Widely used GWAS methods can be poorly suited to SNP-level localization under diffuse polygenic architecture in livestock
TL;DR
The results show that, under a highly polygenic architecture with livestock-like LD, GWAS tool choice has major consequences for biological interpretation, and methods that control long-range LD spillover should be prioritized.
Abstract
In livestock populations, genome-wide association studies (GWAS) can produce strong, apparently localized associations even when no truly discrete nearby causal effect exists. This occurs because small effective population sizes, strong family structure, long-range linkage disequilibrium (LD), and diffuse polygenic architecture can cause the effects of many variants to accumulate and be captured jointly across broad genomic intervals, making variant-level associations difficult to interpret biologically. Using real pig genotypes, we constructed a benchmark in which phenotypes were simulated under diffuse polygenic architecture across a genome partitioned into alternating effect and null windows, with central-null regions (at least 1 Mb away from effect-containing regions) positioned to detect long-range LD-driven signal propagation. We evaluated nine configurations of six GWAS methods (BOLT-LMM, REGENIE, fastGWA, FarmCPU, BLINK, and SLEMM) under this architecture. The central finding is that strong associations, of the kind normally read as evidence of nearby moderate- or large-effect variants, are produced by many of these methods even though the simulated signal is distributed across many tiny effects and cannot be localized to any single variant. The methods differed sharply in the extent of locus-level spillover: several produced large numbers of genome-wide significant loci within central-null regions, whereas the full-GRM mixed-model benchmark (SLEMM) produced no genome-wide significant loci in central-null regions. These results show that, under a highly polygenic architecture with livestock-like LD, GWAS tool choice has major consequences for biological interpretation. When the goal is to localize biologically meaningful signals rather than to flag association peaks that may merely reflect tiny effects accumulated through LD across a broad block, methods that control long-range LD spillover should be prioritized.