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

GLM-Missense, a fine-tuned genomic language model, adds the most complementary signal to seven missense variant impact predictors on held-out ClinVar variants and independent RAD51D saturation genome editing data.

• Why it matters: Missense predictors often focus on protein-level effects, leaving nucleotide sequence context—such as splicing-related signals—insufficiently captured.
• What they did: The team fine-tuned a genomic language model specifically for missense variant impact prediction, then evaluated complementarity against seven established predictors on held-out ClinVar variants and RAD51D saturation genome editing data.
• The result: The added signal is partly linked to splicing disruption, including exonic splicing regulatory elements largely missed by SpliceAI, and may also reflect codon-level effects on mRNA stability.

AI summary of the abstract · may contain errors · not medical advice · read the paper.