Generative Models for Novel Enzymatic Reactions: How Far Do They Go?

A structure-conditioned generative model predicts exact, stereo-respecting novel enzymatic reactions on EnzymeMap_x, recovering the complete correct reaction for 25.0% of eligible test enzyme–reaction pairs within 400 candidates.

• Why it matters: Most enzyme deep learning focuses on known reactions, leaving unknown biochemical reaction prediction underexplored, especially when test reactions are absent from training.
• What they did: The model generates reactants and products from enzyme structure and is evaluated on EnzymeMap_x, a reaction-disjoint benchmark using exact scoring; it also compares against BLASTp, DIAMOND, RxnPred, and Horizyn.
• The result: It ranks the correct reaction first for 5.6% (top ten: 12.4%) under a “not already known” assumption, and without re-ranking a sequence-only variant recovers correct products for 9.5% within five candidates.

AI summary of the abstract · may contain errors · not medical advice · read the paper.