Fig. 4

Comparison of differential expression methods on simulated scRNA-seq datasets. Differential expression methods are compared based on FDP-TPR curves for data simulated from a 10x Genomics PBMC single-cell RNA-seq dataset (n=1200). Zoomed versions of the FDP-TPR curves are shown here and full curves are in Additional file 1: Figure S12. Circles represent working points on a nominal 5% FDR level and are filled if the empirical FDR (i.e., FDP) is below the nominal FDR. 10x Genomics sequencing typically involves high-throughput and massive multiplexing, resulting in very shallow sequencing depths and thus, low counts, making it extremely difficult to identify excess zeros. Unweighted and ZINB-WaVE-weighted EDGER are tied for best performance, followed by ZINB-WaVE-weighted DESEQ2. In general, bulk RNA-seq methods perform well in this simulation, probably because the extremely high zero abundance in combination with low counts can be reasonably accommodated by the negative binomial distribution. The behavior in the lower half of the curve for NODES is due to a smooth increase in true positives with an identical number of false positives over a range of low FDR cut-offs. FDP false discovery proportion, FDR false discovery rate, PBMC peripheral blood mononuclear cell, TPR true positive rate