Building the mega single-cell transcriptome ocular meta-atlas

Abstract

Background: The development of highly scalable single-cell transcriptome technology has resulted in the creation of thousands of datasets, >30 in the retina alone. Analyzing the transcriptomes between different projects is highly desirable because this would allow for better assessment of which biological effects are consistent across independent studies. However it is difficult to compare and contrast data across different projects because there are substantial batch effects from computational processing, single-cell technology utilized, and the natural biological variation. While many single-cell transcriptome-specific batch correction methods purport to remove the technical noise, it is difficult to ascertain which method functions best.

Results: We developed a lightweight R package (scPOP, single-cell Pick Optimal Parameters) that brings in batch integration methods and uses a simple heuristic to balance batch merging and cell type/cluster purity. We use this package along with a Snakefile-based workflow system to demonstrate how to optimally merge 766,615 cells from 33 retina datsets and 3 species to create a massive ocular single-cell transcriptome meta-atlas.

Conclusions: This provides a model for how to efficiently create meta-atlases for tissues and cells of interest.

Figure 1:

A. Schematic of the retina with major cell types delineated. B. Simplified directed workflow of major steps in scEiaD creation from raw counts to gene counts, benchmarking optimal integration methods (SnakePOP) to produce batch-corrected latent dimensions (Latent Dims), then downstream analysis outputs such as clustering, differential gene testing (Diff Testing), and 2D UMAP visualization. C. Counts of published articles and batches (unique biological samples) for each scRNA technlogy, split by organism. D. Cell type counts extracted from published studies for the more common retina cell types, split by species. Count of study accessions for each species overlaid on bar plot. HS: Homo sapiens; MF: Macaca fascicularis; MM: Mus musculus.

Figure 2:

A. Example of a method (combat) that has a high level of batch blending but poor separation of cell types (colored by cell type). B. No batch correction cleanly separates cell types but does not mix batches (colored by study). C. sumZScale (higher is better) for each method across a variety of data normalizations. All methods shown here use 2,000 HVG, Louvain clustering with knn 20, and 8 or 30 latent dimensions. Each color is a different method. D. Box plot of 4 different clustering resolutions for 1,000–10,000 HVG numbers and 4–50 scVI latent dimensions. Open boxes are using scVI-standard and gray boxes are scVI-projection (human reference with the remaining data projected). tukey boxplot. center line is median, box is 25/75 percentile, line extends to 1.5 times interquartile range.

Figure 3:

A. Top genes that are differentially expressed across the major cell types of the retina. PR: photoreceptor. Genes are colored by cell type in which they are differentially expressed. The dot size is proportional to the percentage of those cells that have detectable levels of the gene. The color of the dot is the log₂-scaled counts per million expression. B. 2D UMAP projection of scEiaD, colored by cell type (Tabula Muris data are gray). Arrows indicate scvelo RNA velocity. Longer arrows show cells with higher velocity (relatively more unspliced transcripts). C. Facet plot that demonstrates how each major cell type of the retina is contained within a distinct space. D. Confusion matrix of cell type prediction performance of our xgboost labeler between predicted (x-axis) and known (y-axis) using data withheld from the machine learner. Most of the cell types are indeed labeled as their true type. E. Faceting of 2D UMAP by species and colored by cell type demonstrates how the major cell types of the retina share space with like cell types, despite being from mouse, human, and macaque.

Figure 4:

A. RPE distribution colored by study demonstrates how the most RPE are in 2 locations. iPSC-based RPE that we provided are found to be more enriched in Cluster 47. Tissue RPE are more enriched in Cluster 34. B. Violin plot of 2 functional RPE markers (TTR, RPE65) and vimentin (proliferating RPE marker). RPE, retinal pigmented epithelium, TTR, VIM, RPE65 are the HUGO gene names (genenames.org). The ensembl gene names are given in the parentheses. iPSC is induced puripotent stem cell (so iPSC RPE are RPE derived from iPSC)