Fig. 2

Gene abundance had a large impact on the performance to identify differentially abundant genes. For each method, the receiver operating characteristics curve shows the true positive rate (y-axis) and the false positive rate (x-axis) at each position in the gene ranking list. Panels (a-c) show results for the Qin dataset and panels (d-e) show results for the Yatsunenko dataset. The genes were stratified into three parts based on the average number of DNA fragments, i) ≤500, ii) 500–5000 and iii) >5000 for the Qin dataset and i) ≤10, ii) 10–50 and iii) >50 for the Yatsunenko dataset. The effect size was set to a fold-change of 5 and the group size fixed at 6 + 6 samples. Each curve is based 100 resampled metagenomes. The methods included are edgeR, DESeq2, the overdispersed generalized linear model (oGLM), metagenomeSeq (mSeq), metastats and voom (see Additional file 7: Figure S4 for the additional eight methods)