Genomic foundation models fail to outperform standard annotations in common variant discovery

A study posted as a preprint compared three genomic foundation models — AlphaGenome, Borzoi and GPN-Star — against a standard functional annotation baseline, LD, for discovering genetic variants that influence common traits. The benchmark covered 27 traits drawn from the UK Biobank. Across this comparison, none of the three models surpassed the standard functional annotation baseline, meaning that for common variant discovery, the newer AI approaches have not yet added value over established methods. The work speaks to a central question in the field: whether large deep-learning models trained on genomic sequence can improve the interpretation of results from genome-wide association studies. On the present evidence, for common variants at least, they cannot yet do so.

Written by a model from the reports · updated 3 hours ago