Coevolution-informed Bayesian optimization for sample-efficient protein design

Protein engineering is often limited by the cost of evaluating variant fitness rather than by the ability to generate variants, so methods that can learn from sparse data are valuable. The authors introduce ALSEBO (Active Learning Sequence Exploration via Bayesian Optimization), which connects a generative latent sequence landscape to Bayesian optimization and uses direct-coupling-analysis (DCA) coevolutionary statistics to featurize candidates. This representation imposes a specific inductive bias: it places the dominant organizer of the fitness landscape along a single linear coordinate, creating a smooth funnel-shaped objective that a low-data surrogate model can navigate efficiently.

On a virtual avGFP fluorescence benchmark, ALSEBO reached the optimum in only 40 evaluations and outperformed both protein-language-model embeddings and raw latent coordinates. Control experiments with representation-neutral oracles showed the advantage is intrinsic to the method rather than an artifact of the benchmark. Molecular dynamics simulations of the optimized variant recovered structural hallmarks of fluorescence, and ALSEBO transferred successfully to divergent GFP orthologs and a non-GFP enzyme, indicating a general, data-efficient route to protein design.