Estimating false discovery rate and false non-discovery rate using the empirical cumulative distribution function of p-values in 'omics' studies

Abstract

Large numbers of mRNA transcripts, proteins, metabolites, and single nucleotide polymorphisms can be measured in a single tissue sample using new molecular biological techniques. Accordingly, the interpretation of ensuing hypothesis tests should manage the number of comparisons. For example, cDNA microarray experiments generate large multiplicity problems in which thousands of hypotheses are tested simultaneously. In this context, the false discovery rate (FDR) and false non-discovery rate (FNR) are used to account for multiple comparisons. In this study, we propose non-parametric estimates of FDR and FNR that are conceptually and computationally straightforward. Additionally, to illustrate their properties and use in a procedure for an optimum subset of significant tests, an example from a functional genomics study is presented.