Fig. 2: MetaQ effectively and efficiently infers prototypical metacells.
From: MetaQ: fast, scalable and accurate metacell inference via single-cell quantization
a UMAP visualization of the original 433,495 cells from the human fetal atlas. The red box highlights retina cells. b Classification accuracy of cell classifiers trained with [500, 1000, 2000, 4000] metacells inferred by MetaQ, SEACell, MetaCell V2, SuperCell, and random sub-sampling on five random experiments. Each boxplot ranges from the upper and lower quartiles with the median as the horizontal line and whiskers extend to 1.5 times the interquartile range. Two-sided T-test results: *0.01 < p ≤ 0.05, **0.001 < p ≤ 0.01, ***0.0001 < p ≤ 0.001, ****p ≤ 0.0001. c UMAP visualization of 4000 metacells inferred by the four methods, with cell type colors matching those in b. Retina cells are marked with red boxes. d Agreement between the ground-truth annotations and the labels predicted by classification models trained with 500 metacells. Matrices with a clearer diagonal structure indicate better classification performance. e Compactness of [500, 1000, 2000, 4000] metacells inferred by different methods on five random experiments. Each boxplot ranges from the upper and lower quartiles with the median as the horizontal line and whiskers extend to 1.5 times the interquartile range. Two-sided T-test results: ****p ≤ 0.0001. f Separation of [500, 1000, 2000, 4000] metacells inferred by different methods on five random experiments. Each boxplot ranges from the upper and lower quartiles with the median as the horizontal line and whiskers extend to 1.5 times the interquartile range. g Running times (logged) and memory cost for inferring 1000 metacells from different numbers of original cells. h Running times (logged) and memory cost for inferring different numbers of metacells from 100,000 cells. Source data are provided as a Source Data file.
