Genomic foundation models do not yet boost common variant discovery beyond standard functional annotations

Abstract

Genomic foundation models (GFMs) use large-scale AI to model genetic sequences and their influence on molecular phenotypes. While GFMs have shown promise in rare variant analysis, their utility for common variant discovery remains untested. Using the surrogate functional false discovery rate framework for functionally informed GWAS discovery, we compared standard functional annotations (baselineLD) with annotations derived from three leading GFMs: the sequence-to-function models AlphaGenome and Borzoi, and the genomic language model GPN-Star. Across 27 UK Biobank traits, GPN-Star decreased locus discovery relative to a standard GWAS, whereas AlphaGenome and Borzoi increased discovery but by less than baselineLD (medians 1.8%, 3.8%, and 5.1%). When we combined the GFM and baselineLD annotations, there was no increase over baselineLD alone (median changes -0.1% to 0.2%), suggesting that AlphaGenome and Borzoi mostly capture information already in baselineLD. Thus, GFMs may require larger corpora or trait-specific feature engineering to improve discovery beyond standard functional annotations.