A fine-tuned genomic language model adds complementary nucleotide-context information to missense variant interpretation

Missense variant interpretation remains a central challenge in clinical genomics. Missense pathogenicity predictors achieve strong performance, but many emphasize protein-level consequences or overlapping annotation priors. Whether genomic language models add non-redundant nucleotide-context signal to missense interpretation remains unclear. Here, we systematically adapted genomic language models to ClinVar missense pathogenicity prediction across back-bone architectures, representation strategies, classifier heads, and adaptation regimes. In our analysis, variant-position embeddings consistently outperformed pooled sequence representations, multi-species pretraining provided the strongest backbone-level advantage, and low-rank adaptation generalized better than full fine-tuning. The resulting fine-tuned model, GLM-Missense, substantially outperformed zero-shot scoring from the same pretrained model.

To test whether GLM-Missense contributes information beyond existing methods, we built MetaMissense, an XGBoost ensemble combining GLM-Missense with AlphaMissense, ESM1b, REVEL, CADD, SIFT, and PolyPhen-2. GLM-Missense showed the lowest concordance with other predictors, retained the strongest partial association with pathogenicity after controlling for the other predictors, and ranked as the most informative non-ensemble input to MetaMissense. MetaMissense achieved the best performance in both cross-validation and held-out testing. Analyses of variants correctly classified by GLM-Missense but misclassified by several established predictors suggested two patterns. First, part of the GLM-Missense signal may reflect splice-relevant exonic context. Second, GLM-Missense appears to add value in settings where other predictors may overweight allele frequency, gene-level constraint, or amino-acid-change severity. However, these features explained only about 10% of the distinction between the GLM-Missense-correct subset from the background. Together, our results demonstrate that fine-tuned genomic language models contribute complementary nucleotide-context information to missense variant interpretation.