Adaptive Representations for Real-World Enzyme Optimization

Enzyme fitness landscapes are highly rugged, giving rise to complex optimization landscapes. Moreover, engineering campaigns rarely explore them uniformly: initial screens might saturate the active site before extrapolating to distal positions or to higher-order combinations of the hits found, so the mutational context in which a surrogate must predict shifts from one round to the next. Bayesian optimization (BO) with Gaussian processes (GPs) is widely used to guide such campaigns for its sample efficiency and calibrated uncertainty, but the representation it relies on, whether a one-hot encoding or a protein language model (PLM) embedding, is fixed in advance and cannot adapt as measurements accumulate. Here, we study a joint fine-tuning framework that optimizes the PLM and GP parameters under a single marginal-likelihood objective, across seven enzyme landscapes spanning the mutation regimes encountered in real-world campaigns. Joint fine-tuning yields a median relative improvement in top-5% coverage of 45% over one-hot encoding and 120% over mean-pooled static ESM-C embeddings, and among the pooling strategies we examine, mutation-site pooling, which localizes the representation to the mutated residues, is the most consistent across landscapes. Most consequentially, the adapted representation transfers under mutational shift: when the campaign moves from active-site to distal mutations, or from single mutants to combinatorials, it attains up to twice the top-5% coverage of static representations and recovers top-10 variants that one-hot and frozen embeddings fail to identify. Adapting the representation during a campaign therefore preserves the value of data acquired under earlier mutational regimes, and is the better design choice for BO-guided enzyme engineering.