Structure-Aware Heterogeneous Information Fusion Framework for Protein–Ligand Binding Affinity Prediction

Accurate prediction of protein-ligand binding affinities (PLAs) is essential for drug discovery and development. Recent advancements suggest that transforming protein-ligand complexes into heterogeneous graph representations may offer a viable solution. However, existing methods ignore the importance of heterogeneous graph augmentation and the complementary information provided by sequence and protein-ligand complex structure modalities, which are crucial for enhancing generalization and robustness. In this study, we propose a multimodal data fusion approach GIF-PLA (meta-path-based enhanced hetero<i>g</i>eneous <i>i</i>nformation <i>f</i>usion framework for <i>p</i>rotein-<i>l</i>igand binding <i>a</i>ffinity prediction). Protein-ligand binding complexes are represented as heterogeneous graphs with meta-paths, in parallel with protein sequences and ligand simplified molecular input line entry system (SMILES) strings, which are fed into cascaded deep neural networks, respectively. GIF-PLA effectively captures structure-oriented information, encompassing topological interactions and high-order nonlinear relationships, as well as sequence-oriented information. Finally, a late fusion module is used to integrate multilevel information. Comprehensive evaluations demonstrate that GIF-PLA surpasses state-of-the-art methods, achieving a Pearson's correlation coefficient (Rp) of 0.784 and a root-mean-square error (RMSE) of 1.157 on benchmark data sets. Ablation studies highlight the critical contributions of meta-paths and multimodal fusion. Overall, GIF-PLA shows significant promise in predicting protein-ligand interactions with enhanced reliability.