False discovery rate control: Moving beyond the Benjamini-Hochberg method

Abstract

Bioinformatics studies often involve numerous simultaneous statistical tests, increasing the risk of false discoveries. To control the false discovery rate (FDR), these studies typically apply a statistical method called the Benjamini-Hochberg (BH) method. However, BH can be overly conservative, particularly in small-sample studies, and it does not take advantage of relevant structural information among the hypotheses, such as groupings. Group structures can arise, for example, when genomic features located in close proximity are co-regulated. Recent statistical developments have yielded group-adaptive BH methods that can leverage pre-existing group information to improve statistical power while maintaining FDR control. However, these methods remain underutilized in bioinformatics practice. In this study, we illustrate the practical application of group-adaptive BH methods using a previously published, moderately scaled microRNA (miRNA) dataset. Even under simple groupings based on chromosomal location, these methods identified more miRNAs with significantly deregulated expression (FDR-adjusted p-value < 0.05) compared to the traditional BH method. Most of the new discoveries are supported by prior literature and a related 2017 study. Although sensitivity to grouping strategy varied across methods, our control analysis indicated that, for most methods, the additional detections may be attributable to the incorporation of group information. Our results highlight the potential of specialized BH methods for controlling the FDR in omics studies with pre-defined group structures, and motivate further evaluation of their generalizability across diverse datasets.