus, we hypothesized that the negative regulators in the molecular clock DEC1, DEC2 or both are affected by TNFα. Because of our hypothesis and their central role in immune cell function, their regulation *in vitro* by TNFα and expression *in vivo* in RA were studied. # Materials and Methods ## Subjects The research plan and this study were approved by the ethical committee of the Helsinki University Central Hospital (Dnro 165/E6/03). Written informed consent from each patient was obtained to collect sample for research purposes. Guidelines of the Declaration of Helsinki were followed. RA patients fulfilled the 2010 ACR-EULAR classification criteria of RA. None of the patients were treated with anti-TNF agents or other biologicals. Tissue samples of both RA (n = 6) and OA (n = 5) patients were taken at 10 a.m. ± 2 h during synovectomy or operation for total joint replacement. Samples were formalin fixed and embedded in paraffin. ## Cell culture Primary human fibroblast cultures (n = 6) were established and characterized as previously described. Briefly, tissue samples were minced into small pieces with a sterile scalpel in a laminar flow hood. The explants were left overnight in RPMI-1640 medium containing 10% fetal bovine serum with 1000 U/ml penicillin and 1 mg/ml streptomycin (10×) solution. The next day, the media were changed to basal RPMI with 10% FBS media and 100 U penicillin and 0.1 mg streptomycin (1× solution). The medium was changed twice a week. The explants were removed until roughly 80% monolayer confluence was reached, and the cells were subcultured 1:3 until confluent. The cells were frozen at passage 2 for subsequent experiments. After thawing, the cells were cultured in RPMI-1640 medium (Lonza Group, Basel, Switzerland) containing 10% fetal bovine serum (FBS; Lonza) 100 IU/ml penicillin and 0.1 mg/ml streptomycin and used in passages 4–5. Stimulation and inhibitor experiments were performed with three different donor fibroblasts. Transfection experiments were performed with single donor fibroblasts. In Amaxa Nucleofector II transfection experiments, fibroblasts were cultured in DMEM medium (Thermo Fisher Scientific, Waltham, USA; cat# 41965) containing 10% FBS (Lonza) with 100 IU/ml penicillin and 0.1 mg/ml streptomycin. The synchronization of the molecular clock in cells was performed as described elsewhere with with minor modification. Briefly, cultured human primary fibroblasts were seeded on 24-well plates at 4x10⁴ cells per well in RPMI-1640 containing antibiotics and 1% FBS, cultured for 24 h after which the medium in wells was replaced with RPMI-1640 media containing antibiotics, 1% FBS and TNFα, IMD-0354 or DMSO when indicated. HEK293 cells were cultured in DMEM medium (Thermo Fisher Scientific, cat# 41965) containing 10% FBS (Lonza) with 100 IU/ml penicillin, 0.1 mg/ml streptomycin and 1 mM pyruvate (Lonza, cat# BE13-115E). ## Cell stimulation Human primary fibroblasts were synchronized as described in the previous section. At t = 0, the media was replaced with RPMI-1640 media containing antibiotics, 1% FBS and TNFα (10 ng/ml; R&D Systems, Minneapolis, USA) or with media containing no added stimulants (negative control). At indicated times, the wells were washed with PBS and cells were lysed with 350 μl RLT lysis buffer (RNeasy kit, Qiagen, Hilden, Germany). To study the effect of NF-κB inhibition on DEC2 regulation, IKK-2 inhibitor IMD-0354 (cat# I3159; Sigma-Aldrich Corporation, St. Louis, USA) was used. 24 h after plating the cells, the media was replaced with RPMI-1640 containing antibiotics, 1% FBS, and IMD-0354 in a final concentration of 1 μM or DMSO in the same final concentration as was achieved when IMD-0354 (dissolved in DMSO) containing media were added. After 20 minute incubation (t = 0) TNFα (R&D Systems) was added to the wells to a final concentration of 10 ng/ml. At the indicated times, the wells were washed with PBS and cells were lysed with 350 μl RLT lysis buffer (Qiagen). ## RNA isolation, cDNA synthesis and quantitative real-time PCR RNA was isolated using RNeasy kit (Qiagen) according to the manufacturer’s instructions. RNA concentrations were measured using NanoDrop ND-1000 instrument (Thermo Fisher Scientific). The cDNA synthesis was performed using 500 ng of total RNA and iScript^™cDNA Synthesis Kit (Bio-Rad Laboratories, Hercules, USA) in a 20 μl reaction volume. After cDNA synthesis the cDNA was diluted to 1:5. Quantitative real-time PCR was performed from diluted cDNA in iQ^™ SYBR^® Green Supermix (Bio-Rad) using gene specific primers in 20 μl reaction volume. The PCR was performed in iQ5 real-time PCR detection system (Bio-Rad). RPLP0 was used as a housekeeping gene. ## Plasmids and vectors DEC2 (NM_030762) cDNA was amplified from human primary fibroblast total cDNA and was inserted into pcDNA3.1 V5 hisA vector (Thermo Fisher Scientific). The following primers were used for DEC2 cDNA amplification: sense `5’-AACGAAGGATCCGCCACCATGGACGAAGGAATTCCTCATTTGCA-3’` and antisense `5’-GGACGCCTCGAGTCAGGGAGCTTCCTTTCCTGGCT-3’`. 2 kb part of IL-1β promoter (NG_008851.1) was amplified from Human Genomic DNA (Roche Basel, Switzerland; cat# 11691112001) and inserted into pGL3-Enhancer vector (Promega Corporation, Fitchburg, USA). The following primers were used for amplification: sense `5’-AATTTGGGTACCAATGCTGTCAAATTCCCATTCACCCA-3’` and antisense `5’-TACTTCCTCGAGGGCTGCTTCAGACACTTGAGCA-3’`. The constructs were validated by using nucleotide sequencing. For dual-luciferase assay the control vector was pRL-TK (Promega). Vectors were propagated in competent TOP10 *Escherichia Coli* cells (Thermo Scientific). Ultrapure endotoxin-free plasmid DNA was prepared using NucleoBond^® Xtra Midi EF (Macherey-Nagel, Düren, Germany; cat# 740420) according to the manufacturer’s instructions. Plasmid DNA was diluted in a sterile water. ## Transfection HEK293 cells were seeded on 24-well plates at 4x10⁴ cells per well in 0.5 ml DMEM medium and incubated for 24 h before transfection. For transfection, Fugene HD transfection reagent (Promega, cat# E2311) was used according to manufacturer’s instructions with 500 ng DNA and DNA:Fugene HD ratio of 1:3. All cell manipulations and assays were carried out 48 hours after transfection. Human primary fibroblasts were transfected using Amaxa Nucleofector II (Lonza) and Amaxa Human Dermal Fibroblast Nucleofector Kit (cat# VPD-1001). Transfection was performed according to manufacturer’s instructions using 4x10⁵ cells, 3 μg DNA and transfection program U-O23. Immediately after transfection cells were seeded on 12-well plates at 1x10⁵ cells per well in 1 ml DMEM medium. All cell manipulations and assays were carried out 24 h after transfection. ## Luciferase assay Transfection of HEK293 cells was carried out as described using 500 ng of DEC2 expression plasmid or empty control plasmid, 10 ng of reporter plasmid and 1 ng of Renilla luciferase plasmid. Luciferase assay was done using Dual- Luciferase^® Reporter Assay System (Promega, cat# E1910) according manufacturer’s instructions 48 h after transfection. Luminescence was measured using Plate CHAMELEON V Multilabel Microplate Reader (Hidex, Turku, Finland). ## siRNA transfection Human primary fibroblasts were seeded on 24-well plates at 4x10⁴ cells per well in 0.5 ml RPMI-1640 containing antibiotics and 1% FBS. After 12 h, siRNA transfection using RNAiMAX transfection reagent (Thermo Fisher Scientific, cat# 13778) was performed according to manufacturer’s instructions. Briefly, 1.5 μl of Lipofectamine RNAiMAX diluted in 25 μl OPTI-MEM (Thermo Fisher Scientific, cat# 31985) and 15 pmol of ON-Targetplus Human DEC2 (Thermo Fisher Scientific, cat# 79365) siRNA diluted in 25 μl OPTI-MEM were combined and incubated for 5 min at room temperature (RT) after which 50 μl of transfection mix was added per well. After 12 h (t = 0), the media were replaced with RPMI-1640 containing antibiotics, 1% FBS and 10 ng/ml TNFα (R&D Systems) or no added stimulants (negative control). After 10 h, the wells were washed with PBS and lysed with 350 μl RLT lysis buffer (Qiagen). ## Immunofluorescence Human primary fibroblasts were seeded at 1x10⁵ cells per well on coverslips placed in 12-well plates containing RPMI-1640 supplemented with antibiotics and 1% FBS. Before stimulations the cells were synchronized as described above. For cellular stimulation the media were replaced with RPMI-1640 containing antibiotics and 1% FBS, without or with 10 ng/ml TNFα (R&D Systems). After 24 h cells were washed with PBS and fixed in 4% PFA for 15 min at RT. Fixed cells were permeabilized with 0.1% Triton-X in PBS for 10 min at RT, blocked with 1% BSA-PBS for 1 h at RT, after which slides were incubated with 1 μg/ml rabbit anti-human DEC2 IgG (Santa Cruz Biotechnology, Dallas, USA; cat# sc-32853) or 1 μg/ml non-immune rabbit IgG at 4°C overnight. Next day slides were incubated in 1:100 dilution of Alexa Fluor 568 labeled goat anti-rabbit IgG secondary antibody (Molecular Probes, Leiden, The Netherlands; cat# ab175471) for 1 h at RT, counterstained in 5 μg/ml DAPI and mounted. ## Immunohistochemical staining Formalin-fixed and paraffin-embedded tissue samples of synovial membranes were cut to 3 μm sections, deparaffinized and rehydrated. Antigens were retrieved in citrate buffer using microwaves (Program AR98C-S30M, MicroMED T/T Mega Histoprocessing Labstation; Milestone Srl, Sorisole, Italy) followed by quenching of endogenous peroxidase in 3% H₂O₂ in PBS for 15 min. Sections were incubated in 0.67 mg/ml rabbit anti-human DEC2 IgG (Santa Cruz, cat# sc-32853) at 4°C for overnight. Rabbit IgG at the same concentration was used for negative control staining. Slides were washed with PBS following incubation in biotin-conjugated goat anti-rabbit IgG secondary antibody for 1 h at RT. After washes, slides were incubated for 1h at RT in freshly prepared avidin–biotin–peroxidase complexes (Vector Laboratories, Burlingame, USA; Vectastain Elite ABC kit). Color was developed using H₂O₂ and DAB. Between each step slides were washed at least three times with PBS. Finally, slides were dehydrated, counterstained in haematoxylin and coverslips were mounted using Mountex (Histolab, Västra Frölunda, Sweden). ## Statistical analysis The data of IL-1β or DEC2 expression after TNFα stimulation was analyzed with repeated measures ANOVA. Significance was tested using Bonferroni. Reported p-value is difference of TNFα stimulation and mock group. The means of the experiments with two independent samples were tested using student’s t-test. Tests were performed with SPSS 15.0 for Windows (SPSS Inc. Chicago, IL). All results are expressed as mean ± SEM unless otherwise stated in the figure legend. # Results ## TNFα stimulates the expression of DEC2 but not DEC1 To study the eventual TNFα effects on DEC1 and DEC2, synovial fibroblasts were synchronized by serum starvation after which they were stimulated without or with 10 ng/ml TNFα. TNFα upregulates IL-1β, which was therefore used as a positive control in TNFα stimulation experiments. TNFα-mediated increase of IL-1β (p \< 0.05, F 9.6, df between groups 1,6) confirmed that the stimulation was successful. Samples collected at 1, 2 and 4 hours and then every 4 hours up to 32 hours were analyzed for DEC1 (which was not changed, data not shown) and DEC2 mRNA. TNFα increased DEC2 expression 4-fold (p \< 0.001, F 50.6, df between groups 1,6) already at 2 hours and this effect was maintained until the 32 hour time point. The effect of TNFα on DEC2 was also shown by immunofluorescence staining of TNFα stimulated synovial fibroblasts. DEC2 was increased also at the protein level and mainly localized in nuclei of TNFα stimulated cells. To test if TNFα effect on DEC2 expression is mediated by NF-κB pathway, synovial fibroblasts were stimulated as above but first after 20 min pretreatment with 1 μM IKK-2 inhibitor IMD-0354. Successful inhibition was confirmed by studying the expression of IL-1β (p \< 0.05, t-value 4.1, df 4). Samples collected at 16 hours of stimulation (the highest peak of DEC2 expression) were analyzed for DEC2 mRNA. IMD-0354 significantly (p \< 0.001, t-value 9.0, df 4) inhibited the TNFα-induced DEC2 expression. The 15-fold expression was reduced to only 2-fold when NF-κB pathway was inhibited. ## DEC2 overexpression stimulates IL-1β expression in HEK293 cells and in human fibroblasts Because TNFα increases the expression of DEC2 and IL-1β, it was hypothesized that DEC2 itself might contribute to the upregulation of IL-1β. To test this hypothesis, DEC2 gene was cloned and overexpressed in HEK293 cells. DEC2 downregulates Per1, which was therefore used as a positive control of DEC2 function in HEK293 cells and in synovial fibroblasts. Both experiments demonstrated that DEC2 significantly reduced the expression of PER1 (p \< 0.05, t-value 4.2, df 4 in HEKs and p \< 0.01, t-value 5.6, df 4 in fibroblasts). In addition to this, DEC2 inhibited the expression of DBP and PER3 (not shown) confirming that its overexpression may contribute to the reduction of clock output genes after TNFα stimulation. DEC2 overexpression increased the expression of IL-1β mRNA 8-fold (p \< 0.001, t-value 9.6, df 4) in HEK293 cells and 3-fold (p \< 0.01, t-value 4.6, df 4) in human synovial fibroblasts compared to empty vector controls. Because CCL8 and CXCL5 are regulated by components of the circadian clock, we investigated their regulation by DEC2 in human cells. Indeed they were significantly (p \< 0.05, t-value 3.5, df 4 for CCL8 and p \< 0.01, t-value 6.4, df 4 for CXCL5) regulated by DEC2 in human synovial fibroblasts. ## DEC2 overexpression further increases TNFα responses TNFα induces the expression of IL-1β both in human fibroblasts and in HEK293 cells (hundred fold; data not shown). It may well be that DEC2 only induces IL-1β in unstimulated cells. Thus, we wanted to test the effect of DEC2 during TNFα stimulus. Overexpression of DEC2 in HEK293 cells increased IL-1β mRNA levels in response to TNFα 4-fold. Accordingly, DEC2 also increased TNFα mediated IL-1β promoter activity (p \< 0.05, t-value 4.3 df 4) suggesting that this increase results in part from increased transcription. This effect was also true in human fibroblasts. Overexpression of DEC2 also in these cells increased IL-1β and CCL8 mRNA levels in response to TNFα. ## DEC2 silencing decreases TNFα responses If the results from overexpression experiments were true, silencing of DEC2 should lead to decrease of IL-1β expression. To verify the results, silencing of DEC2 using siRNA was performed. Indeed silencing of DEC2 (p \< 0.005, t-value 6.5, df 4) declined the IL-1β increase (p \< 0.05, t-value 3.0, df 4) in response to TNFα in human fibroblasts. ## DEC2 protein is abundant in the synovial membrane in RA Due to the above described *in vitro* effects of TNFα on upregulation of DEC2, RA and OA synovial tissues were immunostained for the presence of DEC2. DEC2 staining was much more intense and extensive in RA synovitis tissue than in more mildly inflamed OA synovial tissue samples. Negative staining controls confirmed the specificity of the staining. # Discussion BMAL1/CLOCK heterodimer is the major component of the molecular pacemaker responsible for the normal homeostatic circadian rhythm. Its major counter- regulators are PER1-3 and CRY1-2, which in various complexes cyclically oscillate in a fashion reciprocal to that of the BMAL1/CLOCK, regulating the length of the circadian cycle. However, yet another regulatory paralogue pair exists in the negative feedback loop controlling unconstrained and continued effect of the BMAL1/CLOCK. Due to the apparently disturbed circadian rhythm in RA and the upregulated ARNTL2, NPAS2 but paradoxically downregulated DBP and PER3 mRNA expression after TNFα stimulations, the clock counter-regulators DEC1 and DEC2 were analyzed in resting and TNFα stimulated human synovial fibroblasts. Fibroblast was selected as the major target cell because it is an important cellular component of synovial stromal connective tissue, erosive pannus and synovial lining, in which fibroblast-like type B lining cells together with macrophage-like type A lining cells form its two cellular components. Because the central circadian pacemaker at SCN regulates the peripheral clocks in all peripheral cells, fibroblast should in principle be as good indicator of the regulation of the clock components as any other cell type. It was found that the pro-inflammatory cytokine TNFα stimulates DEC2 at both mRNA and protein level in a NF-κB-dependent manner in cultured human synovial fibroblasts. Further studies focused on DEC2 because its paralogue DEC1 was not affected by TNFα. IL-1β displays circadian rhythm in circulation and its expression is rhythmic in fibroblasts. Thus, we wanted to test if DEC2 by itself without upstream TNFα has some independent effects on IL-1β. DEC2 was cloned and first transfected to HEK293 cells. It was shown that IL-1β is increased in both DEC2- and TNFα- dependent manner in HEK293 cell. This was then confirmed for IL-1β via transcription and promoter activation and also for some other pro-inflammatory cytokines by overexpression of DEC2 in human synovial fibroblasts. Thus, TNFα exerts its inflammatory effects in part through DEC2 suggesting that this component of the molecular clock participates in the regulation of inflammatory responses also in human cells. This conclusion was further confirmed by silencing DEC2 with siRNA that significantly decreased TNFα-induced IL-1β expression. Although silencing of DEC2 was quite effective, its effect on TNFα- induced IL-1β was only partial. This suggests that the upregulation of IL-1β by TNFα is only partially DEC2-dependent. There are several signaling pathways and transcription factors that are known to be activated after TNFα stimulus. Thus, it is not surprise that DEC2 is not completely responsible for the regulation of IL-1β. NF-κB pathway is involved in the transcriptional activation of a vast number of inflammatory and apoptotic machinery genes in response to TNFα. DEC2 is involved in the control of apoptosis in cancer cells. Thus, the hypothesis was that TNFα induced DEC2 expression is mediated via the NF-κB pathway. Indeed, the induction of DEC2 was almost completely suppressed by the inhibition of IKK-2. DEC2 protein levels were much higher in RA synovial membrane than in OA synovial membrane. This is in accordance with the higher degree of inflammation and TNFα production in RA compared to that of OA. The high impact of TNFα on the pathomechanisms of RA is supported by the overall effectiveness of anti-TNF drugs in the clinical setting. Many different mechanisms of action of anti-TNF drugs have been suggested, such as diminished expression of vascular endothelial adhesion molecules and therefore diminished recruitment of inflammatory leukocytes to synovitis tissue. The present findings suggest that TNFα and anti- TNF drugs may also affect disease activity and progress via regulation of the circadian clock, which further participates in the regulation of immune responses and fatigue. It can be concluded that DEC2 is aberrantly expressed in RA tissue, it is induced by TNFα and not only affects the expression of genes belonging to molecular clock but also significantly impacts on the expression of IL-1β as well as other inflammatory genes. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JO VPK JH YTK JM. Performed the experiments: JO VPK JH. Analyzed the data: JO VPK JH YTK JM. Contributed reagents/materials/analysis tools: YTK JM. Wrote the paper: JO VPK YTK JM.

