Genomic foundation models fail to add common-variant discoveries: standard annotations still lead across 27 UK Biobank traits

GFM variant scores were run through the same functionally informed GWAS discovery pipeline and compared head-to-head with the standard functional annotation baselineLD on how many independent significant loci each recovers. Earlier GFM evaluations focused on heritability enrichment, rare-variant testing and fine mapping; common-variant discovery had not been systematically tested. Across 27 approximately independent UK Biobank traits, baselineLD raised discovered loci by a median 5.07%, AlphaGenome and Borzoi alone by 1.78% and 3.80%, and GPN-Star and Zoonomia reduced the count.

Adding GFM annotations to baselineLD left locus counts essentially unchanged, indicating the two carry highly overlapping information. Combined annotations found 13,416 to 13,447 loci versus 13,420 for baselineLD alone, a median cross-trait difference of −0.1% to 0.2%. Borzoi principal components correlated strongly with baselineLD regulatory annotations in LD score; after conditioning on the regulatory class, partial correlations with other classes such as conservation, coding and molecular QTL were minimal.

Each SNP's functional annotations are first collapsed by stratified LD score regression into an expected chi-squared value, which enters the surrogate functional false discovery rate (sfFDR) framework as a surrogate variable that reranks SNPs and yields functional p-values. sfFDR was originally built to incorporate GWAS summary statistics from related traits; here it was extended to accept SNP-level functional annotations. In 11 traits with independent FinnGen data, the 657 loci unique to sfFDR replicated at 0.69 (95% CI 0.65–0.72), versus 0.61 (95% CI 0.56–0.66) for the 347 loci unique to the standard analysis.

The evaluation covers 27 approximately independent UK Biobank traits spanning anthropometric, blood cell, disease and other categories, with GWAS summary statistics from up to 459,324 European-ancestry samples. GFM scores were integrated off the shelf with no optimization of the models; in the combined analysis Borzoi scores were distilled by PCA and AlphaGenome scores took the maximum absolute value per output type, so the two-model comparison reflects representations as much as models. The authors propose next steps including PCA or variational autoencoders for AlphaGenome, restricting tissue-specific scores to trait-relevant tissues, and fine-tuning via a lightweight model or low-rank adaptation.

Results concern off-the-shelf GFM outputs without fine-tuning, and the authors explicitly note that fine-tuning or feature engineering could change the conclusion. Gains in signal enrichment and out-of-sample R² did not always translate into more discoveries, which the authors attribute to heterogeneous genetic architecture of common traits and to reranking pushing some SNP classes up while pushing others down. Fully annotating genome features may require much larger data corpora than current training sets. Adding the AlphaGenome variant impact score AVI to baselineLD gave a median locus increase of 5.23% versus 5.07%.