Fig. 2

A Illustration of challenges for cohort-wide DE testing. B Setup for our simulation setup. C Isomap of the cells based on the true DE genes’ expression before introducing the confounding variables. No individuals concentrate tightly in any region on the Isomap manifold, and there is a strong separation between the cases (shades of red) and controls (shades of blue). D Isomap of the observed data based on all the genes. Cells from the same individual concentrate in the embedding, suggesting that confounding covariates additionally drive the difference in expression profiles among individuals. E Downsampling experiment, demonstrating that by pooling information across genes, eSVD-DE outperforms gene-by-gene Negative Binomial regression for regressing out covariate effects. F Illustration of the importance of shrinkage, where the x-axis and y-axis represent each gene’s test statistic with and without posterior correction, respectively. The genes are colored by their true log-fold change, of which the circled genes denote the top 50 genes with the highest true log-fold change. G ROC curve comparing four different methods, illustrating that eSVD-DE has more power than competing methods. The area under the curve (AUC) is shown for each method, where the percentage represents the area between the method’s curve and the diagonal line as a fraction of total possible area. The bolded method denotes the method with highest AUC