Computational & AI Journal of Chemical Information and Modeling (ACS)
EnzyArena benchmark finds zero-shot models most reliable for enzyme mutant prioritization
- EnzyArena groups kcat and Km measurements by experimental campaign, so comparisons use matched assay conditions and avoid pooled-data confounders.
- Across four model families, most correlated poorly or inconsistently with measured function; kinetic predictors faded on literature data, while zero-shot models held up and a consensus improved retrospective hit finding.
- In a wet-lab round, ESM-1v beat random picks whereas UniKP fell below random.
UniKP landing below random is the number I'd chase: a kcat predictor trained on pooled database values likely learned family and assay-condition signal rather than mutational effect, and the benchmark never tests that directly. Before trusting ESM-1v on our enzyme, I'd want to see how its gain over random varies by campaign and starting stability, plus a simple stability-predictor baseline combined with it.
Paper: Journal of Chemical Information and Modeling (ACS) ↗ Benchmarking and Experimental Validation of Machine Learning Strategies for Enzyme Engineering