Datasets:

Dhruvil47
/

bioarxiv_papers

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 86k rows

text stringlengths 7 107k
862 Annual Review of Ecology and Systematics 21: 423–447. 863 864 Charnov, E.L., and W.M. Schaffer. 1973. Life-history consequences of natural selection: Cole’s result 865 revisited. The American Naturalist 107: 791–793. 866 867 Chave, J., D.A. Coomes, S. Jansen, S.L. Lewis, N.G. Swenson, and A.E. Zanne. 2009. Towards a 868 worldwide wood economics spectrum. Ecology Letters 12: 351–366. 869 870 Ciotir, C., W. Applequist, T.E. Crews, N. Cristea, L.R. DeHaan, E. Frawley, S.A. Herron, et al. 2019. 871 Building a botanical foundation for perennial agriculture: Global inventory of wild, perennial herbaceous 872 Fabaceae species. Plants, People, Planet 1: 375–386. 873 874 Cole, L.C. 1954. The population consequences of life history phenomena. The Quarterly Review of 875 Biology 29: 103–137. 876 877 Crews, T.E., and L.R. DeHaan. 2015. The strong perennial vision: A response. Agroecology and 878 Sustainable Food Systems 39: 500–515. 879 880 Csardi, G., and T. Nepusz. 2006. The igraph software package for complex network research. 881 InterJournal, Complex Systems 1695: 1-9. 882 bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 35 883 DeHaan, L.R., D.L. Van Tassel, and T.S. Cox. 2005. Perennial grain crops: A synthesis of ecology and 884 plant breeding. Renewable Agriculture and Food Systems 20: 5–14. 885 886 Delgado-Salinas, A., R. Bibler, and M. Lavin. 2006. Phylogeny of the genus Phaseolus (Leguminosae): A 887 recent diversification in an ancient landscape. Systematic Botany 31: 779–791. 888 889 den Dubbelden, K.C., and R.W. Verburg. 1996. Inherent allocation patterns and potential growth rates of 890 herbaceous climbing plants. Plant and Soil 184: 341–347. 891 892 Díaz, S., J. Kattge, J.H.C. Cornelissen, I.J. Wright, S. Lavorel, S. Dray, B. Reu, et al. 2016. The global 893 spectrum of plant form and function. Nature 529: 167–171. 894 895 Du, G., and W. Qi. 2010. Trade-offs between flowering time, plant height, and seed size within and 896 across 11 communities of a QingHai-Tibetan flora. Plant Ecology 209: 321–333. 897 898 Endo, Y., B. Choi, H. Ohashi, and A. Delgado-Salinas. 2008. Phylogenetic relationships of New World 899 Vicia (Leguminosae) inferred from nrDNA internal transcribed spacer sequences and floral characters. 900 Systematic Botany 33: 356–363. 901 β 902 Frehner, M., and E.E. Conn. 1987. The linamarin glucosidase in Costa Rican wild lima beans 903 (Phaseolus lunatus L.) is apoplastic. Plant Physiology 84: 1296-1300. 904 905 Freytag, G.F., and D.G. Debouck. 2002. Taxonomy, distribution, and ecology of the genus Phaseolus 906 (Leguminosae-Papilionoideae) in North America, Mexico, and Central America. Sida, Botanical 907 Miscellany 23:1-300. 908 bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 36 909 Friedman, J., and M.J. Rubin. 2015. All in good time: Understanding annual and perennial strategies in 910 plants. American Journal of Botany 102: 497–499. 911 912 Friedman, J., A.D. Twyford, J.H. Willis, and B.K. Blackman. 2015. The extent and genetic basis of 913 phenotypic divergence in life history traits in Mimulus guttatus. Molecular Ecology 24: 111–122. 914 915 Friedman, J. 2020. The evolution of annual and perennial plant life histories: ecological correlates and 916 genetic mechanisms. Annual Review of Ecology, Evolution, and Systematics 51: 461–481. 917 918 Gadgil, M., and O.T. Solbrig. 1972. The concept of r- and K-selection: evidence from wild flowers and 919 some theoretical considerations. American Naturalist 106: 14–31. 920 921 Galasso, I., G. Sonnante, D.G. Tota, and D. Pignone. 1997. Comparison of molecular cytogenetic and 922 genetic analyses in accessions of the two biotypes of Vicia benghalensis L. Annals of Botany 79: 311– 923 317. 924 925 Garnier, E. 1992. Growth analysis of congeneric annual and perennial grass species. Journal of Ecology 926 80: 665–675. 927 928 Garnier, E., and G. Laurent. 1994. Leaf anatomy, specific mass and water-content in congeneric annual 929 and perennial grass species. New Phytologist 128: 725–736. 930 931 Garnier, E., and S. Vancaeyzeele. 1994. Carbon and nitrogen content of congeneric annual and perennial 932 grass species: relationships with growth. Plant, Cell & Environment 17: 399–407. 933 bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 37 934 GBIF.org. 2020. GBIF Home Page. Available from: Website https://www.gbif.org [accessed 20 935 November 2020] 936 937 Geber, M.A. 1990. The cost of meristem limitation in Polygonum arenastrum: Negative genetic 938 correlations between fecundity and growth. Evolution 44: 799–819. 939 940 Glover, J.D., J.P. Reganold, L.W. Bell, J. Borevitz, E.C. Brummer, E.S. Buckler, C.M. Cox, et al. 2010. 941 Increased food and ecosystem security via perennial grains. Science 328: 1638–1639. 942 943 González-Paleo, L., and D.A. Ravetta. 2015. Carbon acquisition strategies uncoupled from predictions 944 derived from species life-cycle. Flora 212: 1–9. 945 946 González-Paleo, L., A.E. Vilela, and D.A. Ravetta. 2016. Back to perennials: Does selection enhance 947 tradeoffs between yield and longevity? Industrial Crops & Products 91: 272–278. 948 949 Gosling, P.G. 2003. Viability testing. In R.D. Smith, J.B. Dickie, S.H. Linington, H.W. Pritchard, R.J. 950 Probert [eds. ], Seed conservation: Turning science into practice, 447–481, Royal Botanic Gardens, Kew, 951 UK.
952 953 Grotkopp, E., M. Rejmánek, and T.L. Rost. 2002. Toward a causal explanation of plant invasiveness: 954 Seedling growth and life-history strategies of 29 pine (Pinus) species. American Naturalist 159: 396–419. 955 956 Grime, J.P. 1977. Evidence for the existence of three primary strategies in plants and its relevance to 957 ecological and evolutionary theory. American Naturalist 111: 1169–1194. 958 bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 38 959 Herron, S.A., M.J. Rubin, C. Ciotir, T.E. Crews, D.L. Van Tassel, and A.J. Miller. 2020. Comparative 960 analysis of early life stage traits in annual and perennial Phaseolus crops and their wild relatives. 961 Frontiers in Plant Science 11: 1–14. 962 963 Hodgson, J.G., B.A. Santini, F.R. Pla, G. Jones, A. Bogaard, M. Charles, X. Font, et al. 2017. Trade-offs 964 between seed and leaf size (seed-phytomer-leaf theory): functional glue linking regenerative with life 965 history strategies . . . and taxonomy with ecology? Annals of Botany 120: 633–652. 966 967 Huang, G., S. Qin, S. Zhang, X. Cai, S. Wu, J. Dao, J. Zhang, et al. 2018. Performance, economics and 968 potential impact of perennial rice PR23 relative to annual rice cultivars at multiple locations in Yunnan 969 Province of China. Sustainability 10: 1086. 970 971 Jaaska, V. 2015. Phylogenetic relationships among sections Vicia, Sepium and Lathyroides of Vicia 972 subgenus Vicia: isozyme evidence. Biochemical Systematics and Ecology 62: 186–193. 973 974 Josse, J., and F. Husson. 2016. missMDA: A package for handling missing values in multivariate data 975 analysis. Journal of Statistical Software 70: 1-31. 976 977 Kantar, M., C. Tyl, K.M. Dorn, X. Zhang, J.M. Jungers, J.M. Kaser, R. Schendel, et al. 2016. Perennial 978 grain and oilseed crops. Annual Review of Plant Biology 67: 1–27. 979 980 Kassambara, A., and F. Mundt. 2020. factoextra: Extract and visualize the results of multivariate data 981 analyses. R package version 1.0.7. https://CRAN.R-project.org/package=factoextra 982 983 Kelly, C.K., and M.E. Hanley. 2005. Juvenile growth and palatability in co-occurring, congeneric British 984 herbs. American Journal of Botany 92: 1586–1589. bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 39 985 986 Kenicer, G.J. 2008. An introduction to the genus Lathyrus L. Curtis’s Botanical Magazine 25: 286–295. 987 988 Kitchen, S.G. 1994. Perennial forb life-history strategies on semiarid rangelands: Implications for 989 revegetation.
USDA Forest Service General Technical Report INT-GTR-313, Intermountain Research 990 Station, Provo, Utah, USA. 991 992 Kleyer, M., J. Trinogga, M.A. Cebrián, R. Ejrnæs, A. Trenkamp, C. Fløjgaard, D.C. Albach, et al. 2019. 993 Trait correlation network analysis identifies biomass allocation traits and stem specific length as hub traits 994 in herbaceous perennial plants. Journal of Ecology 107: 829–842. 995 996 Kuznetsova, A., P.B. Brockhoff, and R.H.B. Christensen. 2017. lmerTest Package: Tests in linear mixed 997 effects models. Journal of Statistical Software 82: 1-26. 998 999 Leishman, M.R., M. Westoby, and E. Jurado. 1995. Correlates of seed size variation: a comparison 1000 among five temperate floras. Journal of Ecology 83: 517–530. 1001 1002 Lenth, R. 2020. emmeans: Estimated marginal means, aka least-squares means. R package version 1.4.7. 1003 https://CRAN.R-project.org/package=emmeans 1004 1005 Lewis, G., B. Schrire, B. Mackinder, and M. Lock [eds. ], 2005. Legumes of the world. Royal Botanic 1006 Gardens, Kew, UK. 1007 1008 Lindberg, C.L., H.M. Hanslin, M. Schubert, T. Marcussen, B. Trevaskis, J.C. Preston, and S. Fjellheim. 1009 2020. Increased above-ground resource allocation is a likely precursor for independent evolutionary 1010 origins of annuality in the Pooideae grass subfamily. New Phytologist 228: 318–329. bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 40 1011 1012 Lundgren, M.R., and D.L. Des Marais. 2020. Life history variation as a model for understanding trade- 1013 offs in plant–environment interactions. Current Biology 30: R180–R189. 1014 1015 Marshall, D.L., N.L. Fowler, and D.A. Levin. 1985. Plasticity in yield components in natural populations 1016 of three species of Sesbania. Ecology 66: 753–761. 1017 1018 Mazer, S.J. 1989. Ecological, taxonomic, and life history correlates of seed mass among Indiana Dune 1019 Angiosperms. Ecological Monographs 59: 153–175. 1020 1021 Meyer, R.S., A.E. DuVal, and H.R. Jensen. 2012. Patterns and processes in crop domestication: an 1022 historical review and quantitative analysis of 203 global food crops. New Phytologist 196: 29–48. 1023 1024 Milla, R., J. Morente-Lopez, J. Alonso-Rodrigo, N. Martin-Robles, and F.S. Chapin, III. 2014. Shifts and 1025 disruptions in resource-use trait syndromes during the evolution of herbaceous crops. Proceedings of the 1026 Royal Society B: Biological Sciences 281: 20141429. 1027 1028 Miller, A.J., and B.L. Gross. 2011. From forest to field: Perennial fruit crop domestication. American 1029 Journal of Botany 98: 1389–1414. 1030 1031 Moles, A.T., D.S. Falster, M.R. Leishman, and M. Westoby. 2004. Small-seeded species produce more 1032 seeds per square metre of canopy per year, but not per individual per lifetime.
Journal of Ecology 92: 1033 384–396. 1034 1035 Morishima, H., Y. Sano, and H.I. Oka. 1984. Differentiation of perennial and annual types due to habitat 1036 conditions in the wild rice Oryza perennis. Plant Systematics and Evolution 144: 119–135. bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 41 1037 1038 Murren, C.J. 2002. Phenotypic integration in plants. Plant Species Biology 17: 89–99. 1039 1040 Nabhan, G.P., and R.S. Felger. 1985. Wild desert relatives of crops: their direct uses as food. In G.E. 1041 Wickens, J.R. Goodin, and D.V. Field [eds. ], Plants for arid lands, 19-33, Springer, Dordrecht, 1042 Netherlands. 1043 1044 Pastor-Pastor, A., A.E. Vilela, and L. González-Paleo. 2018. The root of the problem of perennials 1045 domestication: is selection for yield changing key root system traits required for ecological sustainability? 1046 Plant and Soil 435: 161–174. 1047 1048 Patil, V.N., and M. Dadlani. 2009. Tetrazolium test for seed viability and vigour. In J. Renugadevi, P. 1049 Srimathi, R. R. Renganayaki, and V. Manonmani [eds. ], Handbook of Seed Testing, 209–241. Agrobios, 1050 New Delhi, India. 1051 1052 Perez-Harguindeguy, N., E. Garnier, S. Lavorel, H. Poorter, P. Jaureguiberry, W.K. Cornwell, J.M. 1053 Craine, et al. 2013. New handbook for standardised measurement of plant functional traits worldwide. 1054 Australian Journal of Botany 61: 167–234. 1055 1056 Pitelka, L.F. 1977. Energy allocations in annual and perennial lupines (Lupinus: Leguminosae). Ecology 1057 58: 1055–1065. 1058 1059 Pommerening, A., and A. Muszta. 2016. Relative plant growth revisited: Towards a mathematical 1060 standardisation of separate approaches. Ecological Modelling 320: 383–392. 1061 bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 42 1062 Pugliese, J.Y. 2017. Above- and belowground response to managing Kernza (Thinopyrum intermedium) 1063 as a dual-use crop for forage and grain. M.Sc. thesis, Ohio State University, Columbus, Ohio, USA. 1064 1065 R Core Team. 2019. R: A Language and Environment for Statistical Computing. R Foundation for 1066 Statistical Computing, Vienna, Austria. 1067 1068 Schneider, C. A., W. S. Rasband, and K. W. Eliceiri. 2012. NIH Image to ImageJ: 25 years of image 1069 analysis. Nature Methods 9: 671–675. 1070 1071 Rees, M., C.P. Osborne, F.I. Woodward, S.P. Hulme, L.A. Turnbull, and S.H. Taylor. 2010. Partitioning 1072 the components of relative growth rate: How important is plant size variation?
American Naturalist 176: 1073 E152–E161. 1074 1075 Reich, P.B. 2014. The world-wide “fast-slow” plant economics spectrum: A traits manifesto. Journal of 1076 Ecology 102: 275–301. 1077 1078 Reynolds, L.K., J.J. Stachowicz, A.R. Hughes, S.J. Kamel, B.S. Ort, and R.K. Grosberg. 2017. Temporal 1079 stability in patterns of genetic diversity and structure of a marine foundation species (Zostera marina). 1080 Heredity 118: 404–412. 1081 1082 Roumet, C., E. Garnier, H. Suzor, J.L. Salager, and J. Roy. 2000. Short and long-term responses of 1083 whole-plant gas exchange to elevated CO2 in four herbaceous species. Environmental and Experimental 1084 Botany 43: 155–169. 1085 1086 Roumet, C., C. Urcelay, and S. Díaz. 2006. Suites of root traits in species growing differ between annual 1087 and perennial species growing in the field. New Phytologist 170: 357–367. bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 43 1088 1089 Roumet, C., M. Birouste, C. Picon-Cochard, M. Ghestem, N. Osman, S. Vrignon-Brenas, K. fang Cao, 1090 and A. Stokes. 2016. Root structure-function relationships in 74 species: Evidence of a root economics 1091 spectrum related to carbon economy. New Phytologist 210: 815–826. 1092 1093 Ryan, M.R., T.E. Crews, S.W. Culman, L.R. DeHaan, R.C. Hayes, J.M. Jungers, and M.G. Bakker. 2018. 1094 Managing for multifunctionality in perennial grain crops. BioScience 68: 294–304. 1095 1096 SAS Institute. 2017. Base SAS®, version 9.4. SAS Institute Inc., Cary, North Carolina, USA. 1097 1098 Schaefer, H., P. Hechenleitner, A. Santos-guerra, M.M. De Sequeira, R.T. Pennington, G.J. Kenicer, and 1099 M.A. Carine. 2012. Systematics, biogeography, and character evolution of the legume tribe Fabeae with 1100 special focus on the middle-Atlantic island lineages. BMC Evolutionary Biology 12: 1–19. 1101 1102 Schlautman, B., S. Barriball, C. Ciotir, S.A. Herron, and A.J. Miller. 2018. Perennial grain legume 1103 domestication phase I: Criteria for candidate species selection. Sustainability 10: 730. 1104 1105 Segrestin, J., M. Navas, and E. Garnier. 2020. Reproductive phenology as a dimension of the phenotypic 1106 space in 139 plant species from the Mediterranean. New Phytologist 225: 740–753. 1107 1108 Sharma, V., and S. Kumar. 2012. Stipules are the principal photosynthetic organs in the papilionoid 1109 species Lathyrus aphaca. National Academy Science Letters 35: 75–78. 1110 1111 Silvertown, J.W., and M. Dodd. 1996. Comparing plants and connecting traits. Philosophical 1112 Transactions of the Royal Society of London, B, Biological Sciences 351: 1233–1239. 1113 bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 44 1114 Singmann, H., B. Bolker, J. Westfall, F. Aust, and M.S. Ben-Shachar. 2020. afex: Analysis of factorial 1115 experiments. R package version 0.27-2. https://CRAN.R-project.org/package=afex 1116 1117 Smýkal, P., M.N. Nelson, J.D. Berger, and E.J. von Wettberg. 2018. The impact of genetic changes 1118 during crop domestication. Agronomy 8: 119. 1119 1120 Soltis, D.E., M.E. Mort, M. Latvis, E. V. Mavrodiev, B.C. O’meara, P.S. Soltis, J.G. Burleigh, and R.R. 1121 De Casas. 2013. Phylogenetic relationships and character evolution analysis of Saxifragales using a 1122 supermatrix approach. American Journal of Botany 100: 916–929. 1123 1124 Stearns, S.C. 1992. The evolution of life histories. Oxford University Press, Oxford, UK. 1125 1126 Tank, D.C., and R.G. Olmstead. 2008. From annuals to perennials: Phylogeny of subtribe Castillejinae 1127 (Orobanchaceae). American Journal of Botany 95: 608–625. 1128 1129 Thompson, K., and D.J. Hodkinson. 1998. Seed mass, habitat and life history: A re-analysis of Salisbury 1130 (1942, 1974). New Phytologist 138: 163–167. 1131 1132 Turkington, R.A., and P.B. Cavers. 1978. Reproductive strategies and growth in four legumes. Canadian 1133 Journal of Botany 56: 413–416. 1134 1135 Turnbull, L.A., C. Paul-Victor, B. Schmid, and D.W. Purves. 2008. Growth rates, seed size, and 1136 physiology: Do small-seeded species really grow faster? Ecology 89: 1352–1363. 1137 1138 Van Tassel, D.L., L.R. DeHaan, and T.S. Cox. 2010. Missing domesticated plant forms: can artificial 1139 selection fill the gap? Evolutionary Applications 3: 434–452. bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 45 1140 1141 Van Tassel, D.L., K.A. Albrecht, D.E. James, A.A. Boe, Y. Brandvain, T.E. Crews, M. Gansberger, et al. 1142 2017. Accelerating Silphium domestication: an opportunity to develop new crop ideotypes and breeding 1143 strategies informed by multiple disciplines. Crop Science 57: 1–11. 1144 1145 Verboom, G.A., H.P. Linder, and W.D. Stock. 2004. Testing the adaptive nature of radiation: Growth 1146 form and life history divergence in the African grass genus Ehrharta (Poaceae: Ehrhartoideae). American 1147 Journal of Botany 91: 1364–1370. 1148 1149 Vico, G., S. Manzoni, L. Nkurunziza, K. Murphy, and M. Weih. 2016. Trade-offs between seed output 1150 and life span - a quantitative comparison of traits between annual and perennial congeneric species. New 1151 Phytologist 209: 104–114. 1152 1153 Vilela, A.E., L. González-Paleo, D.A. Ravetta, E.G. Murrell, and D.L.
Van Tassel. 2020. Balancing 1154 forage production, seed yield, and pest management in the perennial sunflower Silphium integrifolium 1155 (Asteraceae). Agronomy 10: 1471. 1156 1157 Vu, V.Q. 2011. ggbiplot: A ggplot2 based biplot. R package version 0.55. http://github.com/vqv/ggbiplot 1158 1159 Ward, P.R., J.A. Palta, and H.A. Waddell. 2011. Root and shoot growth by seedlings of annual and 1160 perennial medic, and annual and perennial wheat. Crop and Pasture Science 62: 367–373. 1161 1162 Wei, T., and V. Simko. 2017. R package "corrplot": Visualization of a correlation matrix. R package 1163 version 0.84. https://CRAN.R-project.org/package=corrplot 1164 bioRxiv preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint available under a CC-BY-NC-ND 4.0 International license . 46 1165 Wickham, H. 2016. ggplot2: Elegant graphics for data analysis. Springer-Verlag, New York, New York, 1166 USA. https://ggplot2.tidyverse.org 1167 1168 Wright, I.J., P.B. Reich, M. Westoby, D.D. Ackerly, Z. Baruch, F. Bongers, J. Cavender-Bares, et al. 1169 2004. The worldwide leaf economics spectrum. Nature 428: 821–827. 1170 1171 Wu, Z.Y., P.H. Raven, and D.Y. Hong [eds. ], 2010. Flora of China, vol. 10 (Fabaceae). Science Press, 1172 Beijing, China, and Missouri Botanical Garden Press, St. Louis, Missouri, USA. Website 1173 http://www.efloras.org/browse.aspx?volume_id=2010 [accessed 10 April 2020] 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 Table 1. Sampling summary of seed, germination, and vegetative traits in annual and perennial species of Lathyrus, Phaseolus, and Vicia. Seed size Germination Genus Life span Species a Accessions Seeds Accessions Seeds Lathyrus Annual L. annuus 1 89 1 60 L. aphaca 3 82 (62-99) 2 60 (59-60) L. cicera 3 69 (61-77) 3 57 (52-60) L. hirsutus 4 74 (40-93) 4 50 (37-57) L. odoratus 1 61 1 51 Perennial L. japonicus 3 94 (49-151) 3 54 (44-60) L. latifolius 4 71 (62-80) 4 59 (56-60) L. pratensis 3 155 (121-174) 3 60 (60) L. sylvestris 4 80 (63-100) 4 56 (45-60) L. tuberosus 1 101 1 56 Phaseolus Annual P. acutifolius 6 57 (46-68) 6 52 (42-61) P. filiformis 6 63 (53-74) 6 53 (43-60) P. lunatus 2 29 (25-33) 2 27 (24-29) P. vulgaris 7 35 (25-52) 7 33 (23-49) Perennial P. angustissimus 2 58 (57-59) 2 54 (53-54) Vegetative growth Accessions Plants 1 12 2 13 (11-15) 3 11 (9-12) 4 10 (4-13) 1 12 3 15 (8-19) 4 20 (18-22) 3 18 (18-19) 4 16 (12-19) 1 8 6 13 (9-15) 6 15 (10-20) 2 14 (13-15) 7 10 (7-13) 2 25 (21-29) a v a i l a b e l u n d e r a C C B Y N C N D 4 . 0 I n t e r n a t i o n a l l i c e n s e . ( w h c h w a s i n o t c e r t i f i e d b y p e e r r e v e w i ) i s t h e a u t h o r / f u n d e r , w h o h a s g r a n t e d b o R x v i i a l i c e n s e t o d s p a y i l t h e p r e p r i n t i n p e r p e t u i t y .
I t i s m a d e b o R x v p r e p r i n t d o i : i i h t t p s : / / d o i . o r g / 1 0 . 1 1 0 1 / 2 0 2 1 . 0 2 . 1 7 . 4 3 1 6 5 6 ; t h s i v e r s o n i p o s t e d F e b r u a r y 1 7 , 2 0 2 1 . T h e c o p y r i g h t h o d e r l f o r t h s i p r e p r i n t P. leptostachyus 4 52 (39-68) 4 46 (34-60) 4 P. maculatus 1 37 1 35 1 P. parvulus 1 41 1 38 1 P. polystachios 2 30 (26-34) 2 28 (25-30) 2 Vicia Annual V. benghalensis 1 63 1 57 1 V. ervilia 3 59 (52-64) 3 53 (46-58) 3 V. hirsuta 2 128 (118-138) 2 57 (55-58) 1 V. sativa 3 69 (61-75) 3 59 (57-60) 3 V. villosa 4 64 (53-83) 4 54 (48-60) 4 Perennial V. americana 1 124 1 58 1 V. cracca 4 88 (67-117) 4 59 (56-60) 4 V. dumetorum 1 59 --- --- 1 V. sepium 1 90 1 60 1 V. tenuifolia 2 81 (77-85) 2 60 (60) 2 Notes: “Seeds” and “plants” respectively refer to the average seed number and plant number among accessions studied for each species, with the range of seed/plant number among accessions in parentheses, where applicable. The maximum sampling is shown (after removal of any problematic data and outliers), where seed size collectively includes seed mass and ImageJ seed measurements, germination includes germination proportion and T50, and 19 (17-20) 16 22 21 (19-22) 11 14 (13-15) 10 13 (11-15) 12 (8-14) 31 19 (17-21) 13 20 32 (26-38) a v a i l a b e l u n d e r a C C B Y N C N D 4 . 0 I n t e r n a t i o n a l l i c e n s e . ( w h c h w a s i n o t c e r t i f i e d b y p e e r r e v e w i ) i s t h e a u t h o r / f u n d e r , w h o h a s g r a n t e d b o R x v i i a l i c e n s e t o d s p a y i l t h e p r e p r i n t i n p e r p e t u i t y . I t i s m a d e b o R x v p r e p r i n t d o i : i i h t t p s : / / d o i . o r g / 1 0 . 1 1 0 1 / 2 0 2 1 . 0 2 . 1 7 . 4 3 1 6 5 6 ; t h s i v e r s o n i p o s t e d F e b r u a r y 1 7 , 2 0 2 1 . T h e c o p y r i g h t h o d e r l f o r t h s i p r e p r i n t vegetative growth includes height and/or leaf number measurements on a plant at one or both dates (DAP-21 and DAP-35). A trait-by-trait summary of accession-level sampling can be found in Appendix S1. a Some species’ accessions include subspecific rankings. These include: Lathyrus japonicus subsp. maritimus (1 accession), Phaseolus acutifolius var. acutifolius (3), P. acutifolius var. tenuifolius (3), P. maculatus subsp. ritensis (1), P. polystachios subsp. polystachios (1), P. polystachios subsp. sinuatus (1), P. vulgaris var. aborigineus (all accessions), Vicia tenuifolia subsp. tenuifolia (1), and V. tenuifolia subsp. dalmatica (1). a v a i l a b e l u n d e r a C C B Y N C N D 4 . 0 I n t e r n a t i o n a l l i c e n s e . ( w h c h w a s i n o t c e r t i f i e d b y p e e r r e v e w i ) i s t h e a u t h o r / f u n d e r , w h o h a s g r a n t e d b o R x v i i a l i c e n s e t o d s p a y i l t h e p r e p r i n t i n p e r p e t u i t y . I t i s m a d e b o R x v p r e p r i n t d o i : i i h t t p s : / / d o i . o r g / 1 0 . 1 1 0 1 / 2 0 2 1 . 0 2 . 1 7 . 4 3 1 6 5 6 ; t h s i v e r s o n i p o s t e d F e b r u a r y 1 7 , 2 0 2 1 .
T h e c o p y r i g h t h o d e r l f o r t h s i p r e p r i n t Table 2. Analysis of variance (ANOVA) table of the final, reduced linear mixed models for the full dataset, including accession-level principal components (Figure 1) and seed, germination, and vegetative growth traits. Trait Genus Life span Genus × Life span Species Accession (a) PC1 F2, 47 = 58.61* F1, 47 = 24.89* F2, 47 = 5.33 F21, 47 = 4.33* NA a PC2 F2, 47 = 24.39* F1, 47 = 0.02 F2, 47 = 0.36 F21, 47 = 5.98* NA a (b) Seed mass F2, 49 = 2.04 F1, 49 = 3.49 F2, 49 = 0.95 F23, 49 = 5.38* NA a (c) Seed length F2, 50.8 = 21.80* F1, 50.7 = 0.71 F2, 50.8 = 0.41 F23, 50.7 = 7.84* LRT = 4708.51* Seed width F2, 50.8 = 15.01* F1, 50.7 = 0.27 F2, 50.8 = 1.81 F23, 50.8 = 10.24* LRT = 4029.91* Seed perimeter F2, 50.8 = 21.00* F1, 50.7 = 0.51 F2, 50.8 = 0.19 F23, 50.8 = 8.92* LRT = 4737.46* Seed area F2, 50.8 = 19.09* F1, 50.7 = 0.41 F2, 50.8 = 0.14 F23, 50.7 = 6.93* LRT = 5113.95* Seed circularity F2, 49.9 = 53.58* F1, 49.2 = 1.07 F2, 49.9 = 6.30 F23, 49.5 = 4.97* LRT = 538.28* Seed roundness F2, 50.3 = 27.43* F1, 49.7 = 0.12 F2, 50.3 = 19.08* F23, 49.9 = 4.78* LRT = 674.03* (d) Germination T50 F2, 44.5 = 9.07* F1, 44.5 = 7.22* F2, 44.5 = 2.35 F20, 44.9 = 1.35 LRT = 81.64* (e) Germination proportion F2, 50 = 21.62* F1, 50 = 1.44 F2, 50 = 1.62 F22, 50 = 0.87 NA a (f) Height DAP-21 F2, 47.3 = 1.56 F1, 47.6 = 9.05 F2, 47.3 = 1.63 F22, 47.1 = 0.85 LRT = 222.99* Leaf number DAP-21 F2, 53.0 = 20.83* F1, 60.5 = 22.03* F2, 52.9 = 2.70 F23, 46.9 = 6.06* LRT = 69.38* (g) Height DAP-35 F2, 49.0 = 1.07 F1, 49.5 = 12.74*** F2, 48.9 = 1.73 F23, 48.1 = 1.96* LRT = 158.00* Leaf number DAP-35 F2, 52.0 = 17.16* F1, 54.1 = 13.19* F2, 51.6 = 3.00 F23, 48.5 = 9.37* LRT = 41.05* a v a i l a b e l u n d e r a C C B Y N C N D 4 . 0 I n t e r n a t i o n a l l i c e n s e . ( w h c h w a s i n o t c e r t i f i e d b y p e e r r e v e w i ) i s t h e a u t h o r / f u n d e r , w h o h a s g r a n t e d b o R x v i i a l i c e n s e t o d s p a y i l t h e p r e p r i n t i n p e r p e t u i t y . I t i s m a d e b o R x v p r e p r i n t d o i : i i h t t p s : / / d o i . o r g / 1 0 . 1 1 0 1 / 2 0 2 1 . 0 2 . 1 7 . 4 3 1 6 5 6 ; t h s i v e r s o n i p o s t e d F e b r u a r y 1 7 , 2 0 2 1 . T h e c o p y r i g h t h o d e r l f o r t h s i p r e p r i n t (h) Height AGR F2, 48.1 = 0.26 F1, 48.8 = 1.53 F2, 47.9 = 0.58 F22, 46.8 = 2.51 Height RGR F2, 45.5 = 0.33 F1, 46.9 = 0.17 F2, 45.4 = 1.10 F22, 44.0 = 2.70 Leaf number AGR F2, 49.5 = 1.04 F1, 51.1 = 0.06 F2, 49.1 = 0.23 F22, 46.5 = 3.00* Leaf number RGR F2, 49.3 = 1.31 F1, 52.4 = 3.05 F2, 49.3 = 1.21 F22, 47.1 = 1.64 P < 0.05; P < 0.01; *P < 0.001. Notes: Letters denote separate models with different random effects, while the main effects are the same for all traits. The accession effect and other significant random effects from the reduced model were included; the accession effect is represented by the likelihood ratio test statistic (LRT).