# Introduction Graves’ disease (GD) is an autoimmune disorder and is considered the most common cause of hyperthyroidism, followed by toxic multinodular goiter and toxic adenoma. Stimulation of the thyrotropin receptor by thyrotropin receptor antibodies (TRAb) is the primary mechanism of GD, which results in the production and release of thyroid hormones. The coexistence of thyroid cancer with GD is well known, and the American Thyroid Association (ATA) states that thyroid cancer occurs in up to 2% of patients with GD. In a previous study, among 847 patients with GD who underwent thyroidectomy, the incidence of coexistent thyroid cancer was 4.3%, and 68.2% were papillary microcarcinomas (PTMC). Although not clearly known, the probable mechanism of increased prevalence of thyroid cancer in patients with GD is primarily the binding of TRAb to thyrotropin receptor, which promotes tumor formation, angiogenesis, and further progression of the invasiveness of cancer. Usually, near-total or total thyroidectomy is recommended in patients having thyroid cancer with underlying GD. However, in the absence of underlying GD, thyroid lobectomy is preferred if indicated. The recent ATA guidelines recommended thyroid lobectomy as the initial surgical approach for low-risk PTMCs and for low-risk papillary thyroid carcinoma (PTCs) of size 1–4 cm. As the cases of lobectomy increases, GD diagnosed in the remnant thyroid lobe in some cases has been reported. However, the incidence of newly diagnosed GD after thyroid lobectomy is not well known. Furthermore, due to its rarity, little is known about preoperative factors that can predict the development of GD after lobectomy. Therefore, a cohort study with large sample size is necessary to fill this gap. We aimed to assess the incidence of newly developed GD after lobectomy for thyroid cancer in a retrospective cohort study conducted in a single tertiary center in Korea. Additionally, we aimed to evaluate the clinical and pathological characteristics of these patients and determine factors that might be helpful to predict the occurrence of GD after lobectomy. # Methods ## Patients We retrospectively reviewed the medical records of patients who underwent lobectomy for thyroid cancer between 2010 and 2019 at a tertiary medical center in Korea and were diagnosed with GD after thyroid lobectomy. This study was approved by the Institutional Review Board of the Asan Medical Center (No.: 2021–0621). ## Follow-up protocol for thyroid cancer and diagnosis of GD The data of preoperative thyroid function tests and neck ultrasonography (US) were available for all patients, and tests for preoperative autoantibodies were optionally performed at the discretion of the treating physician. Patients underwent lobectomy with/without prophylactic central neck dissection according to the surgeon’s decision. After thyroid surgery, serum free T4 (fT4) and TSH were measured within the first 2 to 3 months, and patients with overt hypothyroidism or subclinical hypothyroidism with a TSH level of \>10 mIU/L were treated with levothyroxine. During the long-term follow-up, patients were regularly subjected to physical examination, thyroid function tests, serum thyroglobulin, thyroglobulin antibody (TgAb) level measurements every 6–12 months, and neck US every 12–24 months. GD was suspected when the patient complained of symptoms of thyrotoxicosis, or the thyroid function tests showed overt or subclinical hyperthyroidism during routine follow-up. If the patient was taking levothyroxine, the thyroid function tests were performed again after discontinuation of levothyroxine. Serum thyrotropin-binding inhibitory immunoglobulin (TBII) and ^99m-technetium (Tc) thyroid scan uptake studies were also performed in patients with symptoms of thyrotoxicosis to distinguish from painless thyroiditis. The presence of goiter was determined according to the World Health Organization (WHO) goiter classification system via physical examination by a physician. Thyroid-associated orbitopathy (TAO) was diagnosed either by experienced endocrinologists or ophthalmologists. ## Treatment and follow-up for GD Patients with GD were initially treated with antithyroid drugs (ATDs); methimazole (15–30 mg/day), or carbimazole (20–40 mg/day). Serum fT4, TSH, and TBII were regularly measured every 2–3 months from the initiation of ATDs. Radioactive Iodine (RAI) therapy and completion thyroidectomy was considered when the patients failed to achieve a euthyroid state despite ATD treatment or if the patient preferred these treatment options. ## Laboratory measurement Serum TSH levels were measured using the TSH-CTK-3a radioimmunoassay (DiaSorin SpA, Saluggia, Italy) with a functional sensitivity of 0.07 mU/L. Serum fT4 levels were measured by radioimmunoassay (Immunotech, Prague, Czech Republic) with a functional sensitivity of 2.34 pmol/L. The reference ranges of TSH and fT4 were 0.4–4.5 mIU/L and 0.80–1.90 mg/dL, respectively. TBII was measured using the B*·*R*·A·*H*·*M*·*S TRANK human immunoradiometric assay *(*B*·*R*·A·*H*·*M*·*S GmbH, Hennigsdorf /Berlin, Germany) and titer ≥1.5 IU/L were considered positive with a functional sensitivity of 1.0 ± 0.2 IU/L. The thyroid peroxidase antibody (TPOAb) level was determined by radioimmunoassay (BRAHMS anti-TPOn RIA), and a value of ≥60 U/mL was considered positive. The TgAb level was also measured by radioimmunoassay (BRAHMS anti-Tgn RIA), and a value of ≥ 60 U/mL was considered positive. ## Statistical analysis Statistical analyses were performed using the R program (version 3.5.1, R Foundation for Statistical Computing, Vienna, Austria; [http://www.R-project.org](http://www.r-project.org/)). Continuous variables are presented as median and Inter Quartile Range (IQR), and categorical variables are presented as numbers (percentages). # Results ## Baseline characteristics of patients A total of 11043 patients underwent thyroid lobectomy for thyroid cancer between 2010 and 2019. Among them, 26 (0.2%) were newly diagnosed with GD during follow- up. The baseline clinical and pathological characteristics of the 26 patients are described in. The median age was 43.8 years (IQR 34.1–44.4), and 88.5% were female. 3.8% and 7.7% of the patients were current and ex-smokers, respectively. All patients were in the euthyroid state before thyroid surgery with a median TSH level of 2.3 μM/mL and fT4 level of 1.2 ng/dL. Preoperative thyroid autoantibody tests were performed in 18 (69.2%) patients, and TPOAb and TgAb positivity were detected in 11 (61.1%) and 8 (44.4%) patients, respectively. None of the patients had goiter before thyroid surgery or were checked for TBII. All patients were diagnosed with PTC, and the subtypes were as follows: classical type in 23 (88.5%) and follicular variant in 3 (11.5%) patients. The median tumor size was 0.7 cm (0.5–0.9) and 69.2% were PTMCs. Cervical lymph node metastasis was confirmed in 6 (23.1%) patients, and no structural recurrence was observed during a median 6 years of follow-up. Lymphocytic thyroiditis was found in 16 (61.5%) patients, and 11 (42.3%) patients were on levothyroxine after lobectomy due to confirmed hypothyroidism. ## Diagnosis of GD The median duration between thyroid surgery and diagnosis of GD was 3.3 (IQR 2.3–4.9) years; however, the time interval was very variable in different patients. Among 26 patients, 13 (50%) were diagnosed during routine follow-up of the thyroid function tests and had no definite symptoms. Ten (38.5%) patients had an unplanned visit due to symptoms of thyrotoxicosis such as tremors, palpitations, weight loss, and diarrhea. GD was diagnosed later in one patient who had newly developed goiter in the remnant thyroid lobe and in two referral patients who were diagnosed with TAO in the ophthalmology unit. One more patient who initially had no symptoms of thyrotoxicosis was also diagnosed with TAO during the treatment for GD. All patients underwent the autoantibody test when they were diagnosed with GD, and the positivity of TPOAb and TgAb was 76.9% and 53.8%, respectively, which were higher than the preoperative positivity rates. Among patients in whom the autoantibodies were measured preoperatively, 2 and 3 patients respectively showed positive conversion of TPOAb and TgAb, at the time of GD diagnosis. All the patients showed positive TBII at the time of diagnosis of GD. ## Treatment and course of GD For the treatment of GD, all patients received ATDs as the first-line therapy, and 21 (80.8%) continued ATDs for maintenance. RAI therapy and complete thyroid surgery were performed in 2 (7.7%) and 3 (11.5%) patients, respectively. All these five patients were on levothyroxine replacement after RAI therapy or surgery. Among 21 patients who were maintained on ATDs, 11 (52.4%) discontinued ATDs after a median 16 months of treatment and had remission of GD. However, one patient had a relapse of GD after 20 months of remission, and 10 (47.6%) patients were on maintenance ATDs at the last follow-up. # Discussion We found that the incidence of GD after lobectomy for thyroid cancer is 0.2% in a large cohort. Among patient who diagnosed with GD, 61.1% and 44.4% were positive for TPOAb and TgAb, respectively, before surgery. These rates are higher than those seen in the general population of patients with thyroid cancer, considering that autoantibody positivity is reported as approximately 18–23% in such patients. Furthermore, lymphocytic thyroiditis is documented in 61.5% of surgical specimens. These findings indicate that patients who developed GD were more likely to have autoimmune thyroid disease compared with those who did not develop GD. However, preoperatively confirmed autoimmune thyroid disease alone cannot be used as a predictor of the occurrence of GD after surgery, and we could not find other relevant preoperative predictive factors. In previous studies, the incidence of GD diagnosed after thyroid surgery has been reported to be 0.08–0.24%, which is consistent with that in our study. Considering that the median time of 3.3 years that GD occurrence after surgery in this study, patients who underwent thyroid surgery between 2018 and 2019 may have a short follow-up interval to disease occurrence. However, when we assessed GD incidence excluding the patients who underwent lobectomy between 2018 and 2019, the incidence (0.29%) was similar to that of the total cohort. Similarly, Kasuga et al. also emphasized the presence of autoimmune thyroiditis in patients who developed GD. They presented a case series of patients with GD and stated that the incidence of postoperative GD in patients with positive autoantibody (1.5%) was significantly higher than that (0.12%) in the autoantibody negative group, which is also consistent with our results. Our study and previous studies suggested that preoperative measurement of autoantibodies might be helpful in predicting the occurrence of GD in patients undergoing lobectomy. However, the positive conversion of TPOAb and TgAb after the development of GD was seen in several patients in this study, and the development of GD among patients with autoimmune thyroid disease was not common. Thus, prospective studies with large sample sizes are necessary to verify the clinical implications. The clinical course of GD that occurred after lobectomy in this study was similar to that in other patients with GD. In this study, 81% of the patients received ATDs as a definite long-term treatment, and 19% received ATDs as a bridging therapy until RAI or surgery. Among patients who received ATDs, approximately 52% of the patients had a remission, and others were still taking ATDs during a median follow-up period of 1.8 years. The remission rate was similar to that reported in patients with GD reported in Europe and Japan. However, the results for TAO were different from those reported in the previous studies. In general, approximately 5–6.1% of patients with GD have moderate too severe TAO. However, TAO was observed in 3 (11.5%) patients in the present study, and all of them had a severe disease state that required steroid pulse therapy. GD is a complex disease with autoimmune pathophysiology, which results from the interactions between genetic and environmental factors. The pathogenesis of the development of GD after lobectomy is unclear; however, some hypotheses exist. First, the abnormality in antigen-presenting cells that sustain the activation of suppressor and regulatory cells, which then attack the immune system and cause GD. The second hypothesis is that mechanical and biochemical stress from surgery causes neuroendocrine fluctuations, which affect immunological homeostasis. The time to diagnosis of GD after lobectomy ranged from 0.3 to 8.2 years in this study. The variance in the time until onset of the disease was also seen in previous studies, which reported a range of 9 months to 27 years. This might be related to the complex pathogenesis of GD, making it difficult to predict the occurrence of disease after lobectomy. Three patients were diagnosed with GD within 1 year after lobectomy in this study. All of them got surgery due to incidentally found thyroid nodules, and there were no clinical findings that suggest GD preoperatively. Furthermore, all three patients showed a euthyroid state in the preoperative thyroid function test. However, No.1 patient diagnosed with GD at her first follow-up visit, it is impossible to rule out the possibility of GD undetected before the lobectomy, or surgery itself might act as a stress factor to cause GD. This study has some limitations. First, it was a retrospective study from a single tertiary center which makes it difficult to generalize the findings. Second, we could not directly compare patients who developed with GD to those who did not develop GD because the incidence was too low. However, thyroid autoimmunity was the most important characteristic of patients who developed GD after lobectomy in this study. This is obviously higher than that seen in the general population with thyroid cancer, and a direct comparison between the two groups might not be necessary. The strength of this study is the large sample size to evaluate the incidence and clinical course of GD diagnosed after lobectomy for thyroid cancer. # Conclusion In conclusion, although rare, GD can occur in remnant thyroid after lobectomy. Thus, surgeons should consider the possibility of GD when hyperthyroidism is found during the follow-up of patients after lobectomy, and an appropriate workup is required. Preoperative measurements of autoantibodies might be helpful to predict the occurrence of GD; however, more evidence for the same is required. We would like to thank Editage ([www.editage.co.kr](http://www.editage.co.kr/)) for English language editing. 10.1371/journal.pone.0265332.r001 Decision Letter 0 Ahn Byeong-Cheol Academic Editor 2022 Byeong-Cheol Ahn This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 13 Dec 2021 PONE-D-21-23185 Graves’ Disease Diagnosed in Remnant Thyroid After Lobectomy for Thyroid Cancer PLOS ONE Dear Dr. Jeon, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Jan 23 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <[email protected]>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Byeong-Cheol Ahn, M.D., Ph.D. Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at  <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and  <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. We note that the grant information you provided in the ‘Funding Information’ and ‘Financial Disclosure’ sections do not match.  When you resubmit, please ensure that you provide the correct grant numbers for the awards you received for your study in the ‘Funding Information’ section. 3\. Thank you for stating the following in the Acknowledgments Section of your manuscript:  "This study was supported by a grant (2021IL0025) from the Asan Institute for Life Sciences, Asan Medical Center, Seoul, Korea." We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form.  Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows:  "This study was supported by a grant (2021IL0025) from the Asan Institute for Life Sciences, Asan Medical Center, Seoul, Korea." Please include your amended statements within your cover letter; we will change the online submission form on your behalf. 