Incorporating feature reliability in false discovery rateestimation improves statistical power to detect differentially expressed features

Feature selection is a critical step in translational omics research. False discovery rate (FDR) is anintegral tool of statistical inference in feature selection from high-throughput data. It is commonly used to screen features (SNPs, genes, proteins, or metabolites) for their relevance to the specific clinical outcome under study. Traditionally, all features are treated equally in the calculation of false discovery rate. In many applications, different features are measured with different levels of reliability. In such situations, treating all features equally will cause substantial loss of statistical power to detect significant features. Feature reliability can often be quantified in the measurements. Here we present a new method to estimate the local false discovery rate that incorporates feature reliability. We also propose a composite reliability index for metabolomics data. Combined with the new local false discovery rate method, it helps to detect more differentially expressed metabolites that are biologically meaningful in a real metabolomics dataset.