ANOVAs are all type III with the exception of PC1, PC2, seed mass, and germination proportion (type I), due to the data consisting of only accession-level means with no significant random effects. Additional random effect significance is listed in Appendix S9. Significant values are bolded (at least P < 0.05). When non-whole numbers, denominator degrees of freedom were rounded to the first decimal. a Accession could not be used as a random effect in these trait models due to the dataset consisting of accession-level means. LRT = 103.65* LRT = 54.46* LRT = 43.24* LRT = 30.27* a v a i l a b e l u n d e r a C C B Y N C N D 4 . 0 I n t e r n a t i o n a l l i c e n s e . ( w h c h w a s i n o t c e r t i f i e d b y p e e r r e v e w i ) i s t h e a u t h o r / f u n d e r , w h o h a s g r a n t e d b o R x v i i a l i c e n s e t o d s p a y i l t h e p r e p r i n t i n p e r p e t u i t y . I t i s m a d e b o R x v p r e p r i n t d o i : i i h t t p s : / / d o i . o r g / 1 0 . 1 1 0 1 / 2 0 2 1 . 0 2 . 1 7 . 4 3 1 6 5 6 ; t h s i v e r s o n i p o s t e d F e b r u a r y 1 7 , 2 0 2 1 . T h e c o p y r i g h t h o d e r l f o r t h s i p r e p r i n t Figure 1. Principal component analysis for the full dataset accession means. (A) shows the relative contribution of each variable on the first two principal components along with the correlation circle; distance of the arrow from the origin indicates increasing representation of that trait in the PCA in a particular region of PC space. Label abbreviations: S signifies seed (perim. is perimeter), G signifies germination, H signifies height, and L signifies leaf. AGR is absolute growth rate, and RGR is relative growth rate. Variable labels were sometimes adjusted slightly from arrow tips to allow complete visualization. (B) shows the individual accession data points in the same PC space colored by genus and (C) the individual accession data points colored by life span. Figure 2. Correlation networks for (A) the full dataset, and the dataset subgroups: (B) annuals, (C) perennials, (D) Lathyrus, (E) Phaseolus, and (F) Vicia. Presence of lines (edges) between trait nodes indicates a significant correlation between those traits (Pearson; P < 0.05). Blue signifies a positive correlation and red a negative correlation; line thickness corresponds to the strength of the correlation. Node color signifies degree (the number of significant trait connections to that node), which ranges from yellow (low) to red (high); note that color is relative to the maximum number of connections for that subgroup and so is not directly comparable across subgroups. a v a i l a b e l u n d e r a C C B Y N C N D 4 . 0 I n t e r n a t i o n a l l i c e n s e . ( w h c h w a s i n o t c e r t i f i e d b y p e e r r e v e w i ) i s t h e a u t h o r / f u n d e r , w h o h a s g r a n t e d b o R x v i i a l i c e n s e t o d s p a y i l t h e p r e p r i n t i n p e r p e t u i t y . I t i s m a d e b o R x v p r e p r i n t d o i : i i h t t p s : / / d o i .
o r g / 1 0 . 1 1 0 1 / 2 0 2 1 . 0 2 . 1 7 . 4 3 1 6 5 6 ; t h s i v e r s o n i p o s t e d F e b r u a r y 1 7 , 2 0 2 1 . T h e c o p y r i g h t h o d e r l f o r t h s i p r e p r i n t bioRxiv preprint doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license . 5.0 5.0 S massS lengthS widthS per.S areaS circularityS roundnessG T50G proportionH DAP−35H AGRH RGRL DAP−21L DAP−35L AGRL RGR H DAP−21 A B C 2 2.5 2.5 1 ) ) ) % 7 4 2 ( 2 C P % 7 4 2 ( 2 C P % 7 4 2 ( 2 C P 0.0 0.0 . . . 0 −2.5 −2.5 −1 −5.0 −5.0 −2 −2 −1 0 PC1 (42.2%) 1 2 −5.0 −2.5 0.0 2.5 5.0 7.5 −5.0 −2.5 0.0 2.5 5.0 7.5 PC1 (42.2%) PC1 (42.2%) Lifespan Annual Perennial Genus Vicia Figure 1. Principal component analysis for the full dataset accession means. (A) shows the relative contribution of each variable on the first two principal components along with the correlation circle; distance of the arrow from the origin indicates increasing representation of that trait in the PCA in a particular region of PC space. Label abbreviations: S signifies seed (perim. is perimeter), G signifies germination, H signifies height, and L signifies leaf. AGR is absolute growth rate, and RGR is relative growth rate. Variable labels were sometimes adjusted slightly from arrow tips to allow complete visualization. (B) shows the individual accession data points in the same PC space colored by genus and (C) the individual accession data points colored by life span. Lathyrus Phaseolus bioRxiv preprint doi: https://doi.org/10.1101/2021.02.17.431656 ; this version posted February 17, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license . SeedmassSeedlengthSeedwidthSeedperim.SeedareaSeedcirc.Seedround.Germ.T50Germ.prop.HeightDAP−21HeightDAP−35HeightAGRHeightRGRLeaf NDAP−21Leaf NDAP−35Leaf NAGRLeaf NRGRE Phaseolus SeedmassSeedlengthSeedwidthSeedperim.SeedareaSeedcirc.Seedround.Germ.T50Germ.prop.HeightDAP−21HeightDAP−35HeightAGRHeightRGRLeaf NDAP−21Leaf NDAP−35Leaf NAGRLeaf NRGRA Full dataset SeedmassSeedlengthSeedwidthSeedperim.SeedareaSeedcirc.Seedround.Germ.T50Germ.prop.HeightDAP−21HeightDAP−35HeightAGRHeightRGRLeaf NDAP−21Leaf NDAP−35Leaf NAGRLeaf NRGRF Vicia SeedmassSeedlengthSeedwidthSeedperim.SeedareaSeedcirc.Seedround.Germ.T50Germ.prop.HeightDAP−21HeightDAP−35HeightAGRHeightRGRLeaf NDAP−21Leaf NDAP−35Leaf NAGRLeaf NRGRB Annual SeedmassSeedlengthSeedwidthSeedperim.SeedareaSeedcirc.Seedround.Germ.T50Germ.prop.HeightDAP−21HeightDAP−35HeightAGRHeightRGRLeaf NDAP−21Leaf NDAP−35Leaf NAGRLeaf NRGRD Lathyrus SeedmassSeedlengthSeedwidthSeedperim.SeedareaSeedcirc.Seedround.Germ.T50Germ.prop.HeightDAP−21HeightDAP−35HeightAGRHeightRGRLeaf NDAP−21Leaf NDAP−35Leaf NAGRLeaf NRGRC Perennial Figure 2.
Correlation networks for (A) the full dataset, and the dataset subgroups: (B) annuals, (C) perennials, (D) Lathyrus, (E) Phaseolus, and (F) Vicia. Presence of lines (edges) between trait nodes indicates a significant correlation between those traits (Pearson; P < 0.05). Blue signifies a positive correlation and red a negative correlation; line thickness corresponds to the strength of the correlation. Node color signifies degree (the number of significant trait connections to that node), which ranges from yellow (low) to red (high); note that color is relative to the maximum number of connections for that subgroup and so is not directly comparable across subgroups.
bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Evolving a generalist biosensor for bicyclic monoterpenes Simon d’Oelsnitz* 1, Vylan Nguyen 2, Hal S. Alper 3, Andrew D. Ellington* 1 Aﬃliations 1Department of Molecular Biosciences, University of Texas at Austin, Austin, TX, 78712, USA 2Freshman Research Initiative, University of Texas at Austin, Austin, TX, 78712, USA 3McKetta Department of Chemical Engineering, University of Texas at Austin, Austin, TX, 78712, USA To whom correspondence should be addressed: Simon d’Oelsnitz, [email protected] Andrew D. Ellington, [email protected] bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. ABSTRACT Prokaryotic transcription factors can be repurposed as analytical and synthetic tools for precise chemical measurement and regulation. Monoterpenes encompass a broad chemical family that are commercially valuable as ﬂavors, cosmetics, and fragrances, but have proven diﬃcult to measure, especially in cells. Herein, we develop genetically-encoded, generalist monoterpene biosensors by using directed evolution to expand the eﬀector speciﬁcity of the camphor-responsive TetR-family regulator CamR from Pseudomonas putida. Using a novel negative selection coupled with a high-throughput positive screen (Seamless Enrichment of Ligand-Inducible Sensors, SELIS), we evolve CamR biosensors that can recognize four distinct monoterpenes: borneol, fenchol, eucalyptol, and camphene. Diﬀerent evolutionary trajectories surprisingly yielded common mutations, emphasizing the utility of CamR as a platform for creating generalist biosensors. Systematic promoter optimization driving the reporter increased the system’s signal-to-noise ratio to 150-fold. These sensors can serve as a starting point for the high-throughput screening and dynamic regulation of bicyclic monoterpene production strains. Keywords: Biosensors, Protein Engineering, Directed Evolution, Terpenes, Metabolic Engineering 1 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. INTRODUCTION The rapid and facile chemical measurement of biologically produced compounds is crucial for the $14.4B high-throughput screening industry, especially when discrimination is needed for enantiomers or constitutional isomers1. In particular, there remains a pressing need for the development of high-throughput analytical methods for terpenes used commercially in the ﬂavor, fragrance, cosmetic, and pharmaceutical industries.
Although monoterpenes are traditionally sourced from plants, microbes are being increasingly engineered to provide a more reliable, pure and space-eﬃcient means of production2, typically via heterologous expression of a pathway for the production of the common intermediate geranyl pyrophosphate (GPP) followed by a speciﬁc terpene synthase. Since most monoterpenes lack a chromophore and terpene synthases often nonspeciﬁcally produce a range of terpenes with the same exact mass, metabolic screening is typically limited to low-throughput gas chromatography-mass spectrometry. Prokaryotic transcriptional factors can be repurposed as genetic biosensors for high-throughput chemical measurement and metabolic regulation, thereby leveraging speciﬁc protein-ligand interactions to selectively transduce chemical abundance into an easily quantiﬁable ﬂuorescent output3 (Figure 1a). Genetic biosensors for monoterpenes would circumvent the aforementioned analysis challenges and enable high-throughput screening for improved strain variants. However, genetic biosensors are not widely used, largely because the range of natural compounds that can be sensed is far smaller than the repertoire of chemicals and metabolites that see industrial applications. Bioinformatic mining conjoined with high-throughput experimentation has accelerated the pace of natural biosensor discovery4, but only a few natural transcriptional regulators responsive to monoterpenes5,6,7 or derivatives thereof8 have been identiﬁed, and to date no genetic biosensors have been engineered or thoroughly characterized for monoterpene analysis. 2 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. A potential key enabler is the use of directed evolution to reconﬁgure the binding speciﬁcity of natural biosensors for new ligands9,10,11. Even so, directed evolution methods can be time-consuming, and often yield modest changes to eﬀector responsiveness. We have now addressed the discovery gap by using a novel high-throughput selection and screening method (Seamless Enrichment of Ligand-Inducible Sensors, SELIS) to produce generalist biosensors that can be used across many diﬀerent bicyclic monoterpenes. A camphor-responsive Pcamr/CamR system from Pseudomonas putida was repurposed to function in Escherichia coli and subsequent promoter optimization increased the system’s signal-to-noise ratio to 150-fold. SELIS was then used to evolve CamR towards four non-cognate monoterpenes, yielding rapid convergence on a series of broadly useful generalist biosensors. RESULTS Engineering a camphor-responsive biosensor and circuit To identify a generalist monoterpene biosensor, we started from the camphor responsive CamR repressor from P. putida. This transcription factor has been thoroughly characterized in vitro [9], and camphor is structurally similar to other monoterpenes that have been produced microbially, including borneol12, fenchol13, eucalyptol14, and camphene13.
The Pcamr promoter was extracted from the natural P. putida CamR-bearing plasmid (GenBank: D14680.1), and included a portion protected by CamR in a DNase footprinting assay15 upstream from the self-cleaving ribozyme PlmJ16. To test the function of the heterologous Pcamr/CamR system in E. coli, CamR was constitutively expressed on one plasmid (pCamR) in the presence of a co-transformed plasmid bearing the Pcamr promoter upstream of the mScarlet-i gene17 (pPcamr-RFP). The native Pcamr promoter proved to be not very active in E. coli (Figures 1b, 1c). We hypothesized that the -10 region of the promoter was divergent from the E. coli consensus 11, and introduced a single G→A base substitution in the -10 region, which greatly improved activity and the EC50 value (Figures 1b, 1c). Interestingly, while past attempts to repurpose the Pcamr/CamR system for camphor-inducible gene expression in E. coli have failed18, possibly due to an inhibitory eﬀect of the 3 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. native 5’-noncoding region of the native CamR mRNA transcript and insuﬃcient transcription and translation strength, our largely synthetic design produced a reliable and robust dose-dependent signal. The inherent response of CamR to diﬀerent monoterpenes was analyzed. As expected, CamR was found to be highly responsive to its canonical eﬀector, camphor, displaying an EC50 of 36.3 μM and a ~40-fold signal-to-noise ratio (Figure 1c). Responses were also observed to a variety of bicyclic monoterpenes that contained alcohol, ketone, or ether functional groups (such as fenchol, fenchone, and eucalyptol), but not to other bicyclic, monocyclic, or acyclic monoterpenes (such as camphene, limonene, or linalool; Figure 1d). We further explored the speciﬁcities of six diﬀerent CamR homologs from various Pseudomonas species with the aim of identifying the best biosensor for subsequent evolution. The pCamR plasmid was accordingly modiﬁed to replace CamR with one of these six diﬀerent homologs and was then co-transformed with the Pcamr-RFP plasmid prior for monoterpene screening, as described above. This screening eﬀort revealed two other homologs that had a ligand proﬁle similar to the P. putida CamR (mybCamR and thl2CamR), albeit with lower induction, and one homolog (tcuCamR) that was somewhat more speciﬁc for borneol and camphor (Supplementary Figure 1). Three homologs divergent from the P. putida CamR (CamV, my2CamR, bazCamR) generally displayed a poor response to all of the tested monoterpenes, suggesting that they might bind to an alternative operator sequence or have evolved to bind to other, currently unknown eﬀector molecules. Optimization of reporter circuit performance Given that the P. putida CamR functioned as a semi-speciﬁc biosensor, and that no other candidates appeared to be better generalists, we continued towards circuit optimization with this biosensor.
To improve the performance of the circuit for resolving diﬀerences between and improvements with new, relatively inactive eﬀector molecules, the dynamic range of the reporter system was systematically 4 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. optimized via three separate components: the promoter and RBS adjacent to the reporter RFP, and the CamR operator sequence (Supplementary Table 1). Since strong promoters can produce a high background signal, while weak promoters result in a low induced signal (as was previously the case for the unmodiﬁed Pcamr promoter), ﬁne-tuning promoter strength can signiﬁcantly improve the dynamic range of biosensor systems19. Eight synthetic promoters and ﬁve synthetic RBSs that each spanned a range of expression strengths about three orders of magnitude were introduced upstream of RFP. Multiple operator sequences were also considered during the optimization process, including the native operator (WT) and the inverted native operator (INV). In addition, since higher operator symmetry has previously been demonstrated to reduce background signal, possibly by increasing the strength of the regulator-DNA interaction20, two operator variants with greater symmetry than the native operator -- one located upstream of a CamR homolog (WP_145928353.1) in Pseudomonas sp. TCU-HL1 (V2), and the other being our own synthetic design (V3) -- were introduced. The medium-high strength P500 promoter produced the highest signal-to-noise ratio when cells were induced with 1 mM of camphor (Figure 2b), the synthetic operator sequence with the greatest symmetry (V3) surpassed all other tested operator sequences (largely by reducing background signal; Figure 2c), and the two strongest RBS sequences outperformed weaker counterparts (Figure 2d). Overall, after systematic optimization of the promoter, RBS, and operator, the best performing reporter circuit produced a 150-fold signal-to-noise ratio, ~3.7-fold higher than the original construct (Figure 2e). The fact that a largely designed and synthetic circuit was functional in a heterologous system (E. coli) and outperformed the native system, primarily by reducing the background signal, reinforces the notion that synthetic parts may generally have greater orthogonality to the system as a whole, and hence greater predictability21. Evolution of CamR to encompass additional monoterpenes While CamR was responsive to a number of bicyclic monoterpenes, further expanding its eﬀector speciﬁcity would be greatly enabling for serving as a unitary biosensor for monitoring the metabolic 5 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder.
All rights reserved. No reuse allowed without permission. engineering of strains to produce terpene products. We therefore leveraged a previously developed directed evolution approach, Seamless Enrichment of Ligand-Inducible Sensors (SELIS22, (Supplementary Figure 2), to improve the responsivity of CamR towards the terpene synthase products borneol, fenchol, eucalyptol, and camphene, all of which have been biosynthesized in E. coli previously12,13,14. SELIS relies on a positive screen for ligand-dependent regulator function via GFP production, and a conjoined negative selection in which transcriptional repression in the absence of ligand yields antibiotic resistance. Between the positive screen and negative selection, SELIS can quickly deconvolute libraries with over 105 members in under a week and provides substantial genotype and phenotype data that allows optimally performing biosensors to be chosen by the researcher from amongst a range of successful candidates. CamR was introduced into the SELIS pipeline by placing the previously identiﬁed Pcamr promoter construct upstream of a sfGFP gene for the positive screen and having a separate Pcamr promoter drive the expression of the Lambda cI repressor for the negative selection (Supplementary Figure 2). Initial CamR libraries were generated by site-saturating four sets of three, predicted, ligand-proximal residues, resulting in 32,000 unique protein mutants. When these largely failed to yield CamR variants responsive to the target monoterpenes, subsequent libraries were generated via error-prone mutagenesis of the entire CamR coding sequence, introducing an average of two mutations per gene. E. coli co-transformed with CamR libraries and the pSELIScamr plasmid were grown in the presence of zeocin (negative selection) and subsequently plated on solid media containing 1.0 % DMSO and either borneol, fenchol, eucalyptol, or camphene (positive screen; Figure 3a). Highly ﬂuorescent clones were isolated, grown in the presence and absence of the target monoterpene, and ten clones with the high signal-to-noise ratios for each diﬀerent eﬀector were sequenced, a total of 40 CamR variants (Supplementary Table 2). It should be noted that since the ﬂuorescence of clonal isolates is measured both with and without the eﬀector during screening virtually all false positives are eliminated. 6 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. A large fraction of the 40 improved variants shared common substitutions. A Y58F substitution occurred in 25/40 of the recovered clones, while an A139T substitution appeared in 10/40 of the clones. Other shared substitutions, such as M111L, A83T, and G67V, were isolated in multiple screens for diﬀerent monoterpenes (Supplementary Table 1). Given that a majority (22/40) of the clones had only single substitutions, and that 34/40 clones had either the A139T or Y58F substitution, we hypothesized that we could best understand the major contributions to responsivity and speciﬁcity on a substitution-by-substitution basis.
We therefore cloned all 30 unique single amino acid substitutions that were recovered from CamR evolution and screened them for activity against camphor, borneol, fenchol, eucalyptol, and camphene (Figure 3b, Supplementary Figure 3). As might have been expected based on the selection data alone, the three substitutions Y58F, H102N, and A139T independently were found to increase the response to at least one terpene by 6- to 7.5-fold. Ten other substitutions increased responsivity to at least one terpene by 2- to 3-fold, ﬁve increased the response by 20-50%, and 12 variants were found to not signiﬁcantly increase the response to any terpenes. Interestingly, virtually every substitution that increased the response of CamR for one monoterpene also increased its response for all other tested monoterpenes to varying degrees. The two exceptions were Y58F and L75I, which improved responsiveness towards all terpenes except camphor. To further understand the potential contributions of the individual amino acid changes, we mapped all 11 of the most productive substitutions onto a homology model of CamR (Figure 3c). With the exception of Y43 and A139 -- which are located in the DNA-binding and dimerization domains, respectively -- all nine other positions were located in the ligand binding domain. The predominant Y58F (and less ﬁxed M111L) likely faces inward towards the presumed ligand binding cavity and may interact with the ligand directly (Supplementary Figure 4). The other mutated positions within the ligand binding domain are likely facing out towards the solvent, and thus their contributions towards altering terpene response are more diﬃcult to rationalize. 7 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. While we had initially searched for a broadly useful generalist, the apparent lack of any variants that showed greater speciﬁcity for any of the monoterpenes was surprising, especially given previous eﬀorts with SELIS to evolve a similar biosensor (RamR) to recognize benzylisoquinoline alkaloids. Since we and others have previously observed that the substrate- and eﬀector speciﬁcities of evolving proteins often go through generalist intermediates prior to re-specialization23,24, we carried out two additional rounds of SELIS to see if we could increase responsiveness to fenchol (Figure 4a). While these additional generations were successively more responsive to fenchol, they also displayed an increased background signal (Figure 4b; to assist with analysis, CamR variants were expressed auto-inductively using the Pcamr promoter during characterization (Supplementary Figure 5)). Interestingly, even as we attempted to force specialization on fenchol, the improved variants accumulated substitutions that had previously been observed in the CamR variants evolved for borneol, eucalyptol, and camphene, such as Y58F, H102N, and M104L (Supplementary Figure 6 and Supplementary Table 3).
Indeed, screening each ﬁnal CamR generation against camphor, borneol, fenchol, eucalyptol, and camphene revealed that the later, fenchol-targeted generations had nonetheless become more broadly activated by all tested terpenes (Figure 4c, Supplementary Figure 7). Beyond these ﬁve monoterpenes, the ﬁnal CamR generation, FEN3, was observed to respond more evenly to a wider range of bicyclic monoterpenes (Figure 4d). DISCUSSION Genetic biosensors are gaining traction as high-throughput tools for otherwise diﬃcult chemical analyses, such as distinguishing between compounds with identical exact masses and/or lacking a chromophore. However, biosensors often require custom engineering for target analytes, and even with great improvements in bioinformatic mining4 it is diﬃcult to ﬁnd corresponding biosensors. Therefore, we have developed a method (SELIS) that can be broadly used to identify biosensors, and herein we develop a genetic biosensor that is responsive to a range of bicyclic monoterpenes used in the ﬂavor, fragrance, cosmetic, and pharmaceutical industries, which are traditionally limited to analysis using low-throughput 8 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. gas chromatography-mass spectrometry. Indeed, microbes have already been engineered to produce many of the compounds recognized by our CamR sensors, such as borneol12, fenchol13, eucalyptol14, and camphene13, and our biosensors can now be used for metabolic engineering of such pathways, including identifying new monoterpene synthases25, improving strain health (since many monoterpenes are known to be toxic to microbes26), and balancing expression levels to reduce a heterologous pathway’s metabolic load27. While we wanted to make one, broad-ranging sensor for the variety of monoterpene pathways and applications that are even now being developed, we were surprised to ﬁnd that only generalists emerged from our selections, despite the application of both counterselection and additional rounds that attempted to ‘push’ a line towards fenchol speciﬁcity. These results stand in contrast to previous results with benzylisoquinoline alkaloids (BIAs), where we were able to ﬁnd highly speciﬁc sensors for ﬁve diﬀerent compounds starting from a similar TetR family member, RamR. The seeming intransigence of CamR to change, rather than broaden, eﬀector speciﬁcity could be due to several factors. First, it is likely that diﬀerent TetR family members are diﬀerentially evolvable. It is entirely possible that CamR has a narrow evolutionary landscape that is steeped in many, nearby generalist variants; this has previously been observed during the evolution of other transcriptional regulators23,24. Second, the ligand classes we have so far examined are in fact quite diﬀerent, with BIAs occupying a much larger physical space than the monoterpenes explored herein (Supplementary Figure 8).
Thus, there may just be more ‘handles’ for the eﬀector binding site to recognize and distinguish for the BIAs, than for the monoterpenes. In this regard, it is interesting to note that the closest thing to a specialist that has appeared in nature or directed evolution is camphor-mediated activation of CamR itself, possibly because of the unique carbonyl on camphor. Third, the failure to ﬁnd new specialists may be a deﬁciency in SELIS, in which the negative selection step is too stringent, and it is diﬃcult for the protein to ﬁnd an evolutionary path to a new speciﬁcity without going through intermediates that have high background in the absence of eﬀector. This 9 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. would reﬂect diﬃculties that have previously been observed in other coupled negative / positive selection schemes, such as those with tRNA synthetases28. Irrespective of the mechanism, what is somewhat remarkable is that we are able to directly probe the evolvability of CamR via the straightforward directed evolution experiments that constitute SELIS, and hence draw out these hypotheses. Into the future, the evolutionary trajectories of CamR reported herein can be compared with other transcription factors to directly examine even more detailed hypotheses. For example, it has been suggested that protein evolvability correlates with both conformational plasticity within active site residues and rigid stability in core scaﬀold residues29. It is therefore possible that the ligand binding pocket of CamR is trapped in a rigid conformation and that mutations that occurred throughout our evolution experiments introduce ﬂexibility and enable the exploration of a wider diversity of conformational states. The examination of the structures of intermediates and evolved variants can begin to address these questions. Finally, the ability to quickly probe the evolutionary landscape of a target biosensor using SELIS will likely play an increasingly important role in guiding future sensor engineering eﬀorts. Phylogenetic data combined with high-throughput experimentation can be leveraged to predict the capacity of transcription factor scaﬀolds to evolve for diﬀerent ligand chemistries, creating a palette of the most useful starting points for biosensor development. METHODS Strains, plasmids, and media Escherichia coli DH10B (New England BioLabs, Ipswich, MA, USA) was used for all routine cloning and directed evolution. All biosensor systems were characterized in E. coli DH10B. LB-Miller (LB) media (BD, Franklin Lakes, NJ, USA) was used for routine cloning, ﬂuorescence assays, directed 10 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder.
All rights reserved. No reuse allowed without permission. evolution, and orthogonality assays unless speciﬁcally noted. LB + 1.5% agar (BD, Franklin Lakes, NJ, USA) plates were used for routine cloning and directed evolution. The plasmids described in this work were constructed using Gibson assembly, golden gate assembly, and standard molecular biology techniques. Synthetic genes, obtained as gBlocks, and primers were purchased from IDT. Relevant plasmid sequences are provided in Supplementary Table 4 and are available by request from the corresponding authors. Monoterpenes Cells were induced with the following chemicals: Camphor (Tokyo Chemical Industry, CAT#: C0011); borneol (Tokyo Chemical Industry, CAT#: B0525); fenchone (Tokyo Chemical Industry, CAT#: F0164); fenchol (Alfa Aesar, CAT#: L03211); eucalyptol (Acros Organics, CAT#: 110340050); verbenol (Sigma Aldrich, CAT#: 247065-5G); camphene (Tokyo Chemical Industry, CAT#: C0009); carvone (Acros Organics, CAT#: 154590050); alpha-pinene (Acros Organics, CAT#: 131270050); limonene (Tokyo Chemical Industry, CAT#: L0132); linalool (Combi-Blocks, CAT#: QH-3254). All monoterpenes were ﬂushed with argon and stored at -20°C after opening. It should be noted that the commercially available camphene used in this study is >78% pure and contains as much as 20% tricyclene. Chemical transformation For routine transformations, strains were made competent for chemical transformation. 5 mL of an overnight culture of DH10B cells were subcultured into 500 mL of LB media and grows at 37°C, 250 r.p.m. for 3 h. Cultures were centrifuged (3,500 g, 4 °C, 10 min), and pellets were washed in 70 mL of chemical competence buﬀer (10% glycerol, 100 mM CaCl2) and centrifuged again (3,500 g, 4°C, 10 min). The resulting pellets were resuspended in 20 mL of chemical competence buﬀer. After 30 minutes on ice, cells were divided into 250 μL aliquots and ﬂash frozen in liquid nitrogen. Competent cells were stored at −80 °C until use. 11 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Promoter design The original Pcamr promoter (see supplementary ﬁgure 1) was derived from the literature [ref]. For modiﬁed Pcamr promoters, either the promoter, operator, or RBS sequence was swapped using golden gate assembly (Supplementary Table 1). The RiboJ insulator was used in all promoter designs to insulate the variable operator context and increase the overall ﬂuorescent signal 16. Promoter sequences were based on the J23100 Andersen promoter and either the upstream activating sequence, the -35 region (TTGACA), or the -10 region (TATAAT) were modiﬁed to aﬀord variable expression strengths. RBS sequences were designed using the RBS calculator 30. A terminator was placed immediately upstream from the promoter sequence to insulate the reporter output from upstream transcriptional activity (see Supplementary Table 4).
Biosensor response measurement For performing a biosensor response assay, the pCamR and pPcamr-RFP plasmids were co-transformed into DH10B cells and plated on an LB agar plate with appropriate antibiotics. Three separate colonies were picked for each transformation and were grown overnight. The following day, 20 μL of each culture was then used to inoculate 12 separate wells within a 2 mL 96-deep-well plate (Corning, Product #: P-DW-20-C-S) sealed with an AeraSeal ﬁlm (Excel Scientiﬁc, Victorville, CA, USA) containing 900 μL of LB media. After two hours of growth at 37 °C cultures were induced with 100 μL of LB media containing either 10 μL of just DMSO or the target monoterpene dissolved in 10 μL of DMSO. Cultures were grown for an additional 4 hours at 37 °C, 250 r.p.m and subsequently centrifuged (3,500 g, 4°C, 10 min). Supernatant was removed and cell pellets were resuspended in 1 mL of PBS (137 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 1.8 mM KH2PO4. pH 7.4). One hundred μL of the cell resuspension for each condition was transferred to a 96 well microtiter plate (Corning, Product #: 3904), from which the ﬂuorescence (Ex: 566 nm, Em: 596 nm) and absorbance (600 nm) was measured using the Tecan Inﬁnite M1000 plate reader. 12 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. CamR library construction Random mutagenesis libraries were created by amplifying the coding region of the CamR gene with error-prone polymerase chain reaction. To perform this mutagenesis protocol, 50 ng of the template gene was mixed with 2 uL of 2 uM forward and reverse primers, 3 uL of Taq polymerase (New England BioLabs, Ipswich, MA, USA), 1 uM of 50 mM MnCl2, 41 uL of water, and 50 uL of a custom buﬀer (300 uL 10x Taq polymerase buﬀer, 16.5 uL 1 M MgCl2, 15 uL 10 mg/mL bovine serum albumin, 6 uL 100 mM dGTP, 11 uL 100 mM dATP, 12 uL 100 mM dCTP, 40.5 uL 100 mM dTTP, and 1099 uL water). The thermal cycling procedure is as follows: 94 °C for 30 seconds, 55C for 30 seconds, 72C for one minute, repeated for 25 cycles. Libraries were cloned into the pCamR plasmid using Gibson assembly and E. coli DH10B bearing pSELIScamr was transformed with the resulting library. Transformation eﬃciency always exceeded 106 for each round of selection, indicating several fold coverage of the library. Transformed cells were grown in LB media overnight at 37°C in carbenicillin and chloramphenicol. Directed evolution and validation of CamR biosensors Twenty μL of cell culture bearing the sensor library was seeded into 5 mL of fresh LB containing appropriate antibiotics and 100 μg/mL zeocin (Thermo Fisher. CAT#: R25001) and were grown at 37°C for seven hours. Following incubation, 0.5 μL of culture was diluted into 1 mL of LB media, from which 100 μL was further diluted into 900 μL of LB media.
Three hundred μL of this mixture was then plated across three LB agar plates containing carbenicillin, chloramphenicol and the target monoterpene dissolved in DMSO. Plates were incubated overnight at 37 °C. The following day the brightest colonies were picked and grown overnight in 1 mL of LB media containing appropriate antibiotics within a 96-deep-well plate sealed with an AeraSeal ﬁlm at 37°C. A glycerol stock of cells containing pSELIScamr and pCamR bearing the parental CamR variant was also inoculated in 5 mL of LB for overnight growth. 13 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The following day, 20 μL of each culture was used to inoculate two separate wells within a new 96-deep-well plate containing 900 μL of LB media. Additionally, eight separate wells containing 1 mL of LB media were inoculated with 20 μL of the overnight culture expressing the parental CamR variant. A typical arrangement would have 44 unique clones on the top half of the plate, duplicates of those clones on the bottom half of the plate, and the right-most column occupied by cells harboring the parental CamR variant. After 2 hours of growth at 37°C the top half of the 96-well plate was induced with 100 μL of LB media containing 10 μL of DMSO whereas the bottom half of the plate was induced with 100 μL of LB media containing the target monoterpene dissolved in 10 μL of DMSO. The concentration of BIA used for induction is typically the same concentration used in the LB agar plate for screening during that particular round of evolution. Cultures were grown for an additional 4 hours at 37°C, 250 r.p.m and subsequently centrifuged (3,500 g, 4°C, 10 min). Supernatant was removed and cell pellets were resuspended in 1 mL of PBS. One hundred μL of the cell resuspension for each condition was transferred to a 96 well microtiter plate, from which the ﬂuorescence (Ex: 485 nm, Em: 509 nm) and absorbance (600 nm) was measured using the Tecan Inﬁnite M1000. Clones with the highest signal-to-noise ratio were then sequenced and subcloned into a fresh pCamR vector. For single mutation analysis, individual mutations were separately introduced within the CamR coding region of the pCamR plasmid. These plasmids were co-transformed with Pcamr-RFP, plated on solid media containing appropriate antibiotics, and the following day three individual colonies from each transformation were grown overnight. The resulting cultures were then assayed, as described in “Biosensor response measurement”, using 1 mM of each monoterpene dissolved in 10 μL of DMSO. For sensor variant dose response measurement, each CamR variant was ﬁrst subcloned into a fresh pCamR plasmid backbone. The resulting plasmids were then transformed into DH10B cells bearing pPcamr-RFP and three individual colonies from each transformation were subsequently grown overnight.
The resulting cultures were then assayed, as described in “Biosensor response measurement”, using seven 14 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. to eleven diﬀerent concentrations of the target monoterpene along a DMSO-only control. For further evolution, sensor variants that displayed a combination of a low background, a reduced EC50 for the target monoterpene, and a high signal-to-noise ratio were used as templates for the next round of evolution. Homology model The homology model of CamR was constructed using SWISS-MODEL (https://swissmodel.expasy.org/). 6AYI was used as a template for structure generation. Statistical analysis and reproducibility All data in the manuscript are displayed as mean ± s.e.m. unless speciﬁcally indicated. Bar graphs, dose response functions, and orthogonality matrices were all plotted in Python 3.6.9 using matplotlib and seaborn. Dose response curves and EC50 values were estimated by ﬁtting to the hill equation y = d + (a-d)*xb / (cb + xb) (where y = output signal, b = hill coeﬃcient, x = ligand concentration, d = background signal, a = the maximum signal, and c = the EC50), with the scipy.optimize.curve_ﬁt library in Python. Acknowledgements Funding from DARPA Soils (HR00111920019), Welch (F-1654), and AFSOR - (FA9550-14-1-0089) is acknowledged. Author contributions S.D. designed and performed all experiments. V.N. constructed and validated the promoter and RBS series. The manuscript was written by S.D. with support from A.D.E and H.A. S.D., A.D.E., and H.A. supervised all aspects of the study. Corresponding authors 15 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Correspondence to: Simon d’Oelsnitz, [email protected] and Andrew D. Ellington, [email protected] Competing ﬁnancial interests A.D.E has equity in GRO Biosciences, a company developing protein therapeutics. The other authors declare no conﬂict of interest. Supplementary information Additional experimental details, including circuit designs and biosensor genotype and phenotype data. Figures 16 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Figure 1. Design and characterization of the CamR biosensor system. (a) Schematic of the camphor-inducible biosensor circuit. (b) Sequence of the native Pcamr promoter extracted from the native Pseudomonas putida plasmid. A G→A substitution was made in the -10 box (colored red) to increase transcription strength.