4\. In your Data Availability statement, you have not specified where the minimal data set underlying the results described in your manuscript can be found. PLOS defines a study's minimal data set as the underlying data used to reach the conclusions drawn in the manuscript and any additional data required to replicate the reported study findings in their entirety. All PLOS journals require that the minimal data set be made fully available. For more information about our data policy, please see <http://journals.plos.org/plosone/s/data- availability>. Upon re-submitting your revised manuscript, please upload your study’s minimal underlying data set as either Supporting Information files or to a stable, public repository and include the relevant URLs, DOIs, or accession numbers within your revised cover letter. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data-availability#loc-recommended- repositories>. Any potentially identifying patient information must be fully anonymized. Important: If there are ethical or legal restrictions to sharing your data publicly, please explain these restrictions in detail. Please see our guidelines for more information on what we consider unacceptable restrictions to publicly sharing data: <http://journals.plos.org/plosone/s/data-availability#loc- unacceptable-data-access-restrictions>. Note that it is not acceptable for the authors to be the sole named individuals responsible for ensuring data access. We will update your Data Availability statement to reflect the information you provide in your cover letter. 5\. PLOS requires an ORCID iD for the corresponding author in Editorial Manager on papers submitted after December 6th, 2016. Please ensure that you have an ORCID iD and that it is validated in Editorial Manager. To do this, go to ‘Update my Information’ (in the upper left-hand corner of the main menu), and click on the Fetch/Validate link next to the ORCID field. This will take you to the ORCID site and allow you to create a new iD or authenticate a pre-existing iD in Editorial Manager. Please see the following video for instructions on linking an ORCID iD to your Editorial Manager account: <https://www.youtube.com/watch?v=_xcclfuvtxQ> \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: No Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No Reviewer \#2: N/A Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: No Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Jin et al. have demonstrated the prevalence of Graves’ disease after lobectomy in the patients with thyroid cancer. They have also showed high positive rate of anti-TPO and Tg antibodies at preoperative period in the patients with Graves’ disease developed in remnant thyroid after lobectomy for thyroid cancer. As the authors have not performed any statistical analyses to characterize the cases with Graves’ disease after thyroid lobectomy, the data presentation is descriptive and the discussion is weak. To begin with, the authors should demonstrate not only the features of 26 Graves’ cases, but also those of 11017 cases without Graves’ disease in Table1. Then, the authors should simply perform Chi-squared or Fisher’s exact test to explore the factors which is different between the cases with and without Graves’ disease. It is not surprising that high prevalence of Graves’ disease is observed in the cases with positive Tg or TPO antibodies. It is clinically important to clarify whether thyroid lobectomy itself may increase the prevalence of autoimmune thyroid disorders including Graves’ disease. To approach the issue, the authors should provide another data set of age-sex-matched control cases such as benign nodules without operation to perform a case-control study. The contribution of this study to the existing knowledge is insufficient. Reviewer \#2: This study assessed the incidence and clinicopathological characteristic of newly developed Graves’ disease after lobectomy for thyroid cancer. Although it is a single-center retrospective study, the results of analysis using 10 years of data are thought to be a good reference for future prospective studies. This is a well written paper with good syntax and makes an easy reading. And results are presented in a clear manner and easy to understand. However, I would like to address a few Minor issues. 1\. This study was performed on those who underwent thyroidectomy between 2010 and 2019. As it took median 3.3 years (IQR 2.3-4.9) to diagnose GD after surgery, the study interval seems to be short for those who operated 2018-2019 to track whether GD occurred. It seems that the incidence was rather lowered by this, so please give me your opinion. 2\. 3 out of 26 patients were diagnosed with GD within 1 year after surgery. Are these cases among 8 patients who did not check preoperative TPO Ab and Tg Ab? I would like to ask for your opinion on whether the preexisting GD was undetected preoperatively. 3\. In Table 1, please add the related information about ‘None of the patients had goiter before thyroid surgery or were checked for TBII’. 4\. In Table 2, please add a detailed description of ‘Hyperthyroidism symptom’ and the thyroid scan findings. Reviewer \#3: The authors have assessed the incidence and clinicopathological characteristic of newly developed Graves’ disease after lobectomy for thyroid cancer. The manuscript is well written, and the data and the interpretation are technically sound and solid. Nevertheless, minor issues are raised to enhance the strength of the manuscript. As the authors have pointed out, this is not a comparison study between patients with or without Graves’ disease; therefore, the scientific importance is substantially limited. 1\. The legend of Figure 2 does not properly explain the corresponding data. It seems that this is not the “distribution” but “individual data per se” of 26 patients with Graves’ disease. 2\. In addition, the authors need to include each patient No. in left column of the graph in Figure 2. 3\. Please check large alphabets are truly required in the legends in Table 1. Some full names start with large letters whereas the others are not. eg. TSH, Thyroid Stimulating Hormone; TPOAb, Anti-Thyroid Peroxidase Antibody; TgAb, Antithyroglobulin Antibody; PTC, papillary thyroid cancer 4\. Title of Table 2. Characteristics of Patients who Developed Grave's Disease after Thyroid Lobectomy =\> Grave’ disease is a typo of Graves’ disease 5\. Table 2. Before LT4 treatment (yes) -\> This requires English editing. 6\. In general, tables not easy to follow. Please consider revising the format to enhance the quality. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No Reviewer \#3: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0265332.r002 Author response to Decision Letter 0 21 Dec 2021 Response to Reviewer 1 We appreciate your review of our manuscript. Reviewer \#1: Jin et al. have demonstrated the prevalence of Graves’ disease after lobectomy in the patients with thyroid cancer. They have also showed high positive rate of anti-TPO and Tg antibodies at preoperative period in the patients with Graves’ disease developed in remnant thyroid after lobectomy for thyroid cancer. As the authors have not performed any statistical analyses to characterize the cases with Graves’ disease after thyroid lobectomy, the data presentation is descriptive and the discussion is weak. To begin with, the authors should demonstrate not only the features of 26 Graves’ cases, but also those of 11017 cases without Graves’ disease in Table1. Then, the authors should simply perform Chi-squared or Fisher’s exact test to explore the factors which is different between the cases with and without Graves’ disease. It is not surprising that high prevalence of Graves’ disease is observed in the cases with positive Tg or TPO antibodies. It is clinically important to clarify whether thyroid lobectomy itself may increase the prevalence of autoimmune thyroid disorders including Graves’ disease. To approach the issue, the authors should provide another data set of age-sex-matched control cases such as benign nodules without operation to perform a case-control study. The contribution of this study to the existing knowledge is insufficient. → Thank you for your comment, and we agree that our inability to directly compare patients who developed with GD to those who did not is the major limitation of this study. We have added this point as a limitation in the discussion part (Page 12 Line 215-217). However, as the primary purpose of the current study was to assess the incidence of GD after lobectomy and their clinical course, we think this study so far has been sufficient to achieve our purpose. In addition, regarding your comment, “if the thyroid lobectomy itself increases the prevalence of autoimmune thyroid disorder?” we look forward to implementing it in our future studies. Response to Reviewer 2 We appreciate your review of our manuscript. We believe that your comments have helped us to improve our manuscript. In the revised manuscript, changes are shown in bold red text. Reviewer \#2: This study assessed the incidence and clinicopathological characteristic of newly developed Graves’ disease after lobectomy for thyroid cancer. Although it is a single-center retrospective study, the results of analysis using 10 years of data are thought to be a good reference for future prospective studies. This is a well written paper with good syntax and makes an easy reading. And results are presented in a clear manner and easy to understand. However, I would like to address a few Minor issues. 1\. This study was performed on those who underwent thyroidectomy between 2010 and 2019. As it took median 3.3 years (IQR 2.3-4.9) to diagnose GD after surgery, the study interval seems to be short for those who operated 2018-2019 to track whether GD occurred. It seems that the incidence was rather lowered by this, so please give me your opinion. → Thank you for important comment. We agree with your opinion and assessed the incidence after excluding the patients who underwent lobectomy between 2018 and 2019. The incidence was similar to that of the total cohort. We have added this point in the discussion part (Page 10, Line 171–176). 2\. 3 out of 26 patients were diagnosed with GD within 1 year after surgery. Are these cases among 8 patients who did not check preoperative TPO Ab and Tg Ab? I would like to ask for your opinion on whether the preexisting GD was undetected preoperatively. → Among the 3 patients, 2 patients did not check preoperative autoantibody, and one patient had positive autoantibody. All of them got surgery due to incidentally found thyroid nodules, and there were no clinical findings that suggest GD preoperatively. Furthermore, all three patients showed a euthyroid state in the preoperative thyroid function test. However, No.1 patient (Fig2) diagnosed with GD at her first follow-up visit, it is impossible to rule out the possibility of GD undetected before the lobectomy, or surgery itself might act as a stress factor to cause GD. We have added this point in discussion (Page 12, Line 207-213). 3\. In Table 1, please add the related information about ‘None of the patients had goiter before thyroid surgery or were checked for TBII’. → Thank you for your comment. We have added the description of goiter and TBII in Table 1. In addition, we have described the definition of goiter in method part (Page6, Line 83–85) 4\. In Table 2, please add a detailed description of ‘Hyperthyroidism symptom’ and the thyroid scan findings. → Thank you for your comment. We have added the description of hyperthyroidism symptoms in footnote of Table 2. As most of the patients complained of multiple overlapping thyrotoxicosis symptoms, it is difficult to subclassifying them in Table 2. Response to Reviewer 3 We appreciate your review of our manuscript. We believe that your comments have helped us to improve our manuscript. In the revised manuscript, changes are shown in bold red text. Reviewer \#3: The authors have assessed the incidence and clinicopathological characteristic of newly developed Graves’ disease after lobectomy for thyroid cancer. The manuscript is well written, and the data and the interpretation are technically sound and solid. Nevertheless, minor issues are raised to enhance the strength of the manuscript. As the authors have pointed out, this is not a comparison study between patients with or without Graves’ disease; therefore, the scientific importance is substantially limited. 1\. The legend of Figure 2 does not properly explain the corresponding data. It seems that this is not the “distribution” but “individual data per se” of 26 patients with Graves’ disease. → Thank you for your comment. We agree with your opinion and have revised the legend of Fig 2 (Page 9, Line 147–148) 2\. In addition, the authors need to include each patient No. in left column of the graph in Figure 2. → We have revised the figure 2 as your suggestion. 3\. Please check large alphabets are truly required in the legends in Table 1. Some full names start with large letters whereas the others are not. eg. TSH, Thyroid Stimulating Hormone; TPOAb, Anti-Thyroid Peroxidase Antibody; TgAb, Antithyroglobulin Antibody; PTC, papillary thyroid cancer → Legends of tables are revised as your suggestion. Sorry for your inconvenience. 4\. Title of Table 2. Characteristics of Patients who Developed Grave's Disease after Thyroid Lobectomy =\> Grave’ disease is a typo of Graves’ disease → We have corrected the typo in Table 2. Sorry for your inconvenience. 5\. Table 2. Before LT4 treatment (yes) -\> This requires English editing. → Thank you for your comment. We have revised it as “Patients who have previously taken levothyroxine” in Table 2. 6\. In general, tables not easy to follow. Please consider revising the format to enhance the quality. → We have revised the table1 format in accordance with your suggestion (Table 1 and 2). 10.1371/journal.pone.0265332.r003 Decision Letter 1 Ahn Byeong-Cheol Academic Editor 2022 Byeong-Cheol Ahn This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 1 Mar 2022 Graves’ Disease Diagnosed in Remnant Thyroid After Lobectomy for Thyroid Cancer PONE-D-21-23185R1 Dear Dr. Jeon, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Byeong-Cheol Ahn, M.D., Ph.D. Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: (No Response) Reviewer \#2: All comments have been addressed Reviewer \#3: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: No Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No Reviewer \#2: N/A Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: No Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: No Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#2: This study assessed the incidence and clinicopathological characteristic of newly developed Graves’ disease after lobectomy for thyroid cancer. The authors have adequately addressed the comments raised in a previous round of review. Reviewer \#3: The authors have faithfully addressed this reviewer's concerns. Therefore I do not have further comments. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No Reviewer \#3: No 10.1371/journal.pone.0265332.r004 Acceptance letter Ahn Byeong-Cheol Academic Editor 2022 Byeong-Cheol Ahn This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 3 Mar 2022 PONE-D-21-23185R1 Graves’ disease diagnosed in remnant thyroid after lobectomy for thyroid cancer Dear Dr. Jeon: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Professor Byeong-Cheol Ahn Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Dendritic cells (DCs) are potent antigen presenting cells capable of activating naïve T cells. DCs are present in tissues in an immature state and display low levels of maturation or co-stimulatory markers such as CD83, CD80 or CD86. Immature DCs (iDCs) recognise and capture specific antigens, including tumour antigens. DCs undergo a functional maturation process in response to inflammatory mediators such as IFN-α or Toll like receptor (TLR) agonists. As DCs mature they gain the potential of presenting antigen to T cells and activating a specific anti-tumour T cell response. DCs that secrete high levels of bioactive IL-12p70 induce optimal anti-tumour immunity, as they have increased capacity to enhance natural killer cell activity, skew the response to Th1 and prime tumour specific CD8⁺ T cells. However, many tumours evade the immune response by secreting cytokines and other factors that inhibit DC differentiation or the maturation of tumour infiltrating DCs.. One of these ‘pro-tumour’ factors, Vascular Endothelial Growth Factor (VEGF) is known for sustaining tumour growth via its angiogenic properties but can also elicit an inhibitory effect on DC differentiation and maturation, enhancing tumour survival,. VEGF has successfully been targeted by the humanised monoclonal antibody Bevacizumab (Avastin), however response rates are approximately 40% and many patients develop resistance to this treatment. Therefore, it is crucial to explore the potential of other inflammatory mediators present in the tumour microenvironment that may inhibit DC maturation, as these may also be potential therapeutic targets. Several cytokines and chemokines are present at high levels in the tumour microenvironment, compared to normal tissues, such as CCL2 (MCP-1), CXCL1 (GROα) and CXCL5 (ENA-78). CCL2 is known to attract monocytes, T-cells and dendritic cells, while the main function of CXCL1 and CXCL5 is to attract and activate neutrophils. In addition to their chemoattractant functions, CCL2, CXCL1 and CXCL5 also play an important role in angiogenesis, demonstrating the multifunctional nature of these chemokines. It is known that human myeloid DCs express CCR2 and CXCR2, the receptors for CCL2 and CXCL1 and CXCL5, respectively . However the effect of these chemokines on DC