Information Leakage in Enzyme Substrate Prediction.

Models trained on enzyme-small molecule interaction datasets are vulnerable to similarity-induced information leakage, inflating performance metrics in out-of-distribution tests.

• Why it matters: Accurate prediction of enzyme interactions with small molecules is vital for understanding cellular processes, but current models may overestimate their true generalization ability due to data biases.
• What they did: The authors analyzed a popular dataset and four models, revealing that removing ligand similarity drastically reduces predictive performance, highlighting the models' reliance on small-molecule similarity rather than true generalization.
• The result: Findings suggest that current models primarily generalize across enzyme sequences but not across ligand chemical space, emphasizing the need for more robust methods to assess enzyme-small molecule interactions.