Simple baselines rival protein language models in dense combinatorial protein landscapes

In densely sampled combinatorial protein landscapes, protein language models show no consistent advantage over simple mutation-identity and evolutionary baselines for zero-shot scoring, supervised prediction, or library selection.

• Why it matters: Protein design needs models that extrapolate beyond training mutational regimes, but pLM performance is unclear in the dense combinatorial settings used for engineering and design.
• What they did: The study evaluates zero-shot scoring and supervised prediction across mutational orders, comparing widely used pLMs to mutation-identity and evolutionary baselines, then analyzes pLM representations and links sensitivity to measured epistasis.
• The result: pLM variation is largely recoverable from mutation-identity encoding, pLM sensitivity to sequence background poorly matches functional epistasis, and pLMs show no consistent advantage in combinatorial library selection.

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