Benchmarking and Experimental Validation of Machine Learning Strategies for Enzyme Engineering

Enzyme-directed evolution increasingly relies on computational tools to prioritize mutations, yet their practical value is difficult to assess because kinetic data are often aggregated across heterogeneous assay conditions, inflating apparent generalization. Here we introduce EnzyArena, a curated benchmark in which kinetic parameters (kcat, Km, or kcat/Km) are grouped by experimental campaign and measured under matched conditions, thereby minimizing confounders such as pH, temperature, reaction, and batch. Using EnzyArena, we systematically evaluate 20 models spanning four widely used strategies for guiding enzyme evolution: protein-ligand binding affinity prediction, protein stability prediction, zero-shot fitness prediction and enzyme kinetic parameters prediction. Across subsets derived from public databases and 25 literature-derived mutagenesis data sets, most models exhibit weak, fragile, or inconsistent correlation with measured functional outcomes. Kinetic-parameter predictors perform strongly on database-derived subsets but lose their advantage on literature-derived mutagenesis data sets, whereas zero-shot predictors show more consistent generalization. A simple consensus of multiple zero-shot models further improves the precision of identifying beneficial mutants in retrospective analyses. We prospectively validated these findings in a wet-lab campaign (150 mutants) comparing random mutants, UniKP-prioritized mutants, and ESM-1v-prioritized mutants (representing supervised kinetic parameters prediction and zero-shot fitness prediction, respectively), where ESM-1v achieved the highest utility and UniKP underperformed the random baseline. Together, this study establishes realistic baselines for computational mutant prioritization and indicates that zero-shot strategies provide the most consistent, although still modest, performance among the approaches evaluated.