maturation and function has not previously been investigated. In this study, we used explanted human colorectal cancer tissue to model the tumour microenvironment. Explant tissues maintain the complex 3D structure of the tumour, including the stroma, thus allowing the production of many different tumour associated factors, closely mimicking the inflammatory milieu of the tumour *in situ*. Previous studies have shown that supernatants from tumour cell lines can inhibit DC maturation, however the importance of factors secreted by the entire tumour on dendritic cell maturation has not been previously examined. We demonstrate that Tumour Conditioned Media (TCM) from cultured colorectal cancer tumours significantly inhibited LPS induced DC maturation, markedly increasing IL-10 while decreasing IL-12p70 secretion in response to LPS. We found that the TCM contained significant amounts of the chemokines CCL2, CXCL1 and CXCL5 in addition to VEGF. Levels of CCL2, CXCL1 and CXCL5 in the TCM correlated with CD83 expression and IL-12p70 secretion from DCs treated with TCM. Individually, all of these inflammatory mediators significantly inhibited LPS induced IL-12p70 secretion, however IL-10 secretion remained unaffected. Interestingly, the effects of CXCL1 and VEGF on the inhibition of LPS induced IL-12p70 secretion by DCs were additive. In addition, CXCL1 also had a marked inhibitory effect on LPS induced up-regulation of HLA-DR. # Results ## Tumour conditioned media inhibits DC maturation and IL-12p70 secretion while augmenting IL-10 secretion in response to LPS Monocyte derived DCs obtained from 2 healthy volunteers were incubated with different volumes of TCM from 4 colorectal cancer patients for 4 hours prior to LPS simulation for a further 18 hours, to determine the optimal concentration of TCM. We examined the effect of TCM on the ability of DCs to secrete IL-10 and IL-12p70 in response to stimulation by LPS. Secretion of IL-12p70 by DCs augments and directs the expansion and differentiation of tumour specific Th1 responses while high levels of IL-10 secretion preferentially leads to the expansion of immunosuppressive T regulatory cells. Treatment of DCs with a 1 in 2 dilution of TCM showed to have the most potent effect on IL-10 increase and IL-12p70 inhibition (p = 0.028) in response to LPS. No significant difference was observed in expression of CD80, CD86, CD54, CD83 and HLA-DR (MHC II), markers associated with DC maturation, between different dilutions of TCM (data not shown). Therefore we used a 1 in 2 dilution of TCM for further analysis. Monocyte derived DCs obtained from 6 healthy donors were treated with 1 in 2 TCM of 21 colorectal patients for 4 hours before treatment with LPS for a further 18 hours, and interestingly, TCM treated DCs secreted significantly higher levels of IL-10 in response to LPS (p\<0.0001), but significantly lower levels of IL-12p70 (p\<0.0001). In addition IL-12p70 showed a negative correlation with CD83 expression Spearman r = −0.71, p\<0.001 (data not shown), further suggesting that TCM treated DCs with increasing CD83 expression have decreasing levels of IL-12p70 secretion. Treatment of DCs with TCM also significantly inhibited LPS-induced DC maturation, with reduced expression of CD54 (*p*\<0.0001), CD86 (*p* = 0.0035), HLA-DR (*p* = 0.0474) and CD83 (*p* = 0.0018) observed. Treatment of DCs with conditioned media taken from normal tissue adjacent to the tumour tissue (NCM) did not inhibit LPS induced maturation or cytokine secretion from DCs. ## Levels of CCL2, CXCL1, and CXCL5 present in TCM correlate with CD83 expression and IL-12p70 secretion from DCs TCM obtained from 21 patients with colorectal cancer were analysed by ELISA for several known tumour associated factors. Levels of CCL2, CXCL1 and CXCL5 were consistently higher than those of VEGF (a current therapeutic target in colorectal cancer) in all patients examined. We found that levels of CCL2 (Spearman r = 0.59, *p* = 0.0053) CXCL1 (Spearman r = 0.50, *p* = 0.0202) and CXCL5 (Spearman r = 0.50, *p* = 0.0199) present in the TCM positively correlated with CD83 expression on DCs treated with TCM. This implies that with increasing levels of CCL2, CXCL1 and CXCL5 present in the TCM, CD83 expression is elevated on DCs treated with TCM. Even though VEGF is present at detectable levels in the TCM, it showed no correlation with DC maturation. In addition we found that IL-12p70 secretion from DCs treated with TCM negatively correlated with CCL2 (Spearman r = −0.64, *p* = 0.0019), CXCL1 (Spearman r = −0.59, *p* = 0.0049) and CXCL5 (Spearman r = −0.57, *p* = 0.0075). This suggests that with increasing levels of CCL2, CCL1 and CXCL5 in the TCM, IL-12p70 secretion from DCs decreases. There was no significant correlation between the four cytokines and the other DC markers, CD80, CD54, CD86, and HLA-DR, and IL-10 secretion. ## Effect of CCL2, CXCL1, CXCL5 and VEGF on DC maturation marker expression and cytokine secretion in response to LPS Having established that TCM inhibits DC maturation in response to LPS and that TCM contains high levels of the chemokines CCL2, CXCL1 and CXCL5 we next determined if these individual components had a direct effect on DCs. Recombinant VEGF significantly inhibited LPS-induced expression of DC maturation markers CD80 (*p* = 0.0391) and CD54 (*p* = 0.0078) and CXCL1 significantly inhibited the upregulation of HLA-DR (*p* = 0.0156) (Data not shown). Although CCL2 and CXCL5 showed a reduction of LPS-induced maturation of DCs, this did not reach statistical significance (Data not shown). While the pre-treatment of DCs with CCL2, CXCL1, CXCL5 or VEGF failed to augment LPS-induced IL-10 secretion, a significant decrease in the levels of LPS-induced IL-12p70 secretion was observed: CCL2 (*p* = 0.048), CXCL1 (*p* = 0.0068), CXCL5 (*p* = 0.0255), VEGF (*p* = 0.0402). IL-1β, TNF-α, IL-8 and IL-6 production by DCs in response to LPS in the absence or presence of pre-treatment with VEGF, CCL2, CXCL1 or CXCL5 were also measured; however there were no significant differences in the levels of these cytokines determined under these conditions (data not shown). Treatment of DCs with CCL2, CXCL1, CXCL5 and VEGF did not affect their ability to induce T cell proliferation or production of IFNγ (data not shown). In addition we investigated whether CCL2, CXCL1, CXCL5 and VEGF could affect DC migration. CCL19 was used as a positive control since it is known to bind to the chemokines receptor CCR7, which is expressed by DCs. We found that they did not increase migration compared to CCL19 (data not shown). ## CXCL1 and VEGF have an additive inhibitory effect on IL-12p70 secretion by DCs in response to LPS The TCM has a potent inhibitory effect on DC maturation and function, but CCL2 and CXCL5, which are present at high levels in TCM, do not have a significant effect on dendritic cell maturation marker expression, while CXCL1 only significantly inhibited HLA-DR and VEGF significantly inhibited CD80 and CD54 (data not shown). Therefore, we investigated whether combining the recombinant cytokines might have a more potent effect on dendritic cell maturation. We examined seven different combinations of the four cytokines: CCL2+CXCL1, CCL2+CXCL5, CCL2+VEGF, CXCL1+CXCL5, CXCL1+VEGF, CXCL5+VEGF, CCL2+CXCL1+CXCL5+VEGF. The combination of CXCL1 and VEGF did not significantly inhibit LPS-induced maturation marker expression by DCs (CD80 *p* = 0.0625, CD54 *p* = 0.6, CD86 *p* = 0.6, HLA-DR *p* = 0.1, CD83 *p* = 0.8). However combining CXCL1 and VEGF significantly reduced LPS-induced IL-12p70 (*p* = 0.0028), but did not alter the levels of IL-10. IL-1β, TNFα, IL-8 or IL-6 secretions were not altered (data not shown). No other combination had a direct effect on LPS- induced DC maturation or cytokine secretion of DCs in response to LPS when compared to the individual cytokines (data not shown). In addition inclusion of neutralising antibodies to CXCL1 and VEGF, alone and in combination, did not correct TCM induced alteration of DC maturation, IL-10 or IL-12p70 secretion from DCs in response to LPS (data not shown). # Discussion In this study we have shown for the first time that tumour conditioned media from colorectal cancer tumour explants can significantly inhibit LPS-induced maturation of monocyte derived DCs. In addition, treatment of DCs with TCM prior to LPS stimulation significantly increased production of IL-10, an anti- inflammatory cytokine that inhibits a Th1 response, while decreasing the secretion of IL-12p70, a pro-inflammatory cytokine required for Th1 responses: a DC phenotype associated with tolerance. Subsequently, we found that CCL2, CXCL1 and CXCL5 were expressed at high levels in the TCM compared to VEGF, and we determined if these inflammatory mediators could alter the function of DCs. VEGF was included in our analyses as it had previously been shown to inhibit DC maturation. The addition of recombinant VEGF to iDCs had an inhibitory effect on the LPS-induced upregulation of CD80 and CD54, while adding recombinant CXCL1 to iDCs decreased HLA-DR expression following LPS treatment. However, treatment of iDCs with human recombinant CCL2, CXCL1, CXCL5 or VEGF, at doses in line with other studies, resulted in significantly reduced levels of IL-12p70 secretion in response to LPS stimulation. These data suggest that inflammatory mediators present in the tumour microenvironment reduce the capacity of local DCs to secrete IL-12p70 and thus the induction of an effective anti-tumour response. Similar findings were reported in a study investigating the effect of conditioned media taken from the pancreatic cancer cell line BxPC-3 (BxCM), where DC differentiation and maturation was inhibited by BxCM. The study showed that BxCM reduced the expression of CD83, CD1a, CD1c, CD80 and CD86 by DCs in response to TNFα. Consistent with our findings, this study found a significant increase in IL-10 and a reduction in IL-12p70 production in response to stimulation with TNFα following BxCM treatment. However, the treatment of DCs with supernatant of RCC-10 (renal cell carcinoma cells), resulted in a reduction in the expression of maturation markers, but did not affect IL-10 production in response to stimulation with TNFα, IFNα and poly I:C. These studies, together with the data presented here, suggest that many tumour types secrete inflammatory mediators that can inhibit the functionality of DCs; but that there is likely to be variation in the inflammatory milieu and the resultant effect on DC function according to tumour type and stage. While it would have been interesting to assess and isolate the actual levels of DC infiltration in the tumour explant tissue, this could not be performed due to the limitation of explant tissue size we received from surgery. Our results indicate that NCM does not significantly affect DC maturation or cytokine secretion, indicating that the normal adjacent tissue may have functional immunity. There was no evidence of an immunosuppressive field effect. VEGF is established as an important factor contributing to tumour growth by inducing angiogenesis. VEGF is a therapeutic target in several cancers, including colorectal cancer, where the humanised anti-VEGF mAb, Bevacizumab (Avastin) is employed. Previous studies have demonstrated that VEGF inhibits DC maturation and DCs ability to activate T cells,. In this study, we demonstrate that VEGF inhibits the expression of CD80 and CD54, but not CD86 or HLA-DR, data consistent with Alfaro *et al.*. Interestingly, we found that VEGF treated DCs secreted significantly reduced levels of IL-12p70 in response to LPS; a finding that has not been previously documented. The use of LPS to mature the DCs, versus the maturation cocktail consisting of TNFα, IFNα and poly I:C used by Alfaro *et al.* might explain the differences observed in the effect of VEGF on DC maturation and IL-12p70 secretion. Our finding that VEGF stimulation of DCs does not affect T cell proliferation and cytokine secretion is consistent with Alfaro *et al.*, however they found an inhibitory effect of VEGF when present at the differentiation stage of DC development, indicating that differentiated DCs might be less responsive to the effects of VEGF. Our study also showed that our colorectal cancer patients explant tissue secreted high levels of CCL2, CXCL1 and CXCL5, compared to VEGF. Interestingly, CCL2, CXCL1 and CXCL5 levels in TCM correlated with CD83 expression on TCM treated DCs. This is the first time this correlation has been observed, and suggests that patients with higher levels of CCL2, CXCL1 or CXCL5 might have increased levels of CD83 on DCs. CD83 is very important for naive T-cell and B cell activation, however we also found that enhanced expression of CD83 on DCs treated with TCM correlated with reduced IL-12p70 secretion from TCM treated DCs. It has been previously reported that while different stimuli might mature DCs to similar levels, with similar expression of CD83, CD80 and CD86, the profile of cytokines that they secrete differ; stimuli such as IFN-γ can result in IL-12 secreting CD83+ DCs that stimulate Th1 responses while other stimuli such as prostaglandin E2 result in CD83+ DCs that secrete no IL-12 but efficiently promote a Th2 response. While it has been reported that CCL2, CXCL1 and CXCL5 are important for tumour growth, proliferation, and angiogenesis, the effect of these chemokines on DCs has not previously been documented. CCL2, CXCL1 and CXCL5 levels in TCM correlated inversely with IL-12p70 secretion, this is the first time this correlation has been observed, and suggests that with increasing levels of CCL2, CXCL1 and CXCL5 in the TCM, IL-12p70 secretion from DCs is reduced. CCL2 is a chemoattractant for monocytes, memory T-cells and dendritic cells. CCL2 has also been previously reported to induce angiogenesis, which is important for tumour growth and metastasis. Both pro- and anti-tumour effects of CCL2 have been observed through its effect on monocytes and macrophages, while here we show that CCL2 may also influence the anti-tumour immune response by affecting the ability of DCs to secrete IL-12p70 but not other cytokines including IL-10, IL-1β, IL-6, IL-8 or TNF-α in response to LPS stimulation. In contrast, Braun *et al.* have previously shown that CCL2 can inhibit IL-12p70 production from monocytes and this is pertussis toxin sensitive, but they did not observe this reduction in IL-12p70 secretion from dendritic cells treated with either SAC+IFNγ or CD40L+IFNγ. Alternatively, Omata *et al.* found DCs differentiated from monocytes in the presence of CCL2 had a reduced capacity to secrete IL-12p70 following stimulation with CD40 ligand, an effect not sensitive to pertussis toxin. These authors reported no effect of CCL2 on the differentiation or maturation of DCs.. Therefore, the precise mechanism by which CCL2 exerts its effect on DCs remains to be fully elucidated. While the effect of CXCL1 and CXCL5 on DC maturation and cytokine secretion has not been previously investigated, their known function is to attract and activate neutrophils, which in turn have been implicated with tumour cell growth, angiogenesis and metastasis. Since DCs express CXCR2 which is the receptor for CXCL1 and CXCL5, it is not surprising that CXCL1 and CXCL5 could have an effect on DC function. Chemokines such as CCL2, CXCL1 and CXCL5 signal through G-protein coupled receptors, which control intracellular cAMP levels and increased cAMP levels correlates with decreased IL-12 production. However, the mechanism by which these chemokines reduce IL-12p70 secretion from DCs is not known, and it will be subject of our future investigations. Interestingly CXCL1 and VEGF have an additive inhibitory effect on IL-12p70 secretion. They act on different receptors which have previously been shown to be present on DCs, therefore they might activate different signalling pathways simultaneously, or they could activate the same pathway more effectively. VEGF has been shown to enhance phospho-ERK1 & ERK2 in DCs and this pathway negatively regulates monocyte derived DC maturation. Which signalling pathways are activated by CXCL1 through CXCR2 in dendritic cells is yet unknown, however studies in different cell types, such as cancer cells, show that CXCR2 signalling can activate NF-κB, STAT3 and the ERK pathways. Hence, there are a variety of potential mechanisms by which CXCR2 ligands may inhibit the secretion of IL-12p70 by DCs. However, immune depleting either of these cytokines singly or in combination with neutralising antibodies did not reverse the inhibitory effect of the TCM. This proves that it is probably multiple different factors acting together that cause DC inhibition. This is consistent with previous research showing that blocking of a single factor, VEGF in supernatants of RCC-10 cells did not have a significant effect on DC activation. We observed that CCL2, CXCL1, CXCL5 and VEGF do not block DC maturation, nor do they affect DC migration or T-cell proliferation. Unlike DCs treated with TCM, the individual inflammatory mediators do not enhance IL-10 however they all inhibited IL-12p70 secretion; an effect not previously shown for CCL2, CXCL1, CXCL5 and VEGF. The explant tumour tissues secrete many different soluble factors, and it is not very surprising that a few isolated factors on their own do not have an effect comparable to TCM. It is very likely that the many different factors secreted into the TCM by the explant tumour act together in inhibiting DC maturation and function. However the ability of CCL2, CXCL1, CXCL5 and VEGF to significantly reduce IL-12p70 production by DCs is important, as this may potentially result in a reduced anti-tumour Th1 response *in vivo*. In conclusion we found that TCM inhibits DC maturation, and induces IL-10 while inhibiting IL-12p70 secretions from DCs. The TCM components CCL2, CXCL1, CXCL5 and VEGF seem to play