Evo-EquiGPS: Synergizing Dynamic Geometry, Global Topology, and Explicit Evolution for High-Precision Enzyme Active Site Prediction

Abstract

Accurate identification of enzyme active sites is a prerequisite for elucidating protein functions and guiding enzyme engineering. Driven by the exponential growth of protein sequence data from next-generation sequencing, numerous novel enzymes with low sequence similarity have been discovered. While protein structure prediction models have provided highly accurate 3D structural data for these novel enzymes, the precise identification of their active sites remains a significant challenge for existing computational methods. Existing methods are constrained by static geometric representations, limited local receptive fields of graph encoders, and evolutionary semantic dilution. To address these limitations, this study presents Evo-EquiGPS, a multimodal graph neural network framework that synergizes multidimensional features for precise enzyme active site prediction. The model incorporates a three-branch parallel encoding architecture consisting of a dynamic geometric flow, a global topological flow, and an explicit evolutionary flow. It comprehensively integrates sequence semantics, 3D structures, and explicit evolutionary constraints of enzymes. Empirical evaluations demonstrate that Evo-EquiGPS exhibits superior performance across data sets with high structural diversity. On the TS124 data set, its area under the precision-recall curve (AUPRC) surpassed that of leading models such as SCREEN and GraphEC by a significant margin (exceeding SCREEN by 13.1% and GraphEC by 15.1%). Furthermore, the model demonstrated strong generalization capabilities on the highly diverse independent test set CSA112. Overall, the Evo-EquiGPS framework significantly enhances the precision of enzyme active site prediction. This provides a solid foundation for computational protein functional annotation.