Leveraging protein language models to identify complex trait associations with previously inaccessible classes of functional rare variants

Protein language models (PLMs) improve variant effect predictions, but their role in gene discovery for complex traits remains unclear. We introduce an allelic series-based regression test that uses PLM-derived variant effect predictions as proxies for effect sizes, identifying ∼46% more associations than standard burden tests. Extending this to isoform-level analysis, we find 26 gene-trait pairs with stronger associations in non-canonical versus canonical transcripts, highlighting isoform-specific effects. Finally, we identify evolutionary plausible variants (EPVs), missense variants assigned higher likelihoods than the wild-type alleles by PLMs, representing 0.45% of missense variants. EPVs show higher allele frequencies than synonymous variants, consistent with differential selection pressures, and are linked to nine traits, including protective associations with low-density lipoprotein (LDL) and bone mineral density. Together, our results demonstrate how PLMs can enhance rare-variant interpretation and gene-trait association discovery in exome data.