a role in modulating the inflammatory response through inhibition of IL-12p70 secretion by DCs, possibly to protect the tumour from a potent immunologic response against it. In addition, CXCL1 and VEGF act together in the inhibition of IL-12p70 secretion from DCs. Even though CCL2, CXCL1, CXCL5 and VEGF are present in the TCM at high levels, they are not the main components of the TCM that affect DC maturation and function. In conclusion, we have demonstrated the importance of tumour conditioned media in regulating dendritic cell function in colorectal cancer patients, however the underlying mechanisms still need to be elucidated. # Materials and Methods ## Ex vivo tumour explant culture All tissue was obtained with the informed written consent of the patient, and the protocol was approved by the Ethics Committee of St. Vincent's University Hospital. Surgically resected colorectal cancer tumour tissue (n = 21) and normal tissue (\>10cm from the tumour; n = 5) was obtained from patients from the Centre for Colorectal Disease's explant tissue bio-bank at St. Vincent's University Hospital, Dublin (12 male, 9 female, median age 67, 1 Stage II, 5 Stage III, 15 Stage IV colorectal cancer). The explanted tissue was cut into at least 4 equal- sized pieces of approximately 5 mm³ and cultured as previously described. Briefly, the explanted tumour tissues were cultured for 72 h (in 24 well plates) in 2 mL RPMI 1640 containing 100 U/mL Penicillin, 100 µg/mL Streptomycin, 4 µg/mL Fungizone, 30 µg/mL gentamicin (Invitrogen; Carlsbad, California) and supplemented with 20% foetal bovine serum (Invitrogen). Following 72 hours in culture, Tumour Conditioned Media (TCM) and normal conditioned media (NCM) was collected and stored at −20°C until used for analyses. ## Dendritic cell isolation and culture Human monocyte-derived immature dendritic cells (iDCs) were generated from peripheral blood mononuclear cells (PBMCs) obtained from buffy coat preparations (National Blood Centre, St. James Hospital, Dublin). Monocytes were isolated by positive selection using CD14-magnetic beads (Miltenyi Biotec, Bergisch Gladbach, Germany) and seeded at a density of 1×10⁶ cells/mL in 6-well plates in 3 mL of RPMI 1640 containing 100 U/mL Penicillin, 100 µg/mL Streptomycin, 4 µg/mL Fungizone and supplemented with 10% defined Hyclone foetal bovine serum (Thermo Fisher Scientific, Waltham, MA, US), human granulocyte- macrophage colony-stimulating factor (GM-CSF; 50 ng/mL, Immunotools, Friesoythe, Germany), and human IL-4 (70 ng/mL; Immunotools). Cells were fed at day 3 by replacing half the medium and adding fresh cytokines. At day 6 the cells exhibited an immature DC phenotype (CD14^-, CD11c⁺, CD86^-, CD54^low, CD83^-, CD80^-, and HLA-DR^low). ## Stimulation of monocyte derived DC Freshly isolated iDC were plated in triplicate in 96 well plates at 1×10⁵ cells/200 µl in RPMI 1640 media supplemented with 10% defined Hyclone FBS (Thermo Fisher Scientific), and stimulated with 1 in 2, 1 in 4 and 1 in 10 dilution of Tumour Conditioned Media (TCM) from 4 cultured explant tissues for 4 hours before adding 1 µg/mL *Escherichia coli* lipopolysaccharide (LPS; Alexis Biochemicals, Lausen, Switzerland) to determine the best dose of TCM. Subsequently DCs were treated with a 1 in 2 dilution TCM of all 21 patient explant tissues for 4 hours before adding 1 µg/mL LPS. In addition, DCs were treated with a 1 in 2 dilution of NCM of 5 patient explant tissues for 4 hours before 1 µg/ml LPS was added. In separate experiments 1×10⁵ iDC/200 µl media were treated with human recombinant 50 ng/mL CCL2 (MCP-1), 100 ng/mL CXCL1 (GROα), 100 ng/mL CXCL5 (ENA-78) and 20 ng/mL VEGF (R&D Systems, Abingdon, UK) for 4 hours before 1 µg/mL LPS was added to the samples. Cultures were incubated for a further 18 h at 5% CO₂ and 37°C. In addition, iDCs were cultured with a combination of cytokines (at the concentrations described above): CCL2+CXCL1, CCL2+CXCL5, CCL2+VEGF, CXCL1+CXCL5, CXCL1+VEGF, CXCL5+VEGF, CCL2+CXCL1+CXCL5+VEGF for 4 hours before LPS was added for a further 18 hours. Supernatants were harvested and levels of IL-10 and IL-12p70 secretion analysed and cells were assessed for expression of maturation markers by flow cytometry as described below. ## Neutralising CXCL1 and VEGF in TCM iDC were treated with TCM of 4 colorectal cancer patients, as described above, with the addition of neutralising antibodies to VEGF (Avastin,100 µg/mL), CXCL1 (15 µg/ mL, R&D systems) and corresponding isotype control (15 µg/mL, R&D Systems) for 4 hours before LPS was added for 18 hours. Cells were assessed for expression of maturation markers as described below, and levels of IL-10 and IL-12p70 in the supernatant were analysed using ELISA. ## Flow cytometry Dendritic cells were stained with the following monoclonal antibodies (mAb): fluorescein isothiocyanate (FITC)-conjugated anti-CXCR2 (R&D systems) anti-CD14 and anti-CD80; phycoerythrin (PE)–conjugated anti-CD54; Phycoerythrin-Cy5 (PeCy5)–conjugated anti-CD86 and anti-CD83; allophycocyanin (APC)–conjugated anti-CD11c and anti-HLA-DR (all from BD Biosciences, Oxford, UK). Cells were also stained with corresponding isotype control mAbs (BD Biosciences). Cells were acquired on FACScalibur flow cytometer and the data were analysed with CellQuest Pro (BD Biosciences), or Flowjo software (Tree Star Inc., Ashland, OR). ## Quantification of cytokines by ELISA Levels of CCL2, CXCL1, CXCL5, and VEGF in TCM; IL-10 and IL-12p70 in DC supernatant; IFNγ in T cell supernatant were quantified by sandwich ELISA according to the manufacturer's protocol (R&D Systems). IL-1β, IL-6, IL-8 and TNF-α in DC supernatants were measured by MSD multiplex assays as per manufacturer's instructions (Meso Scale Discovery, Sector Imager 2400, Gaithersburg, Maryland 20877). ## Dendritic cell migration assay Monocyte derived dendritic cells were seeded at 5×10⁵ cells/ml in the upper chambers of 96 well plate (Corning B.V. Life Sciences, Amsterdam, The Netherlands). CCL2, CXCL1, CXCL5, VEGF and CCL19 (Immunotools) were added to the lower chambers at 50 ng/ml and 150 ng/ml (n = 3). Plates were incubated for 4 hours at 37°C. Cells were counted using a Z1 Coulter Particle counter (Beckman Coulter Diagnostics Limited, Lismeehan, O'Callahans Mills, Ireland). ## T cell proliferation assay CD3⁺ T cells were isolated from PBMCs of healthy donors using CD3-labelled magnetic beads (Miltenyi Biotec). The T cells were labelled with CFSE (Molecular Probes, Eugene, OR). Briefly, cells were incubated with 0.5 µM CFSE in PBS for 2.5 min. Heat inactivated FBS (Invitrogen) was added to stop the reaction and cells were washed prior to resuspension in complete RPMI. T cells were incubated with DCs in round-bottomed 96-well plates for 5 days in a 10∶1 ratio (T cells:DCs). Cells were harvested for flow cytometric analysis and levels of IFNγ in the supernatant were determined by ELISA. ## Intracellular IFNγ PBMC's form healthy donors were incubated with DCs in a 10∶1 ratio (PBMC: DCs) for 5 days. Cells were harvested and treated for intracellular staining of IFNγ as follows: Brefeldin A (10 µg/ml, Sigma-Aldrich, Dublin, Ireland) phorbol myristate acetate (PMA) (25 ng/ml, Sigma-Aldrich) and Ionomycin (1 µg/ml, Sigma- Aldrich) was added to the cells for 4 hours to stimulate cytokine production and prevent secretion of these cytokines. The cells were stained with CD3-APC (BD biosciences) before fixing the cells with 4% paraformaldehyde (Sigma-Aldrich) for 10 minutes and adding 0.2% w/v saponin (Sigma-Aldrich) for another 10 minutes to permeabilise the cell membrane. 5 µl of PE conjugated anti-IFNγ (BD Biosciences) was added and cells were analysed by flow cytometry. ## Statistical Analysis Statistical analyses were performed using GraphPad Prism version 5.00 for Windows (GraphPad software, La Jolla, CA). The Wilcoxon signed rank test, Mann- Whitney U test or ANOVA were used to compare groups as appropriate. For correlations the Spearman rank test was used. A *p-*value of \<0.05 was considered to be significant. We thank Michelle Corrigan and Aoibhlinn O'Toole for technical assistance. [^1]: Conceived and designed the experiments: JO ER AH JH DO AM. Performed the experiments: AM AH JM MT FC. Analyzed the data: AM ER JO AH HM. Wrote the paper: AM JO ER. Assessment of pathology of tissue: KS. [^2]: The authors have declared that no competing interests exist.

# Introduction Depression is the third leading cause of illness and disability among adolescents and suicide is the third leading cause of death in adolescents aged 15–19 years. Although adolescents living with HIV are commonly exposed to multiple risk factors associated with depression in adolescents, there has been a significant lack of attention to the prevalence, manifestation, impact and management of depression in adolescents living with HIV. Several studies in Zimbabwe and elsewhere have found that adolescents living with HIV are at risk of depression, which in turn correlates with poor adherence to antiretroviral therapy (ART). Recent research in Zimbabwe has found high rates of virological failure (viral load ≥1000 copies/ml) among adolescents living with HIV. Despite the global success of antiretroviral treatment programmes and an overall decrease in AIDS-related deaths, mortality rates continue to increase in the adolescent age group. International guidelines now call for the integration of mental health within HIV service delivery. Evidence-based, adolescent-focused interventions which prevent and manage depression are urgently needed. If HIV services are to effectively meet adolescents’ needs, it is necessary to understand the experience and manifestation of depression in adolescents. Here we report on a study that explored the experience and manifestation of depression in adolescents living with HIV in Zimbabwe in order to inform intervention development. # Methods ## Participants Between January and June 2015, in-depth interviews were conducted with 21 HIV positive adolescents aged 15–19 years and diagnosed with major depressive disorder (MDD). Participants were recruited through purposive sampling from the Zvandiri ('As I am') programme, a model of differentiated clinical service delivery for HIV positive children and adolescents in Zimbabwe ([www.africaid- zvandiri.org](http://www.africaid-zvandiri.org/)). In the Zvandiri programme, adolescents are routinely screened for common mental disorders and those at risk are referred to a registered clinical psychologist or psychiatrist for further assessment and management. Adolescents who received the diagnosis of major depressive disorder by the psychologist or psychiatrist using Diagnostic Statistical Manual of Mental Disorders, 4^th edition (DSM-IV) criteria were provided with information about the study and invited to participate. All participants were seen within two weeks of their diagnosis of depression. None had been initiated on anti-depressant medication at the time of the interview. Participants were recruited until thematic saturation. ## Procedure The interviews took place at an adolescent treatment centre in Harare and were structured around a body mapping process, a creative arts technique. Participants were asked to create a painted map of their body to assist them in externalizing their somatic and emotional experiences. An interview and body mapping guide was used to facilitate the interview (NW, AM and a research assistant), which aimed to engage the participant in a creative dialogue around the research questions. It contained open-ended questions to explore their subjective experiences of depression and perceptions of care received. At the beginning of the interview, an outline of the participant was drawn on a large sheet of paper. The interview was then conducted and with each question, the participant was invited to add words, colours or pictures to their body map, to assist them in responding to the questions. The questions focused first on socio-demographics (name, age, gender, HIV and ART history), then asked the participant to think about the word "depression", its meaning and how depression has affected them. Depression was not a new word to them as they had all been informed of their diagnosis of depression when assessed by the psychologist or psychiatrist prior to enrolment in the study. Participants were then asked to focus on what may have contributed to the depression, their thoughts about the future and the type of support they desired and needed. Probes were used to engage the participant in a deeper exploration and to elicit descriptive statements and expressions of their experience and perceptions. Interviews were conducted by professional counsellors who had extensive experience in body mapping as a therapeutic process with HIV positive adolescents. Each interview lasted between 1 and 1.5 hours and was conducted in English or Shona, according to the preferences of each participant. All interviews were audio-recorded. ## Data processing and analysis Each in-depth interview was transcribed and translated into English, if in Shona, and the body maps were photographed. Pseudonyms were used and any personal identifiers were removed from the transcripts before being entered into NVivo 10 (QSR International, Melbourne, Australia), a qualitative data storage and retrieval program. Verbal and visual data were then separately coded by two individuals (NW and CW) and then compared. Discrepancies were resolved by discussion. Once the coding was consistent for both the transcripts, they were single-coded. Codes were then grouped into categories and emerging themes were identified. Common themes from the different interviews and body maps were then identified and illustrated with quotes and images. The body maps were stored in a locked cupboard which was only accessible by the lead researcher. Once coded, participants were offered their body map but all preferred a photograph rather than a life size original. The original body maps were then destroyed. ## Ethical considerations Ethics approval was given by the Medical Research Council of Zimbabwe (MRCZ/B/665) and the Human Research Ethics Committee at Stellenbosch University. Prior to participation in the study, written informed consent was obtained from all participants 18 years and above. For minors, written consent was obtained from guardians and written assent was obtained from the participants themselves. Names used throughout this paper are pseudonyms chosen by the participants. Participants were given a stipend for participating in the study as well as being provided with lunch. No current acute suicidal ideation was expressed. However, participants who suggested on-going passive suicidal ideation were referred to a psychologist / psychiatrist at the treatment centre. All participants were offered follow up counselling and all readily consented. # Results ## Demographic characteristics of participants Of the 21 study participants, 11 (52%) were male and 19 (90%) were orphans, with 12/19 (63%) having lost both parents, 3/19 (16%) having lost a mother and 4/19 (21%) having lost a father. Only 3/21 (14%) participants lived with a biological parent. The majority of adolescents (18/21, 86%) stated during the body mapping process that they had acquired HIV from their mother whilst the remaining 3 (14%) were horizontally infected with HIV following sexual abuse as children. About three quarters (16/21, 76%) were on first line antiretroviral therapy whilst the remainder (5/21, 24%) were on a second line therapy regimen. Below, we describe the themes that emerged from qualitative data. One adolescent declined to participate in the study. ## Idioms of distress Participants’ idioms of distress were conveyed through both their verbal narratives and the words and images which they painted on their body maps. The most commonly used terms to describe depression were *‘thinking deeply’ (kufungisisa)* or being ‘*lost in thought’* as result of the events in their lives. ‘*Darkness*’, ‘*pain*’, ‘*stress*’ and *‘hopelessness’* were also common. Suicidal ideation was commonly referred to, including ‘slow suicide’ by five participants who described a desire to intentionally default on their antiretroviral treatment as a means of ending their lives. One female participant stated *“Sometimes I feel like giving up on life by not taking my medication \[antiretroviral\] so that people stop talking about my status \[HIV\]” (Melissa*, *18 years)*. Some participants also referred to somatic symptoms of stomach ache and headaches but this information only emerged after probing. ## Contributing factors Participants described a range of negative and traumatic experiences contributing to their depression. They generally attributed their experiences of depression to the expressed thoughts, behaviours and attitudes of the most significant people in their lives, specifically their relatives and peers. This was conveyed through both their verbal narratives and the evocative imagery which they chose to paint on their body maps to illustrate their experiences and emotions. Seven main themes emerged from the data which all related to their emotional experiences of depression: ### 1. Being different from others Participants commonly described feeling different from others due to their HIV status, particularly in relation to their parents, guardians or siblings they lived with and were HIV negative. They narrated their distress at being physically different (due to stunting and delayed puberty), orphaned and failing at school. They referred to the confusion and pain they felt as a result of being the only one to be HIV infected, asking “*How did I get it*?