Black arrows indicate the location of the inverted repeat. (c) Dose response function of the CamR biosensor system with the native promoter (tatgct) and the modiﬁed promoter (tatAct). EC50 values are listed and color coded (blue: tatAct, pink: tatgct) (d) Response of the CamR system to a panel of monoterpenes. All cultures were grown in the presence of 1% DMSO and ﬂuorescence values are the averages of three biological replicates for both c and d. 17 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Figure 2. Systematic optimization of the Pcamr-RFP reporter. (a) Schematic of separate components of the Pcamr reporter plasmid. (b) Fold induction of Pcamr promoter variants. Grey bars and green bars represent the ﬂuorescent signal of uninduced cells and cells induced with 1 mM camphor, respectively. Blue dots represent the fold induction of each variant in response to 1 mM camphor (c) Fold induction of Pcamr operator variants. Same color scheme is used as for b. (d) Fold induction of Pcamr RBS variants. Same color scheme is used as for b. (e) Fold induction of original Pcamr promoter compared to the optimal Pcamr promoter (P500, V3, RBS5) when induced with 1 mM of camphor. 18 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Figure 3. Evolution of CamR towards alternative monoterpenes produces generalists (a) Workﬂow for CamR evolution towards alternative monoterpenes. (b) Response of all recovered CamR single mutation variants to camphor, borneol, fenchol, eucalyptol, and camphene compared to the WT CamR protein. All cultures were grown in the presence of 1% DMSO and ﬂuorescence values are the averages of three biological replicates. Asterisks indicate variants that produced a four-fold or greater change in ﬂuorescent response. (c) Homology structure of CamR with single mutations recovered from evolution labelled on one dimer (left). Residues colored red correspond to mutants that increase the fold change in ﬂuorescence four-fold or more, whereas residues colored orange correspond to mutants that increase the response by two to four-fold. 19 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Figure 4. Evolution of CamR towards fenchol yields generalist monoterpene sensors (a) Workﬂow for multi-generation evolution of CamR towards fenchol. The genotypes of each variant are indicated in bold lettering. (b) Dose response function for each CamR generation with fenchol.
(c) Response of each CamR generation to 1 mM of camphor, borneol, fenchol, eucalyptol, and camphene. (d) Fluorescent response of FEN3 to a panel of monoterpenes. All cultures were grown in the presence of 1% DMSO and ﬂuorescence values are the averages of three biological replicates for panels b, c, and d. 20 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. References 1. Anuj Pathak. High-Throughput Screening: Technologies and Global Markets. BCC Research. PHM205A (2019). 2. Soares-Castro, P., Soares, F. & Santos, P. M. Current Advances in the Bacterial Toolbox for the Biotechnological Production of Monoterpene-Based Aroma Compounds. Molecules 26, 91 (2021). 3. Lin, J.-L., Wagner, J. M. & Alper, H. S. Enabling tools for high-throughput detection of metabolites: Metabolic engineering and directed evolution applications. Biotechnol. Adv. 35, 950–970 (2017). 4. Hanko, E. K. R. et al. A genome-wide approach for identiﬁcation and characterisation of metabolite-inducible systems. Nat. Commun. 11, 1213 (2020). 5. Siu, Y., Fenno, J., Lindle, J. M. & Dunlop, M. J. Design and Selection of a Synthetic Feedback Loop for Optimizing Biofuel Tolerance. ACS Synth. Biol. 7, 16–23 (2018). 6. Phoenix, P. et al. Characterization of a new solvent-responsive gene locus in Pseudomonas putida F1 and its functionalization as a versatile biosensor. Environ. Microbiol. 5, 1309–1327 (2003). 7. Fujita, M., Aramaki, H., Horiuchi, T. & Amemura, A. Transcription of the cam operon and camR genes in Pseudomonas putida PpG1. J. Bacteriol. 175, 6953–6958 (1993). 8. Eaton, R. W. p-Cymene catabolic pathway in Pseudomonas putida F1: cloning and characterization of DNA encoding conversion of p-cymene to p-cumate. J. Bacteriol. 179, 3171–3180 (1997). 9. Yao, J. et al. Developing a highly eﬃcient hydroxytyrosol whole-cell catalyst by de-bottlenecking rate-limiting steps. Nat. Commun. 11, 1515 (2020). 10. Snoek, T. et al. Evolution-guided engineering of small-molecule biosensors. Nucleic Acids Res. 48, e3–e3 (2020). 11. Xiong, D. et al. Improving key enzyme activity in phenylpropanoid pathway with a designed biosensor. Metab. Eng. 40, 115–123 (2017). 12. Lei, D. et al. Combining Metabolic and Monoterpene Synthase Engineering for de Novo Production of Monoterpene Alcohols in Escherichia coli. ACS Synth. Biol. 10, 1531–1544 (2021). 21 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 13. Leferink, N. G. H. et al. A ‘Plug and Play’ Platform for the Production of Diverse Monoterpene Hydrocarbon Scaﬀolds in Escherichia coli. ChemistrySelect 1, 1893–1896 (2016). 14. Mendez-Perez, D. et al.
Production of jet fuel precursor monoterpenoids from engineered Escherichia coli. Biotechnol. Bioeng. 114, 1703–1712 (2017). 15. Aramaki, H., Sagara, Y., Hosoi, M. & Horiuchi, T. Evidence for autoregulation of camR, which encodes a repressor for the cytochrome P-450cam hydroxylase operon on the Pseudomonas putida CAM plasmid. J. Bacteriol. 175, 7828–7833 (1993). 16. Lou, C., Stanton, B., Chen, Y.-J., Munsky, B. & Voigt, C. A. Ribozyme-based insulator parts buﬀer synthetic circuits from genetic context. Nat. Biotechnol. 30, 1137–1142 (2012). 17. Bindels, D. S. et al. mScarlet: a bright monomeric red ﬂuorescent protein for cellular imaging. Nat. Methods 14, 53–56 (2017). 18. Aramaki, H., Fujita, M., Sagara, Y., Amemura, A. & Horiuchi, T. Heterologous expression of the cytochrome P450cam hydroxylase operon and the repressor gene of Pseudomonas putida in Escherichia coli. FEMS Microbiol. Lett. 123, 49–54 (1994). 19. Chen, Y. et al. Tuning the dynamic range of bacterial promoters regulated by ligand-inducible transcription factors. Nat. Commun. 9, 64 (2018). 20. Ruegg, T. L. et al. Jungle Express is a versatile repressor system for tight transcriptional control. Nat. Commun. 9, 3617 (2018). 21. Temme, K., Zhao, D. & Voigt, C. A. Refactoring the nitrogen ﬁxation gene cluster from Klebsiella oxytoca. Proc. Natl. Acad. Sci. 109, 7085–7090 (2012). 22. d’Oelsnitz, S. et al. Using structurally fungible biosensors to evolve improved alkaloid biosyntheses. (2021) doi:10.1101/2021.06.07.447399. 23. Collins, C. H., Arnold, F. H. & Leadbetter, J. R. Directed evolution of Vibrio ﬁscheri LuxR for increased sensitivity to a broad spectrum of acyl-homoserine lactones. Mol. Microbiol. 55, 712–723 (2005). 22 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.20.457167 ; this version posted August 21, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. 24. Taylor, N. D. et al. Engineering an allosteric transcription factor to respond to new ligands. Nat. Methods 13, 177–183 (2016). 25. Leferink, N. G. H. et al. An automated pipeline for the screening of diverse monoterpene synthase libraries. Sci. Rep. 9, 11936 (2019). 26. Zhang, L. et al. Chassis and key enzymes engineering for monoterpenes production. Biotechnol. Adv. 35, 1022–1031 (2017). 27. Hartline, C. J., Schmitz, A. C., Han, Y. & Zhang, F. Dynamic control in metabolic engineering: Theories, tools, and applications. Metab. Eng. 63, 126–140 (2021). 28. Thyer, R. et al. Directed Evolution of an Improved Aminoacyl-tRNA Synthetase for Incorporation of L-3,4-Dihydroxyphenylalanine (L-DOPA). Angew. Chem. 133, 14937–14942 (2021). 29. Tóth-Petróczy, Á. & Tawﬁk, D. S. The robustness and innovability of protein folds. Curr. Opin. Struct. Biol. 26, 131–138 (2014). 30. Salis, H. M., Mirsky, E. A. & Voigt, C. A. Automated design of synthetic ribosome binding sites to control protein expression. Nat. Biotechnol.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Transcranial photoacoustic characterization of neurovascular physiology during early-stage photothrombotic stroke in neonatal piglets in vivo Jeeun Kanga,b,, Xiuyun Liuc,, Suyi Caoc, Steven R. Zeilerd, Ernest M. Grahame,f, Emad M. Boctora,b,, Raymond C. Koehlerc, a Laboratory for Computational Sensing and Robotics, Whiting School of Engineering, Johns Hopkins University, MD 21218, United States. b Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, MD 21287, United States. c Department of Anesthesiology and Critical Care Medicine, Johns Hopkins University School of Medicine, MD 21287, United States. d Department of Neurology, Johns Hopkins University School of Medicine, MD 21205, United States. e Division of Maternal-Fetal Medicine, Department of Gynecology-Obstetrics, Johns Hopkins University School of Medicine, Baltimore, MD 21205, United States. f Neuroscience Intensive Care Nursery Program, Johns Hopkins University School of Medicine, Baltimore, Maryland, MD 21205, United States These authors equally contributed. **Corresponding authors: Emad M. Boctor, Ph.D. [email protected] Raymond C. Koehler, Ph.D. [email protected] 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Abstract Perinatal ischemic stroke is estimated to occur in 1/2300–1/5000 live births, but early differential diagnosis from global hypoxia-ischemia is often difficult. In this study, we tested the ability of a hand-held transcranial photoacoustic (PA) imaging to non-invasively detect a focal photothrombotic stroke (PTS) within 2 hours of stroke onset in a gyrencephalic piglet brain. 17 stroke lesions of approximately 1-cm2 area were introduced randomly in anterior or posterior cortex via the light/dye PTS technique in anesthetized neonatal piglets (n = 11). The contralateral non-ischemic region served as control tissue for discrimination contrast for the PA hemoglobin metrics: HbO2 saturation, total hemoglobin (tHb), and individual quantities of oxygenated and deoxygenated hemoglobin (HbO2 and HbR). The PA-derived tissue HbO2 saturation at 2 hours yielded a significant separation between control and affected regions-of-interest (p < 0.0001), which were well matched with 24-hr post- stroke cerebral infarction confirmed in the triphenyltetrazolium chloride (TTC)-stained image. The quantity of HbO2 also displayed a significant contrast (p = 0.021), whereas tHb and HbR did not.
The analysis on receiver operating characteristic curves and multivariate data analysis also agreed with the results above. This study shows that a hand-held transcranial PA neuroimaging can detect a regional thrombotic stroke in cerebral cortex of a neonatal piglet. In particular, we conclude that the HbO2 saturation metric can be used alone to identify regional stroke lesions. The lack of change in tHb may be related to arbitrary hand-held imaging configuration and/or entrapment of red blood cells within the thrombotic stroke. Keywords: Transcranial; photoacoustic imaging; photothrombotic stroke; hemoglobin; oxygen. 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Introduction Perinatal arterial ischemic stroke is estimated to occur in 1/2300–1/5000 live births, and can result in long-term deficits in motor, cognitive, attention, and executive functions and persistent seizures [1– 4]. Despite of all the efforts, there is still no effective treatment available to improve the functional recovery after perinatal stroke, and “damage control” based on a rapid detection and following neuroprotective treatment has been the best strategy in clinical use [5,6]. The traditional strategy is mainly based on a supportive care [7,8]. Several fetal monitoring technologies have been developed to enable the prompt treatment but have great limitations. Electronic fetal heart rate monitoring (EFHM) has been a standard of care in perinatal monitoring since 1970s. However, no benefit has been shown in l cerebral palsy by using EFHM, while the cesarean delivery has increased from 5% to more than 30% since using this technology [9]. Meanwhile, several studies showed that EFHM has a critical false-positive rate, even as high as 99.8 % [10], and not sensitive for perinatal stroke detection in fetal brain [11]. Magnetic resonance imaging (MRI) is the definitive diagnostic tool for ischemic stroke after birth, but its application in the newborn is questionable due to the inaccessibility in the first day after birth [12]. As alternatives, several biophotonic modalities have been proposed to allow cost-effective and continuous monitoring of neonatal brain in this period. Near-infrared spectroscopy (NIRS) and diffused optical tomography (DOT) have shown their capability of non-invasive, continuous, and fast monitoring of cerebral oxygenation level and hemodynamic changes in the neonatal brain [13,14]. However, both methods suffer from low spatial specificity and incapability of distinguishing the arterial and venous compartments, leading to low clinical specificity. Therefore, there is a critical need for effective and prompt monitoring of the perinatal brain preferably with quantitative metrics of physiological dynamics to enable early-stage decision and immediate commencement of the treatments.
Photoacoustic (PA) sensing modality, a combination of ultrasound and optical modalities, has been highlighted recently due to its high spatiotemporal resolution and deep imaging depth, which may play an alternative role for the unmet clinical need. The technology emits the pulse laser light on intact skin, and the light penetrates into deep tissue determined by the distribution of tissue scattering coefficients. After being absorbed at localized tissue with unique absorption coefficients, the conversion of absorbed energy to instantaneous heat would generate the thermal elastic expansion, which leads to pressure which is detected by an external ultrasound transducer [15]. Several clinical applications using PA technology have been proposed [16–20], and hemodynamic imaging has been a primary clinical application with stark hemoglobin absorbance, enabling quantifications of blood oxygen saturation and cerebral blood volume [21–23]. Transcranial PA sensing of physiological change at the superior sagittal sinus (SSS) has been demonstrated in the neonatal piglet and ovine models through intact scalp [24–26], but there is still room for further development to have physiological measures in brain tissue, wherein the concentration of hemoglobin is an order of magnitude less than in the sagittal sinus vein. J. Lv, et al. also demonstrated transcranial PA imaging of structural and functional dynamics over rodent brain during ischemic stroke at a very early stage (from 5-min to 6-hour onset) [27], but the implication for clinical translation is still limited by the use of the rodents with impractical imaging scale and small signal attenuation with its very thin skull layer. Therefore, development of a preclinical stroke model that has a scale and maturity level comparable to human newborns is an important component for the evaluation of a novel fetal brain monitoring modality. We chose to study newborn piglets with a 40 g gyrencephalic brain in which neuroanatomical and neurophysiological features well resembling those in term human infants can be achieved [28–30] with similar brain growth and developmental patterns [30–32]. A variety of adult 3 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. rodent ischemic stroke models have been developed, including internal carotid arterial suture occlusion of the middle cerebral artery (MCAO), direct surgical occlusion through a craniotomy, direct application of endothelin-1 on the vessels, injection of an embolus, and photothrombosis in the light/dye model. However, because the piglet has a rete mirablis, the internal carotid arterial suture model and embolic model are not feasible [33].
On the other hand, the photothrombosis (PTS) stroke model induces ischemia in a well-defined region, with minimal surgical intervention, low mortality, and high reproducibility [34]. The PTS technique induces a cortical infarction through the photo- activation of a light-sensitive dye (e.g., Rose bengal, erythrosin B) previously delivered into the circulation, resulting in local vessel thrombosis in the areas exposed to the light [35]. This model has been used successfully in the newborn piglet [28,29]. When the circulating dye is illuminated at the effective wavelength, it generates free radicals that lead to endothelial damage, platelet activation, and thrombosis in both pial and intraparenchymal vessels within the irradiated area [34]. Herein, we present transcranial PA imaging of regional ischemic stroke in the neonatal piglet model. We specifically assumed PA imaging metrics that can be achieved in a bedside diagnostic setup using a hand-held PA imaging device in order to achieve a critical translational milestone towards its clinical application. There are three aims of this study: (1) To establish stable regional stroke models in the neonatal piglet; (2) to investigate whether ischemic features in the PA imaging, such as oxyhemoglobin (HbO2) saturation, total hemoglobin concentration (tHb), and HbO2 and HbR quantities, can contribute to an accurate localization of PTS lesions; (3) to validate whether those early features obtained in 2 hours post-stroke period in a hand-held PA imaging form factor can predict cerebral infarction at a later time point. Results Photothrombotic stroke in vivo. Figure 1 shows the experimental setup to induce 17 focal infarctions in 11 piglets (weight 1.35 kg – 1.80 kg). In detail, 5 piglets received one PTS lesion for the controlled characteristics of PA imaging PTS lesions, while the other 6 piglets received two lesions to mimic the unpredictable nature of stroke induction more accurately. For the latter group, the lesions were randomly selected among pre-determined four candidate positions at frontal and posterior cortex regions (Figure 1c). The single-lesion group only considered the posterior cortex regions. We anticipated no difference in the generation of PTS lesions depending on the cortical position, following the mechanism of PTS protocol generating thrombosis in direct region of light illumination. The cortical infarction width (lateral) and length (sagittal) was identified by contouring the white surface area in TTC-stained brain. The PTS protocol yielded well controlled mean cortical infarction area of 1 cm2 with 0.89 ± 0.20 cm and 1.03 ± 0.41 cm of width and length (mean ± SD), respectively. Transcranial PA imaging of PTS lesions in neonatal piglet in vivo. Figures 1a and 1b show the experimental setup and protocol for the induction of regional ischemic stroke and transcranial PA imaging within 2 hours, followed by a triphenyltetrazolium chloride (TTC) staining to confirm eventual development of cortical infarction at 24 hours from the stroke induction.
Figure 1c shows a representative TTC image, illustrating clear development of infarction via PTS technique. Dotted circles indicate the candidate PTS regions, and blue and white colors indicate the affected and control regions-of-interest (ROIs). Apparent regional developments of cerebral infarction were identified at the regions where the PTS protocol was applied (blue dotted circles), compared to the control ROIs (white dotted circles). Figure 1d shows the representative transcranial PA imaging in neonatal piglet in vivo obtained within 2 hours from the PTS onset. The transverse image of HbO2 saturation and tHb shows the maximal and minimal intensity projection maps over 3-mm depth range in the superficial 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. cortical region. In particular, HbO2 saturation clearly localized PTS lesions corresponding to where they manifested in the TTC image (Figure 1c). Figure 1e presents HbO2 and HbR distributions in the corresponding transverse PA image. In visual assessment, interestingly, low HbO2 and higher HbR were introduced in PTS lesions when compared to control ROIs in the contralateral sides. Figure 2 and Table 1 present and summarize the quantitative measurements of HbO2 saturation, tHb, HbO2, and HbR in the PTS model. In the piglets with a single lesion (Figure 2b), the HbO2 saturation was the most obvious metric giving robust statistical significance between control and affected ROIs (p < 0.0001, n = 5): 54.53 ± 6.16% vs. 21.19 ± 4.59% (mean ± SD). The metrics of tHb and HbO2 displayed trends for lower levels in the affected ROIs, but these did not attain statistically effective separation with this small sample size. The values of HbR were not different between the control side (1.63 ± 1.85 a.u.) and affected side (1.14 ± 0.67 a.u), although one outlier value was evident with high HbR on the control side of the brain. Rejection of the case with an outlier out of standard deviation still did not yield statistical significance: 0.94 ± 0.58 vs. 0.82 ± 0.45 a.u. In the piglets with the induction of two PTS lesions (Figure 2c), the HbO2 saturation was still the most effective single metric (60.01 ± 13.46% vs. 31.31 ± 14.43% in control and affected ROIs, respectively), although the decrease in HbO2 still did not attain statistical significance despite the higher sample size (12 lesions in 6 piglets). The HbR concentration in the double-lesioned piglets did display the expected increase in the affected ROIs when compared to those at the control ROIs: 1.27 ± 0.47 vs. 1.94 ± 0.58 in arbitrary unit (a.u. ), respectively (p = 0.0045).
Finally, we analyzed the pooled data to have a generalized perspective how accurate PA imaging can detect PTS lesions (Figure 2a). The HbO2 saturation again yielded strong statistical significance: 58.40 ± 11.86% vs. 28.33 ± 13.08% in control and affected ROIs, respectively. Notably, HbO2 quantity presented a statistically significance decrease with PTS induction (p = 0.0211). Overall, the HbR quantity did not present a statistically significant separation; however, with rejection of the one outlier control value, a significant increase was observed (p = 0.0187). The tHb concentration for the pooled data did not convey statistical significance. Figure 3 shows the critical threshold for the PA imaging metrics to identify PTS lesions from control ROIs. Note that the solid and dotted black lines indicate the sensitivity and specificity, and the blue line indicates the accuracy. First of all, HbO2 saturation yielded effective separation between control and affected ROIs (Figure 3a), as expected from the clear statistical separation seen in Figure 2. We here define the effective threshold range giving sensitivity and specificity higher than 80% at the same time. In the pooled data (n = 17), HbO2 saturation successfully identified PTS lesions at an effective threshold range between 32.48 % and 53.19%; the maximal accuracy was 91.18%. Also, the PTS induction with single lesion presented with a wide effective threshold range between 23.31% and 53.91% (n = 5). Encouragingly, the maximal accuracy at 100 % was achieved between 25.95 % and 44.89 % of threshold range. Introduction of two PTS lesions showed a somewhat narrower effective threshold range between 36.55 % and 54.40 % (n = 12) with the maximal accuracy of 91.67%. Figure 3b shows the corresponding analysis on tHb concentration threshold. In general, there were no cases giving an effective threshold range. Even though the maximal accuracy of 80% was achieved with the single PTS lesion cases in the range between 3.57 and 5.50 a.u., the sensitivity in that range only reached 60%. The maximal accuracy was 66.67% in the cases with two PTS lesions. Pooled data also showed low accuracy at 70.59%. The individual analysis on HbO2 and HbR quantities presented moderate elevation of their accuracies (Figure 3c) but still did not yield meaningful improvement compared to HbO2 saturation. The maximal accuracy ranged from 60% - 80%, and very narrow effective threshold ranges were identified only with HbO2 quantity with single-lesion case (1.62 – 1.72 a.u). Figure 4 shows the receiver operating characteristic (ROC) curves for a threshold-invariant evaluation of PA imaging on PTS lesion detection. Note that the dotted and solid black lines indicate 5 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021.
The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. the ROC curves of the cases with a single-lesion (n = 5) and two-lesions (n = 12), respectively. The blue line presents the ROC curve using the pooled data (n = 17). Table 1 also contains the area-under- the-curve (AUC) measurements from the ROC curves. The HbO2 saturation metric provided creditable classification of PTS lesions with AUC values at 0.92, 1.00, and 0.89 in pooled, single-lesion, and two- lesion cases, respectively. In hemoglobin quantity metrics, AUC value over 0.80 only appeared with HbO2 in the single-lesion cases and HbR in the two-lesion cases. Supervised classification of PTS lesions. Table 3 summarizes the supervised machine learning- based classification accuracy on the control and affected ROIs. Firstly, a single-variable classification scenario was evaluated using individual PA imaging metrics. In general, the classification accuracy followed the statistical significance presented in Figure 3, as expected. The pooled data presented the maximal accuracy at 91.2% when only HbO2 saturation was included. Meanwhile, other hemoglobin quantity metrics (i.e., tHb, HbO2, and HbR) did not present any useful classification accuracy. The reduced data with single and two PTS lesions followed the same trends. Secondly, multivariable SVM models were analyzed with different number and combinations of variables. However, the addition of hemoglobin quantity metrics did not benefit the accuracy over the single-variable scenario solely using HbO2 saturation metric. The maximal accuracy in the multivariable models was 88.2% when used the pooled data and included HbO2 saturation as one of the variables. The cases without HbO2 saturation only provided up to 73.5% of accuracy. The use of other classifiers in the multivariable classification scenario (e.g., tree, discriminant, logistic regression, naïve bayes, quadratic/cubic/Gaussian SVMs, nearest neighbor, and ensemble) did not present any improvement over the single-variable classification accuracy on HbO2 saturation (< 91.2 %), which implies that none of them take advantage from tHb, HbO2 and HbR metrics. Discussion In this paper, we demonstrate the feasibility of transcranial PA imaging of regional ischemic stroke in neonatal piglets in vivo. We found that the HbO2 saturation metric was sufficient for effective detection in a hand-held PA imaging system. A hand-held monitoring system has clinical translation potential for diagnosis of perinatal arterial ischemic stroke at the bedside. Moreover, the near-infrared light could penetrate the scalp and skull of the piglet and provide a PA signal with spatial resolution that was adequate for detecting an ischemic lesion on a centimeter scale. The transcranial PA imaging of HbO2 saturation may address limitations of other perinatal monitoring modalities: having direct cerebral hemodynamic contrast (over EFHM); continuous and prompt monitoring (over MRI); portable bedside monitoring (over MRI); and sub-millimeter-scale spatial specificity (over biophotonic modalities).
However, the tissue quantity of total hemoglobin did not provide additional effectiveness in the context of the PTS model. There are many foreseeable variables in hemoglobin quantity; for example, inconsistent distance and angle of the PA probe from the fetal head would affect light fluence and acoustic attenuation. These will directly affect the estimation of hemoglobin quantity in individual patients, leading to patient-to-patient variability. This limitation could be mitigated by assuming that the ischemia is limited to a focal region of brain, thereby providing a within-patient contrast with non-ischemic tissue. It should be noted that the PTS model used here has specific hemodynamic consequences that differ from embolic models. In particular, the measures of hemoglobin quantity suggested complex physiologic features within the affected ROIs. The affected ROIs exhibited inconsistent decreases in tHb, with as the decreases inHbO2 coincided with increases in HbR. Thrombosis evoked by endothelial damage from the light/dye interaction likely starts in venules and capillaries, where shear stress is lower than in arterioles. Furthermore, because venules and capillaries have no smooth 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. muscle layer, damaging their endothelium will more readily expose underlying collagen in the extracellular matrix, which activates platelets and causes leukocytes to stick to the endothelium. The resulting thrombosis will cause entrapment of red blood cells within the microcirculation. This entrapment may account for the inconsistent decrease in tHb reported by PA imaging in this model. In contrast, embolic occlusion of a major cerebral artery reduces inflow into the microcirculation but does not block outflow. Indeed, PET images of brain in adult ischemic stroke show a decrease in cerebral blood volume that correlates with the decrease in cerebral blood flow [36,37]. Thus, the tHb response may differ with a perinatal stroke in humans, in which a clot typically occludes large cerebral arteries rather than vessels in the microcirculation. Unfortunately, the rete mirabilis between the pharyngeal artery and the intracranial internal carotid artery in the pig does not permit induction of a focal stroke in only a portion of the cerebral hemisphere by insertion of a blunted tip filament or blood clot or by embolization with microspheres. Therefore, further investigation of PA imaging capabilities in a model of embolic stroke in the piglet would likely require a full craniotomy to directly occlude a major cerebral artery. Several engineering strategies could be employed to improve and supplement the proposed framework.
The spatial resolution of the transcranial PA imaging is subject to the degree of phase aberration when imaging through skull layers, having heterogenous thickness with sound propagation speeds significantly different from that of soft biological tissue (4,080 m/s vs. 1,540 m/s). Even though the human newborn skull is thinner than adults and more transparent at the open fontanelles, the use of effective phase aberration correction will further improve the spatial resolution [38]. In addition, spectral attenuation in scalp and skull layers also needs to be corrected for accurate spectral unmixing [39]. ‘Spectral coloring’ artifacts are generated by light absorbance at fetal scalp and superficial vessels containing abundant melanin and blood, resulting in lower spectral unmixing accuracy due to the discrepancy between a measured PA spectrum and reference absorption spectrum of hemoglobin. Even though we compensated the artifacts with ex vivo calibration, more advanced fluence estimation and correction at individual imaging wavelengths will further improve the imaging quality for accurate stroke identification. Our long-term goal is to transform our transcranial PA imaging system into a safe, compact, and cost-effective form factor to facilitate its clinical translation. The investigation should be based on promising progress in compact light source (pulse laser diode [40], light-emitting diode [25], etc.) and ultra-portable ultrasound imaging systems [41–43]. For example, monitoring of a vulnerable neonate may also necessitate a compact yet effective monitoring device integrated in the neonatal intensive care unit (NICU). The careful tradeoff between image quality and compactness for fluent workflow would be a key for a successful clinical translation. In adult stroke, early imaging with CT or MRI provides a differential diagnosis of ischemic versus hemorrhagic stroke and selecting patients for treatment with tissue plasminogen activator and endovascular thrombectomy [44–46]. Early imaging results also allows for selection of patients in clinical trials that are most likely to benefit from therapeutic interventions. In contrast, clinical trials of newborn stroke are hampered by constraints of imaging unstable newborns in the first few hours after birth and the limited ability to differentiate intrapartum global hypoxia-ischemia from a focal stroke based on the early neurological exam. Newborns with stroke can present with seizures by 12 hours after birth [11], but neuroprotectants are unlikely to have a major benefit when initiated after seizure onset. Therefore, the ability to diagnose a focal stroke within a few hours after birth could have a major impact by enabling early enrollment and stratification in clinical trials for early treatments. Hence, noninvasive PA imaging, by differentiating focal stroke from global hypoxia- 7 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021.
The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. ischemia, has great potential for advancing clinical trials of neonatal stroke, along with promising technological progresses [47,48]. In conclusion, a transcranial PA imaging is capable of identifying a focal thrombotic stroke in a neonatal piglet model with intact scalp and skull at an early stage before the infarction is formed. Furthermore, the PA-derived HbO2 saturation is sufficient for providing robust sensitivity and specificity for detecting 1 cm focal cortical strokes. This technology has great potential for rapid and early diagnosis of perinatal ischemic stroke in clinics. Materials and Methods Photothrombotic model in neonatal piglets in vivo. All procedures were approved by the Johns Hopkins University Animal Care and Use Committee. Aseptic surgery was performed on male piglets anesthetized with isoflurane (approximately 1.5%) during the whole procedure. The position of piglet head was stabilized by using a custom stereotaxic equipment. The lungs were ventilated via intubation (endotracheal tube Tube, 2.5 mm, cuffed, Medsource Labs, MN, USA) with approximately 25% O2 and 75% air. A femoral artery and a jugular vein were catheterized, and the mean arterial blood pressure was continuously monitored. Rectal temperature was maintained near 38.0~38.5 °C with a warm blanket. Arterial pH, PCO2, PO2 , Hb, HbO2 saturation, and glucose were measured with a blood gas analyzer (ABL800 FLEX blood gas analyzer, Radiometer America Inc. Brea, CA, USA). A fluid of 5% dextrose and 0.45% sodium chloride was infused through jugular vein at the rate of 10 ml/h. Eleven male piglets, weighing 1.57 ± 0.18 kg were included in this study, with 5 in the single-lesion and 6 in the two-lesion PTS groups. Their mean arterial blood pressure at baseline was 61.3 ± 12.8 mmHg, and mean temperature was at 37.5 ± 0.7 oC. Blood sample data are summarized in Table 2. Note that we did not make any attempt to study equal numbers of male and female piglets because we are not aware of any plausible biological reason for sex differences in neurovascular physiology for direct photothrombosis. A solution of 20 mg/kg Rose Bengal (4,5,6,7-tetrachloro-2',4',5',7'-tetraiodofluorescein, RB) s in sterile saline was injected via jugular vein. The dye has an absorbance peak at approximately 560 nm [49], and the skull was thinned with a drill to allow sufficient light to penetrate the cortex. The focal skull areas overlying the frontal (5 mm anterior and 8 mm lateral from bregma) and posterior brain tissue regions (5 mm posterior and 8 mm lateral from bregma) was thinned with 5-mm width. The skull thinning was performed bilaterally, but only half of the focal areas were targeted for photothrombosis. In those targeted for photothrombosis, the thinned skull was illuminated with a 5- mm aperture cold-white light at the intensity of 60 – 75 mW/cm2 for 20 minutes.
The light interaction with the dye generates free radicals that damage the endothelium sufficiently to activate platelets and produce a thrombus. During the light illumination, 1 ml sterile saline solution was perfused on the thinned skull every 5 minutes to prevent overheating and damage on target surface. No signs of thermal damage due to the light illumination were found with light illumination in the absence of Rose Bengal dye with regular saline flushing. Photoacoustic imaging system and protocol. Figure 1a shows the experimental system configuration. For PA recording, 5.2-MHz clinical ultrasound linear array transducer (L7-4, ATL Ultrasound Inc.) was connected to a research package (Vantage 256, Verasonics Inc., USA). Detailed imaging parameters are as follows: elevation focal depth, 25 mm; element pitch, 0.30 mm; element height, 7.5 mm for transducer, 18 and 24 dB low-noise amplification (LNA) settings during ultrasound (US) and PA imaging. Those parameters and system configuration are available in modern clinical US imaging platforms, maximizing the translationability of the research outcome 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. into clinics. To induce the PA signals, pulsed laser light was generated by a second-harmonic (532 nm) Nd: YAG laser pumping an optical parametric oscillator (OPO) system (Phocus Inline, Opotek Inc., USA). The tunable range is 690-900 nm and the maximum pulse repetition frequency is 20 Hz. Light was delivered into the probe through bifurcated fiber optic bundles, each with 40-mm long and 0.88-mm wide aperture. The PA probe is located on the center of piglet head along with the coronal plane to simultaneously monitor the PA signal changes originated from superior sagittal sinus and cerebral brain tissue. The distance between fiber bundle output and piglet head was maintained at 20 mm, leading to the 5 mJ/cm2 of energy density on the skin surface. This is far below the maximum permissible exposure (MPE) of skin to laser radiation guided by the ANSI Z136.1 standard, Safe Use of Lasers (i.e., 20 mJ/cm2). [54] In addition, 64 subsequent frames were averaged to overcome the limited signal-to-noise ratio (SNR) at cerebral brain tissue regions with naturally less blood content than the sagittal sinus and to alleviate the energy fluctuation in the Nd:YAG OPO laser source. Figure 1b shows in vivo experimental protocol. At 20-min after inducing the last PTS, translational scanning was initiated from frontal to posterior regions for volumetric PA imaging within 20-30 min. The translational scanning was performed from front to the back of the brain in 1 mm intervals with a motorized stage (MTS50/M-Z8 and TDC001, Thorlabs Inc., USA).
