Hierarchical Enzyme Function Prediction Based on Structural Confidence and Active-Site-Aware Attention

Accurate enzyme function prediction is important for protein functional annotation, drug discovery, metabolic-pathway analysis, and synthetic biology. Although recent structure-based methods have improved Enzyme Commission (EC) number prediction, they still have several limitations. First, predicted protein structures may contain unreliable regions, but residue-level structural confidence is often not fully exploited. Second, most methods extract features from the whole protein structure while paying insufficient attention to catalytic residues and active-site regions. Third, the hierarchical dependency among EC-level labels is not always explicitly modeled, which may lead to inconsistent predictions across EC levels. To address these issues, this paper proposes a hierarchical enzyme function prediction framework based on structural confidence and active-site-aware attention. Enzyme structures are represented as residue-level point-cloud representations constructed from Cα coordinates. A structural-confidence-aware geometric encoder incorporates residuelevel confidence values, such as AlphaFold pLDDT, into local geometric aggregation to reduce the influence of unreliable regions. An active-site-aware attention mechanism is further introduced to emphasize function-determining residues and their local catalytic microenvironments. In addition, a hierarchical EC decoder enforces parent-child consistency among EC labels. Experiments on the RCSB and HECNet datasets show that the proposed method outperforms representative sequence-based and structurebased baselines. The proposed model achieves Macro-F1 scores of 96.48% and 95.37% on the RCSB and HECNet datasets, respectively. Ablation and interpretability analyses further demonstrate the effectiveness of the proposed components.