Codon language model scores provide information beyond protein language models for missense variant interpretation

Abstract

Predicting variant effects remains a central challenge in genomics. Protein language models (PLMs) capture amino-acid-level sequence constraints, whereas codon language models operate on coding sequences and may retain information that is lost upon translation. Here, we tested whether scores from the codon language model CaLM provide predictive information beyond protein-level representations for missense-variant interpretation. Across 71,436 ClinVar missense variants from 11,554 genes, adding CaLM to PLM baselines produced modest but reproducible improvements under gene-held-out cross-validation. PLM-only ensemble controls and explicit mutational-context analyses indicated that this improvement could not be explained solely by generic ensembling or simple codon-substitution features. Aggregating CaLM probabilities across synonymous codons attenuated codon-degeneracy-associated discordance while preserving most of the broader differences between CaLM and PLM scores. Gene-level analyses further showed that CaLM contribution varied continuously across genes and depended partly on the protein-model background. Across ClinMAVE functional assays, however, improvements were less consistent, indicating that codon-protein complementarity is context dependent rather than universal. Together, these results identify a modest but reproducible component of variant-effect information in CaLM-derived codon-level scores that is not fully captured by protein-level language-model representations.