The HbO2 saturation and tHb were decomposed from spectral PA data (700–900 nm in 10 nm interval), and transverse planes were projected in axial direction at every 1 mm depth interval. From the data, spectral unmixing was conducted by a constrained linear least-squares problem solver in MATLAB software (Mathworks Inc., USA). Optical absorbance of HbO2 and deoxy-Hb was obtained from spectrophotometric measurement (brought from [50]). The mean spectral attenuation was measured ex vivo and compensated before the spectral unmixing process [24]. Triphenyltetrazolium chloride (TTC) staining and imaging. After the PA scanning, the isoflurane was stopped, and the piglet was carefully removed from the stereotaxic apparatus. The piglet was kept on a pre-warmed heating pad until it was fully awake The pig was extubated when it regained a cough reflex.. The piglet was returned to the cage and fed formula milk. On the next day, the piglet was euthanized with an overdose of pentobarbial, and the brain was removed for staining with the vital dye TTC (2% solution) on fresh tissue. TTC stains intact mitochondrial enzymes red whereas the infarcted region in pale (Figure 1c). The surface contour of cerebral infarction carefully was drawn by one of the authors (X. Liu). We correlated the PA images with 24-hr post-stroke focal infarction developments on cortex detected through TTC image. Statistical analysis. PA metrics were compared between the affected ROIs and the contralateral control ROIs by paired t-test. ROC curves and corresponding AUC were reconstructed and calculated in MATLAB software (R2020a, Mathworks, Inc., USA) using perfcurve function. Supervised machine learning. The pooled measurements from control and affected ROIs (n = 34) were fed into a supervised machine learning classification learner. Various combinations of PA imaging metrics (i.e., HbO2 saturation, tHb, HbO2, and HbR) were trained in a linear SVM model using MATLAB software (R2020a, Mathworks, Inc., USA), and tested with 34-fold cross validation scheme (i.e., 33 training data and 1 test data; repeat 34 times). 9 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Acknowledgements The financial support was provided by NIH, NHLBI (R01HL139543); Maryland Innovation Initiative, TEDCO; Louis B. Thalheimer Fund for Translational Research. Conflict of Interests The authors declare no competing financial interests. References [1] Bosenbark D D, Krivitzky L, Ichord R, Jastrzab L and Billinghurst L 2016 Attention and executive functioning profiles in children following perinatal arterial ischemic stroke Child Neuropsychol 24 1–18 [2] Chabrier S, Peyric E, Drutel L, Deron J, Kossorotoff M, Dinomais M, Lazaro L, Lefranc J, Thébault G, Dray G, Fluss J, Renaud C, Tich S N T, Group A V C du nouveau-né (AVCnn; [Neonatal S S, Darteyre S, Dégano C, Delion M, Groeschel S, Hertz-Pannier L, Husson B, Presles E, Ravel M and Vuillerot C 2016 Multimodal Outcome at 7 Years of Age after Neonatal Arterial Ischemic Stroke J Pediatrics 172 156-161.e3 [3] Fluss J, Dinomais M, Kossorotoff M, Vuillerot C, Darteyre S and Chabrier S 2017 Perspectives in neonatal and childhood arterial ischemic stroke Expert Rev Neurother 17 135–42 [4] Vuillerot C, Dinomais M, Marret S, Chabrier S and Debillon T 2016 Recommendations for clinical practice after neonatal arterial ischemic stroke: Clinical monitoring and early rehabilitation intervention Ann Phys Rehabilitation Medicine 59 e1 [5] Basu A P 2014 Early intervention after perinatal stroke: opportunities and challenges Dev Medicine Child Neurology 56 516–21 [6] Gonzalez F F and Ferriero D M 2009 Neuroprotection in the Newborn Infant Clin Perinatol 36 859–80 [7] Worp H B van der, Sena E S, Donnan G A, Howells D W and Macleod M R 2007 Hypothermia in animal models of acute ischaemic stroke: a systematic review and meta-analysis Brain 130 3063– 74 [8] Harbert M J, Tam E W Y, Glass H C, Bonifacio S L, Haeusslein L A, Barkovich A J, Jeremy R J, Rogers E E, Glidden D V and Ferriero D M 2011 Hypothermia Is Correlated With Seizure Absence in Perinatal Stroke J Child Neurol 26 1126–30 [9] Nelson K B, Sartwelle T P and Rouse D J 2016 Electronic fetal monitoring, cerebral palsy, and caesarean section: assumptions versus evidence Bmj 355 i6405 [10] Hirsch E 2019 Electronic Fetal Monitoring to Prevent Fetal Brain Injury: A Ubiquitous Yet Flawed Tool Jama 322 611–2 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021.
The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. [11] Adami R R, Grundy M E, Poretti A, Felling R J, Lemmon M and Graham E M 2016 Distinguishing Arterial Ischemic Stroke From Hypoxic–Ischemic Encephalopathy in the Neonate at Birth Obstetrics Gynecol 128 704–12 [12] Gunn A J, Bennet L, Gunning M I, Gluckman P D and Gunn T R 1999 Cerebral Hypothermia Is Not Neuroprotective When Started after Postischemic Seizures in Fetal Sheep Pediatr Res 46 274–80 [13] Lee C W, Cooper R J and Austin T 2017 Diffuse optical tomography to investigate the newborn brain Pediatr Res 82 376–86 [14] Semyachkina-Glushkovskaya O V, Kurths J, Pavlov A N, Borisova E G, Abdurashitov A S, Zhu D, Li P, Luo Q and Tuchin V V 2016 Silent Vascular Catastrophes in the Brain in Term Newborns: Strategies for Optical Imaging Ieee J Sel Top Quant 22 88–101 [15] Wang L V and Hu S 2012 Photoacoustic Tomography: In Vivo Imaging from Organelles to Organs Science 335 1458–62 [16] Yao J, Maslov K I and Wang L V 2012 In vivo Photoacoustic Tomography of Total Blood Flow and Potential Imaging of Cancer Angiogenesis and Hypermetabolism Technol Cancer Res T 11 301– 7 [17] Kang J, Chang J H, Kim S M, Lee H J, Kim H, Wilson B C and Song T-K 2017 Real-time sentinel lymph node biopsy guidance using combined ultrasound, photoacoustic, fluorescence imaging: in vivo proof-of-principle and validation with nodal obstruction Sci Rep-uk 7 45008 [18] Kang J, Kim E, Kim G R, Yoon C, Song T and Chang J H 2015 Photoacoustic imaging of breast microcalcifications: A validation study with 3‐dimensional ex vivo data and spectrophotometric measurement J Biophotonics 8 71–80 [19] Kang J, Zhang H K, Kadam S D, Fedorko J, Valentine H, Malla A P, Yan P, Harraz M M, Kang J U, Rahmim A, Gjedde A, Loew L M, Wong D F and Boctor E M 2019 Transcranial Recording of Electrophysiological Neural Activity in the Rodent Brain in vivo Using Functional Photoacoustic Imaging of Near-Infrared Voltage-Sensitive Dye Front Neurosci-switz 13 579 [20] Kang J, Kadam S D, Elmore J S, Sullivan B J, Valentine H, Malla A P, Harraz M M, Rahmim A, Kang J U, Loew L M, Baumann M H, Grace A A, Gjedde A, Boctor E M and Wong D F 2020 Transcranial photoacoustic imaging of NMDA-evoked focal circuit dynamics in the rat hippocampus J Neural Eng 17 025001 [21] Li M, Tang Y and Yao J 2018 Photoacoustic Tomography of Blood Oxygenation: A Mini Review Photoacoustics 10 65–73 [22] Hu S and Wang L V 2010 Neurovascular Photoacoustic Tomography Frontiers Neuroenergetics 2 10 [23] Hu S 2016 Listening to the Brain With Photoacoustics Ieee J Sel Top Quant 22 117–26 11 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder.
All rights reserved. No reuse allowed without permission. [24] Kang J, Boctor E M, Adams S, Kulikowicz E, Zhang H K, Koehler R C and Graham E M 2018 Validation of noninvasive photoacoustic measurements of sagittal sinus oxyhemoglobin saturation in hypoxic neonatal piglets J Appl Physiol 125 983–9 [25] Kang J, Koehler R C, Adams S, Graham E M and Boctor E M 2020 Light-emitting diode-based transcranial photoacoustic measurement of sagittal sinus oxyhemoglobin saturation in hypoxic neonatal piglets Biorxiv 262451 2020.08.22.262451 [26] Petrova I Y, Petrov Y Y, Esenaliev R O, Deyo D J, Cicenaite I and Prough D S 2009 Noninvasive monitoring of cerebral blood oxygenation in ovine superior sagittal sinus with novel multi-wavelength optoacoustic system Opt Express 17 7285 [27] Lv J, Li S, Zhang J, Duan F, Wu Z, Chen R, Chen M, Huang S, Ma H and Nie L 2020 In vivo photoacoustic imaging dynamically monitors the structural and functional changes of ischemic stroke at a very early stage Theranostics 10 816–28 [28] Armstead W M, Riley J, Kiessling J W, Cines D B and Higazi A A-R 2010 Novel plasminogen activator inhibitor-1-derived peptide protects against impairment of cerebrovasodilation after photothrombosis through inhibition of JNK MAPK Am J Physiology-regulatory Integr Comp Physiology 299 R480–5 [29] Armstead W M, Ganguly K, Riley J, Zaitsev S, Cines D B, Higazi A A-R and Muzykantov V R 2012 RBC-coupled tPA Prevents Whereas tPA Aggravates JNK MAPK-Mediated Impairment of ATP- and Ca-Sensitive K Channel-Mediated Cerebrovasodilation After Cerebral Photothrombosis Transl Stroke Res 3 114–21 [30] Dobbing J and Sands J 1979 Comparative aspects of the brain growth spurt Early Hum Dev 3 79–83 [31] Conrad M S, Dilger R N and Johnson R W 2012 Brain Growth of the Domestic Pig (Sus scrofa) from 2 to 24 Weeks of Age: A Longitudinal MRI Study Dev Neurosci-basel 34 291–8 [32] Sheng H Z, Rosbo N K, Carnegie P R and Bernard C C A 1989 Developmental Study of Myelin Basic Protein Variants in Various Regions of Pig Nervous System J Neurochem 52 736–40 [33] McColl B W, Carswell H V, McCulloch J and Horsburgh K 2004 Extension of cerebral hypoperfusion and ischaemic pathology beyond MCA territory after intraluminal filament occlusion in C57Bl/6J mice Brain Res 997 15–23 [34] Kleinschnitz C, Fluri F and Schuhmann M 2015 Animal models of ischemic stroke and their application in clinical research Drug Des Dev Ther Volume 9 3445–54 [35] Labat-gest V and Tomasi S 2013 Photothrombotic Ischemia: A Minimally Invasive and Reproducible Photochemical Cortical Lesion Model for Mouse Stroke Studies J Vis Exp [36] Powers W J, Grubb R L and Raichle M E 1984 Physiological responses to focal cerebral ischemia in humans Ann Neurol 16 546–52 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder.
All rights reserved. No reuse allowed without permission. [37] Powers W J, Grubb R L, Darriet D and Raichle M E 1985 Cerebral Blood Flow and Cerebral Metabolic Rate of Oxygen Requirements for Cerebral Function and Viability in Humans J Cereb Blood Flow Metabolism 5 600–8 [38] Wang L V 2008 Prospects of photoacoustic tomography Med Phys 35 5758–67 [39] Cox B, Laufer J G, Arridge S R and Beard P C 2012 Quantitative spectroscopic photoacoustic imaging: a review J Biomed Opt 17 0612021–06120222 [40] Esenaliev R O 2017 Optoacoustic diagnostic modality: from idea to clinical studies with highly compact laser diode-based systems J Biomed Opt 22 091512–091512 [41] Kang J, Yoon C, Lee J, Kye S-B, Lee Y, Chang J H, Kim G-D, Yoo Y and Song T-K 2016 A System-on-Chip Solution for Point-of-Care Ultrasound Imaging Systems: Architecture and ASIC Implementation Ieee T Biomed Circ S 10 412–23 [42] Kim G-D, Yoon C, Kye S-B, Lee Y, Kang J, Yoo Y and Song T-K 2012 A Single FPGA-Based Portable Ultrasound Imaging System for Point-of-Care Applications Ieee Transactions Ultrasonics Ferroelectr Freq Control 59 1386–94 [43] Kang J, Kim P, Yoon C, Yoo Y and Song T-K 2020 Efficient Parallel-Beamforming Based on Shared FIFO for Ultra-Compact Ultrasound Imaging Systems Ieee Access 8 80490–501 [44] French B R, Boddepalli R S and Govindarajan R 2016 Acute Ischemic Stroke: Current Status and Future Directions. Mo Med 113 480–6 [45] Saver J L, Fonarow G C, Smith E E, Reeves M J, Grau-Sepulveda M V, Pan W, Olson D M, Hernandez A F, Peterson E D and Schwamm L H 2013 Time to Treatment With Intravenous Tissue Plasminogen Activator and Outcome From Acute Ischemic Stroke Jama 309 2480–8 [46] Meretoja A, Keshtkaran M, Saver J L, Tatlisumak T, Parsons M W, Kaste M, Davis S M, Donnan G A and Churilov L 2018 Stroke Thrombolysis Stroke 45 1053–8 [47] Nie L, Cai X, Maslov K, Garcia-Uribe A, Anastasio M A and Wang L V 2012 Photoacoustic tomography through a whole adult human skull with a photon recycler J Biomed Opt 17 110506– 110506 [48] Ovsepian S V, Olefir I and Ntziachristos V 2019 Advances in Optoacoustic Neurotomography of Animal Models Trends Biotechnol 37 1315–26 [49] Watson B D, Dietrich W D, Busto R, Wachtel M S and Ginsberg M D 1985 Induction of reproducible brain infarction by photochemically initiated thrombosis Ann Neurol 17 497–504 [50] Prahl S 1999 Optical Absorption of Hemoglobin 13 486 487 488 489 490 491 492 493 494 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Tables Table 1. Photoacoustic (PA) imaging metrics of photothrombotic (PTS) model in vivo. Asterisk marks indicate the statistical significance between the control and affected regions-of-interest (ROIs). Asterisk marks indicate statistical significance: * (p ≤ 0.05), (p ≤ 0.01), * (p ≤ 0.001), and **** (p ≤ 0.0001).
Pooled data (n = 17 from 11 animals) Control Affected p HbO2 saturation (%) 58.40 ± 11.86 28.33 ± 13.08 < 0.0001 (***) Total Hb (tHb) concentration (a.u.) 5.83 ± 3.04 4.44 ± 2.10 0.0733 HbO2 concentration (a.u.) 4.46 ± 2.46 2.72 ± 1.61 0.0066 () HbR concentration (a.u.) 1.38 ± 1.02 1.71 ± 0.70 0.2314 (p = 0.0118 when n = 16 excluding an outlier, ) 0.7128 AUC 0.9204 0.6540 0.7093 Single PTS lesion (n = 5 from 5 animals) Control Affected p AUC 54.53 ± 6.16 21.19 ± 4.59 0.0019 () 1.0000 5.67 ± 4.42 2.44 ± 1.22 0.1070 0.7200 4.03 ± 2.84 1.29 ± 0.64 0.0720 0.8000 1.63 ± 1.85 1.14 ± 0.67 0.4817 0.5200 Two PTS lesions (n = 12 from 6 animals) Control Affected p AUC 60.01 ± 13.46 31.31 ± 14.43 < 0.0001 (*) 0.8889 5.90 ± 2.51 5.27 ± 1.80 0.4103 0.6111 4.64 ± 2.40 3.32 ± 1.52 0.0586 0.6875 1.27 ± 0.47 1.94 ± 0.58 0.0128 () 0.8403 Table 2. The arterial blood sample analysis of the piglets during stroke induction experiment (Data format: Mean ± SD). Abbreviations: pCO2, Arterial carbon dioxide partial pressure; pO2, Arterial oxygen partial pressure; cBase, Base excess; ctHb, Total hemoglobin; Glu, Glucose; Lac, Lactate. pH 7.37 ± 0.09 K+ pCO2 (mmHg) 37.1 ± 5.5 Na+ pO2 (mmHg) 143.5 ± 23.6 Ca2+ cBase (mmol/L) -3.5 ± 3.3 Cl- HbO2 saturation (%) 99.7% ± 0.5% Glu ctHb (g/dL) 11.0 ± 2.9 Lac (mmol/L) (mmol/L) (mmol/L) (mmol/L) (mg/dL) (mmol/L) 3.7 ± 0.6 141.9± 3.4 1.3 ± 0.2 108.7 ± 5.1 141.9 ± 49.6 4.5 ± 8.6 495 496 497 498 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Table 3. Supervised machine learning-based differential classification accuracy using a linear support vector machine (SVM) classifier. 34 data points (17 ROIs for both controlled and PTS lesions) were used with 34- fold cross-validation. Gray shadings indicate the variables used. #Variables Variables used Accuracy (%) 2 O b H n o i t a r u t a s b H t 2 O b H R b H Pooled data Single PTS lesion Two PTS lesions e l g n i S e l b a i r a v - i t l u M e l b a i r a v o i r a n e c s o i r a n e c s 1 2 3 91.2 55.9 61.8 61.8 88.2 88.2 88.2 73.5 73.5 73.5 88.2 100.0 50.0 80.0 0.0 90.0 90.0 100.0 80.0 80.0 80.0 90.0 87.5 54.2 58.3 75.0 87.5 87.5 87.5 79.2 75.0 75.0 87.5 88.2 88.2 90.0 90.0 87.5 87.5 4 73.5 88.2 80.0 90.0 75.0 87.5 15 499 500 501 502 503 504 505 506 507 508 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Figures Figure 1. The transcranial PA imaging of photothrombotic stroke in neonatal piglets in vivo. (a) Experimental setup and (b) protocol for PTS induction and PA imaging. (c) Triphenyltetrazolium chloride (TTC) image of harvested whole brain, indicating cerebral infarction developed at 24-hr post-stroke time point.
Blue and white dotted circles indicate the PTS lesions and control regions-of-interest. (d) Representative transverse planes of HbO2 saturation, total hemoglobin (tHb) overlaid on the TTC image. (e) Corresponding transverse maps of HbO2 and HbR quantities. Black bars indicate 1 cm. Asterisk marks indicate the cortical regions presented cerebral infarction in TTC image. 509 510 511 512 513 514 515 516 517 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. a ) 7 1 = n ( d e l o o P ) % ( n o i t a r u t a s 2 O b H 100 80 60 40 20 ** ) . u . a ( b H t 10 5 ) . u . a ( 2 O b H 10 8 6 4 2 ) . u . a ( R b H 5 4 3 2 1 ☓ 0 0 0 0 C ontrol Affected C ontrol Affected C ontrol Affected C ontrol Affected b 100 10 5 ) 5 = n ( n o i s e l e l g n i S ) % ( n o i t a r u t a s 2 O b H 80 60 40 20 ) . u . a ( b H t 10 5 ) . u . a ( 2 O b H 8 6 4 2 ) . u . a ( R b H 4 3 2 1 0 0 0 0 c 100 10 5 ) 2 1 = n ( s n o i s e l o w T ) % ( n o i t a r u t a s 2 O b H 80 60 40 20 ) . u a ( b H . t 10 5 ) . u a ( . 2 O b H 8 6 4 2 ) . u a ( . R b H 4 3 2 1 0 0 0 0 C ontrol Affected C ontrol Affected C ontrol Affected C ontrol Affected Figure 2. Quantitative measurements of PA imaging metrics, HbO2 saturation, total hemoglobin (tHb), oxyhemoglobin (HbO2), and deoxyhemoglobin (HbR). (a) Pooled group (n = 17). (b) Single-lesion group (n = 5). (c) Two-lesion group (n = 12). Red indicators show the individual data points from the regions-of- interest (ROI). Red lines indicate the control and affected ROIs from individual piglets. Asterisk marks indicate the degree of statistical significance in paired t test: * (p ≤ 0.05), (p ≤ 0.01), * (p ≤ 0.001), and **** (p ≤ 0.0001). ☓ indicates a statistical significance in a specific case when removed one outlier animal data (n = 16, p ≤ 0.05). 17 518 519 520 521 522 523 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. a b c 100 100 100 100 ) % ) 7 1 = n ( d e l o o P ( n o i t a r u t a s 2 O b H 80 60 40 20 ) . u . a ( b H t 80 60 40 20 ) . u . a ( 2 O b H 80 60 40 20 ) . u . a ( R b H 80 60 40 20 0 0 50 100 0 0 5 10 0 0 5 10 0 0 5 10 Threshold (%) Threshold (a.u.) Threshold (a.u.) Threshold (a.u.) ) 5 = n ( n o i s e l e l g n i S ) % ( n o i t a r u t a s 2 O b H 100 80 60 40 20 ) . u . a ( b H t 100 80 60 40 20 ) . u . a ( 2 O b H 100 80 60 40 20 100 80 ) . u . a ( 60 R b H 40 20 0 0 0 0 0 50 100 0 5 10 0 5 10 0 5 10 100 100 100 100 ) 2 1 = n ( s n o i s e l o w T ) % ( n o i t a r u t a s 2 O b H 80 60 40 20 ) . u . a ( b H t 80 60 40 20 ) . u . a ( 2 O b H 80 60 40 20 ) . u . a ( R b H 80 60 40 20 0 0 50 100 0 0 5 10 0 0 5 10 0 0 5 10 Threshold (%) Threshold (a.u.)
Threshold (a.u.) Threshold (a.u.) Figure 3. Sensitivity, specificity, and accuracy to classify the photothrombotic (PTS) lesions for the varying threshold values. (a) HbO2 saturation, (b) total hemoglobin (tHb), and (c) oxy- and deoxyhemoglobin (HbO2 and HbR) in pooled, single-lesion, and two-lesion groups. Solid and dotted black lines indicate specificity and sensitivity, respectively. Blue line indicates the accuracy. 524 525 526 527 528 bioRxiv preprint doi: https://doi.org/10.1101/2021.07.08.451613 ; this version posted July 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. HbO2 saturation tHb 1 1 0.5 0.5 e t a R e v i t i s o P e u r T 0 1 0 0.5 HbO2 1 0 1 0 0.5 HbR 1 0.5 0.5 0 0 0 0.5 1 0 0.5 1 False Positive Rate Figure 4. Receiver operating characteristic (ROC) curves for a threshold-invariant evaluation of PA imaging on PTS lesion detection. Blue line indicates the ROC curve with the pooled data. Dotted and solid black lines indicate the curves in the single-lesion and two-lesion groups, respectively. 19
5 10 15 20 25 30 35 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . Structure-guided glyco-engineering of ACE2 for improved potency as soluble SARS-CoV-2 decoy receptor Tümay Capraz,1† Nikolaus F. Kienzl,2† Elisabeth Laurent,3† Jan W. Perthold,1 Esther Föderl- Höbenreich,4 Clemens Grünwald-Gruber,5 Daniel Maresch,5 Vanessa Monteil,6 Janine Niederhöfer,7 Gerald Wirnsberger,7 Ali Mirazimi,6,8 Kurt Zatloukal,4 Lukas Mach,2* Josef M. Penninger,9,10* Chris Oostenbrink,1* Johannes Stadlmann5,9* 1 Institute for Molecular Modeling and Simulation, University of Natural Resources and Life Sciences (BOKU), Muthgasse 18, 1190 Vienna, Austria. 2 Institute of Plant Biotechnology and Cell Biology, Department of Applied Genetics and Cell Biology, University of Natural Resources and Life Sciences (BOKU), Muthgasse 18, 1190 Vienna, Austria. 3 Institute of Molecular Biotechnology, Department of Biotechnology and Core Facility Biomolecular & Cellular Analysis, University of Natural Resources and Life Sciences (BOKU), Muthgasse 18, 1190 Vienna, Austria. 4 Diagnostic and Research Institute of Pathology, Medical University of Graz, Neue Stiftingtalstrasse 6, 8010 Graz, Austria. 5 Institute of Biochemistry, Department of Chemistry, University of Natural Resources and Life Sciences (BOKU), Muthgasse 18, 1190 Vienna, Austria. 6 Karolinska Institute, Department of Laboratory Medicine, 17177 Stockholm, Sweden. 7 Apeiron Biologics, Campus Vienna Biocenter 5, 1030 Vienna, Austria. 8 National Veterinary Institute, 75189 Uppsala, Sweden. 9 IMBA - Institute of Molecular Biotechnology of the Austrian Academy of Sciences, Dr. Bohr Gasse 3, 1030 Vienna, Austria. 10 Department of Medical Genetics, Life Sciences Institute, University of British Columbia, Vancouver Campus, 2350 Health Sciences Mall, Vancouver, BC V6T 1Z3, Canada. †equal contributions. Correspondence to: [email protected], [email protected], [email protected], [email protected] Running title: Glycoengineering of ACE2 improves SARS-CoV-2 neutralization. 1 40 45 50 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . Abstract Infection and viral entry of SARS-CoV-2 crucially depends on the binding of its Spike protein to angiotensin converting enzyme 2 (ACE2) presented on host cells. Glycosylation of both proteins is critical for this interaction. Recombinant soluble human ACE2 can neutralize SARS-CoV-2 and is currently undergoing clinical tests for the treatment of COVID-19.
We used 3D structural models and molecular dynamics simulations to define the ACE2 N-glycans that critically influence Spike-ACE2 complex formation. Engineering of ACE2 N-glycosylation by site- directed mutagenesis or glycosidase treatment resulted in enhanced binding affinities and improved virus neutralization without notable deleterious effects on the structural stability and catalytic activity of the protein. Importantly, simultaneous removal of all accessible N-glycans from recombinant soluble human ACE2 yields a superior SARS-CoV-2 decoy receptor with promise as effective treatment for COVID-19 patients. 2 55 60 65 70 75 80 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . Introduction The rapid spread of Severe Acute Respiratory Syndrome Coronavirus-2 (SARS-CoV-2), the causative pathogen of human coronavirus disease 2019 (COVID-19), has resulted in an unprecedented pandemic and worldwide health crisis. Similar to the beta-coronaviruses SARS- CoV and Middle Eastern Respiratory Syndrome (MERS)-CoV, SARS-CoV-2 is highly transmissible and can lead to lethal pneumonia and multi-organ failure.1 For infection and viral entry, the Spike surface protein of SARS-CoV-2 binds to angiotensin converting enzyme 2 (ACE2) on host cells.2,3 Recombinant soluble human ACE2 (rshACE2) has been shown to bind Spike,4 can effectively neutralize SARS-CoV-2 infections,5,6 and the corresponding drug candidate APN01 has undergone a phase 2 clinical trial for the treatment of hospitalized cases of COVID-19 (ClinicalTrials.gov Identifier: NCT04335136). A first case study of its use in a patient has been reported recently.7 Additionally, an aerosol formulation of APN01 has been developed and is currently undergoing Phase I clinical studies. Multiple other therapeutic strategies attempt to target the Spike-ACE2 interaction, e.g. by development of neutralizing antibodies blocking the ACE2-binding site8 or lectins that bind to glycans on the Spike surface.9,10 Using soluble ACE2 as a decoy receptor for Spike is particularly attractive, as it minimizes the risk that variants of concern may evade the treatment through mutations as has been observed for antibodies.11-13 Furthermore, protein engineering has yielded ACE2 variants with substantially improved affinities for Spike.14,15 Hence, soluble ACE2 based therapeutics offer considerable advantages over other therapeutic formats that aim to hamper the Spike-ACE2 interaction sterically. Modern structural biology has been amazingly fast to respond to this pandemic. A mere three months after identification of SARS-CoV-2 as the etiologic agent of COVID-19, structures of the complex between ACE2 and the receptor binding domain (RBD)4,16,17 and of the ectodomain of trimeric Spike18,19 were already solved by X-ray crystallography or cryo-electron microscopy.
While this provided unprecedented insight into the protein-protein interactions between Spike and ACE2, the structural impact of protein-bound glycans on the Spike-ACE2 interface could not be assessed experimentally so far due to their compositional diversity and conformational flexibility. Here, in silico modeling of the glycans offers a powerful alternative to study the 3 85 90 95 100 105 110 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . effects of individual Spike and ACE2 glycans on the molecular interactions between these two proteins. The SARS-CoV-2 Spike protein is heavily glycosylated with both complex and oligo- mannosidic type N-glycans,1,10,20,21 thereby shielding a large portion of the protein surface.22-24 Similarly, ACE2 is a glycoprotein with up to seven highly utilized sites of N-glycosylation.21 Recent computational studies started to investigate protein glycosylation in the context of the interaction between Spike and ACE2.21,22,25 Extensive all-atom molecular dynamics (MD) simulations indicated that Spike N-glycans attached to N165 and N234 could be important stabilizers of the ligand-accessible conformation of the receptor binding domain (RBD).22 Furthermore, it has been proposed that the N-glycan at position N343 acts as a gate facilitating RBD opening.26 Other MD studies concluded that the glycans attached to N90 and N322 of ACE2 could be major determinants of Spike binding,25 while yet other simulation works postulate that glycosylation does not affect the RBD-ACE2 interaction significantly.27,28 Genetic or pharmacological blockade of N-glycan biosynthesis at the oligomannose stage in ACE2- expressing target cells was found to dramatically reduce viral entry,29 even though several glycoforms of ACE2 were found to display comparatively moderate variation with respect to Spike binding.30 Hence, a detailed understanding on how individual glycans on both Spike and ACE2 influence their interaction and a comprehensive experimental validation of the MD findings is crucial for the rational design of novel therapeutic soluble ACE2 variants with enhanced Spike binding affinity and the capacity to block viral entry more efficiently than the native enzyme.21 The identification of the Spike glycans essential for efficient association with ACE2 will be also critical to guide rational design of improved SARS-CoV-2 vaccines. Results We started our research by creating 3D models of the trimeric Spike in complex with human ACE2 (hACE2). The RBD of Spike exists in two distinct conformations, referred to as “up” and “down”.18,19 The “up” conformation corresponds to the receptor-accessible state with the RBD of one monomer exposed. By superimposing the RBD from the RBD-hACE2 complex17 with the single RBD in the “up” conformation (monomer 3) of the trimeric Spike,18 an initial model was obtained.
To assess the impact of all seven individual N-glycosylation sites of hACE2 on its interaction with Spike, we first elucidated the entire glycome of rshACE2 (Fig. S1). This also 4 115 120 125 130 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . provided information on the glycans attached to N690, a glycosylation site not covered in previous glycoproteomic studies of soluble hACE2.21,30 For recombinant trimeric Spike the glyco-analysis has been reported elsewhere.10,20,21,31 Based on the site-specific glycosylation profiles we added complex or oligo-mannosidic glycan trees to the respective sites of Spike and ACE2 (Table S1). We hence constructed fully glycosylated atomistic models of the trimeric Spike glycoprotein, free dimeric ACE2 and of the Spike glycoprotein in complex with dimeric hACE2 (Fig. 1). Using these fully glycosylated structures, we performed molecular dynamics simulations of the Spike-ACE2 complex (Video S1), and of free hACE2. Inspection of the most important interacting residues on Spike and ACE2, their average distances and the electrostatic potential of the interface area identified critical contact sites (Figs. S2 and S3). a b d c e Figure 1. A 3D structural model of the glycosylated Spike-hACE2 complex. (a) 3D model of the Spike trimer (in green, with RBD of monomer 3 in dark green) binding to ACE2 (in grey) with complex glycosylation in magenta, Man5 glycans in light blue and Man9 glycans in orange. (b) Close-up view of the glycans at N122 (orange sticks) and N165 (dark green sticks) on monomer 3 of Spike. (c) Close-up view of the glycans at N331 (yellow sticks) and N343 (purple sticks) on monomer 3 of Spike. (d) Close-up view of the glycans at N53 (blue sticks) and N90 (yellow sticks) on ACE2. (e) Close-up view of the glycans at N53 (blue sticks), N90 (yellow sticks), N322 (black sticks) and N546 (red sticks). 5 135 140 145 150 155 160 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . We next quantified the complete solvent-accessible surface area (SASA) of the Spike protein in complex with ACE2, both with and without glycans. The average accessible area of protein atoms for non-glycosylated and glycosylated Spike was 1395 nm² and 864 nm², respectively, indicating that glycans shield about 38% of the protein surface of Spike, a value that is comparable to what was previously found in simulations of Spike alone.23,32 The area of protein atoms that are shielded by the individual glycans are shown in Fig.
2a and Fig. S4. Further analysis showed that glycans at N122, N165 and N343 on Spike directly interact with ACE2 or its glycans (Fig. 1b, c, Fig. 2b, c). It has been reported that Spike mutants lacking the glycans at N331 and N343 display reduced infectivity, while elimination of the glycosylation motif at N234 results in increased resistance to neutralizing antibodies, without reducing infectivity of the virus.33 The equilibrium between the “up” and “down” conformations of Spike involves various stabilizing and destabilizing effects, with possible roles for the glycans at N165, N234, N331 and N343.22,26,34 Removing the glycans at N165, N234 and N343 was experimentally seen to reduce binding to ACE2 by 10%, 40% and 56%, respectively.22,26 In our MD simulations, the glycan at position N343 interacts directly with ACE2 (Fig. 2), while the glycan at N331 interacts with a neighboring Spike monomer (Fig. 1c, Fig. S5), indicating that the N331 glycosylation site only indirectly affects the interaction of Spike with ACE2. In our model, the glycan at N234 also does not interact directly with ACE2, but seems to stabilize the "up" conformation. Its removal could favor the “down” conformation of the RBD, possibly explaining the observed more effective shielding against neutralizing antibodies. In agreement with previous simulations22 the Man9 glycan at N234 of Spike partially inserts itself into the vacant space in the core of the trimer that is created when the RBD of monomer 3 is in the "up" conformation (Fig. S6). In our simulations, the free space created by the "up" conformation seems slightly smaller for Spike in complex with ACE2, suggesting that binding to ACE2 has a stabilizing effect on the Spike monomer. The N165Q mutant was experimentally found to be more sensitive to neutralization.33 In our models, the glycan at N165 is positioned directly next to the RBD (Fig. 1b) and thus could shield important antigenic sites. These data highlight the complex impact of Spike glycosylation on the intramolecular interactions of the Spike monomers and, critically, the interaction with ACE2, posing a challenge to design SARS-CoV-2 neutralizing moieties. 6 165 170 175 180 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . a b c ce n rre ccu O ce n rre ccu O ce n rre ccu O Figure 2. Role of Spike glycosylation in shielding the protein surface and interactions with hACE2. (a) Normalized distribution of the area of Spike protein atoms that is shielded by each of its glycans on monomer 1 (see Fig. S4 for monomers 2 and 3). (b) Normalized distribution of the number of atoms in contact with ACE2 and (c) the number of hydrogen bonds with ACE2 for glycans on monomer 3 of Spike (see Fig.
