Local-Global Structure-Aware Geometric Equivariant Graph Representation Learning for Predicting Protein-Ligand...

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Accurate prediction of protein-ligand binding affinities is essential for efficient drug discovery and design.
• Existing computational methods often struggle to effectively utilize the geometric properties of protein-ligand complexes.

Purpose of the Study:

• To develop a novel geometric equivariant graph representation learning framework, Geo-protein-ligand binding affinity (PLA), for enhanced prediction of binding affinities.
• To capture both local and global structural information within protein-ligand complexes.

Main Methods:

• Utilized an equivariant graph neural network (EGNN) to extract local 3D structural information while preserving coordinate transformation equivariance.
• Employed a graph transformer to capture long-range atomic interactions for a global perspective.
• Integrated multiscale information from both EGNN and graph transformer channels.

Main Results:

• Geo-PLA demonstrated superior performance on two benchmark datasets for predicting binding affinities.
• The model successfully captured geometric information crucial for accurate predictions.
• Visual interpretation provided valuable biological insights into protein-ligand interactions.

Conclusions:

• Geo-PLA offers a powerful new approach for predicting protein-ligand binding affinities by leveraging geometric deep learning.
• The framework provides a more accurate and insightful method for computational drug discovery.
• The model's ability to offer biological insights supports its application in virtual screening and drug repositioning.