Fast and covariate-adaptive method amplifies detection power in large-scale multiple hypothesis testing.

Area of Science:

• Statistical genomics
• Bioinformatics
• Computational biology

Background:

• Multiple hypothesis testing is crucial in data science, especially when covariates like variant annotations are available.
• Existing methods, such as the Benjamini-Hochberg procedure (BH), often ignore valuable covariate information.
• This leads to missed discoveries and reduced statistical power in complex analyses.

Purpose of the Study:

• Introduce AdaFDR, a novel method for multiple hypothesis testing that leverages covariates.
• Enhance detection power by adaptively learning optimal p-value thresholds.
• Provide a computationally efficient and flexible tool for broad scientific applications.

Main Methods:

• AdaFDR adaptively learns p-value thresholds using multi-dimensional covariates (numeric and categorical).
• The method is designed to control the false discovery rate (FDR) and false discovery proportion.
• Theoretical guarantees for FDR control are provided.

Main Results:

• AdaFDR discovered 32% more associations than the BH procedure in eQTL analysis of GTEx data at the same FDR.
• Extensive experiments demonstrate AdaFDR's ability to make substantially more discoveries while maintaining FDR control.
• The method proved computationally efficient and robust across various datasets.

Conclusions:

• AdaFDR significantly improves detection power in multiple hypothesis testing by incorporating covariate information.
• The method offers a powerful and flexible alternative to existing procedures, applicable across diverse scientific fields.
• AdaFDR provides a computationally efficient solution for maximizing discoveries while controlling error rates.