Fig. 1

Schematic summaries of the workflow used in this study. a Filtering of Tabula Muris and Tabula Sapiens atlases to retain cell types with ≥ 300 cells for subsequent data sampling. b Sampling from Tabula Muris and Tabula Sapiens atlases for creating scRNA-seq datasets with varying number of cell types; number of cells in each cell type; and ratios of cells in the major and minor cell types. c Deep learning feature selection methods applied for scRNA-seq data analysis in this study were grouped by their category as either perturbation-based and gradient-based methods. Popular differential distribution-based and machine-learning-based methods were included for comparison. d Genes selected by each feature selection method were evaluated for their utility in cell type classification. Reproducibility of feature selection results and computational time were also assessed for each feature selection method