JustEvo: A Prior-Guided Adaptive Mixture-of-Experts Framework for Multifunctional Protein Directed Evolution

Abstract Machine-learning-guided directed evolution can reduce the experimental burden of protein engineering by prioritizing variants before they are tested. In practice, however, many campaigns begin with few or no target-specific measurements, face choices about how to use heterogeneous protein-model priors, and require functional improvement to be balanced with developability properties such as expression and stability. Here, we present JustEvo, a closed-loop framework for low-data and multifunctional protein evolution. Before experimental measurement data are available, JustEvo nominates an informative first experimental batch using a mutation-prior model or a property-specific predictor. After measurements are obtained, it trains surrogate experts on multiple protein representations and adaptively combines their predictions according to their agreement with the current experimental fitness ranking. We evaluated these components using an EVOLVEpro-derived MoE validation data set and a ProteinGym-derived DMS collection split into a cold-start validation data set and a reserved benchmark data set. On the 81-landscape benchmark data set, JustEvo achieved higher final-round top 5 average fitness than EVOLVEpro and RelaVDEP under the same finite-budget acquisition protocol, with normalized scores of 0.862 for JustEvo, 0.849 for EVOLVEpro, and 0.818 for RelaVDEP. We further adapted JustEvo for fluorescent-protein optimization as a multiobjective design task. In this setting, objective-specific surrogate models were used to rank candidates by their predicted joint performance across coupled properties. We tested this extension on four in silico GFP landscapes and an experimental EYFP closed-loop campaign, where JustEvo expanded the fluorescence-stability Pareto front and increased Pareto hypervolume by 872.1%. These results show that coordinating prior-guided initialization, adaptive mixture-of-experts integration and objective-aware candidate ranking can improve protein variant selection under limited assay budgets.