Sampling protein language models for functional protein design

Protein language models have emerged as powerful tools for learning rich protein representations, improving tasks like structure prediction, mutation effect estimation, and homology detection. Their ability to model complex sequence distributions also holds promise for designing novel, functional proteins with applications in therapeutics, materials, and sustainability. Given the vastness of sequence space, efficient exploration is essential. However, most existing design approaches rely on single-mutant sampling strategies borrowed from natural language processing, which fail to capture epistatic interactions critical for function. Here, we develop an in silico framework to systematically compare sampling methods and introduce several approaches tailored for protein design. We show that sampling multiple mutations simultaneously substantially outperforms single-mutant approaches by better capturing epistatic effects. We evaluate these strategies across three protein families spanning eukaryotic, prokaryotic, and viral origins, examine key hyperparameters, and validate our findings on a de novo binder design task against a major therapeutic target.