Excluded causal variants produce false discoveries in genetic fine-mapping

Abstract

Statistical fine-mapping seeks to distinguish causal genetic variants from correlated variants in linkage disequilibrium (LD). Although fine-mapping algorithms perform well in simulations, they are often applied in settings that violate their assumptions. In particular, genome-wide association study (GWAS) and quantitative trait locus (QTL) workflows routinely omit variants with minor allele frequency (MAF) below 1% and most insertions and deletions (indels), excluding many causal variants before fine-mapping begins. To assess this empirically, we compared filtered and unfiltered fine-mapping at 499 All of Us loci. Twenty percent of high-confidence calls after a 1% MAF filter, and 32% when indels were also omitted, lacked support in the unfiltered analysis. Errors were particularly enriched when tested variants tagged an omitted variant, so fine-mapping misassigned its signal to a proxy variant. At ABO, omitting the frameshift that defines blood group O gave near-certain posterior inclusion probabilities (PIPs) to linked single-nucleotide polymorphisms (SNPs). To address this disconnect, we developed SuSiE-OUT, which enables fine-mapping in scenarios with possibly omitted causal variants. Modeling infinitesimal effects did not reliably prevent these errors, whereas imputing omitted variants or modeling them with our SuSiE-OUT method prevented most errors. Overall, we find that common analysis workflows substantially increase false discoveries in fine-mapping analyses, and we demonstrate two approaches to mitigate this challenge.