Optimality of Gene Ranking Based on Univariate P-values for Detecting Differentially Expressed Genes

Abstract

Ranking significant genes based on the P-value in multiple testing is a simple and common practice in microarray data analysis, and its theoretical optimality is of particular interest. McLachlan et al. (Bioinformatics 2006; 22: 1608-1615) presented a method for calculating the local FDR under normal mixture models and provided a theoretical optimality of the local FDR as a ranking statistic. In this article, we show that the optimal gene ranking based on the local FDR calculated by the McLachlan et al.’s method perfectly accords with that based on P-value under certain conditions. We argue that these conditions are generally satisfied for significant genes with small P-values. We demonstrate it using several real examples.