Controlling error rates for gene signatures. A: Realized level and power for the bootstrap test at 5% nominal level. B: Realized FDR, power and stability for signatures selected by the bootstrap test after Benjamini-Hochberg (BH) correction. Here the nominal FDR was set at 5%. C: Same as (B) for signatures selected by recursive feature elimination (RFE). D: Same as (B) for signatures selected as the top 200 genes. Acc, classifier accuracy.