Borzoi-informed fine mapping improves causal variant prioritization in complex trait GWAS

Abstract

1Genome-wide association studies (GWAS) have identified thousands of trait-associated loci. Prioritizing causal variants within these loci is critical for characterizing trait biology. Statistical fine mapping identifies causal variants at trait-associated loci, but linkage disequilibrium (LD) and limited GWAS sample sizes prevent the resolution of many associations. Functionally informed approaches augment fine mapping by estimating variant prior causal probabilities based on overlap with trait-relevant functional annotations. However, functional enrichment provides only an indirect proxy for variant functional impact. Sequence-to-function models directly estimate variant effects on molecular phenotypes from underlying sequence context. Borzoi is a long-context model that predicts sequence determinants of transcription, splicing, and polyadenylation across diverse tissues and cell types. Here we present Sniff, a Borzoi-informed fine-mapping approach that integrates broad genomic functional annotations with Borzoi-predicted variant effects via PolyFun to estimate variant prior causal probabilities. Applied to 15 UK Biobank traits, Sniff identifies 9.45% additional fine-mapped variants compared to PolyFun-Baseline at posterior inclusion probability (PIP) > 0.8. Sniff-prioritized variants exhibit allele-specific activity in reporter assays and are predicted to have tissue-specific activity in trait-relevant tissues. For most traits, genes nominated by Sniff receive higher scores from the orthogonal gene prioritization method PoPS compared to genes nominated using functional annotations alone. Because differentially prioritized variants are driven by Borzoi predictions, we leverage attribution techniques to characterize sequence features underlying fine mapping and generate mechanistic hypotheses for GWAS associations.