A center-anchored adaptive hierarchical graph neural network with application in structure-aware recognition of enzyme catalytic specificity

Non-Euclidean modeling of biomolecules such as enzymes has been advanced by utilizing graph neural networks (GNNs). While existing whole-molecule modeling approaches offer significant advantages, they present two primary challenges: optimizing long-range interactions within the enzyme molecule and ensuring model interpretability. Furthermore, current GNN-based enzyme modeling methods pay insufficient attention to the hierarchical structure of biological functions. To address these challenges, we propose a novel Center-Anchored Adaptive Hierarchical GNN (CAAH-GNN) model that integrates a center-anchored method with an adaptive hierarchical strategy for structure-aware recognition of enzyme catalytic specificity. Our model features three key characteristics: (1) It autonomously identifies and integrates local key residues within enzyme molecules while embedding global graph information into each node's features, significantly enhancing prediction accuracy. (2) The spherical spatial radius is adaptively learned using a two-step hierarchically sampling method. Compared to full-molecule models, this approach reduces spatial complexity more than 80 %. (3) Our model exhibits strong generalization performance and interpretability, as it is not tailored to specific enzyme families and does not remove any nodes during training. Extensive experiments based on two enzyme datasets and various graph convolution kernels demonstrate that CAAH-GNN outperforms baseline models and other hierarchical strategy models in prediction accuracy. In summary, our research not only introduces a novel GNNbased enzyme modeling approach but also provides a new perspective and a tool for enzyme engineering.