DecoyOpt: A two-stage pipeline for decoy-aware enzyme discrimination and EC number prediction

DecoyOpt, a two-stage decoy-aware pipeline, achieves MCC 0.899 and a 6.17% D0 decoy false-positive rate on the GraphPart 40% split for enzyme EC prediction.

• Why it matters: Similarity and deep learning methods can misclassify active enzymes as closely related decoys created by local functional-site mutations, inflating false positives.
• What they did: DecoyOpt first generates putatively inactive functional-site decoys using ESM3, then trains an MLP with classification and ranking losses to separate enzymes from unannotated proteins and decoys before a contrastive model predicts EC numbers.
• The result: Across splits, it improves EC number prediction on the full test set versus retrained baselines while staying close to CLEAN on the enzyme subset, correctly annotating all four wild-type enzymes and rejecting all eight inactive or nearly inactive variants.

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