*”* and “*Why me when others do not have it*?” One boy stated his use of the colour red *“represents the pain that I go through each day I go to school because of my body…It’s small so people always tease me” (Persevere*, *18 years)*. ### 2. Learning of their HIV status Participants described feelings of ‘*sadness’*, *‘stress’ and ‘great pain’* in their lives at the time of learning their HIV status, as articulated by one female participant. *“After hearing* (the) *results … when you are told that you have it (HIV)*. *That is when you find all the stresses*. *That is when you start to think of bad ideas… that’s when depression comes” (Tarisai*, *18 years)*. ### 3. Isolation and rejection A sense of isolation and rejection was common, whether by others, self-imposed or anticipated. Participants commonly described being ridiculed, laughed at, talked about and excluded by peers as a result of their HIV status or the fact that they took medicines. A female participant also referred to the concerns of her peers around HIV transmission when she explained “*They will not play with me because if they touch me*, *they will be "infected"*. *(Paradzai*, *15 years)*. Participants commonly drew themselves playing alone. They described their ‘pain’ and ‘hurt’ as being worse when the isolation or rejection in their lives was imposed by the people they expected should support them, and commonly depicted this with paintings of relatives and peers. A male participant stated *“They hurt me when they say*, *'This one drinks pills (for HIV)*, *so there is no need to bother with him because he is not my child'*. *(Wangaa*, *17 years)*. Some participants referred to the way in which they isolated themselves from others in order to conceal their HIV status and their fears of what would happen if they were to be rejected. One male participant stated “*You will not be drinking the tablets on time*, *because you will be saying*, *'If l drink the tablets and people see me*, *they will laugh at me' (Tambudzai*, *18 years)*. ### 4. Loss and grief As stated earlier, the majority of respondents had been orphaned and demonstrated signs of profound, unresolved grief. Participants commonly described the death of a parent/parents as contributing to their depression and expressed a yearning to have known their parents through their words, paintings, body language and emotions. *“You need that love*, *but you won’t get it*. *I cry myself to sleep every day*. *My mother died when l was three*. *I never knew her*, *l don’t have a picture of her” (Janet*, *19 years)*. Grief was often exacerbated where participants were living in unsupportive households or where a loving, caregiver relationship was lacking. A female participant symbolised her stepmother by drawing a snake and going on to explain, “*I stay with my stepmother who is very cruel to me*. *She ill-treats me because she has her own child and so it is done on the basis that I am not her child*. *My mother passed away a long time ago…I wish she was around” (Paradzai*, *15 years)*. ### 5. Low self-worth Participants described a longing to be important or to matter to the people in their lives, specifically family members and peers. Whilst a few described elements of support from caregivers, the majority narrated traumatic relationships with their primary caregiver who lacked care or concern for them compared with their HIV negative siblings or other children in the household. Others stated that they were moved from household to household, without being cared for or loved. One male participant explained *“When I was sick*, *some relatives would say I was to be left to die just the same way my mother died*. *I was moved from relative to relative" (Kudzanai*, *17 years)*. ### 6. Lack of protection Participants reported that they had not been protected by the people closest to them, who not only failed to protect them, but also inflicted abuse. A sense of betrayal and disillusionment was evident in their narratives, particularly by those who had been sexually abused. They castigated both the perpetrator and family members who did not seek justice against those who had abused them. One female participant painted her entire body black and explained, “M*y mother died…My father l have but he doesn’t care about me*… *He was sexually abusing me (Janet*, *19 years)*. ### 7. The future Most participants expressed high expectations for a successful, brighter future, illustrated by drawings of getting married, having children, achieving academically and gaining employment. The need for independence from unsupportive caregivers was significant for some participants and was linked to the desire to be employed and economically stable. They also stated that marital status and having children of their own, along with completing their education and becoming employed, would improve their sense of self-worth. However, these expectations were often accompanied by an uncertainty regarding whether they could actually achieve these, given their poor health and lack of finances, academic qualifications and support. One male participant stated, “F*or now there is nothing \[in his future\]*. *It’s just hazy and a bit complicated to understand where it’s headed to*. *At times it’s just sorrowful and just so sad"* (*John*, *18 years*). Another participant referred to his fear of the future when he painted a hen as a symbol for ‘depression’. He described himself as a ‘hen’ *“because anytime it can be killed*. *So just like me*, *l am like a hen*. *l don’t know when l will die but I just know that l will die because of the situation that l am in” (Prince*, *16 years)*. ### Supportive factors A range of factors were identified by participants as supporting them with depression. Participants who had accessed peer-led and psychosocial support services through the Zvandiri programme stated that this support had played a central role in improving their sense of self-worth and confidence, and reducing their sense of isolation and rejection. This was particularly noted among those who lacked support from caregivers or peers in their daily lives. Supportive counsellors were also highlighted by some participants as being important in their care. Educational assistance and skills training for future employment were identified as being critical interventions for their well-being. Participants described feeling relieved as a result of being able to share experiences they had not shared before. # Discussion The results from this study confirm the profound significance of family and peer relationships in the lives of HIV positive adolescents with depression. Although learning their HIV status was significant, participants commonly attributed their negative experiences and subsequent depression to the expressed thoughts, behaviours and attitudes of the most significant people in their lives. Stable, supportive families are critical in promoting normal childhood development and there is a strong causal relationship between parental relationships and adolescent depression. Yet the results of this study suggest that stability and support from parents were critically lacking in participants’ lives, with 63% being double orphans and only 14% living with a biological parent. Participants were coping with complicated grief, from losses that included not only the death of one or both parents, but siblings and grandparents, as well as the consequences of those losses. Participants reported that they yearned to be loved, accepted, valued and supported by their immediate relatives and the absence of this further compounded their sense of grief and loss for their biological parents. Interventions to support caregivers have been found to improve the mental health of children living with HIV and could be adapted and scaled up for adolescents. The role of grief and loss in their narratives of depression concurs with the literature indicating that children who experience the death of a parent demonstrate lower self-esteem and higher rates of psychological problems than non-bereaved children. Multiple losses can negatively affect the development of a sense of self-worth, interfere with a person’s ability to trust or depend on others, and can lead to an avoidance of close interpersonal relationships. The results from this study and the literature suggest that there is a critical need for therapeutic grief interventions for children and adolescents living with HIV. Yet grief and loss have been largely neglected in this group of young people in Sub-Saharan Africa, despite the magnitude of losses in the lives of children and adolescents living with HIV in the region. A grief intervention for adolescent girls in South Africa was found to significantly reduce the incidence of complicated grief and depression. This intervention model could be replicated for adolescents living with HIV. The study results also demonstrate the critical role of peer relationships on the confidence, self-esteem and identity of HIV positive adolescents with depression. Participants demonstrated that in their own constructed realities, they felt worthless, of no value and thought they had no future when compared with their peers. Participants expressed a yearning to be accepted and valued by their peers and to have opportunities to socialise with them, and to also identify with them physically, socially and developmentally. They also commonly referred to the way in which their peers do not need to take medication. In the same way that peers were identified as contributing to their negative experiences, participants clearly described them as being central to the support that they required. Adolescence is a unique stage of life characterised by rapid growth and development and characterised by increasing autonomy, independence and an intense desire to associate and identify with peers. Yet the participants described multiple challenges which they perceived as making them different from their peers. Skin disfiguration, stunted growth and pubertal delay were clearly identified as contributing to stigmatising behaviour by peers, setting them aside from their peers and affecting their confidence and self-esteem. A fear of peers or partners finding out their HIV status was common in all narratives. Participants also described being different due to poor academic achievement in school and narrated a deep yearning to succeed, like their peers. Although skin disfiguration, growth and pubertal delay, cognitive impairments and fear of disclosure to others are described in the literature, there has been a lack of attention to the impact on adolescents’ mental health. The World Health Organisation now recommends peer-led interventions for adolescents living with HIV. Several models of group-based and peer-led interventions exist for adolescents living with HIV and there is some evidence that these have contributed to improved retention, psychosocial well-being and virological suppression. However, studies that specifically investigate the role of HIV positive young people in supporting their peers diagnosed with depression are lacking. Lay counsellors have been found to be effective in addressing the mental health of adolescents living with HIV and there is now need to investigate the effectiveness of adolescents and young people in this role. Similarly, despite the growing number of studies that show HIV-infected adolescents are at increased risk of mental health problems including depression, current models of adolescent HIV service delivery worldwide do not integrate mental health services. This lack of attention to the importance of mental health services for young people with HIV leads to challenges in ensuring early, accurate diagnosis and treatment, resulting in the mental health needs and HIV treatment and care being unmet. This is not only essential for their own HIV outcomes but also for their development, and survival in to adulthood. There is need to ensure that national HIV policies and guidelines are inclusive of mental health and that health care workers are trained to respond to the multi-faceted mental health needs of adolescents living with HIV. Since most sub-Saharan African countries (Zimbabwe included) have a critical shortage of psychiatrists, it is important that mental health interventions be adapted for health workers and lay workers so that mental health services can be effectively rolled out and integrated within the national HIV programme. This approach has been effectively implemented in Zimbabwe by the Friendship Bench which engages lay counsellors as therapeutic counsellors for adults with depression. Also in Zimbabwe, HIV positive adolescents and young people trained as Community Adolescent Treatment Supporters (CATS) have been found to be effective in improving adherence and psychosocial well-being among their HIV positive peers and current studies are investigating their effectiveness in improving common mental disorder in adolescents living with HIV. A potential limitation of this study is that the sample size was small and it may therefore not be possible to make generalizations about the larger population of adolescents living with HIV and depression. However, these in- depth data from young people in urban Zimbabwe provide important evidence towards an improved understanding of the needs and experiences of HIV positive adolescents with depression and their perceptions of the care they have received. # Conclusions and recommendations An understanding of the narratives of depression among adolescents living with HIV is necessary to inform the development of services which are responsive to their mental health needs, and can have a positive impact on the long-term HIV treatment and care outcomes, and transmission risk to partners and children. The findings from this research suggest that family and peer-led interventions may be effective in preventing and responding to depression among adolescents living with HIV and should be a key component of differentiated service delivery models for this age group, in order to improve their mental health as well as adherence to ART. Efforts should focus on the development and scale up of family interventions and equipping peer counsellors with skills to integrate mental health interventions within their work with adolescents living with HIV. However, studies are needed to evaluate the effectiveness, acceptability and feasibility of such family and peer-led mental health interventions in preventing and managing depression and improving adherence to ART. # Supporting information The authors would like to acknowledge the following for their invaluable contributions: The Ministry of Health (MoH) for its leadership in advancing this research, commissioning this study and supporting its ethical approval; the MoH EIMC Task Force members for their technical guidance; the adolescent participants for participating in the study and providing the research team with their time and valuable information; and the 2 supervisors \`(AK WM), 3 research assistants (AM, EG, SB), 1 transcriber (AM) whose dedication contributed significantly to study execution and completion. [^1]: The authors have declared that no completing interests exist.

# Introduction Engaging stakeholders early in the translational spectrum could help to advance public trust and understanding of science and the impact of scientific research on human health. The National Center for Advancing Translational Science (NCATS) and other research funders have made substantial investments in developing approaches and resources to support engagement in translational research. Established frameworks for stakeholder identification and involvement in research exist for clinical and outcomes research, but it is reasonable to question the extent to which these apply to early stage translational science. Four Clinical Translational Science Award (CTSA) hubs launched a collaboration to explore how to engage with stakeholders in the setting of T0 (basic biomedical research) and T1 (translation to humans) research. The “T’s” in the translation research spectrum represent the transitions between the phases of research. We set out to answer three research questions: (1) Who are the stakeholders in early stage translational science? (2) How can CTSA institutions and researchers improve stakeholder engagement in early stage research? (3) What are the barriers and facilitators to engaging stakeholders in early stage translational research? Clinical and translational research (CTR) is the process of turning scientific observations into interventions that improve and enhance the health and well- being of individuals and populations; basic science is research that addresses foundational questions in the earliest stages of translation. We define a *stakeholder* as an individual or group who is responsible for or affected by health- and healthcare-related decisions that can be informed by research evidence. We define *engagement* as a bi-directional relationship between the stakeholder and researcher that results in informed decision-making about the selection, conduct, and application of research findings. # Methods This study was approved by the Tufts University/ Tufts Medical Center Health Sciences IRB. IRB approval \#12224—This study was deemed exempt. The Clinical and Translational Science Institute at Tufts University (Tufts CTSI) convened its own and three additional CTSA hubs–the Institute for Translational Health Sciences at the University of Washington (ITHS), the Clinical and Translational Science Institute at New York University Langone Health (NYU CTSI), and the Translational and Clinical Sciences Institute at the University of North Carolina, Chapel Hill (NC TRaCS)–to investigate the views of early stage translational science researchers on the involvement of stakeholders in their work. Data were collected via focus groups and semi-structured interviews. Each CTSA recruited discussants from lists of T0 and T1 researchers who had accessed resources in their own hubs. A focus group discussion guide was developed from a simple logic that ties research studies directly to