Scalable Enzyme–Reaction Retrieval Enables Metagenomic Biocatalyst Discovery

Enzyme-catalyzed reactions drive metabolism and enable diverse applications in biotechnology and sustainable synthesis, but assigning catalytic functions to uncharacterized enzymes and finding enzymes for target reactions remain major challenges. Despite recent advances in machine learning, translating enzyme function prediction into systematic discovery across the vast space of uncharacterized proteins and biochemical reactions remains challenging. Here we present EzHit, a lightweight framework for bidirectional enzyme–reaction retrieval. Across enzyme annotation, orphan-reaction retrieval and family-specific fine-tuning benchmarks, EzHit achieved strong top-ranked performance and outperformed representative enzyme–reaction matching baselines. We further show that Mahalanobis-distance-based geometric confidence provides more useful reliability estimates than probability-derived uncertainty metrics or evidential learning, enabling confidence-aware prioritization of high-scoring candidates. Owing to its efficiency, EzHit was applied to a nearly trillion-scale OMG_Prot50 × RHEA screen, yielding 8.20 billion retained enzyme–reaction associations across 30.27 million predicted enzyme sequences. Experimental validation further confirmed the activity of EzHit-prioritized metagenomic enzyme candidates for a sparsely annotated reaction. Together, these results establish EzHit as a scalable framework for enzyme annotation, reaction-guided mining and biocatalyst discovery.