Fig. 6

Evaluation of methods for identifying expression differences in single-cell expression data sets in which cells were simulated with partial membership to 2 topics (A–D) or 6 topics (E–H). Methods compared are the Kullback-Leibler (K-L) divergence score of [47] and GoM DE with adaptive shrinkage (s-values, lfsr) and without adaptive shrinkage (p-values). The left-most panels (A, E) show the distribution of K-L divergence scores for all candidate expression differences (approximately half of 10,000 genes \(\times\) 2 or 6 topics \(\times\) 20 simulated data sets), shown separately for true expression differences (dark blue) and non-differences (orange). K-L divergence scores smaller than \(10^{-8}\) are plotted as \(10^{-8}\). Similarly, B, C, F, and G show the distribution of GoM DE p-values or s-values with or without adaptive shrinkage, separately among differences and non-differences. D and H summarize performance in identifying expression differences; it shows power and FDR as the GoM DE p-value or lfsr are varied from 0 to 1 or as the K-L divergence score is varied from large to small. Note that in E and G, some bar heights are actually larger than 25,000 but are cut off at 25,000 for better visualization