GMSF: A Dual-Path Multimodal Framework for Enzyme Function Prediction via Difference Graph Encoding and Multiscale Semantic Fusion

Abstract

Enzymes play a central role in green chemistry and biomanufacturing. However, precise enzyme selection via Enzyme Commission (EC) number prediction remains a critical bottleneck between the retrosynthetic pathway planning and experimental implementation. Existing computational methods predominantly rely on single-modality representations that either lose spatial topological structures or treat reactions as static snapshots, failing to capture the explicit atom-level state transitions in chemical bond evolution. To address these limitations, we propose GMSF, a dual-path multimodal deep learning framework. GMSF features two core innovations: Difference Graph Encoding, which leverages atom–atom mapping (AAM) to compute node feature differences and extract structural difference signatures of reaction centers; and Multiscale Sequence Encoding, which uses AAM-aligned node mappings to hierarchically capture chemical semantics. Experimental results on the ECREACT dataset demonstrate that GMSF achieves an accuracy of 93.17% in the highly challenging level 3 (subsubclass) enzyme function prediction, outperforming the state-of-the-art (SOTA) baseline by a margin of 1.36%. Furthermore, its successful application to real-world metabolic networks in the BioCyc database validates its immense potential for deciphering complex biochemical reactions, serving as a powerful computational screening tool to narrow down catalytic mechanisms and coenzyme dependencies prior to the exact enzyme selection.