PAMP: Continuous-to-Discrete Mutation Projection for Enzyme Engineering

Efficient enzyme engineering requires translating quantitative turnover predictions into specific mutations that im- prove catalytic activity while reducing experimental search. Here we introduce Proximal Adversarial Mutation Projection (PAMP), a gradient-guided framework that maps a differentiable turnover objective directly to single-amino-acid substitutions by combining multi-point gradient averaging with distance-penalized projection in protein embedding space. Using a substrate-conditioned Resid- ual Predictor trained on CataPro-derived data, PAMP achieved a mean first-round predicted gain of 0.1795 log2 units across 2,754 enzyme substrate pairs, with positive changes in 84.64% of cases; among the top five ranked substitutions, 99.49% of records contained at least one positive candidate. Five sequential rounds increased the mean cumulative predicted gain to 0.5631 log2 units, and re-scoring with an independent, non-differentiable predictor left the ordering of the methods intact while attenuating the absolute gain. PAMP provides a computationally efficient strategy for prioritizing experimentally testable enzyme mutations under small candidate budgets.