S5 for monomers 1 and 2). Since our modeling clearly confirmed that ACE2 glycosylation plays a significant role in its binding to Spike (Fig. 1d,e), we also determined the area of the Spike-ACE2 interface region, by subtracting the SASA of the complex from the SASA of the individual proteins and dividing by two. The total interface area was 24.6 nm², with glycans accounting for up to 51% of the interface area, i.e. 12.6 nm², contributed by the four most relevant glycans at positions N53, N90, N322 and N546 of ACE2 (Fig. 3). Furthermore, we scored the number of atoms of each ACE2 glycan in contact with Spike. A contact was defined as a distance of less than 0.4 nm between two atoms. This allowed us to identify the glycans at N53, N90, N322 and N546 as interacting with Spike, with the glycan at position N53 having the weakest interaction. Notably, N546 interacted with Spike for a significant amount of time only in one of the two independent simulations. The degree of interaction correlated with the spatial proximity between the glycans and the RBD (Fig. 1e, f). Assessing the number of hydrogen bonds that formed during the simulations, the glycans at N90 and N322 appear most prominent (Fig. 3b). Interestingly, the glycans at N90 and N322 interact directly with Spike protein atoms, while the glycan at N546 (red sticks in Fig. 1f) interacts with the glycans at N122 and N165 of Spike (dark green and orange sticks in Fig. 1b). These findings are in agreement with previously reported simulations of the complexes.21,25 7 185 190 195 200 205 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . a b c ce rren ccu O ce rren ccu O ce rren ccu O Figure 3. The role of hACE2 glycosylation in the interaction with Spike. (a) Normalized distribution of the number of atoms of glycans at N53, N90, N322 and N546 of ACE2 that are in contact with Spike (distance < 0.4 nm). (b) Normalized distribution of the number of hydrogen bonds between glycans at N53, N90, N322, N546 of ACE2 and Spike. (c) Normalized distribution of the interface area between Spike and glycans at N53, N90, N322 and N546 of ACE2. Next, we assessed the conformational freedom of ACE2 glycans upon binding to Spike and compared their respective density maps in the simulations of free ACE2, and ACE2 in complex with Spike (Fig. 4). The density map of the unbound ACE2 (Fig. 4a) shows a continuous density of glycans, largely covering the interface area. Formation of the ACE2-Spike complex significantly reduces the conformational freedom of the glycans, in particular the ones at N90 and N322 (Fig. 4b). We predict that the glycans at N90 and N322 hamper binding to Spike, either sterically or through an entropic penalty upon binding due to a loss of conformational freedom.
These glycans have been implicated as being relevant for binding before,21, as well as the glycan at N53,35 but no conclusions were drawn if they contribute positively or negatively to binding. Mehdipour and Hummer predicted the glycan at N322 to contribute favorably to binding, because of the favorable interactions of this glycan with the Spike surface.25 We did not observe a significantly more pronounced interaction with Spike for the glycan at N322, compared to the one at N90 (Fig. 3). Based on conformational considerations, we therefore rather predict a negative impact on binding for both glycans (Fig. 4). 8 210 215 220 225 230 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . a b Figure 4. Average location density maps of glycans on ACE2. (a) The density map (grey mesh) of the glycans at N53, N90, N322 and N546 as observed in the simulations of unbound ACE2 are superimposed onto the ACE2 – Spike complex. (b) The density map of the same glycans, as observed in the simulation of the ACE2 – Spike complex. ACE2 in grey, Spike in green. Single, randomly selected conformations of the glycans are shown in blue (N53), yellow (N90), black (N322) and red (N546). Since only the glycans at N90 and N322 directly interact with the protein atoms of the Spike proteins, while the glycan on N546 forms hydrogen bonds with glycans present on Spike, we set out to confirm the negative influence of N90 and N322 glycosylation on the interactions with Spike experimentally. First, we ablated N-glycosylation at N90 and N322 individually using the ACE2-Fc fusion constructs ACE2-T92Q-Fc (ref. 15) and ACE2-N322Q-Fc. Note that ref. 15 indeed suggests that removal of the glycan at N90 through a mutation of T92 leads to enhanced interaction with Spike. The same data set, however, suggests that removal of the glycan at N322 through a mutation of T324 most likely leads to reduced affinity to Spike. However, T324 is itself part of the interface with Spike (Fig. S7), and any mutation of this residue could easily disrupt ACE2 – Spike binding directly, rather than through its effect on the N322 glycosite. We therefore decided to mutate N322 into glutamine to prevent glycosylation at this position. The wild-type and mutant ACE2-Fc constructs were expressed in HEK293-6E cells and purified from the culture supernatants by protein A affinity chromatography to apparent homogeneity (Fig. S8). Analysis by size-exclusion chromatography combined with detection by multi-angle light scattering (SEC-MALS) demonstrated that all purified proteins were dimers of the expected native molecular mass (Fig. S9). The impact of the introduced mutations on the overall fold of ACE2-Fc was tested with differential scanning calorimetry (DSC), a sensitive biophysical method for the assessment of the thermal stability of proteins.
Three thermal transitions could be discriminated. The first midpoint of transition (Tm1) is due to the unfolding of ACE2, whereas 9 235 240 245 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . the second and third midpoints of transitions (Tm2 and Tm3) reflect the thermal denaturation of the CH2 and CH3 domains of the Fc part of the fusion proteins.36 The Tm1 midpoint transition temperatures of the ACE2-Fc glycomutants (53.3-54.0°C) were slightly higher than for the wild- type protein (52.2°C), while Tm2 and Tm3 remained unchanged (Fig. 5). This indicates that removal of the N90 and N322 glycans does not compromise the structural integrity of ACE2. Figure 5. Analysis of ACE2 variants by differential scanning calorimetry (DSC). Raw data (black) were smoothened (red) and then fitted using a non-two-state thermal unfolding model (grey). Data are presented as mean ± SEM of three independent experiments. Cp, heat capacitance; rshACE2, clinical-grade recombinant soluble human ACE2; deglyco-rshACE2, enzymatically deglycosylated rshACE2; deglyco-ACE2-wt-Fc, enzymatically deglycosylated wild-type ACE2-Fc; desialo-ACE2-wt-Fc, enzymatically desialylated wild-type ACE2-Fc. The Spike-binding properties of the purified ACE2-Fc variants were characterized by biolayer interferometry (BLI). For this, ACE2-wt-Fc, ACE2-T92Q-Fc and ACE2-N322Q-Fc were 10 250 255 260 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . biotinylated, immobilized on streptavidin biosensor tips and dipped into serial dilutions of trimeric Spike. Since we did not observe appreciable dissociation of ACE2-Fc/trimeric Spike complexes in our analyses (Fig. S10), we evaluated the association rates (kobs; Fig. 6a). To determine equilibrium affinity constants (KD), we analyzed the interactions between the immobilized ACE2-Fc constructs and monomeric RBD (Fig. 6b, Fig. S11). The BLI data are in good agreement with our computational models, confirming that the removal of protein N- glycosylation at either N90 or N322 results in up to 2-fold higher binding affinities, when compared to ACE2-wt-Fc (ACE2-wt-Fc: KD = 16.2 ± 0.7 nM; ACE2-T92Q-Fc: KD = 8.0 ± 0.7 nM; ACE2-N322Q-Fc: KD = 11.4 ± 0.3 nM; Fig. 6b; Fig. S11). Thus, structure-guided glyco- engineering at N90 and N322 results in ACE2 forms with increased affinity for SARS-CoV-2 Spike binding. Figure 6. Binding of Spike and RBD to glyco-engineered ACE2 variants. (a) Binding of Spike to glyco- engineered ACE2 variants as determined by biolayer interferometry (BLI).
Plots of kobs (observed association rate) as a function of Spike concentration were generated by fitting the association data to a 1:1 binding model. Binding analysis was performed by dipping ACE2-loaded biosensors into 2-fold serial dilutions of purified Spike (1.6-50 nM). All measurements were performed in triplicates. rshACE2, clinical-grade recombinant soluble human ACE2; deglyco-rshACE2, enzymatically deglycosylated rshACE2; deglyco-ACE2-wt-Fc, enzymatically deglycosylated 11 265 270 275 280 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . wild-type ACE2-Fc. (b) KD values for the interaction of the indicated glyco-engineered ACE2 variants with monomeric RBD. Data are presented as mean ± SEM of 3 independent experiments. Desialo-ACE2-wt-Fc, enzymatically desialylated wild-type ACE2-Fc. Next, we tested the virus neutralization properties of ACE2-wt-Fc, ACE2-T92Q-Fc and ACE2- N322Q-Fc. For this, we infected Vero E6 cells with 60 plaque-forming units (PFU; multiplicity of infection (MOI): 0.002) of SARS-CoV-2 in the presence of 10-50 µg/mL ACE2-wt-Fc, ACE2-T92Q-Fc, or ACE2-N322Q-Fc. The extent of SARS-CoV-2 infection and replication was quantified by RT-qPCR detection of viral RNA present in the culture supernatants. Untreated SARS-CoV-2 infected cells released up to 10 times more viral RNA than ACE2-wt-Fc treated cells. Importantly, co-incubation of cells with SARS-CoV-2 and ACE2-T92Q-Fc resulted in significant further reduction of the viral load when compared to ACE2-wt-Fc. Enhanced SARS- CoV-2 neutralization was also observed for ACE2-N322Q-Fc. However, this mutant was less effective in promoting virus neutralization than ACE2-T92Q-Fc (Figs. 7 and S12). Similar results were obtained when SARS-CoV-2 neutralization assays were performed with much larger amounts of inoculated virus (MOI: 20) and concomitantly increased ACE2-Fc concentrations (Fig. S13). Hence, in line with our structural glycan interaction map, the removal of either of the N-glycans attached to N90 and N322 gives rise to ACE2 decoy receptors with improved SARS- CoV-2 neutralization properties. 12 285 290 295 300 305 310 315 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . Figure 7. Critical role of ACE2 glycosylation for SARS-CoV-2 infectivity. Inhibition of SARS-CoV-2 infection of Vero E6 cells using wild-type ACE2-Fc and the indicated glyco-engineered ACE2-Fc variants at final concentrations of 10-50 µg/mL.
The viral RNA content of the culture supernatants was quantified by RT-qPCR and expressed as fold change reduction relative to untreated controls. Data are presented as mean ± SEM of 2-3 independent experiments each performed in triplicates. Deglyco-ACE2-wt-Fc, enzymatically deglycosylated wild- type ACE2-Fc; desialo-ACE2-wt-Fc, desialylated wild-type ACE2-Fc. P<0.05; P<0.01; **P<0.001 (Kruskal- Wallis). To investigate a potential additive effect of simultaneous elimination of N-glycosylation at N90 and N322, we generated a double mutant ACE2-T92Q-N322Q-Fc construct. We also digested ACE2-wt-Fc with peptide-N4-(N-acetyl-beta-glucosaminyl)asparagine amidase F (PNGase F) to remove all accessible N-glycans (deglyco-ACE2-wt-Fc) and neuraminidase to release terminal sialic acid residues (desialo-ACE2-wt-Fc). Purity and homogeneity of these additional ACE2-Fc variants was ascertained by SDS-PAGE and SEC-MALS (Figs. S8 and S9). The absence of N- glycans attached to N90 and/or N322 in ACE2-T92Q-N322Q-Fc and the respective single mutants was demonstrated by LC-ESI-MS (Fig. S14). Quantitative release of sialic acids and complete removal of N-glycans from all ACE2-wt-Fc N-glycosylation sites with the exception of N546 was also confirmed (Figs. S15 and S16). The glycans at N546 of ACE2-wt-Fc exhibited partial resistance (40%) to PNGase F treatment (Fig. S15). Combined introduction of the mutations T92Q and N322Q as well as enzymatic desialylation did not reduce the thermal stability of ACE2-Fc as assessed by DSC, while close-to-complete removal of N-glycans by PNGase F led to a slightly decreased midpoint transition temperature of the ACE2 domain (Fig. 5). Studies of the interaction between ACE2-T92Q-N322Q-Fc and deglyco-ACE2-wt-Fc with RBD by BLI analysis yielded KD values similar to those determined for the single mutant ACE2- T92Q-Fc (ACE2-T92Q-N322Q-Fc: KD = 8.2 ± 0.2 nM; deglyco-ACE2-wt-Fc: KD = 7.6 ± 0.3 nM). The affinity of desialo-ACE2-wt-Fc for RBD (KD = 11.3 ± 0.4 nM) was also higher than that of native ACE2-wt-Fc (Fig. 6b). The increased affinities of these ACE2-Fc variants for Spike correlate well with their potencies to neutralize SARS-CoV-2, with deglyco-ACE2-wt-Fc followed by ACE2-T92Q-N322Q-Fc displaying the highest neutralization potencies (Figs. 7 and S13). The effect of desialo-ACE2-wt-Fc on SARS-CoV-2 infections of Vero E6 cells was less pronounced and comparable to that of the single mutant ACE2-N322Q-Fc (Fig. 7), in good agreement with the almost identical RBD-binding affinities of these two ACE2-Fc variants (Fig. 13 320 325 330 335 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . 6b). Taken together, these data identify critical glycans at position N90 and N322 of ACE2 that structurally and functionally interfere with Spike-ACE2 binding; ablation of these glycans via site-directed mutagenesis or enzymatic deglycosylation generated ACE2 variants with improved Spike-binding properties and increased neutralization strength.
The results presented above uncover the critical importance of N-glycans located at the ACE2- Spike interface for the infection of host cells by SARS-CoV-2. This prompted us to test the feasibility of removing all N-glycans from clinical-grade rshACE2, which has undergone placebo-controlled phase 2 clinical testing in 178 COVID-19 patients (ClinicalTrials.gov Identifier: NCT04335136), and to test for its SARS-CoV-2 neutralization properties. To this end, we generated enzymatically deglycosylated clinical-grade rshACE2 (deglyco-rshACE2) using PNGase F. The quantitative release of all N-glycans, with the exceptions of those attached to the N432 and N546 glycosites, was confirmed by LC-ESI-MS/MS (Fig. S15), and the integrity and homogeneity of dimeric deglyco-rshACE2 was demonstrated by SEC-MALS (Fig. S9). Paralleling our observations with deglyco-ACE2-wt-Fc, we found the binding affinity of deglyco-rshACE2 to RBD (KD = 5.1 ± 0.5 nM) to be two times higher than for native rshACE2 (KD = 10.5 ± 0.4 nM; Fig. 6b). Furthermore, deglyco-rshACE2 displayed improved SARS-CoV- 2 neutralization properties in Vero E6 cell infection assays. At a final concentration of 200 µg/mL deglyco-rshACE2, we observed a significant reduction in SARS-CoV-2 replication when compared to treatment with the native form of the protein (Fig. 8). 14 340 345 350 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . Figure 8. Deglycosylated rshACE2 is a potent SARS-CoV-2 decoy receptor. (a) Inhibition of SARS-CoV-2 infection of Vero E6 cells at an MOI of 20 using native and enzymatically deglycosylated rshACE2 at final concentrations of 200 µg/mL. The viral RNA content of the infected cells was quantified by RT-qPCR and expressed as neutralization efficiency relative to native rshACE2 (set to 100%). Data are presented as mean ± SEM of 4 independent experiments. Deglyco-rshACE2, enzymatically deglycosylated rshACE2; P = 0.0026 (Student’s t- test). (b) Inhibition of SARS-CoV-2 infection of Vero E6 cells at an MOI of 20 using deglyco-rshACE2 at final concentrations of 50-200 µg/mL. The viral RNA content of the infected cells was quantified by RT-qPCR and expressed as fold change reduction relative to untreated controls. Data are presented as mean ± SD of triplicates. Besides serving as a soluble decoy receptor to prevent SARS-CoV-2 infection of ACE2- expressing host cells, rshACE2 also regulates blood pressure and protects multiple organs such as the heart, kidney and lung as well as blood vessels via enzymatic degradation of angiotensin II.37 In contrast to other recently described ACE2 mutants displaying improved Spike binding concomitant with inadvertently or intentionally impaired enzymatic activity,14,15 the catalytic 15 355 360 365 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021.
The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . activities of ACE2-T92Q-Fc, ACE2-N322Q-Fc and ACE2-T92Q-N322Q-Fc were found to be only modestly reduced as compared to ACE2-wt-Fc (ACE2-T92Q-Fc: 65 ± 11 %; ACE2- N322Q-Fc: 69 ± 7 %; ACE2-T92Q-N322Q-Fc: 79 ± 11 %; Fig. 9). Figure 9. Enzymatic activity of ACE2-Fc mutants. Hydrolysis of 100 µM 7-methoxycoumarin-4-yl-acetyl-Ala- Pro-Lys-2,4-dinitrophenyl was continuously monitored by spectrofluorimetry. Hydrolytic activity is plotted as relative fluorescence units (RFU) over ACE2-Fc concentration (in nM). All assays were performed in technical triplicates. One representative experiment out of two is shown. Interestingly, deglyco-ACE2-wt-Fc (149 ± 1 %) and desialo-ACE2-wt-Fc (160 ± 2 %) exhibited higher enzymatic activities than native ACE2-wt-Fc (Figs. 10 and S17). A similar observation was made for deglyco-rshACE2, although the enhancing effects of enzymatic deglycosylation on catalytic efficiency were less pronounced (113 ± 2 % as compared to native rshACE2; Fig. 10). These results show that enzymatic removal of N-glycans from ACE2-Fc and clinical-grade 16 370 375 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . rshACE2 results in increased Spike binding and enhanced SARS-CoV-2 neutralization while preserving its potentially critical enzymatic activity. Figure 10. Enzymatic activity of enzymatically deglycosylated ACE2. Hydrolysis of 100 µM 7- methoxycoumarin-4-yl-acetyl-Ala-Pro-Lys-2,4-dinitrophenyl was continuously monitored by spectrofluorimetry. Hydrolytic activity is plotted as relative fluorescence units (RFU) over ACE2 concentration (in nM). All assays were performed in technical triplicates. One representative experiment out of two is shown. Discussion Our data demonstrate that structure-guided glycoengineering is a powerful means to develop ACE2 variants with improved SARS-CoV-2 neutralization properties without compromising the structural stability and catalytic activity of the enzyme. Our in silico models of the Spike-ACE2 complex combined with simulations of its spatial and temporal dynamics rationalized previously published data and led to predictions that were confirmed by in vitro binding studies and cell- 17 380 385 390 395 400 405 410 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
It is made available under a CC-BY-ND 4.0 International license . based SARS-CoV-2 neutralization assays. It is of note that the moderately enhanced affinity of ACE2 glycovariants for monomeric RBD observed in biolayer interferometry experiments relates to a far more pronounced increase of their virus-neutralization potency. This may be explained by multiple cooperative effects. First, a cooperative effect may be expected for the association of trimeric Spike molecules present in the viral envelope with membrane-bound ACE2 dimers. In this supramolecular setting, a subtle increase in the affinity of ACE2 for RBD can lead to a dynamic equilibrium of binding and unbinding events with up to six potential interactions, leading to an overall much stronger avidity effect. Second, a slight advantage of the soluble ACE2 decoy receptor over endogenous native ACE2 may be sufficient to tip the balance between SARS-CoV-2 attachment and shedding of viral particles from the host cell surface. Third, blocking of initial binding prevents viral propagation and hence spread of the virus to surrounding cells. Finally, it is possible that the N-glycan moiety of ACE2 also modulates other aspects of viral entry besides promoting the docking of Spike to the cell surface.29 It has been reported that the sialylation status of ACE2 affects its interactions with SARS-CoV-2 Spike.30 We have found that enzymatic desialylation of ACE2 results in a reproducible increase of its affinity to RBD without detectable structural penalties. Importantly, desialylated ACE2 is more efficient in neutralizing SARS-CoV-2 than its native counterpart. Molecular simulations suggest that the terminal sialic acids of the N-glycans attached to ACE2 residues N90 and N322 mask parts of the Spike-ACE2 interface and thus could interfere with Spike binding through steric clashes and/or electrostatic effects (Fig. S18). This provides a structural rationale how sialic acids present on ACE2 might dampen interactions with Spike during SARS-CoV-2 attachment to host cells.38 In line with other reports,15,38 our results indicate that the elimination of the N-glycans attached to N90 is largely responsible for the improved Spike-binding properties of enzymatically deglycosylated ACE2. As proposed14,15 and corroborated by our mutational analysis, substitution of ACE2 residues N90 or T92 could indeed provide an alternative approach for the development of ACE2 variants with improved SARS-CoV-2 sequestering properties. Our data indicate that ablation of N90 glycosylation could be combined with mutations of N322 and possibly other ACE2 N-glycosylation sites to achieve an even higher SARS-CoV-2 neutralizing potency. However, expression of an ACE2 variant lacking all potential N-glycosylation sites in ACE2- negative host cells led to reduced rather than enhanced susceptibility of the cells to SARS-CoV-2 18 415 420 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
It is made available under a CC-BY-ND 4.0 International license . as compared to transduction with wild-type ACE2.38 This was attributed to the much lower cellular content of the mutant protein relative to the native enzyme, thus demonstrating that the importance of N-glycosylation for proper folding of glycoproteins during their biosynthesis39 also applies to ACE2. Given the inferior expression yields of glycan-free ACE2 and the potential of unwanted immunological side effects when non-natural mutations are introduced into a therapeutic glycoprotein, we believe that the clinical potential of enzymatically deglycosylated rshACE2 is superior to that of any of our ACE2 glycomutants. In our opinion, treatment of clinical-grade rshACE2 with deglycosylation enzymes such as PNGase F followed by a final polishing step represents a straightforward, Good Manufacturing Practice (GMP)-compliant and industrially feasible alternative to generate a potent therapeutic drug for the treatment of SARS- CoV-2 infected persons and patients. 19 425 430 435 440 445 450 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . Materials and Methods Modeling of the Spike-hACE2 complex To model the fully glycosylated SARS-CoV-2 Spike-human ACE2 (hACE2) complex, a protein model was created using partial experimental structures deposited in the protein databank (PDB). The Spike RBD domain in complex with hACE217 (PDB: 6M17) was superimposed with the opened RBD domain in a Spike structure with one open RBD domain18 (PDB: 6VYB). Alternative Spike structures have been published, which show very similar conformations.19 Similarly, further structures of the Spike RBD-hACE2 complex4,16 have been reported which show very similar conformations to the templates used. Missing residues in Spike were modeled using SWISS-MODEL40 and the superimposed structure as template based on the complete SARS-CoV-2 S sequence (GenBank QHD43416.1). Different types of glycans were added to Spike and hACE2. For Spike, the assignments of Watanabe et al.20 were followed, selecting oligomannosidic (Man5 or Man9) or complex (bi- antennary di-sialylated core fucosylated; NaNaF) glycans according to the majority of the glycans detected at the respective site. This was largely confirmed by our own analysis.10 For hACE2, complex (i.e. bi-antennary di-sialylated core fucosylated) N-glycans were added. See Table S1 for the exact assignments. Initial conformations of the glycans were selected following previously derived procedures.41 In brief, molecular dynamics simulations were performed of mini-peptides with the glycans attached. Local Elevation42 was used to enhance the sampling of all glycosidic linkages, during simulations of 100 ns.
The entire glycan trees were clustered based on the conformations of the individual glycosidic linkages.43 This resulted in conformational bundles containing 1301, 1340 and 2413 distinct conformations of Man5, Man9 and NaNaF, respectively. These conformations were fitted onto the respective glycosylation site in the Spike-hACE2 complex using a superposition of the backbone of the asparagine residues and the non-bonded interaction energy between the glycan and protein atoms or previously added glycans was computed. The lowest energy conformation was retained. Topologies and initial conformations were generated using the gromos++ suite of pre- and post-MD tools.44 Glycans were added to the complex sequentially, to avoid collisions between individual glycans. A few modeled glycans were incompatible with loops of the Spike protein not resolved in the experimental structures. Loops involved in these structural incompatibilities (residues 141-165 20 455 460 465 470 475 480 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . and 471-490) were partially re-modeled in the fully glycosylated model using the RCD+ loop modeling server.45,46 The final model was energy-minimized with the GROMOS 54A8 protein force-field,47,48 the GROMOS 53A6glyc glycan force-field41,49,50 and the GROMOS simulation software using the steepest decent algorithm.51 Molecular dynamics simulations Molecular dynamics simulations were performed using the simulation package Gromacs (Version 2019.5) and the indicated force field parameters. hACE2 was reduced to residues 21 to 730 in the models, to reduce its overall size prior to simulation. The models were placed in rhombic dodecahedron simulation boxes and solvated by explicit SPC water molecules.52 This resulted in simulation systems of 5.9 × 105 and 2.2 × 106 atoms for hACE2 and the Spike-hACE2 complex, respectively. Two independent 100-ns molecular dynamics simulations were performed for hACE2 and for the Spike-hACE2 complex each. The equations of motion were integrated using a leapfrog integration scheme53 with a time-step of 2 fs. Non-bonded interactions were calculated within a cutoff sphere of 1.4 nm and electrostatic interactions were computed using a particle-particle particle-mesh (P3M) approach.54 Bond-lengths were constrained to their optimal values using the Lincs algorithm.55 Temperature was maintained at a constant value using a velocity-rescaling algorithm56,57 with a relaxation time of 0.1 ps. Pressure was maintained constant using a Parrinello-Rahman barostat58,59 with a relaxation time of 2.0 ps and an estimated isothermal compressibility of 4.5 × 10–5 bar–1. Configurations were stored every 10 ps for subsequent analyses. Hydrogen bonds were identified using a geometric criterion.
A hydrogen bond was logged if the donor-acceptor distance is within 0.25 nm and the donor- hydrogen-acceptor angle was larger than 135 degrees. The solvent-accessible surface area was determined by rolling a probe with diameter 0.14 nm over the surface of the protein, using slices of 0.005 nm width. An atom contact was assigned if the distance between two atoms were within 0.4 nm. The distributions of atom contacts, hydrogen bonds and solvent-accessible surface area were estimated using a kernel density estimator with gaussian kernels. Distributions obtained from the first and second half of the simulations were compared to ensure convergence. Glycan densities were calculated using the program GROmaps.60 21 485 490 495 500 505 510 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . Recombinant expression of proteins Soluble recombinant human ACE2 (rshACE2) was provided by Apeiron Biologicals (Vienna, Austria). Recombinant expression of all other proteins was performed by transient transfection of HEK293-6E cells, licensed from National Research Council (NRC) of Canada, as previously described.36,61 Cells were cultivated in FreeStyle F17 expression medium supplemented with 0.1% (v/v) Pluronic F-68 and 4 mM L-glutamine (all from Thermo Fisher Scientific, United States) in shaking flasks at 37°C, 8% CO2, 80% humidity and 130 rpm in a Climo-Shaker ISF1- XC (Adolf Kühner AG, Switzerland). pCAGGS vector constructs containing either the sequence of the SARS-CoV-2 RBD (residues R319-F541) or the complete luminal domain of Spike, modified in terms of removal of the polybasic furin cleavage site and introduction of two stabilizing point mutations (K986P and V987P), were kindly provided by Florian Krammer, Icahn School of Medicine at Mount Sinai (New York, United States).62,63 Plasmid constructs pcDNA3-sACE2(WT)-Fc(IgG1) and pcDNA3-sACE2-T92Q-Fc(IgG1) were obtained from Addgene (United States). The N322Q mutation was introduced into ACE2-wt-Fc and ACE2- T92Q-Fc using the QuikChange Lightning Site-Directed-Mutagenesis kit (Agilent Technologies, United States) according to the manufacturer’s instructions and the respective parental vector as template. High quality plasmid preparations for expression of ACE2-Fc variants were prepared using the PureYield Plasmid Midiprep System (Promega, United States). Transient transfection of the cells was performed at a cell density of approximately 1.7×106 cells mL-1 culture volume μ μ using a total of 1 g of plasmid DNA and 2 g of linear 40-kDa polyethylenimine (Polysciences Inc., Germany) per mL culture volume. 48 h and 96 h after transfection, cells were supplemented with 0.5% (w/v) tryptone N1 (Organotechnie, France) and 0.25% (w/v) D(+)-glucose (Carl Roth, Germany).
Soluble proteins were harvested after 120-144 h by centrifugation (10 000 g, 15 min, 4°C). Purification of recombinantly expressed proteins μ After filtration through 0.45 m membrane filters (Merck Millipore, Germany), supernatants containing RBD or soluble Spike were concentrated and diafiltrated against 20 mM sodium phosphate buffer containing 500 mM NaCl and 20 mM imidazole (pH 7.4) using a Labscale TFF system equipped with a 5 kDa cut-off Pellicon XL device (Merck Millipore). The His-tagged proteins were captured using a 5 mL HisTrap FF crude column connected to an ÄKTA pure 22 515 520 525 530 535 540 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . chromatography system (both from Cytiva, United States). Bound proteins were eluted by applying a linear gradient of 20 to 500 mM imidazole over 20 column volumes. ACE2-Fc variants were purified by affinity chromatography using a 5 mL HiTrap Protein A column (Cytiva) according to the manufacturer’s instructions and 0.1 M glycine-HCl (pH 3.5) for elution. Eluate fractions were immediately neutralized using 2 M Tris (pH 12.0). Fractions containing the protein of interest were pooled, concentrated using Vivaspin 20 Ultrafiltration Units (Sartorius, Germany) and dialyzed against PBS (pH 7.4) at 4°C overnight using SnakeSkin Dialysis Tubing (Thermo Fisher Scientific). The RBD was further purified by size exclusion chromatography (SEC) using a HiLoad 16/600 Superdex 200 pg column (Cytiva) eluted with PBS. All purified proteins were stored at -80°C until further use. Enzymatic deglycosylation and desialylation of ACE2 For deglycosylation of ACE2-wt-Fc and rshACE2, proteins (2 mg mL-1) were incubated with 180000 U mL-1 PNGase F (New England Biolabs, Unites States) in PBS (pH 7.4) for 24 h at 37°C. Desialylation of ACE2-wt-Fc was performed with 2500 U mL-1 neuraminidase (New England Biolabs) in 50 mM sodium citrate (pH 5.0) under otherwise identical conditions. The deglycosylated or desialylated ACE2 variants were purified by preparative SEC using a HiLoad 16/600 Superdex 200 pg column eluted in PBS. The extent of enzymatic deglycosylation and desialylation was assessed by SDS-PAGE (Fig. S8), SEC-MALS (Fig. S9) and ESI-LC-MS/MS (Figs. S15 and S16). Bio-Layer Interferometry (BLI) measurements Interaction studies were performed on an Octet RED96e system using high precision streptavidin (SAX) biosensors (both from ForteBio, United States). Thus, all capture molecules (ACE2-wt- Fc, ACE2-T92Q-Fc, ACE2-N322Q-Fc, ACE2-T92Q-N322Q-Fc, deglyco-ACE2-wt-Fc, desialo- ACE2-wt-Fc, rshACE2 and deglyco-rshACE2) were biotinylated using the EZ-Link Sulfo-NHS- LC-Biotin kit (Thermo Fisher Scientific). Excess sulfo-NHS-LC-biotin was quenched by adding Tris-HCl buffer (800 mM, pH 7.4) to a final concentration of 3 mM.
Biotinylated proteins were further purified using PD-10 desalting columns (Cytiva) according to the manufacturer’s protocol. All assays were conducted in PBS supplemented with 0.05% (v/v) Tween 20 and 0.1% (w/v) BSA (PBST-BSA) at 25°C with the plate shaking at 1000 rpm. The SAX biosensors were first equilibrated in PBST-BSA and then dipped into a 34 nM solution of the respective 23 545 550 555 560 565 570 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . biotinylated capture molecule until a signal threshold of 0.8 nm was reached. Subsequently, the biosensors were dipped into PBST-BSA for 90 sec to record a baseline, before they were submerged into different concentrations of RBD or the Spike protein to record association rates. For binding analysis of trimeric Spike, all biosensors were dipped into 2-fold serial dilutions of the protein (1.6-50 nM). To determine KD values, titration of RBD was performed at different concentrations to cover a broad concentration range around the respective KD value.64 Biosensors loaded with ACE2 variants were submerged into 2-fold (6.25-200 nM) or 3-fold (0.8-200 nM) serial dilutions of RBD as appropriate for 600 sec. For dissociation, the biosensors were dipped into PBST-BSA for 300 sec (for analysis of Spike) or 100 sec (for analysis of RBD). Each experiment included a baseline measurement using PBST-BSA (negative control) as well as a positive control (RBD). Of note, no unspecific binding of RBD or the Spike protein to SAX biosensors was observed. Data were evaluated under consideration of the limit of detection (LOD) and limit of quantification (LOQ) as reported elsewhere.65,66 Each experiment was performed 3 times. Analysis was performed using the Octet data analysis software version 11.1.1.39 (ForteBio) according to the manufacturer’s guidelines. SDS-PAGE SDS-PAGE was carried out using a 4-15% MINI-PROTEAN TGX Stain-Free Protein Gel, the Mini-PROTEAN Tetra Vertical Electrophoresis Cell (both from Bio-Rad Laboratories Inc., United States) and SDS-PAGE running buffer (20 mM Tris, 200 mM glycine, 0.1% (w/v) SDS). μ 1 g of each purified protein was mixed with SDS sample buffer (62.5 mM Tris/HCl (pH 6.8), 2.5% (w/v) SDS, 10% (w/v) glycerol, 0.01% (w/v) bromophenol blue), heated to 70°C for 10 min and loaded onto the gel. For reducing conditions purified samples were mixed with SDS- β PAGE sample buffer containing 0.75 M mercaptoethanol and heated to 95°C for 5 min prior to loading. The PageRuler Unstained Protein Ladder (Thermo Fisher Scientific) was used as a size marker. Protein bands were visualized with the Gel Doc XR+ Imager (Bio-Rad Laboratories). Size-Exclusion Chromatography - Multi-Angle Light Scattering (SEC-MALS) Size-exclusion chromatography combined with multi-angle light scattering was performed to determine the homogeneity and the native molecular mass of all proteins under study.
Analyses were performed on an LC20 Prominence HPLC equipped with a refractive index detector RID- 24 575 580 585 590 595 600 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . 10A and the photodiode array detector SPD-M20A (all from Shimadzu, Japan). In-line MALS was analyzed either with a miniDAWN TREOS II MALS (for analysis of Spike) or a Heleos Dawn8+ plus QELS apparatus (Wyatt Technology, United States). Prior to analysis, all proteins were centrifuged (16 000 g, 10 min, 20°C) and filtered (0.1 µm Ultrafree-MC filter, Merck Millipore). Proper performance of the MALS detectors was validated with bovine serum albumin. Purified Spike was analyzed by injection of a total of 50 µg onto a Superose 6 Increase 10/300 GL column (Cytiva) at a flow rate of 0.25 mL min-1. The mobile-phase buffer used was PBS supplemented with 10% glycerol (pH 7.4). All other proteins were analyzed by using a Superdex 200 10/300 GL column (Cytiva) equilibrated with PBS plus 200 mM NaCl (pH 7.4). A total of 25 µg of each protein was injected and experiments were performed at a flow rate of 0.75 mL min-1. Data were analyzed using the ASTRA 6 software (Wyatt Technology). Differential scanning calorimetry (DSC) DSC experiments were performed using a MicroCal PEAQ-DSC Automated system (Malvern Panalytical, Germany), using 2.5 µM protein solutions in PBS (pH 7.4). The heating was performed from 20°C to 100°C at a rate of 1°C/min. The protein solution was then cooled in situ and an identical thermal scan was run to obtain the baseline for subtraction from the first scan. All measurements were performed in triplicates. Fitting was done with Origin 7.0 for DSC software using the non-2-state transition model. Peptide mapping and glycopeptide analysis All samples were analysed as in-solution proteolytic digests of the respective proteins by LC- ESI-MS/MS. For this, the pH of the samples was first adjusted to pH 7.8 by the addition of 1 M HEPES (pH 7.8) to a final concentration of 100 mM. The samples were then chemically reduced and S-alkylated, using a final concentration of 10 mM dithiothreitol for 30 min at 56°C, and a final concentration of 20 mM iodoacetamide for 30 min at room temperature in the dark. To maximize sequence coverage, proteins were digested for 18 h at 37°C with chymotrypsin (Roche, Germany), followed by 3 h at 37°C using trypsin (Promega). All proteolytic digests were acidified to pH 2 by addition of 10% formic acid and directly analyzed by LC-ESI-MS/MS, μ using a capillary BioBasic C18 reversed-phase column (BioBasic-18, 150 × 0.32 mm, 5 m, Thermo Fisher Scientific), installed in a Dionex Ultimate U3000 HPLC system (Thermo Fisher Scientific), developing a linear gradient from 95% eluent A (80 mM ammonium formate, pH 3.0, 25 605 610 615 620 625 630 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021.
