Fig. 2

Comparison of CSS to different integration methods to resolve cell type heterogeneity in cerebral organoid scRNA-seq data. a Schematic of the experimental design. 20 2-month-old cerebral organoids were generated from seven different human PSC lines in four batches, and scRNA-seq data was generated using 10x Genomics. b UMAP embedding of cells without integration, with cells colored by RSS-based cell type annotation (left) and organoid of origin (right). c RSS with the fetal BrainSpan RNA-seq data as the reference integrates cells of different organoids and is used for the comparison. The RSS-based UMAP embedding is colored by cell type annotation and organoid, respectively. d UMAP embeddings based on CSS and five other integration methods, colored by RSS-based cell type annotation (top row) and organoid (bottom row). e Proportions of cell neighbors in each annotation category across all cells in the data set. Cell neighbors are defined as the top 50 cells with the shortest Euclidean distances from the cell in PCA (no integration, left) or the seven different integrated representation spaces. Bars on top are colored by the median proportions, with the numbers showing ranks of different methods. (f) Distribution of LISI scores of organoid batches (left) and N-LISI scores of cell types (right) for all cells, without integration or with each of the seven integration. Bars on top are colored by averages across all cells, with the numbers showing ranks of different methods