Fig. 4: EmpirDE rigorously controls FDR.
From: Worm Perturb-Seq: massively parallel whole-animal RNAi and RNA-seq
a, b Benchmarking the performance of EmpirDE analysis framework. The observed False Discovery Proportion (FDP) is compared to target FDR (a) and power (b). The full metabolic-gene WPS dataset (3691 samples) was simulated 10 times with random mean fluctuation (Δμ) to produce the error bars of each metric. FDP and power were measured based on a pooled set of 117,096 simulated DE changes in all conditions (Supplementary Methods). c Benchmarking FDR control of different multiple testing adjustment strategies. The data points and error bars in (a–c) indicate mean ± s.d. from 10 simulations. d Evaluating false discoveries using NTP experiments. We estimated false discoveries using the 90% quantile of the numbers of DEG across 71 NTP conditions (shown in the heatmap color and numbers). The 90% quantile represents the value below which 90% of the data points fall, effectively capturing the upper range of typical DEG counts while excluding the most extreme outliers. The red lines show the threshold boundary for five false positive DEG calls. e Number of DEGs (defined by FDR < 0.1 and FC > 1.5 using EmpirDE) for 36 perturbations that were repeated by a second WPS experiment. f Fraction of unreproducible DEGs for EmpirDE versus DESeq2 analysis. Unreproducible DEGs were defined by genes that are called as DEG in one experiment (FDR < 0.1, FC > 1.5) but confidently non-DEG in the other (FC < 1.1 or show a different FC direction). The red dashed line shows the theoretical FDR (FDR = 0.1). Comparison of log2(FC) measured in two independent experiments for representative RNAi with either high (g) or moderate (h) number of DEGs. The green dashed line indicates the diagonal (y = x).
