Exploratory data analysis for large-scale multiple testing problems and its application in gene expression studies

TL;DR

This paper introduces a novel empirical Bayes-based method with sample splitting for large-scale multiple testing, improving false discovery rate control and discovery reliability in gene expression studies.

Contribution

It proposes a cross-validation approach with detection frequency thresholds to enhance Fdr control and explore relationships among significant findings.

Findings

Effective Fdr control demonstrated on microarray and RNA-seq data

Method reduces overfitting through resampling and cross-validation

Power analysis shows improved detection efficiency

Abstract

In large scale multiple testing problems, a two-class empirical Bayes approach can be used to control the false discovery rate (Fdr) for the entire array of hypotheses under study. A sample splitting step is incorporated to modify that approach where one part of the data is used for model fitting and the other part for detecting the significant cases by a screening technique featuring the empirical Bayes mode of Fdr control. Cases with high detection frequency across repeated random sample splits are considered true discoveries. A critical detection frequency is set to control the overall false discovery rate. The proposed method helps to balance out unwanted sources of variation and addresses potential statistical overfitting of the core empirical model by cross-validation through resampling. Further, concurrent detection frequencies are used to provide visual tools to explore the inter-relationship between significant cases. The methodology is illustrated using a microarray data set, RNA-sequencing data set, and several simulation studies. A power analysis is presented to understand the efficiency of the proposed method.