Fig. 4

False discovery rate and true positive rate of DE tools using simulated data from the Zhang RNA-seq dataset. The actual FDR and TPR (at various nominal FDR) of eight DE tools from joint simulation and DGE analysis of mRNA and lncRNA. These particular results are from simulations with 25% true DE genes among 10,000 genes (constituting approximately 30% lncRNAs and 70% mRNAs) for designs with n = 20 and 40 replicates per group. The curves represent the trade-off between the average TPR and the average actual FDR at different nominal FDR (ranging from 0 to 100%). The points on the curve indicate the actual FDR and TPR values at 5% nominal FDR threshold. Although negative binomial models (edgeR, DESeq2, and QuasiSeq) showed higher sensitivity, in general they tend to lose FDR control for simulated data with lower numbers of replicates. In contrast, DESeq, NOISeq, and PoissonSeq showed better capability of controlling FDR, with actual FDR below the threshold level (5%), but these tools have lower sensitivity than all other DE tools. For simulated data with at least ten replicates per group, SAMSeq and limma tools consistently showed better FDR control and comparable TPR to negative binomial models (more results can be found in Additional file 1). DE pipelines generally exhibited substandard performance (high FDR and low TPR) for lncRNAs than for mRNAs