The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . in HPLC-grade water) to 65% eluent B (80% acetonitrile in 80 mM ammonium formate, pH 3.0) over 50 min, followed by a linear gradient from 65% to 99% eluent B over 15 min, at a constant μ flow rate of 6 L/min, coupled to a maXis 4G Q-TOF instrument (Bruker Daltonics, Germany; equipped with the standard ESI source). For (glyco)peptide detection and identification, the mass spectrometer was operated in positive-ion DDA mode (i.e. switching to MS/MS mode for eluting peaks), recording MS scans in the m/z range from 150 to 2200 Th, with the 6 highest signals selected for MS/MS fragmentation. Instrument calibration was performed using a commercial ESI calibration mixture (Agilent Technologies). Site-specific profiling of protein glycosylation was performed using the dedicated Q-TOF data-analysis software packages Data Analyst (Bruker Daltonics) and Protein Scape (Bruker Daltonics), in conjunction with the MS/MS search engine MASCOT (Matrix Sciences Inc., United States) for automated peptide identification. ACE2 activity assays Enzymatic activity of ACE2 was determined and quantified as described previously,37 using 100 µM 7-methoxycoumarin-4-yl-acetyl-Ala-Pro-Lys-2,4-dinitrophenyl (Bachem, Switzerland) as substrate. SARS-CoV-2 neutralization assays All work with infectious SARS-CoV-2 was performed under BSL-3 conditions. Vero E6 cells (Biomedica, Austria) were grown in Minimum Essential Medium (MEM) containing Earle’s Salts, 1% penicillin/streptomycin stock solution and 2 mM L-glutamine (all from Thermo Fisher Scientific), supplemented with 5% fetal bovine serum (FBS), at 37°C and 5% CO2. A German 2019-nCoV isolate (Ref-SKU: 026V-03883, Charité, Berlin, Germany) was propagated in Vero E6 cells. The TCID50 titer of virus stocks was determined by the Reed-Munch method 67 and converted to plaque-forming units (pfu) using the conversion factor 0.7 (https://www.atcc.org/support/technical-support/faqs/converting-tcid-50-to-plaque-forming- units-pfu). Vero E6 cells were seeded in 48-well cell culture plates (3 × 104 cells per well) in MEM supplemented with 2% FBS overnight to reach approximately 80% confluence on the day of infection. ACE2 variants (final concentrations: 10-100 µg/mL) were preincubated with 60 pfu SARS-CoV-2 for 30 min at 37°C under constant shaking (300 rpm). After preincubation, Vero E6 cells were infected for 1 h at 37°C with samples containing either SARS-CoV-2 and ACE2 variants or solely SARS-CoV-2 (untreated controls) at a multiplicity of infection (MOI) of 0.002. 26 635 640 645 650 655 660 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
It is made available under a CC-BY-ND 4.0 International license . Subsequently, cells were washed two times with MEM to remove unadsorbed virus. After incubation for 24 h at 37°C in MEM supplemented with 2% FBS, viral RNA was extracted from the culture supernatant using the QiaAmp Viral RNA Minikit (Qiagen, Germany), according to the manufacturer’s protocol. SARS-CoV-2 replication was quantified via RT-qPCR using the QuantiTect Multiplex RT-qPCR Kit (Qiagen) with a Rotor Gene Q cycler (Qiagen). The reactions were performed in a total volume of 25 µL at 50°C for 30 min followed by 95°C for 15 min and 45 cycles of 95°C for 3 s and 55°C for 30 s. Forward primer: 2019-nCoV_N1-F 5’- GACCCCAAAATCAGCGAAAT-3’; reverse primer: 2019-nCoV_N1-R 5’- TCTGGTTACTGCCAGTTGAATCTG-3’; probe: 2019-nCoV_N1-P 5’-FAM- ACCCCGCATTACGTTTGGTGGACC-BHQ1-3’. Statistical analyses were conducted using GraphPad Prism 8. Significance was determined by Kruskal-Wallis, comparing the mean rank of the ACE2-wt-Fc group with the mean rank of every other group (, P<0.05; , P<0.01; **, P<0.001). SARS-CoV-2 neutralization assays were also performed independently in another laboratory using a different virus isolate. For these assays, Vero E6 cells (ATCC, United States) were grown in Dulbecco’s Modified Eagle’s Medium (DMEM, Thermo Fisher Scientific) supplemented with 1% non-essential amino acid stock solution (Thermo Fisher Scientific), 10 mM HEPES (Thermo Fisher Scientific) and 10% FBS at 37˚C and 5% CO2. SARS-CoV-2 isolated from a nasopharyngeal sample of a Swedish COVID-19 patient (GenBank accession number MT093571) was propagated in Vero E6 cells. Virus was titered using a plaque assay as previously described68 with fixation of cells 72 h post infection. Vero E6 cells were treated and infected as described previously.5 Briefly, Vero E6 cells were seeded in 48-well plates (5 × 104 cells per well) in DMEM containing 10% FBS. 24 h post-seeding, ACE2 variants (final concentrations: 50-200 µg/mL) were mixed with 106 pfu SARS-CoV-2 (MOI: 20) in a final volume of 100 µl DMEM containing 5% FBS, incubated for 30 min at 37˚C and then added to the cells. 15 h post-infection, cells were washed 3 times with PBS and then lysed using Trizol (Thermo Fisher Scientific) before analysis by RT-qPCR to quantify the content of SARS-CoV-2 RNA as described. 5 Acknowledgments 27 665 670 675 680 685 690 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . We thank Florian Krammer (Icahn School of Medicine at Mount Sinai, NY, United States) for providing the constructs used for production of recombinant Spike and RBD. pcDNA3- sACE2(WT)-Fc(IgG1) and pcDNA3-sACE2-T92Q-Fc(IgG1) were used with the kind permission of Erik Procko (University of Illinois, IL, United States).
Transfection-grade pCAGGS plasmids were provided by Rainer Hahn and Gerald Striedner (University of Natural Resources and Life Sciences Vienna, Austria) in the framework of the BOKU COVID-19 Initiative. The authors thank Irene Schaffner and Jakob Wallner (BOKU Core Facility Biomolecular & Cellular Analysis) for assisting in BLI measurements, Gerhard Stadlmayr (University of Natural Resources and Life Sciences Vienna, Austria) for performing SEC-MALS analysis of Spike and ForteBio for supporting the BOKU COVID-19 Initiative with SAX biosensors. The high-performance computing center (HLRS) of the University of Stuttgart is gratefully acknowledged for providing computational resources. Funding This project is supported by the PhD programme BioToP funded by the Austrian Science Fund (grant No. W1224-B09), Vienna Science and Technology Fund (WWTF; grant No. COV20- 015), and funding from the Innovative Medicines Initiative 2 Joint Undertaking (JU; grant agreement No. 101005026). The JU receives support from the European Union’s Horizon 2020 research and innovation programme and EFPIA. J.W.P. is a recipient of a DOC Fellowship of the Austrian Academy of Sciences (ÖAW) at the Institute for Molecular Modeling and Simulation at the University of Natural Resources and Life Sciences, Vienna (Grant No. 24987). J.M.P. and the research leading to these results has received funding from the T. von Zastrow foundation, the FWF Wittgenstein award (grant No. Z271-B19), the Austrian Academy of Sciences, the Canada 150 Research Chairs Program in Functional Genetics (grant No. F18- 0133), and the Canadian Institutes of Health Research COVID-19 grants F20-02343 and F20- 02015. Author contributions Tümay Capraz: investigation, formal analysis, writing original draft, visualization; Esther Föderl-Höbenreich: investigation, formal analysis, visualization, writing review and editing; 28 695 700 705 710 715 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . Clemens Grünwald-Gruber: investigation; Nikolaus F. Kienzl: investigation, formal analysis, visualization, writing review and editing; Elisabeth Laurent: investigation, formal analysis, writing original draft, visualization; Lukas Mach: investigation, supervision, formal analysis, writing original draft, writing review and editing; Daniel Maresch: investigation; Ali Mirazimi: supervision; Vanessa Monteil: investigation; Janine Niederhöfer: investigation; Chris Oostenbrink: conceptualization, supervision, writing original draft, writing review and editing; Josef M. Penninger: initiation of the project, conceptualization, writing review and editing; Jan W. Perthold: investigation, supervision, writing original draft; Johannes Stadlmann: conceptualization, investigation, supervision, formal analysis, writing original draft, writing review and editing; Gerald Wirnsberger: investigation, supervision, resources; Kurt Zatloukal: supervision, resources, writing review and editing.
Competing interests J.M.P. declares a conflict of interest as a founder, supervisory board member, and shareholder of Apeiron Biologics. J.N. and G.W. are employees of Apeiron Biologics. Apeiron holds a patent on the use of ACE2 for the treatment of lung, heart, or kidney injury and applied for a patent to treat COVID-19 with rshACE2. Data and materials availability Molecular models and simulation trajectories are available through the BioExcel COVID-19 Molecular Structure and Therapeutics Hub (https://covid.bioexcel.eu/simulations/). 29 720 725 730 735 740 745 750 755 760 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . References 1. 2. 3. 4. 5. 6. 7. 8. Zhu, N., Zhang, D., Wang, W., Li, X., Yang, B., Song, J., Zhao, X., Huang, B., Shi, W., Lu, R., Niu, P., Zhan, F., Ma, X., Wang, D., Xu, W., Wu, G., Gao, G. F., Tan, W., China Novel Coronavirus, I., et al. (2020) A Novel Coronavirus from Patients with Pneumonia in China, 2019. N Engl J Med 382, 727-733. doi: 10.1056/NEJMoa2001017 Wu, F., Zhao, S., Yu, B., Chen, Y. M., Wang, W., Song, Z. G., Hu, Y., Tao, Z. W., Tian, J. H., Pei, Y. Y., Yuan, M. L., Zhang, Y. L., Dai, F. H., Liu, Y., Wang, Q. M., Zheng, J. J., Xu, L., Holmes, E. C., and Zhang, Y. Z. (2020) A new coronavirus associated with human respiratory disease in China. Nature 579, 265-269. doi: 10.1038/s41586-020- 2008-3 Zhou, P., Yang, X. L., Wang, X. G., Hu, B., Zhang, L., Zhang, W., Si, H. R., Zhu, Y., Li, B., Huang, C. L., Chen, H. D., Chen, J., Luo, Y., Guo, H., Jiang, R. D., Liu, M. Q., Chen, Y., Shen, X. R., Wang, X., et al. (2020) A pneumonia outbreak associated with a new coronavirus of probable bat origin. Nature 579, 270-273. doi: 10.1038/s41586-020-2012- 7 Wang, Q., Zhang, Y., Wu, L., Niu, S., Song, C., Zhang, Z., Lu, G., Qiao, C., Hu, Y., Yuen, K. Y., Wang, Q., Zhou, H., Yan, J., and Qi, J. (2020) Structural and Functional Basis of SARS-CoV-2 Entry by Using Human ACE2. Cell 181, 894-904 e899. doi: 10.1016/j.cell.2020.03.045 Monteil, V., Kwon, H., Prado, P., Hagelkruys, A., Wimmer, R. A., Stahl, M., Leopoldi, A., Garreta, E., Hurtado Del Pozo, C., Prosper, F., Romero, J. P., Wirnsberger, G., Zhang, H., Slutsky, A. S., Conder, R., Montserrat, N., Mirazimi, A., and Penninger, J. M. (2020) Inhibition of SARS-CoV-2 Infections in Engineered Human Tissues Using Clinical-Grade Soluble Human ACE2. Cell doi: 10.1016/j.cell.2020.04.004 Monteil, V., Dyczynski, M., Lauschke, V. M., Kwon, H., Wirnsberger, G., Youhanna, S., Zhang, H., Slutsky, A. S., Hurtado Del Pozo, C., Horn, M., Montserrat, N., Penninger, J. M., and Mirazimi, A. (2021) Human soluble ACE2 improves the effect of remdesivir in SARS-CoV-2 infection. EMBO Mol Med 13, e13426. doi: 10.15252/emmm.202013426 Zoufaly, A., Poglitsch, M., Aberle, J. H., Hoepler, W., Seitz, T., Traugott, M., Grieb, A., Pawelka, E., Laferl, H., Wenisch, C., Neuhold, S., Haider, D., Stiasny, K., Bergthaler, A., Puchhammer-Stoeckl, E., Mirazimi, A., Montserrat, N., Zhang, H., Slutsky, A. S., et al.
(2020) Human recombinant soluble ACE2 in severe COVID-19. Lancet Respir Med 8, 1154-1158. doi: 10.1016/S2213-2600(20)30418-5 Barnes, C. O., Jette, C. A., Abernathy, M. E., Dam, K. A., Esswein, S. R., Gristick, H. B., Malyutin, A. G., Sharaf, N. G., Huey-Tubman, K. E., Lee, Y. E., Robbiani, D. F., Nussenzweig, M. C., West, A. P., Jr., and Bjorkman, P. J. (2020) SARS-CoV-2 neutralizing antibody structures inform therapeutic strategies. Nature 588, 682-687. doi: 10.1038/s41586-020-2852-1 Chan, J. F. W., Oh, Y. J., Yuan, S., Chu, H., Yeung, M. L., Canena, D., Chan, C. C. S., Poon, V. K. M., Chan, C. C. Y., Zhang, A. J., Cai, J. P., Wen, L., Chik, K. K. H., Shuai, H., Wang, Y., Hou, Y., Luo, C., Chan, W. M., Sit, K. Y., et al. (2021) A molecularly 181, 905-913 e907. 9. 30 765 770 775 780 785 790 795 800 805 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . 10. engineered, broad-spectrum anti-coronavirus lectin inhibits SARS-CoV-2 and MERS- CoV infection in vivo. Submitted to Nature. doi: Hoffmann, D., Mereiter, S., Jin Oh, Y., Monteil, V., Elder, E., Zhu, R., Canena, D., Hain, L., Laurent, E., Grunwald-Gruber, C., Klausberger, M., Jonsson, G., Kellner, M. J., Novatchkova, M., Ticevic, M., Chabloz, A., Wirnsberger, G., Hagelkruys, A., Altmann, F., et al. (2021) Identification of lectin receptors for conserved SARS-CoV-2 glycosylation doi: 10.15252/embj.2021108375 sites. EMBO J 10.15252/embj.2021108375, e108375. 11. Weisblum, Y., Schmidt, F., Zhang, F., DaSilva, J., Poston, D., Lorenzi, J. C., Muecksch, F., Rutkowska, M., Hoffmann, H. H., Michailidis, E., Gaebler, C., Agudelo, M., Cho, A., Wang, Z., Gazumyan, A., Cipolla, M., Luchsinger, L., Hillyer, C. D., Caskey, M., et al. (2020) Escape from neutralizing antibodies by SARS-CoV-2 spike protein variants. Elife 9, e61312. doi: 10.7554/eLife.61312 Li, D., Sempowski, G. D., Saunders, K. O., Acharya, P., and Haynes, B. F. (2021) SARS- CoV-2 Neutralizing Antibodies for COVID-19 Prevention and Treatment. Annu Rev Med 10.1146/annurev-med-042420-113838. doi: 10.1146/annurev-med-042420-113838 Yuan, M., Huang, D., Lee, C. D., Wu, N. C., Jackson, A. M., Zhu, X., Liu, H., Peng, L., van Gils, M. J., Sanders, R. W., Burton, D. R., Reincke, S. M., Pruss, H., Kreye, J., Nemazee, D., Ward, A. B., and Wilson, I. A. (2021) Structural and functional ramifications of antigenic drift in recent SARS-CoV-2 variants. Science 373, 818-823. doi: 10.1126/science.abh1139 Glasgow, A., Glasgow, J., Limonta, D., Solomon, P., Lui, I., Zhang, Y., Nix, M. A., Rettko, N. J., Zha, S., Yamin, R., Kao, K., Rosenberg, O. S., Ravetch, J. V., Wiita, A. P., Leung, K. K., Lim, S. A., Zhou, X. X., Hobman, T. C., Kortemme, T., et al. (2020) Engineered ACE2 receptor traps potently neutralize SARS-CoV-2.
Proc Natl Acad Sci U S A 117, 28046-28055. doi: 10.1073/pnas.2016093117 Chan, K. K., Dorosky, D., Sharma, P., Abbasi, S. A., Dye, J. M., Kranz, D. M., Herbert, A. S., and Procko, E. (2020) Engineering human ACE2 to optimize binding to the spike protein of SARS coronavirus 2. Science 369, 1261-1265. doi: 10.1126/science.abc0870 Lan, J., Ge, J., Yu, J., Shan, S., Zhou, H., Fan, S., Zhang, Q., Shi, X., Wang, Q., Zhang, L., and Wang, X. (2020) Structure of the SARS-CoV-2 spike receptor-binding domain bound to the ACE2 receptor. Nature 581, 215-220. doi: 10.1038/s41586-020-2180-5 Yan, R., Zhang, Y., Li, Y., Xia, L., Guo, Y., and Zhou, Q. (2020) Structural basis for the recognition of SARS-CoV-2 by full-length human ACE2. Science 367, 1444-1448. doi: 10.1126/science.abb2762 14. 12. 13. 15. 16. 17. 18. Walls, A. C., Park, Y. J., Tortorici, M. A., Wall, A., McGuire, A. T., and Veesler, D. (2020) Structure, Function, and Antigenicity of the SARS-CoV-2 Spike Glycoprotein. Cell 181, 281-292 e286. doi: 10.1016/j.cell.2020.02.058 19. Wrapp, D., Wang, N., Corbett, K. S., Goldsmith, J. A., Hsieh, C. L., Abiona, O., Graham, B. S., and McLellan, J. S. (2020) Cryo-EM structure of the 2019-nCoV spike in the prefusion conformation. Science 367, 1260-1263. doi: 10.1126/science.abb2507 20. Watanabe, Y., Allen, J. D., Wrapp, D., McLellan, J. S., and Crispin, M. (2020) Site- specific glycan analysis of the SARS-CoV-2 spike. Science 369, 330-333. doi: 10.1126/science.abb9983 Zhao, P., Praissman, J. L., Grant, O. C., Cai, Y., Xiao, T., Rosenbalm, K. E., Aoki, K., Kellman, B. P., Bridger, R., Barouch, D. H., Brindley, M. A., Lewis, N. E., Tiemeyer, 21. 31 810 815 820 825 830 835 840 845 850 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . 22. 23. 24. 25. Mehdipour, A. R., and Hummer, G. (2021) Dual nature of human ACE2 glycosylation in to SARS-CoV-2 spike. Proc Natl Acad Sci U S A 118. doi: binding 10.1073/pnas.2100425118 Sztain, T., Ahn, S. H., Bogetti, A. T., Casalino, L., Goldsmith, J. A., Seitz, E., McCool, R. S., Kearns, F. L., Acosta-Reyes, F., Maji, S., Mashayekhi, G., McCammon, J. A., Ourmazd, A., Frank, J., McLellan, J. S., Chong, L. T., and Amaro, R. E. (2021) A glycan gate controls opening of the SARS-CoV-2 spike protein. Nat Chem 10.1038/s41557-021- 00758-3. doi: 10.1038/s41557-021-00758-3 Cong, Y., Feng, Y., Ni, H., Zhi, F., Miao, Y., Fang, B., Zhang, L., and Zhang, J. Z. H. (2021) Anchor-Locker Binding Mechanism of the Coronavirus Spike Protein to Human ACE2: Insights from Computational Analysis. Journal of Chemical Information and Modeling 61, 3529-3542. doi: 10.1021/acs.jcim.1c00241 Delgado, J. M., Duro, N., Rogers, D. M., Tkatchenko, A., Pandit, S. A., and Varma, S. (2021) Molecular basis for higher affinity of SARS-CoV-2 spike RBD for human ACE2 receptor.
Proteins 89, 1134-1144. doi: 10.1002/prot.26086 Yang, Q., Hughes, T. A., Kelkar, A., Yu, X., Cheng, K., Park, S., Huang, W. C., Lovell, J. F., and Neelamegham, S. (2020) Inhibition of SARS-CoV-2 viral entry upon blocking N- and O-glycan elaboration. Elife 9. doi: 10.7554/eLife.61552 Allen, J. D., Watanabe, Y., Chawla, H., Newby, M. L., and Crispin, M. (2021) Subtle Influence of ACE2 Glycan Processing on SARS-CoV-2 Recognition. J Mol Biol 433, 166762. doi: 10.1016/j.jmb.2020.166762 Sun, Z., Ren, K., Zhang, X., Chen, J., Jiang, Z., Jiang, J., Ji, F., Ouyang, X., and Li, L. (2020) Mass Spectrometry Analysis of Newly Emerging Coronavirus HCoV-19 Spike Protein and Human ACE2 Reveals Camouflaging Glycans and Unique Post-Translational doi: Modifications. 10.1016/j.eng.2020.07.014 Grant, O. C., Montgomery, D., Ito, K., and Woods, R. J. (2020) Analysis of the SARS- CoV-2 spike protein glycan shield reveals implications for immune recognition. Sci Rep 10, 14991. doi: 10.1038/s41598-020-71748-7 25. Mehdipour, A. R., and Hummer, G. (2021) Dual nature of human ACE2 glycosylation in to SARS-CoV-2 spike. Proc Natl Acad Sci U S A 118. doi: binding 10.1073/pnas.2100425118 Sztain, T., Ahn, S. H., Bogetti, A. T., Casalino, L., Goldsmith, J. A., Seitz, E., McCool, R. S., Kearns, F. L., Acosta-Reyes, F., Maji, S., Mashayekhi, G., McCammon, J. A., Ourmazd, A., Frank, J., McLellan, J. S., Chong, L. T., and Amaro, R. E. (2021) A glycan gate controls opening of the SARS-CoV-2 spike protein. Nat Chem 10.1038/s41557-021- 00758-3. doi: 10.1038/s41557-021-00758-3 Cong, Y., Feng, Y., Ni, H., Zhi, F., Miao, Y., Fang, B., Zhang, L., and Zhang, J. Z. H. (2021) Anchor-Locker Binding Mechanism of the Coronavirus Spike Protein to Human ACE2: Insights from Computational Analysis. Journal of Chemical Information and Modeling 61, 3529-3542. doi: 10.1021/acs.jcim.1c00241 Delgado, J. M., Duro, N., Rogers, D. M., Tkatchenko, A., Pandit, S. A., and Varma, S. (2021) Molecular basis for higher affinity of SARS-CoV-2 spike RBD for human ACE2 receptor. Proteins 89, 1134-1144. doi: 10.1002/prot.26086 Yang, Q., Hughes, T. A., Kelkar, A., Yu, X., Cheng, K., Park, S., Huang, W. C., Lovell, J. F., and Neelamegham, S. (2020) Inhibition of SARS-CoV-2 viral entry upon blocking N- and O-glycan elaboration. Elife 9. doi: 10.7554/eLife.61552 Allen, J. D., Watanabe, Y., Chawla, H., Newby, M. L., and Crispin, M. (2021) Subtle Influence of ACE2 Glycan Processing on SARS-CoV-2 Recognition. J Mol Biol 433, 166762. doi: 10.1016/j.jmb.2020.166762 Sun, Z., Ren, K., Zhang, X., Chen, J., Jiang, Z., Jiang, J., Ji, F., Ouyang, X., and Li, L. (2020) Mass Spectrometry Analysis of Newly Emerging Coronavirus HCoV-19 Spike Protein and Human ACE2 Reveals Camouflaging Glycans and Unique Post-Translational doi: Modifications. 10.1016/j.eng.2020.07.014 Grant, O. C., Montgomery, D., Ito, K., and Woods, R. J. (2020) Analysis of the SARS- CoV-2 spike protein glycan shield reveals implications for immune recognition. Sci Rep 10, 14991. doi: 10.1038/s41598-020-71748-7 26.
27. 28. 29. 30. 31. Engineering (Beijing) 10.1016/j.eng.2020.07.014. 32. 32 855 860 865 870 875 880 885 890 895 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . 33. Li, Q., Wu, J., Nie, J., Zhang, L., Hao, H., Liu, S., Zhao, C., Zhang, Q., Liu, H., Nie, L., Qin, H., Wang, M., Lu, Q., Li, X., Sun, Q., Liu, J., Zhang, L., Li, X., Huang, W., et al. (2020) The Impact of Mutations in SARS-CoV-2 Spike on Viral Infectivity and Antigenicity. Cell 182, 1284-1294 e1289. doi: 10.1016/j.cell.2020.07.012 34. Mori, T., Jung, J., Kobayashi, C., Dokainish, H. M., Re, S., and Sugita, Y. (2021) Elucidation of interactions regulating conformational stability and dynamics of SARS- CoV-2 S-protein. Biophys J 120, 1060-1071. doi: 10.1016/j.bpj.2021.01.012 Barros, E. P., Casalino, L., Gaieb, Z., Dommer, A. C., Wang, Y., Fallon, L., Raguette, L., Belfon, K., Simmerling, C., and Amaro, R. E. (2021) The flexibility of ACE2 in the context doi: infection. Biophys 10.1016/j.bpj.2020.10.036 Lobner, E., Humm, A. S., Goritzer, K., Mlynek, G., Puchinger, M. G., Hasenhindl, C., Ruker, F., Traxlmayr, M. W., Djinovic-Carugo, K., and Obinger, C. (2017) Fcab-HER2 Interaction: a Menage a Trois. Lessons from X-Ray and Solution Studies. Structure 25, 878-889 e875. doi: 10.1016/j.str.2017.04.014 Vickers, C., Hales, P., Kaushik, V., Dick, L., Gavin, J., Tang, J., Godbout, K., Parsons, T., Baronas, E., Hsieh, F., Acton, S., Patane, M., Nichols, A., and Tummino, P. (2002) Hydrolysis of biological peptides by human angiotensin-converting enzyme-related carboxypeptidase. J Biol Chem 277, 14838-14843. doi: 10.1074/jbc.M200581200 Chu, H., Hu, B., Huang, X., Chai, Y., Zhou, D., Wang, Y., Shuai, H., Yang, D., Hou, Y., Zhang, X., Yuen, T. T., Cai, J. P., Zhang, A. J., Zhou, J., Yuan, S., To, K. K., Chan, I. H., Sit, K. Y., Foo, D. C., et al. (2021) Host and viral determinants for efficient SARS-CoV-2 infection of the human lung. Nat Commun 12, 134. doi: 10.1038/s41467-020-20457-w Xu, C., and Ng, D. T. (2015) Glycosylation-directed quality control of protein folding. Nat Rev Mol Cell Biol 16, 742-752. doi: 10.1038/nrm4073 39. 38. 37. 41. 35. of SARS-CoV-2 120, 1072-1084. J 36. 40. Waterhouse, A., Bertoni, M., Bienert, S., Studer, G., Tauriello, G., Gumienny, R., Heer, F. T., de Beer, T. A. P., Rempfer, C., Bordoli, L., Lepore, R., and Schwede, T. (2018) SWISS-MODEL: homology modelling of protein structures and complexes. Nucleic Acids Res 46, W296-W303. doi: 10.1093/nar/gky427 Turupcu, A., and Oostenbrink, C. (2017) Modeling of Oligosaccharides within Glycoproteins from Free-Energy Landscapes. Journal of Chemical Information and Modeling 57, 2222-2236. doi: 10.1021/acs.jcim.7b00351 Huber, T., Torda, A. E., and van Gunsteren, W. F. (1994) Local elevation: A method for improving the searching properties of molecular dynamics simulations.
Journal of Computer-Aided Molecular Design 8, 695 - 708. doi: 10.1007/Bf00124016 Peric-Hassler, L., Hansen, H. S., Baron, R., and Hünenberger, P. H. (2010) Conformational properties of glycose-based disaccharides investigated using molecular dynamics simulations with local elevation umbrella sampling. Carbohydrate research 345, 1781 - 1801. doi: 10.1016/j.carres.2010.05.026 Eichenberger, A. P., Allison, J. R., Dolenc, J., Geerke, D. P., Horta, B. A. C., Meier, K., Oostenbrink, C., Schmid, N., Steiner, D., Wang, D. Q., and van Gunsteren, W. F. (2011) GROMOS plus plus Software for the Analysis of Biomolecular Simulation Trajectories. Journal of Chemical Theory and Computation 7, 3379-3390. doi: 10.1021/Ct2003622 Chys, P., and Chacon, P. (2013) Random Coordinate Descent with Spinor-matrices and Geometric Filters for Efficient Loop Closure. Journal of Chemical Theory and Computation 9, 1821-1829. doi: 10.1021/ct300977f 45. 44. 33 42. 43. 900 905 910 915 920 925 930 935 940 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . 46. 47. 48. 49. 50. 51. 52. 53. 54. 55. 56. 57. 58. 59. 60. Lopez-Blanco, J. R., Canosa-Valls, A. J., Li, Y., and Chacon, P. (2016) RCD+: Fast loop modeling server. Nucleic Acids Res 44, W395-400. doi: 10.1093/nar/gkw395 Reif, M. M., Hunenberger, P. H., and Oostenbrink, C. (2012) New Interaction Parameters for Charged Amino Acid Side Chains in the GROMOS Force Field. Journal of Chemical Theory and Computation 8, 3705-3723. doi: 10.1021/Ct300156h Reif, M. M., Winger, M., and Oostenbrink, C. (2013) Testing of the GROMOS Force- Field Parameter Set 54A8: Structural Properties of Electrolyte Solutions, Lipid Bilayers, and Proteins. Journal of Chemical Theory and Computation 9, 1247-1264. doi: 10.1021/Ct300874c Pol-Fachin, L., Rusu, V. H., Verli, H., and Lins, R. D. (2012) GROMOS 53A6GLYC, an Improved GROMOS Force Field for Hexopyranose-Based Carbohydrates. Journal of Chemical Theory and Computation 8, 4681-4690. doi: 10.1021/ct300479h Pol-Fachin, L., Verli, H., and Lins, R. D. (2014) Extension and validation of the GROMOS 53A6(GLYC) parameter set for glycoproteins. J Comput Chem 35, 2087- 2095. doi: 10.1002/jcc.23721 Schmid, N., Christ, C. D., Christen, M., Eichenberger, A. P., and van Gunsteren, W. F. (2012) Architecture, implementation and parallelisation of the GROMOS software for biomolecular simulation. Computer Physics Communications 183, 890 - 903. doi: 10.1016/j.cpc.2011.12.014 Berendsen, H. J. C., Postma, J. P. M., van Gunsteren, W. F., and Hermans, J. (1981) Interaction models for water in relation to protein hydration, In Intermolecular Forces (Pullman, B., Ed. ), pp 331-342, Reidel, Dordrecht, The Netherlands. doi: Hockney, R. W. (1970) The potential calculations and some applications.
Methods in Computational Physics 9, 136 - 211. doi: Hockney, R. W., and Eastwood, B. J. (1988) Computer simulation using particles doi: 10.1201/9780367806934, 10.1201/9780367806934 Hess, B., Bekker, H., Berendsen, H. J. C., and Fraaije, J. G. E. M. (1997) LINCS: A linear constraint solver for molecular simulations. Journal of Computational Chemistry 10.1002/(SICI)1096-987X(199709)18:12<1463::AID- 1463 18, JCC4>3.0.CO;2-H Berendsen, H. J. C., Postma, J. P. M., van Gunsteren, W. F., DiNola, A., and Haak, J. R. (1984) Molecular-dynamics with coupling to an external bath. Journal of Chemical Physics 81, 3684-3690. doi: Bussi, G., Donadio, D., and Parrinello, M. (2007) Canonical sampling through velocity rescaling. J Chem Phys 126, 014101. doi: 10.1063/1.2408420 Parinello, M., and Rahman, A. (1981) Polymorphic transitions in single crystals: A new molecular dynamic method. J. Applied Physics. 52, 7182 - 7190. doi: 10.1063/1.328693 Nosé, S., and Klein, M. L. (1983) Constant pressure molecular dynamics for molecular systems. Molecular Physics 50, 1055 - 1076. doi: 10.1080/00268978300102851 Briones, R., Blau, C., Kutzner, C., de Groot, B. L., and Aponte-Santamaria, C. (2019) GROmarhos: A GROMACS-Based Toolset to Analyze Density Maps Derived from Molecular Dynamics Simulations. Biophys J 116, 4-11. doi: 10.1016/j.bpj.2018.11.3126 Durocher, Y., Perret, S., and Kamen, A. (2002) High-level and high-throughput recombinant protein production by transient transfection of suspension-growing human 293-EBNA1 cells. Nucleic Acids Res 30, E9. doi: 10.1093/nar/30.2.e9 Physics of Institute Publishing, Bristol. doi: 1472. 61. 34 945 950 955 960 965 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.31.458325 ; this version posted August 31, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-ND 4.0 International license . 62. 63. 64. 65. 66. 67. 68. Amanat, F., Stadlbauer, D., Strohmeier, S., Nguyen, T. H. O., Chromikova, V., McMahon, M., Jiang, K., Arunkumar, G. A., Jurczyszak, D., Polanco, J., Bermudez- Gonzalez, M., Kleiner, G., Aydillo, T., Miorin, L., Fierer, D. S., Lugo, L. A., Kojic, E. M., Stoever, J., Liu, S. T. H., et al. (2020) A serological assay to detect SARS-CoV-2 seroconversion in humans. Nat Med 26, 1033-1036. doi: 10.1038/s41591-020-0913-5 Stadlbauer, D., Amanat, F., Chromikova, V., Jiang, K., Strohmeier, S., Arunkumar, G. A., Tan, J., Bhavsar, D., Capuano, C., Kirkpatrick, E., Meade, P., Brito, R. N., Teo, C., McMahon, M., Simon, V., and Krammer, F. (2020) SARS-CoV-2 Seroconversion in Humans: A Detailed Protocol for a Serological Assay, Antigen Production, and Test Setup. Curr Protoc Microbiol 57, e100. doi: 10.1002/cpmc.100 Hulme, E. C., and Trevethick, M. A. (2010) Ligand binding assays at equilibrium: validation and interpretation. Br J Pharmacol 161, 1219-1237. doi: 10.1111/j.1476- 5381.2009.00604.x Armbruster, D. A., and Pry, T. (2008) Limit of blank, limit of detection and limit of quantitation.
