README.md

---
license: mit
task_categories:
- feature-extraction
tags:
- biology
- medical
---

# Dataset Card for single-cell multiome from bone marrow

<!-- Provide a quick summary of the dataset. -->

Single-cell multiomics data collected from bone marrow mononuclear cells of 12 healthy human donors. 

## Dataset Details
Multimodal data as a basis for benchmarking
"Developing machine learning methods for biological systems is complicated by the difficulty of obtaining ground truth. Typically, machine learning tasks rely on manual annotation (as in images or natural language queries), dynamic measurements (as in longitudinal health records or weather), or multimodal measurement (as in translation or text-to-speech). However, this is more complicated in the context of single-cell biology.
With single-cell data, annotation isn’t feasible. The data is noisy and not fully understood with descriptions of cell types evolving rapidly. Similarly, longitudinal measurement of all the RNA in a cell isn’t possible because the current measurement technologies involve destroying the cell. However, with multimodal single-cell data, we can now directly observe two layers of genetic information in the same cells. This provides an opportunity to use the fact these two sets of data were observed co-occurring in the same cells as ground truth. This is akin to the way that access to the same sentiment expressed in two languages provides ground truth for machine translation.
However, as these technologies are relatively new, most publicly available datasets are designed for exploration, not benchmarking. To set up a competition for multimodal single-cell data integration, we set out to create a fit-for-purpose benchmarking dataset."

### Dataset Description
The study design is as follows:

Multiome
Site 1 - Donors 1, 2, 3
Site 2 - Donors 1, 4, 5
Site 3 - Donors 3, 6, 7, 10
Site 4 - Donors 1, 8, 9

- **Curated by:** Burkhardt DB, Lücken MD, Lance C, Cannoodt R, Pisco AO, Krishnaswamy S, Theis FJ, Bloom JM
- **License:** MIT

### Dataset Sources

<!-- Provide the basic links for the dataset. -->

- **Repository:** https://github.com/openproblems-bio
- **Paper:** https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/hash/158f3069a435b314a80bdcb024f8e422-Abstract-round2.html

## Uses

<!-- Address questions around how the dataset is intended to be used. -->
Challenges included modality prediction, matching profiles from different modalities, and learning a joint embedding from multiple modalities.

## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

The training data is accessible in an AnnData h5ad file. More information can be found on AnnData objects here. You can load these files is to use the AnnData.read_h5ad() function.
The dataset was designed with a nested batch layout such that some donor samples were measured at multiple sites with some donors measured at a single site.

## Dataset Creation
Joint profiling of single-nucleus RNA and chromatin accessibility using the 10X Genomics Single Cell Multiome ATAC + Gene Expression Kit

#### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

To facilitate exploring the data, each dataset has been preprocessed to remove low quality cells and doublets. The following sections detail this process for each data modality.

Preprocessing of gene expression (GEX)
In this dataset, gene expression was measured using 3’ capture of nuclear RNA as described in the 10X Multiome Product Guide. Note, not all RNA is found in the nucleus. Comparisons of nuclear and cytosolic RNA have been previously reported (e.g. Bakken 2018; Abdelmoez 2018) as have comparisons of single-nucleus and single-cell RNA sequencing (Lake 2017).

For gene expression data, cells were filtered based on mitochondrial content, UMI counts per cell, and genes detected per cell. Size factors were then calculated using scran and stored in adata.obs["size_factors"].

Counts were then normalized per cell by divided the UMI counts by the size factors. Original counts are stored in adata.layers["counts"]. The size factor normalized counts are stored in adata.X.

Finally, normalized counts are log1p transformed. These normalized counts are stores in adata.layers["log_norm"].

More information about best practices for single-cell analysis can be found here.

Preprocessing of ATAC
The chromatin accessibility data acquired by ATAC-seq as part of the 10X Multiome protocol was processed using Signac. Quality control, dimensionality reduction and translating peaks to gene activity scores was performed using Signac, following the authors’ instructions. After loading the peak-by-cell matrix, counts were binarized to only represent an accessible versus non-accessible state of each region. Cells were then filtered based on 5 quality control metrics comprising the total number of fragments, the enrichment of fragments detected at transcription start sites (TSS), the fraction of fragments in peak regions compared to peak-flanking regions, the fraction of peaks blacklisted by the ENCODE consortium, and the nucleosome signal, which describes the length distribution of fragments which is expected to follow the length of DNA required span across one nucleosome or multiples of it.

Since ATAC data is sparser than gene expression data, peaks were included if they were accessible in at least 15 cells.

Finally, the data was binarized by setting all values >0 to 1 and stored in adata.X. Raw UMI counts for each peak can be found in adata.layers["counts"].

Preprocessing of protein abundance (ADT)
The protein data was measured using the TotalSeq™-B Human Universal Cocktail, V1.0 of 134 cell surface markers and 6 isotype controls. The isotype controls are stored in adata.obsm["isotype_controls"]. These controls do not target any human proteins and their expression should be considered background.

