Fig. 2

Comparison of Clipper with generic FDR control methods in terms of their FDR control and power in six example simulation studies. a 1vs1 enrichment analysis with 1000 features generated from the Gaussian distribution with a homogeneous background. b 1vs1 enrichment analysis with 10,000 features generated from the Gaussian distribution with a heterogeneous background. c 2vs1 enrichment analysis with 10,000 features generated from the Poisson distribution with a heterogeneous background. d 3vs3 enrichment analysis with 10,000 features generated from the Gaussian distribution without outliers and with a heterogeneous background. e 3vs3 enrichment analysis with 10,000 features generated from the Gaussian distribution with outliers and with a heterogeneous background. f 3vs3 differential analysis with 10,000 features generated from the negative binomial distribution with a heterogeneous background. At target FDR thresholds q∈{1%,2%,⋯,10%}, each method’s actual FDRs and power are approximated by the averages of false discovery proportions (see Additional File 1: Eq. (S14)) and power evaluated on 200 simulated datasets. In each panel, the top row shows each method’s actual FDRs at target FDR thresholds: whenever the actual FDR is larger than the target FDR (the solid line is higher than the dashed line), FDR control is failed; the bottom row shows each method’s actual FDRs and power at the target FDR threshold q=5%: whenever the actual FDR is greater than q (on the right of the vertical dashed line), FDR control is failed. Under the FDR control, the larger the power, the better. Note that BH-pair-correct is not included in a–c because it is impossible to correctly specify the model with only one replicate per condition; locfdr-swap is not included in a and b because it is inapplicable to the 1vs1 design