Reinforcement learning with evolutionary priors unlocks multi-site enzyme optimization from sparse experimental feedback

Industrial enzymes underpin sustainable manufacturing, yet natural proteins rarely combine the activity, robustness and application compatibility required for practical use. Engineering such properties requires identifying beneficial combinations of interacting mutations within vast sequence spaces that are too large to search exhaustively and too sparsely sampled to model reliably. Here we introduce AlphaVariant, a reinforcement-learning (RL) framework that couples a generative evolutionary prior with iteratively updated low-N fitness models to enable feedback-driven search of large combinatorial protein sequence spaces. Across six experimentally measured protein fitness landscapes, AlphaVariant ranked first on the three large, non-enumerable multi-site landscapes. In two successive rounds of savinase activity optimization, AlphaVariant-designed libraries yielded 75.4% and 83.8% activity-improved variants among experimentally characterized candidates even as the accessible mutation space expanded from 25 to 67 positions. The search identified beneficial mutation combinations involving up to 17 structurally dispersed sites and produced a variant with 7.1-fold higher activity than wild type after only two design-test cycles. We then extended AlphaVariant to jointly optimize activity, thermostability and a patent-informed constraint for intellectual-property freedom. This multi-objective search yielded variants that outperformed the commercial protease Uno 100 L in both biochemical assays and application-relevant detergent washing tests. These results demonstrate that generative RL can efficiently navigate vast protein fitness landscapes, progressively incorporate multiple engineering objectives and convert sparse experimental feedback into enzyme variants with industrially relevant performance, effectively resolving the limits of biocatalyst engineering.