decision problems faced by stakeholders: decisions made by stakeholders can be informed by evidence; the need for this evidence can be formed into a topic and question; and research can be developed to address this topic and question. The guide posed three key broad questions: (1) how is your work used in other applications; (2) who uses your work in other applications; (3) who is affected by your work as it is used in other applications? Probes included assessing the barriers and facilitators to achieving the ideal answer to each question. Introductory material, probes, and instructions to the interviewer were included. The guide was pilot tested with two researchers at Tufts CTSI and revised to improve the clarity of the questions. Focus groups were held for one hour, interviews lasted approximately 30 minutes, and both were audio recorded. Focus groups and interviews at non-Tufts sites were administered via WebEx and facilitated by investigators at Tufts with extensive experience in qualitative data collection to ensure consistency across sites. Audio recordings were transcribed verbatim and deleted after transcripts were de-identified. De- identified transcripts were coded using Dedoose™. A codebook was developed deductively from the discussion protocol based on previous literature on stakeholder engagement. Two coders (AL and VK) reviewed each transcript independently and added emergent themes identified using a modified grounded theory approach. Once they finalized the codebook, the coders reanalyzed transcripts and used a comparison and consensus approach to resolve any discrepancies. After coding was complete, we continued to iteratively group categories of codes into the broader categories of “barriers” and “facilitators” until the major themes discussed below crystalized as unique but related concepts. # Results We convened six focus groups (Tufts CTSI = 3; ITHS = 2; NYU CTSI = 1), and, at one site (NC TRaCS) where convening a focus group was not feasible, we conducted two interviews using the same discussion guide. The focus groups ranged in size from two to five participants. Ultimately, we held eight conversations with 24 individuals representing a range of clinical, methodological, career stage characteristics, and previous experience with stakeholder engagement in their work. Participants’ understandings of stakeholder engagement, their views on barriers that stand in the way of engaging stakeholders, and their recommendations for introducing new facilitators to support engagement work are described below. ## Stakeholder engagement in early stage translational science All early stage translational researchers reported some level of engagement, and many had engaged with a broad array of stakeholders, but not all engaged with the same groups of stakeholders as their colleagues. Researchers working on specific diseases or clinical conditions had engaged with federal and local government agencies (the Centers for Disease Control, the Department of Defense, the Department of Agriculture, local branches of Health and Human Services), international government bodies (a foreign department of health, the World Health Organization), private industry, patients and advocacy groups, and many other types of stakeholders. One researcher, whose focus is a parasite common in cows, talked about working with farmers in the country where she conducted her work. The researcher’s passion was children affected by the disease, but the financial impact on the local agricultural community produced more engagement, highlighting the ways in which factors external to the research can shape engagement. Those doing clinically agnostic work, such as the mechanisms of cell death, were more likely to have engaged only with other researchers as stakeholders in their work. As one chemist stated: “My fondest desire for an end user is another basic scientist.” Working with practitioners to define potential clinical applications of early stage research was important to some researchers. A senior basic scientist said, “It’s very important to have the clinicians on board early,” because it helps ensure the work will be of clinical relevance. Another said, “We regularly have medical doctors as trainees in my laboratory to make sure the stuff we do is of medical interest.” ## Barriers and facilitators to engaging stakeholders in early stage research Participants described a wide range of barriers to engaging stakeholders. Barriers were grouped iteratively until three major themes emerged. The three themes and related recommendations to address them–the facilitators–are described below under the following themes with titles drawn directly from transcripts. ### Theme 1. “Poor definitions” and “Translation is in the eye of the beholder” The terms of reference we use in CTR emerged as a barrier. Who qualifies as a stakeholder, what constitutes engagement, and how one understands translation were all points of confusion and contention. While definitions exist, there appeared to be a lack of shared understanding among many participants. Whether research is deemed translational may be a matter of definition. Some participants felt uneasy with the general concept of “translational science” and viewed it as at odds with the fundamentally incremental nature of science. Some worried that a “rush” to translation could lead to skipping steps or avoiding interesting avenues of research that could lead to new discovery. One researcher said, > *I struggle with the definition of what’s ‘translational research’ and > who our work is supposed to affect. I think by using the word > ‘translational,’ you tie it directly to patients, yet work in disease > models is the fundamental first step to this. And so I guess I’m just > somewhat uncomfortable even with the term ‘translational research’ in > this whole need to directly tie everything we do immediately to some > end outcome.* Another researcher suggested a different emphasis is needed to address research in T0 and T1 settings: > *We really have a push at NIH to be translational and I don’t think > that people really understand what that means…It doesn’t mean that you > don’t do basic science or that everything has to be a model of some > disease…\[The\] frustration is \[with\] the idea that every scientific > work has to be transformational and not incremental.* Participants also shed light on what defines a stakeholder and which stakeholders are important in their work. No one disagreed with the definition of stakeholders that we put forward, the same one that is presented in frameworks for identifying stakeholders in T2 through T4 CTR(2)–namely that stakeholders are those who make decisions with evidence or are affected by the decisions that are made with evidence. However, several participants pointed out that this previous framework’s call to scan the “7Ps” (patients, providers, payers, purchasers, product makers, policy makers, and principal investigators) for relevant decision-makers obscures a necessary emphasis on involving one “P” in particular: the principal investigators (researchers, research entities, and research funders) doing similar basic science work in other disciplines. This special emphasis amounts to a call for multi-disciplinary research to broaden the evidence base before basic discoveries begin translation to humans: “There’s a lot of stuff in the ‘Stage 1’s’ of translational research that requires a lot of intensive investigations for a lot of potential applications.” ### Recommendations to address Theme 1 Researchers expressed a need a for better direction from NCATS and their CTSA hubs about the terms of reference surrounding engagement work in early stage translational science: definitions of translation need to allow for discovery research to proceed without being linked strictly to a clinical, intervention, or product pathway; identification of relevant stakeholders should start with an emphasis on researchers and research groups representing multiple disciplines; and guidance on identifying relevant stakeholders should shift to the “7Ps” framework only as clinical applications, interventions, and product pathways emerge from discovery. Building the case for using T0 and T1 evidence in other applications often involves collaboration with a cohort of other principal investigators, often a multi-disciplinary team, before the work can progress from early stage to latter stage research. As careers progress, the more distal end users of T0 and T1 work may become apparent, as the investigator sees their work wending its way into T2 and T3 settings. Here, the end users may be the full complement of stakeholders: “We all want to feel that what we're doing is important and will one day lead to something, whether it’s directly from what we conceive or it might be a step in the process…That maybe we don't have the magic, but what magic we develop will be able to be part of the step, then, to the final process.” ### Theme 2: “No instructions” and “Tell us what to do, in what order” The second theme that emerged was a skills barrier: This is the skills issue: absence of guidance, training, mentorship and skills. For early career investigators the absence of training on practical approaches to stakeholder engagement is a barrier. For instance, a junior investigator said he did not need to be sold on the concept of stakeholder engagement in translational science; he bought it, “hook, line, and sinker.” What he lacked was instructions about which stakeholders to engage and when in the process to engage them: > *I don’t even know what that pathway is. There’s not really a trail > that says, ‘Do this, then do this, then do this. Speak with these > people along the way.’ I don’t know what the path is or who the path > is through, I just know very broadly where it needs to go.* Another echoed this comment: “Everyone wants to translate something, right? It’s a buzz word. Everyone wants to do that thing or say they’ve done that thing, but the *‘hows’* are not as well defined for at least people in my space.” Similarly, some participants talked about a lack of training in the mechanics of translating bench discoveries into clinical applications, interventions, or products. For instance, one participant focused on administrative hurdles they discovered when working with industry on patents and licensing: “\[I\] feel as a junior faculty that I don’t necessarily have the commercial training nor the time to learn this skill as I would perhaps after tenure.” Some researchers reported lacking the skills necessary to communicate scientific ideas to lay audiences in simple but meaningful ways. Good science communication skills were seen as crucial: “No matter how good their ideas, you have to be able to explain it to whoever it is you're talking to if you want them to pay attention;” but lacking: “Scientists aren’t exactly the best spokespeople for their own work much of the time.” Lack of mentoring was also challenging for junior faculty. Variations in the culture of the discipline, department, or institution could also lead to researchers feeling isolated or silo-ed. “I feel a little isolated where I am. I don’t necessarily have a mentor to help guide me in a lot of the decision processes that I’m making.” The clinician-researcher role was seen as better established in some fields (cardiology) than others (gynecology), although this may vary between institutions, and this impacted both the availability of protected research time for clinicians and the availability of mentors. In non- clinical fields (chemistry), the proximity (or lack thereof) of the department to the healthcare setting could also create a barrier: > *Very few of my colleagues are involved with anything that is beyond > foundational basic research. And even fewer are involved in anything > related to healthcare. I don’t necessarily have a mentor to help guide > me in a lot of the decision processes that I’m making. That just adds > to a feeling of isolation and flying by the seat of pants in not being > savvy regarding who I contact about collaborations or grants because > there’s no one telling me, ‘That’s a dumb idea, that’s a good idea.’* ### Recommendations to address Theme 2 Several recommendations were made to address the barriers described by skills- related barriers. These include: training that folds engagement work into researcher education, creation of practical “how to” guides on engagement work, and more opportunities to use mentoring in CTR, including on how to identify and become engaged with potential en-users of the work. > *You have at least two junior faculty members \[here\] that are > begging for this sort of pathway forward. Tell us what to do, in what > order, how to contact people, when’s the right time to engage this > stakeholder and this stakeholder and this stakeholder.* CTSA hub investments in mentorship programs were identified as a way to overcome these barriers. After describing the absence of a clear clinician-research path in her field and institution, one participant said that the mentorship she encountered through her institution’s CTSA from colleagues outside her field had greatly benefited her: “One thing that \[the CTSA\] has given me that’s been a real boon is the mentorship…without that, I couldn’t have done it.” ### Theme 3: “Competing demands on resources” and “lack of resources” Material resources to support research were viewed by our discussion participants as scarce, may have strings, and difficult to come by. All of this puts pressure on researchers to choose among the many competing demands of research, and stakeholder engagement can fall off the radar without funds that are directed specifically to that purpose, or without funding criteria that call for it. Having protected time for research was one of the most basic challenges. Junior faculty without established research portfolios felt this acutely: > *My chair is very morally supportive but nobody gives you startup > money or protected time. So I started out, after my Fellowship, doing > 100% clinical, and I had to rearrange my time in order to have a > little bit of research time that wasn't 9:00 at night. That was very > stressful, because I have the research knowledge, but I didn't have > the time or energy to do it. And, slowly, I've crawled my way to > having more protected time through luck and opportunity and meeting > people–enough protected time that I was able to apply for this award > \[institutional K\], but it's taken me four and a half years.* A tight funding climate was described by all participants as a barrier to adding stakeholder engagement to their list of priorities, but tight funding presents a unique challenge for junior researchers. Some participants described being unable to find a “home” in the National Institutes of Health (NIH), because their research interests do not line up with individual Institutes and Centers of NIH: “I don't really have an institute in the NIH where I can apply very easily.” An early career researcher said, “The mentality of the tenure track is just go, go, go, go, go, which is great. But sometimes I just wish I had the time to just sit back, think logically of the next step.” Some participants described funding challenges in another context: the misalignment of industry profit motives and the patient or public health goals of researchers. Returning to an earlier example of a research who wanted to prioritize the health of humans over livestock, they state: “We have licensed an antibody that we developed to use for treatment of \[this bacteria\] … I’ve sort of hit a wall against finding somebody who is interested in, in developing the technology or marketing it for use in developing countries because there’s no money in it.” ### Recommendations to address Theme 3 To successfully engage stakeholders in the earliest stages of translational research, researchers need material resources that are specifically targeted to that purpose. CTSAs are in a unique position to deliver many of these material resources, and participants recognized this: > *I think that the \[CTSA\] is an important space that values research > and, maybe, could do some more pushing or open up some more > opportunities for people who are in the clinical field that really > would like to pursue research.