Clin Biochem Rev 29 Suppl 1, S49-52. doi: Carvalho, S. B., Moreira, A. S., Gomes, J., Carrondo, M. J. T., Thornton, D. J., Alves, P. M., Costa, J., and Peixoto, C. (2018) A detection and quantification label-free tool to speed up downstream processing of model mucins. PLoS One 13, e0190974. doi: 10.1371/journal.pone.0190974 Ramakrishnan, M. A. (2016) Determination of 50% endpoint titer using a simple formula. World J Virol 5, 85-86. doi: 10.5501/wjv.v5.i2.85 Becker, M. M., Graham, R. L., Donaldson, E. F., Rockx, B., Sims, A. C., Sheahan, T., Pickles, R. J., Corti, D., Johnston, R. E., Baric, R. S., and Denison, M. R. (2008) Synthetic recombinant bat SARS-like coronavirus is infectious in cultured cells and in mice. Proc Natl Acad Sci U S A 105, 19944-19949. doi: 10.1073/pnas.0808116105 35
bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . RPnet: A Reverse Projection Based Neural Network for Coarse- graining Metastable Conformational States for Protein Dynamics Hanlin Gu,a‡ Wei Wang,b‡ Siqin Cao,b‡ Ilona Christy Unarta,c Yuan Yao,a Fu Kit Sheong,b,d* Xuhui Huangb,c* aDepartment of Mathematics, Hong Kong University of Science and Technology, Kowloon, Hong Kong bDepartment of Chemistry, Hong Kong University of Science and Technology, Kowloon, Hong Kong cDepartment of Chemical and Biological Engineering, Hong Kong University of Science and Technology, Kowloon, Hong Kong dInstitute for Advanced Study, Hong Kong University of Science and Technology, Kowloon, Hong Kong To whom correspondence should be addressed. Email: [email protected] or [email protected] ‡ Hanlin Gu, Wei Wang, and Siqin Cao contribute equally to this work. ABSTRACT Markov State Model (MSM) is a powerful tool for modeling the long timescale dynamics based on numerous short molecular dynamics (MD) simulation trajectories, which makes it a useful tool for elucidating the conformational changes of biological macromolecules. By partitioning the phase space into discretized states and estimate the probabilities of inter-state transitions based on short MD trajectories, one can construct a kinetic network model that could be used to extrapolate long time kinetics if the Markovian condition is met. However, meeting the Markovian condition often requires hundreds or even thousands of states (microstates), which greatly hinders the comprehension of conformational dynamics of complex biomolecules. Kinetic lumping algorithms can coarse grain numerous microstates into a handful of metastable states (macrostates), which would greatly facilitate the elucidation of biological mechanisms. In this work, we have developed a reverse projection based neural network (RPnet) method to lump microstates into macrostates, by making use of a physics-based loss function based on the projection operator framework of conformational dynamics. By recognizing that microstate and macrostate transition modes can be related through a projection process, we have developed a reverse projection scheme to directly compare the microstate and macrostate dynamics. Based on this reverse projection scheme, we designed a loss function that allows effectively assess the quality of a given kinetic lumping. We then make use of a neural network to efficiently minimize this loss function to obtain an optimized set of macrostates. We have demonstrated the power of our RPnet in analyzing the dynamics of a numerical 2D potential, alanine dipeptide, and the clamp opening of an RNA polymerase. In all these systems, we have illustrated that our method could yield comparable or better results than competing methods in terms of state partitioning and reproduction of slow dynamics.
We expect that our RPnet holds promise in analyzing conformational dynamics of biological macromolecules. 1 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . I. INTRODUCTION Many biologically relevant events occur at milliseconds or longer timescales.1–3 While molecular dynamics (MD) simulation tools and computation resources had a large advancement in the past few decades, obtaining a reliable sampling of these long-timescale events is still challenging. Markov State Model (MSM) is a mathematical framework that was built from a large number of short MD simulations but allows the estimation of long timescale dynamics, so that conformational sampling could be done in an highly parallelized manner.4–27 MSM works by decomposing the phase space into discrete regions (“states”) and then estimating the interstate transition probabilities at a specified lag time. If the complex system dynamics has a separation of timescale and the resulting transition probability matrix (TPM) satisfies the Markovian (memoryless) condition, the population evolution of the system under study could be calculated through repeated self-propagation of the TPM. In order to construct a Markovian model with an affordable lag time that allows efficient sampling (bound by the length of MD simulations to estimate transition probabilities), hundreds or thousands of microstates are often needed, but the large number of microstates present would hinder the interpretation of biological insights. This is a dilemma often encountered in MSM construction that models with large number of states are often hard to interpret, while models with only a few states are much more challenging to meet the Markovian condition. One popular way to strike the balance and obtain an MSM with only a handful of states is via a two-stage procedure.8,28,29 Collective variables that can properly describe the conformational dynamics of biological macromolecules are first chosen (e.g. by the time-lagged independent component analysis: tICA5,30), all the conformations are then split into a large number of “microstates” using geometric clustering algorithms like k-centers7,31 or k-means32,33 clustering based on the chosen representation, giving rise to a microstate MSM. It is then followed by a kinetic lumping procedure, merging the microstates into a handful coarse-grained metastable states (macrostates), so that we can better interpret the biological mechanisms.2,3,34 One widely used group of kinetic lumping algorithms makes use of the dominant eigenvectors of the TPM to determine the lumping, like the Perron-Cluster Cluster Analysis (PCCA)35 that repeatedly bi- partitions the microstates into groups based on the sign structures of the dominant eigenvectors, or its robust variants.36–38 There are also methods based on Bayesian inferences, as exemplified by the Bayesian agglomerative clustering engine (BACE),39 by repeatedly merging the microstate pairs with smallest BACE Bayes factor.
In another Bayesian kinetic lumping algorithm, a Gibbs sampling algorithm is applied to facilitate the search of the optimal kinetic lumping.40 Methods that explicitly considers the possible transition paths in the microstate transition matrix also exists, an example of which is the Most Probable Path (MPP) lumping algorithm.41 Rather than the two-step procedure discussed above, several deep-learning methods have recently been developed to obtain macrostate models directly from MD simulations trajectories. 2 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . A notable example is the VAMPnet,42 which takes Cartesian coordinates or other physical variables extracted from MD simulation trajectories as input, and directly produces a macrostate assignment and the corresponding Markov model through a neural network. The quality of lumping is determined by the VAMP-2 score based on the variational principle of the conformational dynamics. There is also a related approach with a symmetrization constraint to enforce the detailed balance, known as State-free Reversible VAMPnet (SRV),43 which enforces the detailed balance condition, which has been successfully applied to the study of Trp-cage protein.44 The kinetic lumping can also be viewed as a projection process of conformational dynamics. This view makes use of the projection operator framework in statistical mechanics developed by Zwanzig and Mori45–47. In fact, this idea has been previously realized by Hummer and Szabo in multistate kinetics, where the original dynamics of the system is projected into a reduced system by keeping the exact dynamics of the reduced system in both non-Markovian and Markovian region.48 We have also applied such a projection scheme to understand the conformational dynamics of complex systems through the extrapolation of occupancy-number correlation based on a given macrostate partitioning.29 A similar idea with the projection operator scheme is to examine the transfer operator6 itself as opposed to population correlation, which also provides a useful framework for developing the kinetic lumping algorithm. Based on the above view of the projection scheme for conformational dynamics29, we have developed a novel deep learning-based kinetic lumping method: a reverse projection based neural network (RPnet) method. The key insight of RPnet lies in its loss function that can quantify the quality of lumping via a reverse projection process of conformational dynamics. The reverse projection scheme utilizes the projection operator to evaluate macrostate models, by assessing the ability to match the dynamics of the microstate model in the macrostate resolution. In this scheme, an overlapping matrix between microstate eigenvectors (i.e.
the eigenvectors of the microstate TPM) and a backward projected vector of macrostate eigenvectors is adopted as the scoring function for evaluation. In the RPnet, this eigenvector-based scoring function is used as the loss function, and the optimization of this loss function would improve the matching of dynamics between macrostate models and the microstate model. This design of the loss function makes RPnet different from other deep learning-based lumping methods such as VAMPnet, where the loss function is designed based on the variational principle of conformational dynamics. In our loss functions, we have formulated a reverse projection scheme to allow direct comparisons between the dynamics of lumped macrostate models and the original microstate models to evaluate the quality of lumping, and further optimize the macrostate boundaries. In terms of architectures of deep learning networks, we have constructed a two-lobe encoder neural network. We demonstrate that our RPnet method performs well when applied to study systems ranging from numerical potentials to the complex RNA polymerase (RNAP) system. We anticipate that our RPnet holds promise to be widely applied to study biomolecular dynamics. 3 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . II. METHODS A. Kinetic lumping as a projection operation To understand the basis of our design, we first examine the relationship between populations of microstates (small states, usually in a number of hundreds to thousands) and macrostates (large metastable states, often only a handful of macrostates), and examine how the populations evolve under the projection operator framework. (cid:0)(cid:2)(cid:3)(cid:4) The transition probability matrix (TPM) of the microstate model ( model ( (cid:5)(cid:2)(cid:3)(cid:4) (cid:3) ) for a prespecified lagtime are defined as follows, ) and the macrostate (cid:6)(cid:2)(cid:3)(cid:4) (cid:7) (cid:0)(cid:2)(cid:3)(cid:4)(cid:6)(cid:2)0(cid:4) ( 1) (cid:9)(cid:2)(cid:3)(cid:4) (cid:7) (cid:5)(cid:2)(cid:3)(cid:4)(cid:9)(cid:2)0(cid:4) ( 2) (cid:9)(cid:2)(cid:3)(cid:4) (cid:7) (cid:10)(cid:0)(cid:6)(cid:2)(cid:3)(cid:4) ( 3) (cid:6) (cid:9) are vectors (with dimensions n and N respectively) corresponding to the where is the matrix defining the mapping from microstate and macrostate populations, respectively. the microstates (j) to macrostates (I): reflects the , thus or (cid:10)(cid:0) macrostate lumping of the system. We note that left-multiplication of to the microstate population vector is effectively summing up all the populations of microstates within each of the macrostates. Combining Eq (1)-(3), we show that the microstate and macrostate transition probability matrices can be related by the following equations: and (cid:10) j (cid:13) I (cid:10) (cid:10) (cid:7) 0 j (cid:16) I (cid:7) 1 (cid:10) (cid:2)(cid:3) (cid:2)(cid:3) if if (cid:5)(cid:2)(cid:3)(cid:4)(cid:10)(cid:0) (cid:7) (cid:10)(cid:0)(cid:0)(cid:2)(cid:3)(cid:4) ( 4) (cid:5)(cid:2)(cid:3)(cid:4) (cid:7) (cid:10)(cid:0)(cid:0)(cid:2)(cid:3)(cid:4)(cid:17) (cid:4) (cid:10)(cid:17) (cid:6)(cid:7) (cid:5) ( 5) (cid:18) (cid:8) (cid:2)(cid:3)(cid:4) (cid:7) (cid:10)(cid:0)(cid:18) (cid:9) (cid:2)τ(cid:4)(cid:10) ( 6) (cid:20) (cid:9) correspond to transition count matrices of macrostates and microstates where (cid:17) (cid:5) are diagonal matrices containing the microstate and macrostate respectively, and is always an equilibrium populations.
Eq. (5) is straightforwardly obtained as (cid:5) is related by summing up all the equilibrium identity matrix, following the fact that populations within the corresponding macrostates. We will then examine the “evolution of , respectively). Under the population” at microstate and macrostate level ( times Markovian condition, with the lag time of using the transition matrix (cid:20) (cid:8) and (cid:17) (cid:4) and (cid:10)(cid:0)(cid:17) (cid:4) (cid:10)(cid:17) (cid:6)(cid:7) (cid:21) (cid:22) (cid:5) (cid:17) (cid:17) (cid:4) and (cid:9)(cid:2)(cid:3)(cid:4) (cid:6)(cid:2)(cid:3)(cid:4) and (cid:3) , propagating the microstate population for (cid:0)(cid:2)(cid:3)(cid:4) (cid:23)(cid:3) (cid:23) should give the population at time : 4 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (cid:6)(cid:2)(cid:23)(cid:3)(cid:4) (cid:7) (cid:0)(cid:2)(cid:3)(cid:4)(cid:10)(cid:6)(cid:2)0(cid:4) ( 7) (cid:3) The minimum lag time called the Markovian lagtime time for which the population is well described by the equation above is (cid:3) (cid:11) of the kinetic network model. The coarse-grained macrostate model is obtained by grouping a number of microstates into one macrostate, where the population of a coarse-grained macrostate is the sum of all corresponding microstates, (cid:9)(cid:2)(cid:23)(cid:3) (cid:11) (cid:4) (cid:21) (cid:10)(cid:0)(cid:6)(cid:2)(cid:23)(cid:3) (cid:11) (cid:4) (cid:7) (cid:10)(cid:0)(cid:0)(cid:10)(cid:2)(cid:3) (cid:11) (cid:4)(cid:6)(cid:2)0(cid:4) ( 8 ) (cid:9) This equation also implies that the population distribution of coarse-grained macrostate at time (cid:24) (cid:7) (cid:23)(cid:3) (cid:11) could be obtained through repeated propagation of the original (microstate) population (cid:4) (cid:0)(cid:2)τ (cid:11) . by (cid:6)(cid:2)(cid:24)(cid:4) is Markovian, the evolution of (as opposed to only ) information of . Specifically, even if the macrostate population distribution is known, the given population vector could actually be , because the population distribution mapped to many different microstate distributions within each macrostate could be assigned in different ways. Although these distributions could all be mapped to the same , the intra-macrostate relaxation processes are different, and this would subsequently affect the evolution of macrostate dynamics. Therefore, obtaining a Markovian propagation of One important point to note is that, although the evolution of (cid:9)(cid:2)t(cid:4) (cid:6)(cid:2)0(cid:4) is not necessarily so, because information of all elements in ( , as seen in Eq 8 (cid:9)(cid:2)0(cid:4) (cid:6)(cid:2)0(cid:4) (cid:9)(cid:2)(cid:24)(cid:4) (cid:9)(cid:2)0(cid:4) ) is needed in order to obtain (cid:9)(cid:2)0(cid:4) (cid:9)(cid:2)t(cid:4) is more challenging.
We note that the “lumping” process can be viewed as a projection, by making use of the projection operator framework that was previously developed by Szabo and Hummer.48 (cid:6)(cid:7)(cid:10)(cid:0), (cid:13) (cid:26) (cid:7) (cid:17) (cid:12) (cid:10)(cid:17) (cid:28) (cid:7) 1 (cid:29) (cid:26) ( 9 ) (cid:17) (cid:4) is the diagonal matrix containing the population of the original kinetic network model (cid:5) denotes the coarse-grained one. Because the coarse-grained population is obtained by , we can then easily . (See where (cid:17) and summing the populations of the states in the original model, prove the idempotency Appendix A for the proof). (cid:7) (cid:10)(cid:0)(cid:17) (cid:4) (cid:17) (cid:10) (cid:5) (cid:26)(cid:14) (cid:7) (cid:26) (cid:28)(cid:14) (cid:7) (cid:28) (cid:10)(cid:0)(cid:26) (cid:21) (cid:10)(cid:0) (cid:26)(cid:28) (cid:7) 0 , , , and also the identity (cid:9)(cid:2)(cid:24)(cid:4) The projection operator then relates the population (cid:6)(cid:2)(cid:24)(cid:4) at the macrostate level and population at the microstate level. (cid:9)(cid:2)(cid:24) (cid:30) (cid:23)(cid:3) (cid:11) (cid:4) (cid:7) (cid:10)(cid:0)(cid:6)(cid:2)(cid:24) (cid:30) (cid:23)(cid:3) (cid:11) (cid:4) ( 10 ) 5 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (cid:7) (cid:10)(cid:0)(cid:26)(cid:6)(cid:2)(cid:24) (cid:30) (cid:23)(cid:3) (cid:11) (cid:4) (cid:7) (cid:10)(cid:0)(cid:26)(cid:0)(cid:2)(cid:3) (cid:11) (cid:4)(cid:10)(cid:26)(cid:6)(cid:2)(cid:24)(cid:4) (cid:30) (cid:10)(cid:0)(cid:26)(cid:0)(cid:2)(cid:3) (cid:11) (cid:4)(cid:10)(cid:28)(cid:6)(cid:2)(cid:24)(cid:4) B. Evaluating the Quality of Kinetic lumping via reverse projection To closely connect the dynamics of microstate and macrostate models, we have formulated a “reverse projection” process by mapping eigenvectors of the macrostate TPM (N transition modes) back to the microstate space (n eigenvectors, ). The “reverse projected” transition mode of a given lumping matrix (cid:31) ! (cid:10) is defined as. " (cid:15)(cid:16) (cid:2)#(cid:4) (cid:7) (cid:17) (cid:4) (cid:10)(cid:17) . (cid:6)(cid:7)$(cid:2)#(cid:4) (cid:5) ( 11 ) 1 % # % ! where projection can be found in Appendix A. , meaning that only the top N modes are considered. Properties of this reverse To better illustrate the actual meaning of the “modes” other than its mathematical representations, we have made an illustration using a simple 1D potential (Fig. 1). When there is a good lumping where all four energy minima in the 1D potential are properly identified as metastable states (Green curves in right panels (Left panel of Fig. 1b), the “reverse projected” modes (Fig. 1a). In sharp contrast, when of Fig. 1b) agree well with the original transition modes there is a bad lumping (Left panel of Fig.
1c), the reverse-projected modes (Red curves in right panels of Fig. 1c) largely deviate from the original transition modes. In addition, when we examine closely the reverse projected modes (green lines in Fig. 1b and red lines in Fig. 1c), we notice the reverse-projected modes are “continuous” within each macrostate region, where there are clear discontinuities at the boundaries of macrostates (especially when a “bad lumping” was evaluated in Fig. 1c). This could be understood from the definition of the reverse projection (Eq ) ( ), that the components of the reverse projected modes within each macrostate are 11 proportional to the equilibrium population within the macrostate. This simple 1D potential example clearly illustrates that the resemblance between microstate transition modes and reverse projected modes could serve as a way to examine the quality of state boundaries. (cid:17) (cid:4) (cid:10)(cid:17) " (cid:6)(cid:7)& (cid:5) 6 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (cid:0)(cid:2)(cid:3)(cid:4) FIG. 1. Illustration of reverse projected modes in good and bad lumping. (a) Energy landscape and the corresponding microstate transition modes. (b) Reverse projected modes of good lumping. (c) Reverse projected modes of bad lumping. The dashed line in (b) and (c) are the same as the microstate transition modes in (a), shown for the ease of comparison. It is clear from the figure that the reverse projected modes are smooth within each macrostate region, but at the boundaries between two macrostates, clear discontinuities could be present, especially for the modes corresponding to a bad lumping. To assess the quality of lumping, we quantify the similarity between the transition modes of the microstate models and “reverse projected” modes. Thus, we have defined the following overlap matrix ’ to quantify the similarity, ’ (cid:18)(cid:19) (cid:7) )(cid:20)(cid:21)(cid:22)(cid:23)(cid:24) ()(cid:20)(cid:21)(cid:22)(cid:23)(cid:24)(cid:2)#(cid:4)(cid:0)(cid:10)(cid:0)*+ (cid:9) (cid:2)#(cid:4)-. (cid:2)#(cid:4) (cid:0)$(cid:20)(cid:21)(cid:22)(cid:23)(cid:24) (cid:9) (cid:2),(cid:4) (cid:2)#(cid:4) (cid:0)+ (cid:9) (cid:2)#(cid:4) ( 12 ) $(cid:20)(cid:21)(cid:22)(cid:23)(cid:24) )(cid:20)(cid:21)(cid:22)(cid:23)(cid:24) denotes the right and left eigenvectors of the TPM of the proposed Where + (cid:9) denotes the right and left eigenvectors of the macrostate lumping, respectively, and (cid:0) original microstate transition matrix is essentially the left reverse- ’ projected eigenvector for the trial macrostate partitioning (refer to Appendix A for details), and (cid:18)(cid:19) therefore reflects the overlap between the the denominator is the normalization factor.
reverse projected transition modes and the original microstate transition modes. and . (cid:9) and . Basically, )(cid:20)(cid:21)(cid:22)(cid:23)(cid:24)(cid:2)#(cid:4)(cid:0)(cid:10)(cid:0) 7 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . ’ measures overlap of the original and the reverse projected eigenvectors, and it Because matrix is clear from Fig. 1 that the resemblance between the two would be useful for directly quantifying the quality of state boundaries. Matrix is therefore a good candidate as a loss function for automatic optimization of state boundaries. This loss function will be different from the popular methods such as VAMP-2 score that are based on variational principle of the ’ conformational dynamics. Basically, a perfectly Markovian lumping will be seen with the matrix equals to identity matrix (see Appendix A for details), so we may take the Freboneus norm of ’ ’ the difference between the matrix against an identity matrix Y-loss (cid:7) /(cid:22) (cid:29) ’/ (cid:25) ( 13 ) to serve as a loss function for optimization purpose (referred to as Y-loss hereafter). This proposal is also consistent with the lumpability condition proposed by Kemeny and Snell,49 who have introduced the “lumpability” condition to reduce the size of state space of some continuous- time Markov chains in probability theory. We can see in Fig. 1 that Y-loss can indeed clearly distinguishes the good from bad lumping, as the good lumping in Fig. 1b has a much smaller Y- loss than the bad lumping in Fig. 1c. C. Optimizing Kinetic lumping using an encoder neural network /(cid:22) (cid:29) ’/ (cid:10) (cid:25) with respect to the choice of lumping is a highly non- Note that the minimization of linear problem. We have thus designed a neural network for this optimization. The “lumping” of microstates into macrostates is understood as a “membership assignment”, in which the input of is the number of microstate) is one-hot vector, where microstate assignment (as a is the number of mapped to the macrostate assignment (as a microstate), and we aim to search for a (fuzzy) state assignment by finding a good membership assignment seen as an encoder network. At the end, a crisp lumping is obtained by taking the assignment with maximum membership. 1 0 (cid:31) (cid:31) 1 0 ! ! membership vector, where Our algorithm basically consists of four steps (see Fig. 2(a)): (1) Preprocessing: Computing the microstate Transition Count Matrix (TCM) by counting the number of transitions between prespecified states. And converting every frame of microstate assignment (from one or many trajectories) into 1 0 (cid:31) one-hot vectors. (2) Input: Feeding pairs of one-hot vector assignments (separated by a prespecified lagtime) into two encoder networks that share same architecture and weights.
The output of each network is a macrostates. The output is also used to compute the macrostate TCM. 1 0 ! ! vector, representing the membership of the input microstate to the 8 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (3) Training: Based on the microstate and macrostate TCM, computing the Y-loss (Eq. (13)), , and using backpropagation to optimize the network until convergence. (4) Evaluation: The optimized lumping matrix is computed by enumerating all the microstate one-hot vectors and stacking the corresponding output (resulting in an ed matrix). The quality of the output is also examined by other criteria, including but not limited to metastability or generalized matrix Rayleigh quotient (GMRQ).50 More details of the algorithm can be seen from the pseudocode in Appendix B. er FIG. 2. Architecture of the RPnet. (a) Overall architecture of the RPnet. (b) Details of the encoder network. 9 9 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . D. Architecture of the encoder network of RPnet The architecture of each of the two encoder networks is shown in Fig. 2(b), which includes five fully-connected layers. The dimension of the input and output are, respectively, the number of microstates . For example, in the alanine dipeptide dataset, we aimed to lump 100 microstates into 4 macrostates, and therefore there are 100 nodes on the first layer of the encoder and 4 nodes on the last layer of the encoder. In our design, we try to keep the ratio of the number of the nodes between each layer to be similar (see Sec. II.E.5 for the number of nodes for each example). Rectified linear unit (ReLU) 51,52 is used as the activation function between each layer, and SoftMax is used for the output so that the output vector resembles a membership function to the macrostates that sums to one. Dropout and max-pooling were used in the first and second layers to avoid overfitting. (cid:31) ! and the number of macrostate The lumping matrix encoded in the neural network after the training procedure could be retrieved by feeding in all the microstates) vectors) together. The obtained sequentially into the encoder and stacking all the outputs ( (cid:31) 0 ! (cid:31) (cid:31) possible one-hot vector assignments (representing 1 0 ! (cid:10) . matrix is the fuzzy lumping matrix Transpose symmetrization is used when computing the TCMs, for Y-loss computation to ensure (cid:9) are computed through a singular detailed balance condition is fulfilled.
And the eigenvectors value decomposition procedure and the update is done using the PyTorch functionality.53,54 Scaling and normalization have been done in the singular value decomposition (SVD) step to ensure the eigenvectors of original microstates and reverse projected ones are of the same scale. More details could be found in Appendix B. " E. System setup and simulation details 1. 1D potential To illustrate the reverse projection idea using a simple system, we have performed an MD simulation on the 1-D potential presented in Ref6: &(cid:2)1(cid:4) (cid:7) 4(1(cid:26) (cid:30) 0.85(cid:6)(cid:26)(cid:11)(cid:27)(cid:0) (cid:30) 0.25(cid:6)(cid:26)(cid:11)(cid:28)(cid:27)(cid:6)(cid:11). (cid:29)(cid:30)(cid:0) (cid:30) 0.55(cid:6)(cid:31)(cid:11)(cid:28)(cid:27) (cid:11). (cid:29)(cid:30)(cid:0)*, 1 (cid:16) (cid:2)(cid:29)1,1(cid:4) ( 14 ) which contains 4 minima in the stated region. Instead of the kinetic Monte Carlo simulation used in the original work6, we have performed an NVT MD simulation in the current example. A reflective boundary condition was set on the boundary to prevent the particle from diffusing away from the region. A velocity-Verlet integrator55 coupled with Andersen thermostat56 (T=1, collision frequency = 50 per time unit) was adopted for the NVT simulation. The integration time step is 0.0001 time unit and the trajectories is saved every 100 integration steps. So the 108-step simulation resulted in 106 saved frames in total. In this work, we have split the 1D region into 10 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . 100 equally spaced grids as microstates. The reverse projected eigenvectors are computed based on the two lumping at a lagtime of 50 saving intervals. 2. Alanine Dipeptide The simulation trajectories for alanine dipeptide are the same as those used in Ref40, which is consisted of one hundred 10-ns MD simulations. MD conformations are saved every 0.1 ps and thus there are 100,001 frames per trajectory. In this work, the 10,000,100 MD conformations are split into 100 microstates using k-centers clustering algorithm based on the root-mean-squared distances aligned on all heavy atoms. The microstate Markovian lagtime for this 100-state model is estimated to be 5 ps. The “PCCA+ lumping” is obtained from the crisp assignment of the PCCA+ fuzzy lumping38 result in PyEMMA.33 The hierarchical clustering with Ward linkage57 is done based on the : is the element of the symmetrized TCM. MPP distance matrix lumping41 is done with qMin = 0.7 so that 4 states are obtained. 8 (cid:7) 1/(cid:2)0.01 (cid:30) (cid:24) " (cid:4) ! (cid:2) (cid:24) ! (cid:2) ! (cid:2) , where 3. 2D potential In order to examine in detail the behavior of kinetic lumping when the lag time is relatively long and the separation of timescale is relatively small, we have set up an easy-to-visualize single- particle molecular dynamics simulations on the same 2D potential as in Ref 40 with the following potential form: &(cid:2)1, ;(cid:4) (cid:7) < # 8 = >cos 1 6 (cid:29) 3 sin 1 3 (cid:30) 5F >cos ; 6 (cid:29) sin ; 3 (cid:30) 3F , 1, ; (cid:16) G0,30H ( 15 ) 1, ; (cid:16) G0,30H which contains 4 energy minima, and reflective boundary condition set on the boundary to .
A velocity-Verlet integrator55 prevent the particle from diffusing out of the region: coupled with Andersen thermostat56 (T=1, collision frequency = 5 per time unit) was adopted for the NVT simulation. The integration time step is 0.001 time unit, and the trajectories is saved every 100 steps. As a result, our 109-step MD simulation trajectories is consisted of 107 saved frames. We have further split the 2D region ( equally spaced grids, resulting in 961 microstates. The microstate Markovian lagtime of this microstate partitioning is estimated to be 80 saving intervals (8000 integration steps). The PCCA+ lumping is obtained from the crisp assignment of the PCCA+ fuzzy lumping38 result in PyEMMA.33 1, ; (cid:16) G0,30H 31 0 31 ) into 4. RNAP loading gate dynamics The conformational changes of opening and closing of the Clamp conformation of a holoenzyme with the loading of a promoter RNA has been simulated in Ref 58. The system contains 543,237 atoms, and the simulation dataset is consisted of 306 200-ns MD trajectories. The all-atom MD 11 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . conformations were projected into three tICs obtained from the tICA. Then a 100-state microstate model was constructed via k-centers clustering in the three-dimensional tICA space with the tICA lag time of 10 ns. Please refer to Ref 58 for more details of the microstate MSM construction and validation. For this system, the PCCA+ lumping is obtained from the crisp assignment of the PCCA+ fuzzy lumping38 via the PyEMMA33 software. When performing the kinetic lumping, we have adopted the same microstate assignments as Ref 58 but chose a slightly different lagtime of 90ns rather than the lag time of 60ns adopted in Ref58. The hierarchical clustering with Ward linkage is done : " (cid:4) ! (cid:2) is the element of the symmetrized based on the distance matrix TCM. MPP lumping41 is done with qMin = 0.97 so that 4 states are obtained. 8 (cid:7) 1/(cid:2)0.01 (cid:30) (cid:24) (cid:24) ! (cid:2) ! (cid:2) , where 5. Parameters of the Neural Networks As mentioned in Sec. II.D, we aim to keep the ratio of number of nodes between the n-th and the (n+1)-th layer roughly constant, thus we have set up the networks with number of nodes as the following: (1) alanine dipeptide: 100-50-25-16-8-4; (2) 2D potential: 961-300-100-32-12-4; and (3) RNAP loading gate dynamics: 100-50-25-16-8-4. We made use of the ADAM59 optimizer for training, with the learning rate of the network set to 0.001 and a decay rate of 0.99 in each epoch. Dropout probability for the dropout later is chosen to be 0.2, and 30 training epochs were used for all datasets. The batch size of 6000 is used in alanine dipeptide and 2D potential, and batch size of 2000 is used to study the RNAP loading gate dynamics.
These batch sizes are chosen by preliminary scanning through the batch sizes ranging from 1000 to 10000, and the tests showed that if the batch sizes are too small or too large, the network might be much harder to converge. 6. Assessing the quality of the lumped models To compute the GMRQ and metastability of the lumping results, we first randomly divide the dataset into two equal portions as training and test dataset for 30 different times. The lumping matrix is then computed based only on the training dataset and the GMRQ50 and metastability are computed for both the training and testing dataset separately for each of the 30 random partitions. III. RESULTS AND DISCUSSIONS We demonstrate the performance of RPnet in three systems: the conformational dynamics of alanine dipeptide in water, a single-particle diffusion on a 2D potential, and the conformational dynamics of the DNA loading gate of a bacterial RNA polymerase. A. Alanine Dipeptide 12 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . We first demonstrate the performance of RPnet using the commonly used benchmarking example: the alanine dipeptide in water. The sampled conformations are grouped into 100 microstates (see Methods for details). As there exists a stable gap between the 3rd and 4th slowest implied timescale (Fig. 3(h)), we lumped the 100 microstates into 4 macrostates, which is also consistent with previous studies40,60. As shown in Fig. 3 (a & c), RPnet correctly identifies the (cid:5)). The implied timescales of the resulted lumping (Fig. 3(g)) four metastable states ( also reproduce those of the microstates reasonably well. We can also clearly visualize the optimization process of RPnet in Fig. 3(f). Although the initial Y-loss is high (at 0.18755), after (cid:0) (cid:0) (cid:3)(cid:4), (cid:0) , (cid:3), (cid:4) (cid:2) h around 20 epoc s, the Y-loss has gone down to 0.00127, which is the same as the Y-loss calculated for PCCA+ and MPP (see Fig. S1a for the corresponding Y matrices), consistent with same lumping results shown in Fig. 3 (b,c,e). In this example, RPnet displays comparable performance with other popular kinetic lumping methods (e.g., PCCA+, Hierarchical clustering using Ward linkage, and MPP). Specifically, all of these methods can correctly identify 4 macrostates containing largely similar microstate assignments (Fig. 3(b-e)). Because alanine dipeptide is a well-studied system containing clear separation of timescales, it is expected that all these kinetic lumping methods yield comparable results. In the next example, we will apply RPnet on a more challenging system with less clear separation of timescales, where we show that RPnet can robustly identify the metastable states and outperform other kinetic lumping methods.
