A0957
Title: A data adaptive nonparametric procedure to define and assess reproducibility across high-throughput studies
Authors: Austin Ellingworth - Colorado State University (United States) [presenting]
Wen Zhou - Colorado State University (United States)
Debashis Ghosh - University of Colorado Anschutz Medical Campus (United States)
Zhigen Zhao - Temple University (United States)
Abstract: Reproducibility is a fundamental aspect of research that ensures the validity of findings. Although a consensus on assessing reproducibility remains elusive, in high-throughput studies, reproducibility is often defined as the consistency of test results across experiments. Most existing approaches either rely on stringent parametric assumptions of summary statistics or only focus on the hypothesis-wise alignment of summary statistics but overlook the experiment-wise heterogeneity. Inspired by previous work, a function based on the ranks of summary statistics from each experiment is introduced to define a notion of reproducibility and also to identify reproducible discoveries. The proposed nonparametric procedure takes into account both the signal strength and experiment-wise heterogeneity. Examining the geometry of the space of ranks of summary statistics and utilizing the blessing negative association of ranks, we introduce a novel procedure for identifying reproducible findings while controlling the false discovery rate (FDR). Our method surpasses existing approaches in terms of power, effectively controlling FDR under relatively mild assumptions. We validate our theoretical findings through extensive simulations and apply our approach to two large-scale TWAS datasets, uncovering reproducible features. Overall, our innovative method significantly advances reproducibility research and offers valuable practical implications.