Accurate enzyme function annotation is a grand challenge due to the vast number of uncharacterized proteins and the difficulty of distinguishing subtle functions. We introduce EnzymeHunter, a deep learning framework that achieves fine-grained prediction via a novel, hierarchically-aware contrastive learning strategy. By synergistically integrating sequence and structural information and using the EC hierarchy to guide its loss function, our model learns a functionally coherent embedding space where distances reflect precise levels of catalytic similarity.

EnzymeHunter significantly outperforms state-of-the-art models, particularly in the most challenging scenarios. It excels at achieving fine-grained precision down to the fourth EC level, maintains robust performance in low-homology cases, and accurately predicts rare enzyme classes, effectively overcoming long-standing barriers in the field. We showcase its real-world utility through diverse applications, including the high-confidence completion of partial database annotations and the correct identification of multifunctional enzymes. In a proteome-wide application to Thermus thermophilus, EnzymeHunter discovered novel catalytic functions, one of which was subsequently validated by an independent UniProt update, providing unequivocal evidence of its predictive power. Furthermore, our model is interpretable, with its predictions guided by learned attention on mechanistically critical functional sites.

By learning to decode the complex sequence-structure-function relationship, EnzymeHunter provides a powerful, precise, and interpretable tool to illuminate the functional dark matter of the enzymome, accelerating progress in fields from fundamental genomics to applied synthetic biology.