FIG. 3. Performance of RPnet on the alanine dipeptide system. (a) The Ramachandran plot of the alanine dipeptide. (b-e) Lumping results by 4 different algorithms. (b) PCCA+. (c) RPnet. 13 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (d) Hierarchical clustering with Ward linkage. (e) MPP. (f) Evolution of the Y-loss of RPnet over 30 epochs, with referenced to the Y-loss of the other three lumping. (g-h) Implied timescales of the (g) 4 macrostates resulted from RPnet, and (h) original 100 microstates. B. 2D potential Although the alanine dipeptide provides as a good benchmark system for the development of kinetic lumping algorithms, it is often too simple to differentiate the performance of difference algorithms. In particular, the separation of timescale of the constructed MSM might be less clear in more complex systems, and this may impose challenges to eigenspectrum-based lumping methods such as PCCA35 or its variants36–38 due to the numerical instability. To investigate this situation, we have designed a 2D potential, where its 4 minima can be analytically identified as the gold standard of lumping results (see Eq. (15)). We have divided the XY-space as shown in Fig 4 into equally spaced grids, thus forming a 961-state microstate model. We then tried to lump the 961-microstate model into a 4-macrostate model (see Methods for more details). To create a situation with less clear separations of timescales, we have adopted a relatively long lagtime (3000 saving intervals) to construct the macrostate model when analyzing the single- particle diffusion on this 2D potential. 31 (cid:7) 31 (cid:8) As shown in Fig. 4(b), the state boundary of the PCCA+ becomes fuzzy at this lagtime ( =3000) and some microstates belonging to the free energy minima in blue) were mis-assigned to the yellow macrostate, while our RPnet correctly separate the blue state from the yellow state (see Fig. 4c). The optimization process of our RPnet is presented in Fig. 4(d), which clearly indicates that the RPnet achieved a lower Y-loss (~0.01224) than that of PCCA+ (see Fig. S1b for the corresponding Y matrices and Fig. S2 for the implied timescales). In addition to the Y-loss value, we have also evaluated the quality of lumped macrostate models using two other criterions: the metastability and the GMRQ. Metastability measures the average probability of self-transitions between macrostates, and a high metastability usually indicates a good separation of slow inter- state dynamics and fast intra-state dynamics. GMRQ assesses the quality of the macrostate models via a cross-validation process with an objective function based on the variational principle of conformational dynamics. As shown in Fig.
4(e), the lumped model from RPnet results in higher GMRQ and metastability, indicating that the macrostate model obtained from RPnet is better than that of PCCA+. These observations suggest that PCCA+ could suffer from the instability of eigenvector components and thus generate fuzzy state boundaries and leads to lowered GMRQ and metastability. We anticipate that this situation may be more prominent in complex biological systems, while our RPnet can yield more robust results in kinetic lumping. 14 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (cid:8) (cid:9) 3000 FIG. 4. Performance of RPnet on the 2D potential system. (a) The potential energy surface for (b) PCCA+, and (c) RPnet. (d) used in this test. (b-c) Lumping results at lagtime Evolution of the Y-loss of RPnet over 30 epochs, with referenced to the Y-loss of the PCCA+ lumping. (e-f) Comparison of (e) Generalized Matrix Rayleigh Quotient (GMRQ) and (f) metastability of the two lumping methods, redline denotes the training mean and the box-and- whisker plot denotes the distribution of 30 test sets (see Sec. II.E.6 for details). C. Dynamics of the clamp domain of a bacterial RNA polymerase After applying RPnet to two simple systems, we here proceed to a realistic biological system: i.e. the conformational conformation of bacterial RNAP clamp opening/closing motion that we reported recently58 (see Methods for details). As shown in Fig. 5a, the four macrostates of this system correspond to the open state (State S1, yellow dots), the closed state (State S4, green dots), as well as two partially closed states differing by the switch 2 region; partially closed with α -helix switch 2 (State S2, blue dots) and -helix switch 2 (State S3, red dots). The switch 2 region is a helical structure under the clamp domain which is crucial for the clamp movement58. In the PCCA+ based four-state model reported before, the transitions between S1/S2 or S3/S4 s, while the transitions between the two partially closed state S2/S3 would only takes 10~100 take 2 ms, indicating a huge gap between S2 and S3. We use the tICA projection (tICA lag time = 10 ns) for the visualization of different macrostate lumping. (cid:12) μ 15 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . FIG. 5. Performance of RPnet on the RNAP clamp motion. (a) representative conformations of the four metastable states shown as yellow (state S1), blue (state S2), red (state S3) and green (state S4) dots of (b-e).
The Clamp open/close angles are also presented in the snapshots of metastable states. (b-e) Lumping results at lagtime 90 ns for (b) PCCA+, (c) RPnet, (d) Hierarchical clustering with Ward linkage, and (e) MPP. (f) Evolution of the Y-loss of RPnet over 30 epochs, with referenced to the Y-loss of the other three lumpings. (g) Comparison of metastability of all four lumping methods. The box-and- whisker plot denotes the distribution of 30 test sets (see Sec. II.E.6 for details). (cid:0) (cid:2) As shown in Fig. 5(c), the four macrostates obtained by RPnet can be well separated in the projection of microstate centers onto the top two tICs, indicating that the macrostate boundaries clearly reflect the slowest dynamic modes of the system. Furthermore, the Y-loss score of the RPnet is the lowest (see Fig. 5(f)), and it also has the highest metastability among all the methods (see Fig. 5(g)). This state assignment is fully consistent with the previous study58, which was produced by PCCA+ with a shorter lag time (see Methods for details). Close examination of the Y-loss of the four assignments would reveal that Y-loss does correlate well with the quality of different lumpings (see Fig. S1c for the corresponding Y matrices). For example, the macrostate in green color is totally mis-assigned in the model obtained from the hierarchical clustering with Ward linkage (see Fig 5(d)), and this model also yields the highest Y-loss value. For PCCA+ (see Fig. 5(b)) and MPP (see Fig. 5(e)), both methods generate models in overall good quality, however, a few microstates at the state boundaries are not correctly assigned especially between the green and red macrostate. As a result, PCCA+ and MPP substantially improve in the Y-loss scores compared to the Ward linkage clustering, but they still produce models with higher Y-loss score than our RPnet. 16 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Taking all the results together, we show that RPnet can serve as a powerful kinetic lumping algorithm. The novelty of RPnet lies in its objective function (or loss function), which is based on the reverse projection of conformational dynamics from a macrostate model to a microstate model. In particular, the Y-loss function directly examines the similarity between the transition modes before and after kinetic lumping. This is distinct from the variational principle based objective function, which is focus on optimizing the slowest timescales (i.e., eigenvalues of the TPMs). We show that at situations in which the separation of timescales between intra- and inter-state transitions is not clear, our Y-loss function could robustly identify lumpings with correct and cleaner state boundaries.
In contrast, algorithms that are based on eigen- decompositions of TPMs (e.g. PCCA+) may become more sensitive to numerical instabilities under those conditions (e.g. in the 2D-potential example in Fig. 4, see also Fig S3 for the analysis on robustness). Our design of the loss function is inspired by a previously developed projection operator framework by Hummer and Szabo48. In that study, they make use of the projection operator framework to construct a transition matrix that can best describe the dynamics of a prespecified macrostate lumping, by matching the time-dependent occupation-number correlation function of microstates and macrostates. Although this method would be able to extrapolate the population evolution (in terms of time-dependent occupation-number correlation functions) of multistate kinetics reasonably well, we anticipate that it may not perform well when applied to distinguish good and poor lumpings as all the kinetic lumpings would result in reasonably good approximation of the evolution using their method. Our work gains inspiration from the projection operator framework developed by Hummer and Szabo48, but our aim is not at the extrapolation of kinetics, and instead we further designed a reverse projection to evaluate the quality of lumping at the same time. As a result, our RPnet is not simply a tool to predict long timescale kinetics, but also produce the optimized state partitioning. We note that although the Y-loss in our RPnet scheme itself can already be used to judge the quality of lumping, the use of encoder network provides an efficient approach to search for lumped models. Some early attempts of kinetic lumping11,61 optimizes lumping assignments through repeated trials and Monte-Carlo type optimization, in which the improvement between iterations are often relatively small and the optimization is easily trapped into local minima. With the use of a neural network, not only the combination of SoftMax activation functions could approximate the complex loss function landscape, it also allows a parallelized nonlinear search via backpropagation and quickly identify a good solution.62 In fact, the neural network also encodes the key representation of the complex system dynamics, which not only facilitates the backpropagation update,63 but also allows easy extension of the neural network architecture for other purposes like fuzzy lumpings. IV. CONCLUSION 17 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . In this work, we have developed a kinetic lumping algorithm: RPnet, that combines a physics- based loss function and the optimization using neural network. Inspired by the projection operator framework in statistical mechanics, we have designed a reverse projection scheme for comparing the transition modes among original microstates and lumped macrostates, which allows the quantification of lumping quality that is truly based on the underlying physical process.
Using our proposed Y-loss based on the reverse projection scheme, we have also set up a neural network that allows an automatic optimization of lumping. Combining the two parts, we have obtained our RPnet framework, that allows an automatic method that can give rise to physically sound lumping. As demonstrated by the three systems in the text, this framework can have good performance across a wide range of systems. The projection-operation-based loss function also provides an alternative line of thought for the evaluation of macrostate lumping quality and should give new insight to the identification of slow dynamics in complex systems. We anticipate that our method holds promise to be widely applied in the MSM construction to study protein functional dynamics.28 The source code of RPnet is available for download at: https://github.com/ghl1995/BpNet- lumping . SUPPLEMENTARY MATERIAL See supplementary material for the details of Y-matrices in different cases and implied timescales of the 2D potential example. ACKNOWLEDGMENTS We thank Dr. Gerhard Hummer for the inspiration and fruitful discussions. X.H acknowledges the support from the Padma Harilela Endowment Fund. REFERENCES 1 S. Peng, X. Wang, L. Zhang, S. He, X. S. Zhao, X. Huang and C. Chen, Proc. Natl. Acad. Sci., 2020, 117, 21889–21895. 2 L.-T. Da, F. Pardo-Avila, L. Xu, D.-A. Silva, L. Zhang, X. Gao, D. Wang and X. Huang, Nat. Commun., 2016, 7, 11244. 3 D.-A. Silva, D. R. Weiss, F. Pardo Avila, L.-T. Da, M. Levitt, D. Wang and X. Huang, Proc. Natl. Acad. Sci., 2014, 111, 7665–7670. 4 J. D. Chodera and F. Noé, Curr. Opin. Struct. Biol., 2014, 25, 135–144. 5 C. R. Schwantes and V. S. Pande, J. Chem. Theory Comput., 2013, 9, 2000–2009. 6 J.-H. Prinz, H. Wu, M. Sarich, B. Keller, M. Senne, M. Held, J. D. Chodera, C. Schütte and F. Noé, J. Chem. Phys., 2011, 134, 174105. 7 G. R. Bowman, X. Huang and V. S. Pande, Methods, 2009, 49, 197–201. 8 B. E. Husic and V. S. Pande, J. Am. Chem. Soc., 2018, 140, 2386–2396. 9 R. D. Malmstrom, C. T. Lee, A. T. Van Wart and R. E. Amaro, J. Chem. Theory Comput. , 2014, 10, 2648–2657. 18 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . 10 G. R. Bowman, V. S. Pande and F. Noé, Eds., An Introduction to Markov State Models and Their Application to Long Timescale Molecular Simulation, Springer Netherlands, Dordrecht, 2014, vol. 797. 11 J. D. Chodera, N. Singhal, V. S. Pande, K. A. Dill and W. C. Swope, J. Chem. Phys., 2007, 126, 155101. 12 A. C. Pan and B. Roux, J. Chem. Phys., 2008, 129, 064107. 13 F. Morcos, S. Chatterjee, C. L. McClendon, P. R. Brenner, R. López-Rendón, J. Zintsmaster, M. Ercsey-Ravasz, C. R. Sweet, M. P. Jacobson, J. W. Peng and J. a Izaguirre, PLoS Comput.
Biol., 2010, 6, e1001015. 14 X. Huang, G. R. Bowman, S. Bacallado and V. S. Pande, Proc. Natl. Acad. Sci., 2009, 106, 19765–19769. 15 N.-V. Buchete and G. Hummer, J. Phys. Chem. B, 2008, 112, 6057–6069. 16F. Noé, C. Schütte, E. Vanden-Eijnden, L. Reich and T. R. Weikl, Proc. Natl. Acad. Sci., 2009, 106, 19011–19016. 17 G. R. Bowman, V. A. Voelz and V. S. Pande, Curr. Opin. Struct. Biol., 2011, 21, 4–11. 18 I. Buch, T. Giorgino and G. D. Fabritiis, Proc. Natl. Acad. Sci., 2011, 108, 10184–10189. 19 D.-A. Silva, G. R. Bowman, A. Sosa-Peinado and X. Huang, PLoS Comput. Biol., 2011, 7, e1002054. 20 F. Noé, I. Horenko, C. Schütte and J. C. Smith, J. Chem. Phys., 2007, 126, 155102. 21 G. R. Bowman, D. L. Ensign and V. S. Pande, J. Chem. Theory Comput., 2010, 6, 787–794. 22 M. Sarich, F. Noé and C. Schütte, Multiscale Model. Simul., 2010, 8, 1154–1177. 23 F. Noé and F. Nüske, Multiscale Model. Simul., 2013, 11, 635–655. 24 H. Wu and F. Noé, J. Nonlinear Sci., 2020, 30, 23–66. 25J. Weng, M. Yang, W. Wang, X. Xu and Z. Tian, J. Am. Chem. Soc., 2020, 142, 21344–21352. 26 X. Zeng, L. Zhu, X. Zheng, M. Cecchini and X. Huang, Phys. Chem. Chem. Phys., 2018, 20, 6767–6776. 27 B. W. Zhang, W. Dai, E. Gallicchio, P. He, J. Xia, Z. Tan and R. M. Levy, J. Phys. Chem. B, 2016, 120, 8289–8301. 28 X. Wang, I. C. Unarta, P. P.-H. Cheung and X. Huang, Curr. Opin. Struct. Biol., 2021, 67, 69–77. 29 W. Wang, S. Cao, L. Zhu and X. Huang, Wiley Interdiscip. Rev. Comput. Mol. Sci., 2018, 8, e1343. 30 G. Pérez-Hernández, F. Paul, T. Giorgino, G. D. Fabritiis and F. Noé, J. Chem. Phys., 2013, 139, 015102. 31 D. S. Hochbaum and D. B. Shmoys, Math. Oper. Res., 1985, 10, 180–184. 32 J. A. Hartigan and M. A. Wong, J. R. Stat. Soc. Ser. C Appl. Stat., 1979, 28, 100–108. 33 M. K. Scherer, B. Trendelkamp-Schroer, F. Paul, G. Pérez-Hernández, M. Hoffmann, N. Plattner, C. Wehmeyer, J.-H. Prinz and F. Noé, J. Chem. Theory Comput., 2015, 11, 5525– 5542. 34 L.-T. Da, D. Wang and X. Huang, J. Am. Chem. Soc., 2012, 134, 2399–2406. 35 P. Deuflhard, W. Huisinga, A. Fischer and Ch. Schütte, Linear Algebra Its Appl., 2000, 315, 39–59. 36 P. Deuflhard and M. Weber, Linear Algebra Its Appl., 2005, 398, 161–184. 37 M. Weber and S. Kube, Comput. Life Sci. Proc., 2005, 3695, 57–66. 38 S. Röblitz and M. Weber, Adv. Data Anal. Classif., 2013, 7, 147–179. 39 G. R. Bowman, J. Chem. Phys., 2012, 137, 134111. 19 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . 40 W. Wang, T. Liang, F. K. Sheong, X. Fan and X. Huang, J. Chem. Phys., 2018, 149, 072337. 41 A. Jain and G. Stock, J. Chem. Theory Comput., 2012, 8, 3810–3819. 42 A. Mardt, L. Pasquali, H. Wu and F. Noé, Nat. Commun., 2018, 9, 5. 43 W. Chen, H. Sidky and A. L. Ferguson, J. Chem.
Phys., 2019, 150, 214114. 44 H. Sidky, W. Chen and A. L. Ferguson, J. Phys. Chem. B, 2019, 123, 7999–8009. 45 R. Zwanzig, Nonequilibrium Statistical Mechanics, Oxford University Press, 2001. 46 R. Zwanzig, J. Stat. Phys., 1983, 30, 255–262. 47 H. Mori, Prog. Theor. Phys., 1965, 33, 423–455. 48 G. Hummer and A. Szabo, J. Phys. Chem. B, 2015, 119, 9029–9037. 49 J. G. Kemeny and J. L. Snell, Finite Markov Chains: With a New Appendix “Generalization of a Fundamental Matrix,” Springer-Verlag, New York, 1976. 50 R. T. McGibbon and V. S. Pande, J. Chem. Phys., 2015, 142, 124105. 51 R. L. T. Hahnloser, Neural Netw., 1998, 11, 691–697. 52 V. Nair and G. E. Hinton, in Proceedings of the 27th International Conference on International Conference on Machine Learning, Omnipress, Madison, WI, USA, 2010, pp. 807–814. 53 B. Hu, R.-J. Song, X.-S. Wei, Y. Yao, X.-S. Hua and Y. Liu, in Proceedings of the 28th ACM International Conference on Multimedia, Association for Computing Machinery, New York, NY, USA, 2020, pp. 4461–4464. 54 G. H. Golub and C. Reinsch, Numer. Math., 1970, 14, 403–420. 55 L. Verlet, Phys. Rev., 1967, 159, 98–103. 56 H. C. Andersen, J. Chem. Phys., 1980, 72, 2384–2393. 57 J. H. Ward Jr, J. Am. Stat. Assoc., 1963, 58, 236–244. 58 I. C. Unarta, S. Cao, S. Kubo, W. Wang, P. P.-H. Cheung, X. Gao, S. Takada and X. Huang, Proc. Natl. Acad. Sci., , DOI:10.1073/pnas.2024324118. 59 D. P. Kingma and J. Ba, ArXiv14126980 Cs. 60 J. Apostolakis, P. Ferrara and A. Caflisch, J. Chem. Phys., 1999, 110, 2099–2108. 61 F. K. Sheong, D.-A. Silva, L. Meng, Y. Zhao and X. Huang, J. Chem. Theory Comput., 2015, 11, 17–27. 62 I. Goodfellow, Y. Bengio and A. Courville, Deep Learning, MIT Press, Cambridge, MA, USA, 2016. 63 Z. Allen-Zhu, Y. Li and Z. Song, in International Conference on Machine Learning, PMLR, 2019, pp. 242–252. 20 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Appendix A. Theory of Reverse Projection Prove: idempotency (cid:13)(cid:14) (cid:9) 0 (cid:14)(cid:6) (cid:9) (cid:14) (cid:13)(cid:6) (cid:9) (cid:13) , , , and also the identity (cid:15)(cid:7)(cid:13) (cid:16) (cid:15)(cid:7) (cid:14) (cid:19) is (cid:12)(cid:13) (cid:13) (cid:13)(cid:14) (cid:9) 0 (cid:14)(cid:6) (cid:9) (cid:14) (cid:13)(cid:6) (cid:9) (cid:13) (cid:9) (cid:15)(cid:7)(cid:20) , , (cid:15)(cid:7)(cid:17) (cid:15)(cid:17) A well-known property of the projection operators (cid:8) (cid:15)(cid:17) (cid:9) (cid:15)(cid:7)(cid:17) (cid:17) could be proved by (cid:8) (cid:9) the matrix form: (cid:13)(cid:6) (cid:9) (cid:17) (cid:8) (cid:10)(cid:11) (cid:16) (cid:18) (cid:9) (cid:15) and (cid:10)(cid:11) (cid:16) (cid:18) (cid:15)(cid:7)(cid:17) . From Eq. (3) we have (cid:8) (cid:9) (cid:10)(cid:11)(cid:15)(cid:7) (cid:9) (cid:17) (cid:8) (cid:9) (cid:14)(cid:6) (cid:9) (cid:21)(cid:18) (cid:22) (cid:13)(cid:23)(cid:6) (cid:9) (cid:18) (cid:22) 2(cid:13) (cid:25) (cid:13)(cid:6) (cid:9) (cid:18) (cid:22) (cid:13) (cid:9) (cid:14) .
This (cid:12)(cid:13), which can be written as , therefore (cid:10)(cid:11)(cid:15)(cid:7)(cid:17) (cid:15)(cid:17) (cid:8) (cid:9) . Then we will have, (cid:15)(cid:17) (cid:15)(cid:17) (cid:10)(cid:11)(cid:15)(cid:7) (cid:9) (cid:13) (cid:9) (cid:13)(cid:14) (cid:9) (cid:13)(cid:21)(cid:18) (cid:22) (cid:13)(cid:23) (cid:9) 0 (S1) (cid:15)(cid:7)(cid:13) (cid:9) (cid:15)(cid:7)(cid:17) (cid:8) (cid:15)(cid:17) (cid:10)(cid:11)(cid:15)(cid:7) (cid:9) (cid:15)(cid:7) (cid:9) which proves the proposition. (cid:27) (cid:26) (cid:26) (cid:14) (cid:28)(cid:21)(cid:29)(cid:23) (cid:15). In the A simple understanding of Eq. (11) can be made with the variational principle, (cid:21)(cid:31)(cid:23)(cid:7) (cid:28)(cid:30)(cid:21)(cid:31)(cid:23) (cid:9) (cid:30)(cid:21)(cid:31)(cid:23)(cid:26) is written as: spectral space that used for variational principles, the eigen decomposition of $(cid:21)(cid:31)(cid:23)(cid:16)" (cid:9) (cid:26) . At the same time, the eigen (cid:14) (cid:10)(cid:11)#(cid:21)%(cid:23) (cid:7)(cid:21)%(cid:23) (cid:9) (cid:15)(cid:17) , where (cid:9) (cid:21)(cid:31)(cid:23)(cid:16)(cid:28) (cid:9) (cid:26) "(cid:21)(cid:29)(cid:23) (cid:21)(cid:31)(cid:23) (cid:9) ! (cid:10)(cid:11)(cid:30)(cid:21)(cid:31)(cid:23) (cid:8) "#(cid:21)(cid:31)(cid:23) (cid:9) #(cid:21)(cid:31)(cid:23)(cid:26) (cid:14) and (cid:15) (cid:15) and $(cid:21)(cid:31)(cid:23)(cid:7) , where (cid:30)(cid:21)%(cid:23) (cid:9) (cid:17) (cid:8) $(cid:21)(cid:31)(cid:23) (cid:9) ! (cid:10)(cid:11)#(cid:21)(cid:31)(cid:23) decomposition of (cid:9) $(cid:7)(cid:21)%(cid:23)(cid:15)(cid:7) is written as: . Simply, when the top eigenvectors satisfy and , we will achieve the variational principle: (cid:28)(cid:21)(cid:29)(cid:23) (cid:16) (cid:30)(cid:26) (cid:15) (cid:26) (cid:21)(cid:29)(cid:23) (cid:9) $(cid:7)"(cid:21)(cid:29)(cid:23)# (cid:9) $(cid:7)(cid:15)(cid:7)(cid:28)(cid:21)(cid:29)(cid:23)(cid:17) (cid:8) (cid:10)(cid:11)# (cid:9) &(cid:26) (cid:9) (cid:15)(cid:17) (cid:14) (cid:21)(cid:29)(cid:23) (cid:7) & . is the overlapping matrix defined in Eq. (12). (cid:15) (cid:21)(cid:29)(cid:23)&(cid:15) (cid:9) (cid:26) (cid:15) (cid:21)(cid:29)(cid:23) (S2) A rigorous proof is shown as follows. First, straightforwardly from Eq. (5), (cid:21)(cid:29)(cid:23) (cid:10)(cid:11)# (cid:9) (cid:17) (cid:8) (cid:9) (cid:10)(cid:11)#, (cid:9) (cid:13)(cid:28)(cid:21)(cid:29)(cid:23)(cid:17) (cid:8) (cid:15)(cid:17) (cid:15)(cid:17) (cid:10)(cid:11)"(cid:21)(cid:29)(cid:23)# (cid:9) (cid:17) (cid:8) (cid:9) (cid:15)(cid:17) (cid:10)(cid:11)#(cid:26) (cid:14) (cid:9) (cid:9) (cid:17) (cid:8) (cid:7) (cid:9) $(cid:7)(cid:15)(cid:7), (cid:17)(cid:7) (cid:26) (cid:15)(cid:17) ’ (cid:21)(cid:29)(cid:23) (cid:21)(cid:29)(cid:23) (cid:9) (cid:26) (cid:17)(cid:7) (cid:14) (cid:10)(cid:11)# (cid:9) (cid:17)(cid:7) (cid:13)(cid:28)(cid:21)(cid:29)(cid:23) (cid:15)(cid:17) (cid:17) (cid:8) and Indicating the right eigenvector and eigenvalue of Secondly, when both microstate and macrostate model have master equations, (cid:10)(cid:11) ( (cid:15)(cid:7)(cid:20)(cid:21)0(cid:23) (cid:15)(cid:7)(cid:28)(cid:21)(cid:29)(cid:23)(cid:20)(cid:21)0(cid:23) (cid:9) (cid:15)(cid:7)(cid:20)(cid:21)(cid:29)(cid:23) (cid:9) (cid:19)(cid:21)(cid:29)(cid:23) (cid:9) (cid:15)(cid:7)(cid:28)(cid:21)(cid:29)(cid:23)(cid:17) (cid:8) (cid:9) (cid:10)(cid:11) ( (cid:15)(cid:7) (cid:9) are (cid:15)(cid:17) (cid:15)(cid:7)(cid:28)(cid:21)(cid:29)(cid:23) (cid:9) (cid:15)(cid:7)(cid:28)(cid:21)(cid:29)(cid:23)(cid:17) (cid:8) (cid:15)(cid:17) (cid:26) (cid:14) (cid:21)(cid:29)(cid:23) (S3) , respectively.
(S4) (cid:13)(cid:28)(cid:21)(cid:29)(cid:23) (cid:9) (cid:13)(cid:28)(cid:21)(cid:29)(cid:23)(cid:13) 21 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Then make use of the relation (cid:13)(cid:28)(cid:21)(cid:29)(cid:23) (cid:9) (cid:13)(cid:28)(cid:21)(cid:29)(cid:23)(cid:13) , (cid:13)(cid:28)(cid:21)(cid:29)(cid:23)(cid:13)(cid:30) (cid:9) (cid:13)(cid:28)(cid:21)(cid:29)(cid:23)(cid:30) (cid:9) (cid:13)(cid:30)(cid:26) (cid:21)(cid:29)(cid:23) (cid:30) (cid:9) (cid:30) (cid:9) (cid:13)(cid:30), (cid:26) (cid:21)(cid:29)(cid:23) (cid:9) (cid:26) (cid:15) (cid:21)(cid:29)(cid:23) (cid:9) (cid:26) (cid:21)(cid:29)(cid:23) (S5) (cid:17)(cid:7) (cid:17)(cid:7)(cid:17) (cid:17)(cid:7) (cid:17)(cid:7)(cid:17) (cid:15) (cid:13)(cid:28)(cid:21)t(cid:23)(cid:13) (cid:13)(cid:28)(cid:21)t(cid:23) (cid:13)(cid:30) λ Indicating that the right eigenvector and eigenvalues of (cid:7) (cid:21)t(cid:23) or respectively. Now combine the conclusion Eq. (S3) and (S5), we have: should be and (cid:17) (cid:8) (cid:15)(cid:17) (cid:10)(cid:11)# (cid:9) (cid:13)(cid:30), (cid:9) (cid:26) (cid:21)(cid:29)(cid:23) (cid:9) (cid:26) # (cid:9) (cid:15)(cid:7)(cid:30) (cid:21)(cid:29)(cid:23) (S6) (cid:14) (cid:15) (cid:30) (cid:9) (cid:30) Then we will prove (cid:17)(cid:7): (cid:7) (cid:21)%(cid:23)(cid:30)(cid:21)+(cid:23) (cid:9) $(cid:7)(cid:21)%(cid:23)(cid:15)(cid:7)(cid:30)(cid:21)+(cid:23) (cid:9) $(cid:7)(cid:21)%(cid:23)#(cid:21)+(cid:23) (cid:9) , (cid:17)(cid:7) (cid:3)(cid:18) (cid:7) (cid:21)%(cid:23) (cid:22) (cid:15)(cid:21)%(cid:23). (cid:30)(cid:21)+(cid:23) (cid:16) 0, (cid:17)(cid:7) (cid:15)(cid:21)%(cid:23)/(cid:30) (cid:17)(cid:7) (cid:21)+(cid:23) (cid:22) (cid:30)(cid:21)+(cid:23)0 (cid:16) 0 (S7) (cid:7) (cid:21)%(cid:23) (cid:9) (cid:15)(cid:21)%(cid:23), (cid:17)(cid:7) (cid:30) (cid:17)(cid:7) (cid:21)+(cid:23) (cid:9) (cid:30)(cid:21)+(cid:23) 0 1 %, + (cid:27) 2 Where the above equations only apply for top eigenmodes: (cid:10)(cid:11)# (cid:9) (cid:30) (cid:9) (cid:30) (cid:9) (cid:17) (cid:8) (cid:15)(cid:17) (cid:17)(cid:7) . Therefore, (S8) Which is Eq. (11). Appendix B. RPnet algorithm The following Algorithm 1 shows how we implement the RPnet in the network. In neural network, back propagation cannot support eigen decomposition, thus, we need to use SVD to replace it. The key idea is to borrow the relation of singular vectors of transition count matrix and eigenvector of transition probability matrix. Specifically, according to Eq. (S3) the relation between singular vector 3 (cid:9) (cid:17)(cid:10)(cid:19). (cid:21)3 : and eigenvector is: (cid:17)(cid:10)(cid:19).(cid:21)4(cid:17)(cid:10)(cid:22). (cid:23) (cid:9) 3(cid:26) 3(cid:15) (cid:15) (cid:28)(cid:7) (cid:16) (cid:26) (cid:15) (cid:10)(cid:11) (cid:9) (cid:17)(cid:10)(cid:11)4 (cid:9) (cid:17)(cid:10)(cid:19).
(cid:21)3(cid:26) (cid:15) 3(cid:15)(cid:17)(cid:19). (cid:21) (S9) 3(cid:7)3 (cid:9) (cid:15)(cid:17)(cid:19). (cid:21) ( (cid:17)(cid:19). (cid:21) (cid:9) (cid:15)(cid:30) (cid:9) (cid:18) 22 bioRxiv preprint doi: https://doi.org/10.1101/2021.08.04.455071 ; this version posted August 6, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . 23
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 bioRxiv preprint doi: https://doi.org/10.1101/2020.06.16.154187 ; this version posted June 17, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. ppGpp is a bacterial cell size regulator Ferhat Büke1,2, Jacopo Grilli3, Marco Cosentino Lagomarsino4,5, Gregory Bokinsky1, Sander Tans1,2 (1) Department of Bionanoscience, Kavli Institute of Nanoscience, Delft University of Technology, Delft, The Netherlands (2) AMOLF, Amsterdam, The Netherlands. (3) The Abdus Salam International Centre for Theoretical Physics (ICTP), Strada Costiera 11, 34014 Trieste, Italy (4) IFOM, FIRC Institute of Molecular Oncology, Via Adamello 16, 20143, Milan, Italy (5) Physics Department, University of Milan, and I.N.F.N., Via Celoria 16, 20133, Milan, Italy Corresponding authors. These authors contributed equally to the manuscript. Email: [email protected]; [email protected] 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 bioRxiv preprint doi: https://doi.org/10.1101/2020.06.16.154187 ; this version posted June 17, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Summary Growth and division are central to cell size. Bacteria achieve size homeostasis by dividing when growth has added a constant size since birth, termed the “adder” principle, by unknown mechanisms [1–4]. Growth is well known to be regulated by ppGpp, which controls anything from ribosome production to metabolic enzyme activity and replication initiation, and whose absence or excess can induce the stress response, filamentation, and yield growth-arrested miniature cells [5–8]. These observations raise unresolved questions about the relation between ppGpp and size homeostasis mechanisms during normal exponential growth. Here, to untangle effects of ppGpp and nutrients, we gained control of cellular ppGpp by inducing the synthesis and hydrolysis enzymes RelA and Mesh1. We found that ppGpp not only exerts control over the growth rate, but also over cell division and hence the steady state cell size. The added size responds rapidly to changes in the ppGpp level, aided by transiently accelerated or delayed divisions, and establishes its new constant value while the growth rate still adjusts. Moreover, the magnitude of the added size and resulting steady-state birth size correlate consistently with the ppGpp level, rather than with the growth rate, which results in cells of different size that grow equally fast. Our findings suggest that ppGpp serves as a critical regulator that coordinates cell size and growth control. Keywords: ppGpp; cell size; growth rate; cell size homeostasis; adder mechanism; Escherichia coli; microfluidics; microscopy; single cell analysis; lineage tree analysis. 2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 bioRxiv preprint doi: https://doi.org/10.1101/2020.06.16.154187 ; this version posted June 17, 2020.
The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Results Control of ppGpp synthesis and hydrolysis To study the relation between ppGpp and cell growth and division, two enzymes were used: the catalytic domain of the E. coli ppGpp synthesis enzyme RelA (RelA) [9,10], and the ppGpp hydrolysis enzyme Mesh1 from Drosophila melanogaster [11,12], which were fused to YFP and CFP respectively. The former was inducible by doxycycline (Dox) and the latter by isopropyl-β-D-thiogalactopyranoside (IPTG) (Fig. 1A). As a control, we showed that the expression of a fluorescent protein alone did not significantly affect cell size or growth rate (Supp. Fig. 1E-F). We characterized this co-expression system in a ppGpp0 strain (DrelA, DspoT) that cannot produce ppGpp. In minimal medium lacking amino acids, growth was undetectably low in absence of RelA induction, consistent with previous reports [8,12]. ppGpp is then required to activate amino acid biosynthesis operons [13,14]. However, growth became exponential if both RelA* and Mesh1 we co- expressed (Supp. Fig. 2A). These findings confirm that balanced synthesis and hydrolysis can achieve the constant ppGpp levels that are critical to normal exponential growth. If RelA* and Mesh1 indeed counteract in ppGpp production, then the same growth rates should be achievable by increasing both in parallel, as the additional synthesis by RelA* can then be canceled by the additional hydrolysis by Mesh1. The data indeed showed similar growth profiles for different combinations of RelA* and Mesh1 expression; with both either at lower levels, or both at higher levels (Supp. Fig. 2A). ppGpp exerts cell size control We studied the effects of ppGpp at the single-cell level using a microfluidic chip that allowed media exchange, phase contrast and fluorescence microscopy, and cell- tracking algorithms [15,16]. We determined the length at birth (LB) and division (LD), the cycle duration (Tcyc) and exponential growth or elongation rate (µ) for each cell cycle, and RelA* and Mesh1 enzyme concentrations, as quantified by the mean fluorescence per pixel (Fig. 1B, Supp. Fig. 1A-D). Here, we expressed either RelA*-YFP or Mesh1-CFP at moderate levels in the WT background (relA+ and spoT+), in order to produce minor 3 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 bioRxiv preprint doi: https://doi.org/10.1101/2020.06.16.154187 ; this version posted June 17, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. deviations in the ppGpp concentration, from above to below basal levels, while maintaining balanced exponential growth (Fig. 1C, D). The (population-mean) trend in the growth rate µ showed an optimum while the birth size LB went up monotonically, as ppGpp decreased from above to below basal levels (Fig.

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.