Adaptive Partial Conjunction Hypothesis: Identifying Pleiotropy Across Heterogeneous Effect Units

Abstract

Pleiotropy arises when a single SNP, gene, or locus affects multiple traits, and in many applications the key scientific question is which specific subset of traits shares a non-null effect. We introduce the Adaptive Partial Conjunction Hypothesis (APCH), an empirical Bayes framework that uses summary statistics to test, for each effect unit, which subsets of features are jointly non-null while controlling the false discovery rate for subset selection. APCH combines adaptive shrinkage priors with a hierarchical model that accounts for correlation among effect estimates and computes local false discovery rates for all candidate subsets. A top-down search over subset size then selects at most one maximally informative jointly significant subset per effect unit. In simulations, APCH maintained well-calibrated FDR and, in most settings, achieved higher power than both empirical Bayes methods that learn across effects and subset-based procedures that test each effect unit separately. It remained reliable with heterogeneous effect sizes, correlated estimation errors across features, and sparse jointly significant patterns. We apply APCH to GWAS summary statistics for five type 2 diabetes-related traits and identify co-occurring trait subsets that go beyond what single-trait genome-wide scans reveal. In particular, APCH brings in traits that fall short of genome-wide significance on their own but repeatedly appear in the same jointly significant subsets, revealing stable cross-trait co-association patterns and sharpening our overall picture of how loci act across these traits. More broadly, APCH is directly applicable to other multi-feature settings, such as multi-tissue or multi-omic association analyses, using only summary statistics.