.png){kind=icon}
.png){kind=icon}
Return
Exploring Multi-Scale Interaction Features through a Physics-Aware Graph Network for Enhanced Binding Affinity Prediction
.png){kind=icon}
PRE
AI
DOI：10.1021/acs.jcim.5c02480.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Protein-ligand binding affinity plays a central role in molecular recognition and drug discovery, yet accurate prediction remains challenging due to the complexity of three-dimensional interactions. Conventional computational approaches, including docking, physics-based free energy calculations, and graph neural networks (GNNs), have advanced this field but still suffer from oversimplified interaction modeling, high computational cost, insufficient angular feature utilization, and limited generalization. In this work, we present MSPANet (Multi-Scale Physics-Aware Network), which employs the Kolmogorov-Arnold Network (KAN) to learn multiscale spatial interactions under physics-aware guidance, thereby improving both accuracy and generalization in protein-ligand binding affinity prediction. MSPANet represents protein-ligand complexes as atomic graphs enriched with distance and angular features through radial and spherical basis functions, enabling the explicit encoding of both local interaction details and long-range dependencies. A multiscale information propagation mechanism is introduced to capture dynamic interactions across different levels without requiring dihedral angle calculations. Leveraging adaptive spline-based nonlinear transformations, KAN layers enhance model expressiveness and interpretability, while an attention-weighted fusion module integrates local and global structural contexts for robust prediction. Extensive evaluations on the PDBbind v.2020 benchmark demonstrate that MSPANet achieves superior performance, showing relative improvements of 8-12% over state-of-the-art methods across key metrics including RMSE, MAE, standard deviation, and Pearson's R. Consistent gains on the CSAR-HiQ data set further highlight its robustness and generalizability. Overall, this work establishes a scalable and interpretable framework that combines physics-aware geometric encoding with KAN-driven representation learning, offering a powerful tool for protein-ligand affinity prediction and related biomolecular modeling tasks.
Keywords:
PROTEIN
MACHINE
Journal
IF:
5.3
Papers:
9.1K
Citations:
4.0W