Relational biological structure improves fine-mapping of causal GWAS variants under weak signal

Linkage disequilibrium (LD) makes causal GWAS variants indistinguishable from correlated neighbours; resolving them is the fine-mapping problem, and the challenge is species-specific: humans face dense ancestry-imbalanced LD, yeast and Arabidopsis exceptionally long LD, and crop germplasm sparse and fragmented annotations that defeat human-biobank curation pipelines. Bayesian fine-mappers integrate annotations as flat per-variant priors, discarding the relational structure linking variants to tissue-specific eQTLs, pathways and protein–protein interactions. Hierarchical belief propagation (HBP) on a variant– gene–pathway factor graph matches Bayesian baselines at 5–40× speed; an annotation-adaptive complement, graph-augmented fine-mapping (GAFM), wins 27–2 against SuSiE at weak signal and recovers LDLR, APOE, LPL, GCKR and ANGPTL3 at single-variant resolution across four Pan-UK Biobank ancestries. On the 3,000 Rice Genomes grain weight + shape panel, mixture-prior posterior reweightings of GAFM/HBP and their ensemble (GAFM-MX, HBP-MX, ENS) reach 47.6% top-1-PIP exact-position recovery of 21 panel-matched stable QTNs — the highest of any method, exceeding SuSiE (28.6%) and SBayesRC (14.3%) —at 200–700× SuSiE’s per-locus speed. Across 692 leads in four species, a non-uniform per-variant prior, not uniform high coverage, lets the graph break LD ties: adding a regulatory-element flag to an otherwise uniform human cache flips HBP narrower than GAFM from 0% to 88% on 321 Pan-UKB leads. These results recast multi-omics fine-mapping as a non-uniform-prior-curation problem rather than a uniform-coverage problem, and reframe post-GWAS analysis as message passing over biological structure rather than weighted regression on flattened annotations.