Protein language models learn the underlying accessibility of mutations alongside protein fitness landscapes

Virus-specific fine-tuning of base protein language models (ESM-2, ESM-C) amplifies their implicit learning of nucleotide-level mutational constraints, improving observed amino-acid change prediction.

• Why it matters: PLM pseudo-probabilities mix protein structural constraints with nucleotide mutation biases, limiting accurate evolutionary forecasting and synthetic fitness inference.
• What they did: Using the rapid emergence of the influenza A H3N2 K lineage, the authors sweep parameters to couple or decouple empirical nucleotide mutational supply from PLM amino-acid substitution pseudo-probabilities across forecasting tasks.
• The result: Explicit mutational accessibility improves binary prediction of observed amino-acid changes, but degrades prediction of final circulating variant frequency after emergence; epistatic scanning links I160K to co-occurring S144N and N158D.

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