Mechanistically Informed Open-World Enzyme Retrieval with a Dual-Tower Graph-Sequence Model

Identifying enzymes capable of catalyzing specific chemical transformations across large sequence databases remains a major challenge in biocatalyst discovery. Conventional fingerprint-based methods capture global molecular structure but fail to represent bond-breaking and bond-forming events, limiting generalization to structurally novel reactions. We introduce a dual-track evaluation framework to distinguish true generalization from memorization, assessing retrieval on structurally isolated reactions (n = 50) within a 63,259-sequence enzyme pool. The strongest fingerprint baseline achieves R@10 = 0.020. To address this limitation, we develop GATv2-ECR, a heterogeneous dual-tower model integrating reaction-center graph encoding, a frozen ESM-2 sequence encoder, contrastive learning, and EC-aware soft reranking. GATv2-ECR achieves R@10 = 0.160 on isolated queries and R@10 = 0.308 on an out-of-distribution subset (n = 39), capturing mechanistically relevant features and supporting generalizable enzyme retrieval under open-world conditions.