PRGNet: a Parallel Residual Graph Network for enhanced drug-target binding affinity prediction
J W Liu, Aamir Mehmood, Haiqi Liu, Yanjie Liu, Dongqing Wei, Daixi Li · BMC Genomics · 2026
Predicting drug-target binding affinity (DTA) remains a cornerstone of structure-based drug discovery but is still constrained by fundamental methodological trade-offs. Classical molecular docking relies on high-resolution protein structures and is sensitive to conformational uncertainty. In contrast, sequence-driven deep learning approaches largely disregard the three-dimensional interaction geometry that governs binding thermodynamics. These limitations hinder robust and generalizable affinity estimation across diverse chemical and biological spaces. To overcome these challenges, we introduce PRGNet (Parallel Residual Graph Network), a dual-branch graph neural architecture designed to simultaneously capture large-scale structural context and detailed atomic-level interactions. PRGNet employs a dual-branch graph neural architecture consisting of a GCN branch and a GATv2 branch, which leverage different inductive biases—degree-normalized aggregation versus learned attention weights—to extract complementary features from the same local neighborhoods. Residual connections are systematically incorporated to facilitate stable deep propagation, mitigate over-smoothing, and preserve hierarchical interaction semantics across network depths. Residual connections are systematically incorporated to facilitate stable deep propagation, mitigate over-smoothing, and preserve hierarchical interaction semantics across network depths. Furthermore, PRGNet integrates multimodal fusion of learned graph embeddings with physicochemical ligand descriptors, yielding a richer representation of protein-ligand complexes. Comprehensive evaluation on the CASF-2016 benchmark demonstrates competitive performance (RMSE = 1.2966, MAE = 1.0168, Pearson’s r = 0.8118, CI = 0.8029), consistently outperforming widely used baselines. Ablation studies substantiate the complementary and synergistic contributions of the GCN and GATv2 branches, and confirm the critical role of residual pathways in stabilizing optimization and enhancing representational depth. Importantly, external testing on the CSAR-HiQ dataset reveals strong out-of-distribution generalization. Overall, PRGNet provides a robust and generalizable framework for DTA prediction, well-suited for structure-based virtual screening and AI-assisted lead optimization.