IHGCN-PLA: An interpretable heterogeneous graph convolutional network for protein-ligand binding affinity prediction with multimodal interaction fusion.

Abstract

Protein-ligand binding affinity prediction is fundamental to computer-aided drug discovery, enabling accelerated therapeutic development at reduced costs. Despite advances in graph neural network architectures, existing methods inadequately integrate heterogeneous molecular representations, limiting their ability to model the complex multimodal nature of protein-ligand interactions. We introduce an interpretable heterogeneous convolutional network whose core innovation lies in an early-stage multimodal interaction fusion mechanism. This approach integrates complementary molecular representations-including structural topology, physicochemical properties, and interaction dynamics-at the initial stage of feature extraction. Unlike traditional late-fusion strategies, our early-stage multimodal interaction fusion achieves cross-modal feature learning through heterogeneous graph convolution, enabling residue nodes to aggregate structural signals from the protein and binding information from the protein-ligand interface within each convolutional layer. Utilizing only binding pocket-ligand data, this framework achieves performance comparable to or exceeding full-protein multimodal models. Evaluation on the PDBbind benchmark demonstrates that our framework achieves state-of-the-art performance with reduced data requirements. Ablation studies validate the contribution of the multimodal interaction fusion mechanism, with interpretability analysis revealing that multimodal interaction edges play a role in identifying key molecular determinants of binding affinity. Application to non-small cell lung cancer demonstrates practical value. Open-source code: https://github.com/trybestxk/IHGCN-PLA.