A fine-tuned genomic language model captures nucleotide-level information overlooked by missense variant impact predictors

Abstract

Established missense variant impact predictors primarily capture protein-level consequences, yet missense-annotated variants can also act through nucleotide sequence context. We developed GLM-Missense by systematically optimizing the fine-tuning of a genomic language model for missense variant impact prediction. GLM-Missense contributed the most complementary information to seven established predictors on held-out ClinVar variants and independent RAD51D saturation genome editing data. The complementary signal partly reflects splicing disruption, including disrupted exonic splicing regulatory elements largely missed by SpliceAI, and possibly codon-level effects on mRNA stability. These findings suggest fine-tuned genomic language models capture nucleotide-level information overlooked by current missense variant impact predictors.