Simple baselines rival protein language models in dense combinatorial protein landscapes

Computational protein design demands generally applicable models that reliably predict or generate new variants with superior functional properties. A central test of this capability is whether models can extrapolate beyond the mutational regimes represented in their training data. Recent benchmarks have begun to address this question, but the performance of protein language models (pLM) remains less well characterized in densely sampled combinatorial landscapes of protein function that are commonly explored in protein engineering and design. Here, we evaluate zero-shot scoring and supervised prediction across mutational orders, comparing widely used pLMs with simple mutation-identity and evolutionary baselines. We find no consistent advantage associated with pLM architecture or scale, and that pLMs do not consistently outperform conventional baselines. We examined pLM representations and found that much of their variation could be recovered from mutation-identity encoding alone, helping explain the strong performance of such encodings without pretraining. We further found little correspondence between pLM sensitivity to sequence background and experimentally measured functional epistasis. In combinatorial library selection, pLMs likewise showed no consistent advantage over evolutionary scoring. Together, these results indicate that greater pLM complexity does not necessarily improve prediction or experimental prioritization in combinatorial protein design.