Exploring Multi-Scale Interaction Features through a Physics-Aware Graph Network for Enhanced Binding Affinity Prediction | JoVE Visualize

Area of Science:

Computational chemistry
Structural biology
Drug discovery

Background:

Protein-ligand binding affinity is crucial for molecular recognition and drug discovery.
Current computational methods face challenges in accurately modeling complex 3D interactions, leading to limitations in prediction accuracy and generalization.
Existing approaches often oversimplify interactions, are computationally expensive, or fail to fully utilize angular features.

Purpose of the Study:

To develop an accurate and generalizable computational framework for predicting protein-ligand binding affinity.
To address the limitations of existing methods by incorporating physics-aware principles and advanced neural network architectures.
To improve the understanding of molecular recognition through enhanced modeling of protein-ligand interactions.

Main Methods:

Introduction of MSPANet (Multi-Scale Physics-Aware Network), utilizing Kolmogorov-Arnold Networks (KAN) for learning multiscale spatial interactions.
Representation of protein-ligand complexes as atomic graphs with enriched distance and angular features using radial and spherical basis functions.
Implementation of a multiscale information propagation mechanism and an attention-weighted fusion module for capturing dynamic and contextual interactions.

Main Results:

MSPANet demonstrated superior performance on the PDBbind v.2020 benchmark, achieving 8-12% relative improvements in RMSE, MAE, standard deviation, and Pearson's R.
The model showed consistent performance gains on the CSAR-HiQ dataset, highlighting its robustness and generalizability.
KAN layers enhanced model expressiveness and interpretability, while physics-aware geometric encoding improved prediction accuracy.

Conclusions:

MSPANet provides a scalable and interpretable framework for protein-ligand binding affinity prediction.
The integration of physics-aware geometric encoding with KAN-driven representation learning offers a powerful tool for biomolecular modeling.
This approach advances the field of molecular recognition and drug discovery by enabling more accurate and reliable affinity predictions.