The ADT protein measurements were run through quality control based on the total number of ADTs (ranging from 1100-1200 to 24000 across samples), the number of proteins captured in each cell (with a lower limit of 80) and the ADT count of the 6 isotype controls summed up in each cell (ranging from 1 to 100).

Since the total number of captured ADTs is limited, absolute ADT counts appear to be lower if highly abundant proteins are present. To account for this effect, normalization was performed using the centered log ratio (CLR) transformation. CLR counts are stored in adata.X and the raw counts are stored in adata.layers["counts"].

#### Annotation process

<!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->
Metadata
More information about the features are available in the .var and .obs DataFrames of each object.

Gene expression observation metadata
The GEX adata objects have the following columns:

.obs.index - The cell barcode for that observation with the batch label appended.
.obs["n_genes_by_counts"] - The number of genes with at least 1 count in a cell.
.obs["pct_counts_mt"] - Percent of UMI counts mapped to mitochondrial genes.
.obs["n_counts"] - Number of UMIs detected in the cell
.obs["n_genes"] - Number of genes detected in the cell
.obs["size_factors"] - The estimated size factor for the cell. See OSCA Ch. 7 - Normalization
.obs["phase"] - The cell cycle phase for each cell as calculated by scanpy.tl.score_genes_cell_cycle
.obs["leiden_final"] -
.obs["atac_ann"] - The cell type annotation of the cell from the joint ATAC data
.obs["cell_type"] - The cell type annotation of the cells from the GEX data
.obs["pseudotime_order_GEX"] - The diffusion pseudotime annotation for the developmental trajectories annotated in the data.
.obs["batch"] - The batch from which the cell was sampled. Format is s1d1 for Site 1 Donor 1.
For more info on how the QC metrics were calculated, consult scanpy.pp.calculate_qc_metrics

Gene expression feature metadata
The GEX adata.var DataFrames have the following columns:

.var.index - Ensembl Gene Names for each gene
.var["gene_ids"] - Ensembl Stable IDs used to uniquely track genes whose Gene Names may change over time.
.var["feature_types"] - Denotes the each feature as a gene expression feature. Should be GEX for all genes
.var["genome"] - The Genome Assembly used for read mapping.
.var["n_cells-[batch]"] - The number of cells in [batch] in which the gene was detected.
.var["highly_variable-[batch]"] - Whether the gene was determined to be highly variable in [batch]
ATAC observation metadata
The ATAC adata.obs DataFrames have the following columns:

.obs.index - The cell barcode for that observation with the batch label appended.
.obs["nCount_peaks"] - The number of peaks detected in the cell.
.obs["atac_fragments"] - Number of UMI counts in the cell (both in and not in peaks)
.obs["reads_in_peaks_frac"] - Fraction of UMIs in peaks
.obs["blacklist_fraction"] - Fraction of UMIs in Encode Blacklisted regions
.obs["nucleosome_signal"] - The nucleosome signal, which describes the length distribution of fragments which is expected to follow the length of DNA required span across one nucleosome or multiples of it
.obs["phase"] - The cell cycle phase for each cell as calculated by scanpy.tl.score_genes_cell_cycle
.obs["leiden_final"] -
.obs["rna_ann"] - The cell type annotation of the cell from the joint RNA data
.obs["cell_type"] - The cell type annotation of the cells from the ATAC data
.obs["pseudotime_order_ATAC"] - The diffusion pseudotime annotation for the developmental trajectories annotated in the data.
.obs["batch"] - The batch from which the cell was sampled. Format is s1d1 for Site 1 Donor 1.
For more info on how the QC metrics were calculated, consult the Signac documentation.

ATAC feature metadata
The ATAC adata.var DataFrames have the following columns:

.var.index - Genomic coordinates for each ATAC peak that are directly related to the reference genome, and include the chromosome name*, start position, and end position in the following format: chr1-1234570-1234870.
.var["feature_types"] - Denotes the each feature as a gene expression feature. Should be ATAC for all peaks
.var["n_cells-[batch]"] - The number of cells in [batch] in which the peak was detected.
*For the curious, chromosome names like KI270726.1 represent scaffold that are either unlocalized or unplaced (see Genome Assemblies from Ensembl)

There is also information about the observations in the .obs DataFrame of each AnnData object.

## Potential biases
Cell type identification and doublet removal were already performed. Donors varied by age (22 - 40), sex, and ethnicity
(details in the associated datasheet).

#### Who are the annotators?

<!-- This section describes the people or systems who created the annotations. -->

Burkhardt DB, Lücken MD, Lance C, Cannoodt R, Pisco AO, Krishnaswamy S, Theis FJ, Bloom JM

## Citation 

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->
	https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/hash/158f3069a435b314a80bdcb024f8e422-Abstract-round2.html