* Participants had the fewest suggestions for how to address funding barriers. Everyone recognized it as a problem, but no one purported to have the solution. It may be this is where institutions, such as NCATS, need to take the lead in helping to create structures that support protected time for research and access to funding which prioritizes stakeholder engagement especially for researchers in the earliest stages of CTR. # Discussion In mathematics, a lemma is known as a "helping theorem," a rule that may be used to develop or support some result. We use the term “three lemmas” to call attention to three helping theorems by which barriers described in this article may be eliminated, and we describe the organizations who may take charge of this effort: the National Center for Advancing Translational Sciences (NCATS), the consortium of CTSAs, and the individual CTSA “hubs” are three organizational entities that can help T0 and T1 researchers develop shared terms of reference, build the necessary skills, and assemble the appropriate resources for engaging stakeholders in basic science research. Getting this right will involve a coordinated push by all three entities; therefore, this is a triple helping theorem with three entities playing important roles. There is tremendous variation in the level of stakeholder engagement among researchers in early stage (T0-T1) translational science. Participants identified a number of barriers to engaging stakeholders. We heard also, in these conversations, about the facilitators–the resources that help to remove those barriers. Below we summarize those barriers and make recommendations for how to address them. ## Addressing definitions There seems to be a problem with the operational definition of translational research. We have identified numerous definitions of translational research, and while there is overlap, there is not total consensus. Some researchers we spoke with are conflating “translational” with “patient centered,” and while we might agree that good translational research *is* patient-centered, focusing *solely* on patients as end-users comes at the exclusion of multiple other stakeholder groups and may represent a cognitive barrier to stakeholder engagement. CTSAs can help with emphasizing the *spectrum* of translational research and the full range of stakeholders that can be engaged along that spectrum. The multi—nodal, concentric rings presented in the Translational Research Framework put forth by the National Institute of Environmental Health Sciences (NIEHS) is a useful diagram for visually representing the complexity and possibilities of CTR. Translational can occur by moving along one’s own circle, to a different discipline (node), for example. Reaching the outer circle of population health can be the ultimate goal, but there can be plenty of movement within the other circles. Translational science can be a marathon or a relay–every researcher or research team does not have to go the full distance on their own. ## Instruction manual Even for those for whom the definition of translational research is not an impediment, the process can seem impenetrable. The *what* is not the charge, but the *how* is. The good news is that frameworks exist \[–, \]. We need to do a better job of disseminating these guidelines and actively targeting researchers in the earliest stages of CTR. In the Clinical and Translational Science graduate program at Tufts CTSI, graduate students are required to have a stakeholder on their thesis committee. The Clinical and Translational Science Institute at New York University’s Langone Health prioritizes funding pilot studies that demonstrate strong stakeholder engagement. At the NC TraCS Institute at the University of North Carolina, the Community and Stakeholder Engagement (CaSE) unit has developed a range of tools and trainings to promote stakeholder engagement, build researcher capacity and provide technical assistance to investigators across the translational spectrum. ## Focusing the CTSA hubs to deliver material support for engagement work in basic science research When it comes to resources, the good news is that there already exists a mechanism designed specifically to address the issue of providing resources to researchers engaging in translational research. This is, in fact, the very mandate of NCATS. While CTSAs cannot replenish the coffers at NIH, they can provide resources–pilot grants, grant writing workshops, mentorship–that can help support researchers in their pursuit of funding and achieving the research once it is funded. Of the recommendations we have made here, 1 (definitions) and 2 (instructions) are necessary, but not sufficient. Even with crystal clear definitions and guidelines, it is ultimately 3, the resources, that make the difference. These issues are particularly challenging for young researchers because they are the most strapped for resources. While an initial challenge may be to convince people of the value of translational research, this may require less of a paradigm shift for early career investigators and researchers. CTSAs should embrace and encourage their enthusiasm by supporting them via focused training and resources. CTSAs are charged with facilitating the translation of science. Stakeholder engagement is a key component of this mission. Paying special attention to the unique needs of research working in at the earliest stages of the translational spectrum and providing them with resources to overcome the barriers they encounter could facilitate the advancement of the science itself as well as public trust and understanding of science and the impact of scientific research on human health. # Supporting information The authors would like to thank the following individuals for their assistance with this project: Lisa Quarles, BA, Training Coordinator and Community Engagement Specialist for the North Carolina Translational and Clinical Sciences Institute, home of the CTSA at the University of North Carolina at Chapel Hill; Smiti Kapadia Nadkarni, MPH, Program Manager for Community Engagement and Population Health Research Program (CEPHR), NYU-H+H Clinical & Translational Science Institute. 10.1371/journal.pone.0235400.r001 Decision Letter 0 Edward Karen-Leigh Academic Editor 2020 Karen-Leigh Edward This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 26 Feb 2020 PONE-D-19-35056 Facilitating Stakeholder Engagement in Early Stage Translational Research PLOS ONE Dear Dr LeClair, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. ACADEMIC EDITOR: I found the study interesting. Please make overt the framework used to inform the maintainence of rigour throughout the qualitative data collection and analysis. Also further detail about how the themes emerged is required. Please disregard the reviewer 2 comment about quantitative research as this does not apply to your paper. We would appreciate receiving your revised manuscript by 20/05/2020. When you are ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. To enhance the reproducibility of your results, we recommend that if applicable you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission- guidelines#loc-laboratory-protocols> Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). This letter should be uploaded as separate file and labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. This file should be uploaded as separate file and labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. This file should be uploaded as separate file and labeled 'Manuscript'. Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out. We look forward to receiving your revised manuscript. Kind regards, Karen-Leigh Edward Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements: 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <http://www.plosone.org/attachments/PLOSOne_formatting_sample_main_body.pdf> and <http://www.plosone.org/attachments/PLOSOne_formatting_sample_title_authors_affi liations.pdf> 2\. We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For information on unacceptable data access restrictions, please see <http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data- access-restrictions>. In your revised cover letter, please address the following prompts: a\) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially identifying or sensitive patient information) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. b\) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. Please see <http://www.bmj.com/content/340/bmj.c181.long> for guidelines on how to de- identify and prepare clinical data for publication. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data- availability#loc-recommended-repositories>. We will update your Data Availability statement on your behalf to reflect the information you provide. 3\. PLOS requires an ORCID iD for the corresponding author in Editorial Manager on papers submitted after December 6th, 2016. Please ensure that you have an ORCID iD and that it is validated in Editorial Manager. To do this, go to ‘Update my Information’ (in the upper left-hand corner of the main menu), and click on the Fetch/Validate link next to the ORCID field. This will take you to the ORCID site and allow you to create a new iD or authenticate a pre-existing iD in Editorial Manager. Please see the following video for instructions on linking an ORCID iD to your Editorial Manager account: <https://www.youtube.com/watch?v=_xcclfuvtxQ> 4\. Thank you for stating the following in the Competing Interests section: "The authors have declared that no competing interests exist." We note that one or more of the authors are employed by a commercial company: RAND Corporation. 1.     Please provide an amended Funding Statement declaring this commercial affiliation, as well as a statement regarding the Role of Funders in your study. If the funding organization did not play a role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript and only provided financial support in the form of authors' salaries and/or research materials, please review your statements relating to the author contributions, and ensure you have specifically and accurately indicated the role(s) that these authors had in your study. You can update author roles in the Author Contributions section of the online submission form. Please also include the following statement within your amended Funding Statement. “The funder provided support in the form of salaries for authors \[insert relevant initials\], but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The specific roles of these authors are articulated in the ‘author contributions’ section.” If your commercial affiliation did play a role in your study, please state and explain this role within your updated Funding Statement. 2\. Please also provide an updated Competing Interests Statement declaring this commercial affiliation along with any other relevant declarations relating to employment, consultancy, patents, products in development, or marketed products, etc. Within your Competing Interests Statement, please confirm that this commercial affiliation does not alter your adherence to all PLOS ONE policies on sharing data and materials by including the following statement: "This does not alter our adherence to  PLOS ONE policies on sharing data and materials.” (as detailed online in our guide for authors <http://journals.plos.org/plosone/s/competing- interests>). If this adherence statement is not accurate and  there are restrictions on sharing of data and/or materials, please state these. Please note that we cannot proceed with consideration of your article until this information has been declared. Please include both an updated Funding Statement and Competing Interests Statement in your cover letter. We will change the online submission form on your behalf. Please know it is PLOS ONE policy for corresponding authors to declare, on behalf of all authors, all potential competing interests for the purposes of transparency. PLOS defines a competing interest as anything that interferes with, or could reasonably be perceived as interfering with, the full and objective presentation, peer review, editorial decision-making, or publication of research or non-research articles submitted to one of the journals. Competing interests can be financial or non-financial, professional, or personal. Competing interests can arise in relationship to an organization or another person. Please follow this link to our website for more details on competing interests: <http://journals.plos.org/plosone/s/competing-interests> 5\. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: <http://journals.plos.org/plosone/s/supporting-information>. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: N/A Reviewer \#2: N/A \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The terms CTSA, T0, T1, T2, should be explained the first time each is used in the manuscript. On page 2 where it says 'are three actors can help" should read "are three factors". On page 11 where it says "by the set of pants" should probably read "by seat of pants" On page 14 "we describe the actors who may take charge". Actors should be changed to a more scientific term such as participants or entities. I thought the parts "Addressing definitions" and "Instruction manual" were both accurate and insightful. The section "Focusing on CTSA hubs...research", clearly outlines hoe limited access to funding affects all aspects of the scientific research process. Reviewer \#2: The major concern about the current study is qualitative and the findings were from interviews and subjective. The topic is interesting, if conduct the study by design a survey with questionnaire and collect the data from the identified stakeholders, and then analyze the data, according to the response rate, the reliability and validity would allow the study more informative and objective. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: Yes: Frank Abbruscato Reviewer \#2: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at <[email protected]>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0235400.r002 Author response to Decision Letter 0 29 May 2020 Dear Karen-Leigh Edward and reviewers, Thank you for the thoughtful reviews and opportunity to respond. We have organized our responses below: 1\. Style requirements – We have reviewed our files to ensure they meet PLOS ONE’s style requirements. 2\. Data availability – While we understand data sharing has become more common for qualitative data, we are new to the concept for the sharing of qualitative data. We realize, in retrospect, we should have addressed this when submitting the manuscript, and the lead author – Dr. Amy LeClair – consulted with the Qualitative Data Repository (QDR) group for guidance. What we should have said in the original submission is that the data cannot be shared. We appreciate the opportunity to learn from this experience, both for the submission of future manuscripts and for the design of qualitative studies going forward. a\. This study was deemed exempt by Tufts Health Sciences Institutional Review Board, therefore a written informed consent was not used. However, participants were given an information sheet and verbal consent was obtained prior to beginning the focus groups/interviews. During this process, participants were explicitly told that their data would not be shared outside of the research team, and that only de-identified excerpts of the transcripts would be used in publication and other forms of dissemination. Furthermore, the complete data set from all four sites was only available to three members of the team for coding. 3\. ORCID ID – The corresponding author now has an ORCID ID, and it has been validated in Editorial Manager. 4\. Competing Interests – The RAND Corporation is not a commercial company. RAND has the same nonprofit status as the academic institutions and medical centers the other authors’ are affiliated with. It would not be appropriate to declare this as a competing interest and is not the policy of RAND or its researchers and staff. We have, therefore, not updated any of our Funding Statements or Competing Interests Statements 5\. Supporting Information – We have including captions for our supporting Information files. 6\. Reviewers’ Comments: Comment Response Academic Editor: “Please made overt the framework used to inform the maintenance of rigour throughout the qualitative data collection and analysis. Also further detail about how the themes emerged is required.” We have made several edits to the 3rd paragraph of the methods section to clarify the framework that guided our data collection and analysis, including adding citations to the literature that guided our methods. We hope this clarifies the methodology. Reviewer \#1: The terms CTSA, T0, T1, T2, should be explained the first time each is used in the manuscript. We have made sure all acronyms and abbreviations are spelled out in their first use and have also added a citation for our definitions of T0 and T1 in the 2nd paragraph of the article. On page 2 where it says 'are three actors can help" should read "are three factors". We have changed the word “actors” to “organizational entities” to more clarify t our meaning. On page 11 where it says "by the set of pants" should probably read "by seat of pants" Yes – thank you for picking up that type. We have made the correction. On page 14 "we describe the actors who may take charge". Actors should be changed to a more scientific term such as participants or entities. We have changed “actors” to “organizations.” I thought the parts "Addressing definitions" and "Instruction manual" were both accurate and insightful. The section "Focusing on CTSA hubs...research", clearly outlines hoe limited access to funding affects all aspects of the scientific research process. Thank you. Reviewer \#2: The major concern about the current study is qualitative and the findings were from interviews and subjective. The topic is interesting, if conduct the study by design a survey with questionnaire and collect the data from the identified stakeholders, and then analyze the data, according to the response rate, the reliability and validity would allow the study more informative and objective. Per the Academic Editor – “Please disregard the reviewer 2 comment about quantitative research as this does not apply to your paper.” 10.1371/journal.pone.0235400.r003 Decision Letter 1 Edward Karen-Leigh Academic Editor 2020 Karen-Leigh Edward This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 16 Jun 2020 Facilitating Stakeholder Engagement in Early Stage Translational Research PONE-D-19-35056R1 Dear Dr. LeClair We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <[email protected]>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <[email protected]>. Kind regards, Karen-Leigh Edward Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: 10.1371/journal.pone.0235400.r004 Acceptance letter Edward Karen-Leigh Academic Editor 2020 Karen-Leigh Edward This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 22 Jun 2020 PONE-D-19-35056R1 Facilitating Stakeholder Engagement in Early Stage Translational Research Dear Dr. LeClair: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <[email protected]>. If we can help with anything else, please email us at <[email protected]>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Professor Karen-Leigh Edward Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.