Fig. 3: The performance of CellScope in distinguishing similar cell types, detecting rare types, and performing multi-level clustering.
From: CellScope: high-performance cell atlas workflow with tree-structured representation
a UMAP visualizations of human brain cells (NHGRI dataset) generated by CellScope, Seurat, and Scanpy. Colors represent different brain regions. b UMAP visualizations of mouse pancreatic cells (Keller (P) dataset) generated by CellScope, Seurat, and Scanpy. Colors represent different cell types. c Comparison of visualization quality using Silhouette coefficients for all cell types across 36 benchmark datasets. One-sided Wilcoxon signed-rank test (alternative: greater) comparing CellScope versus Scanpy, p = 9.60 × 10−8. d Silhouette coefficients for rare cell types (defined as < 5% of total population) comparing CellScope and Scanpy across 36 datasets (one-sided Wilcoxon signed-rank test, alternative: greater, p = 4.26 × 10−5). e Silhouette coefficients for non-rare cell types (≥5% of total population) comparing CellScope and Scanpy across 36 datasets (one-sided Wilcoxon signed-rank test, alternative: greater, p = 8.82 × 10−6). c–e The sample size for each boxplot is n = 36. Boxplots display the 25%, 50% (median), and 75% percentiles, where the whiskers extend to the most extreme data points within 1.5 times the interquartile range. f Hierarchical clustering visualization of mouse lumbar sensory neurons (Usoskin dataset) showing major cell types at resolution level 2. g Extended hierarchical clustering of the same dataset showing cell subtypes at resolution level 5. h Confusion matrix comparing CellScope’s level-2 clustering results with annotated major cell types. Perfect classification shown by diagonal values. i Confusion matrix comparing CellScope’s level-5 clustering results with annotated cell subtypes. j Expression heatmap of key marker genes identified by CellScope for different cell types and subtypes in the Usoskin dataset.
