Equivariant Line Graph Neural Network for Protein-Ligand Binding Affinity Prediction.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Predicting protein-ligand binding affinity is crucial for drug discovery and screening.
• Current methods often convert 3D complexes to 2D graphs, losing global spatial and topological information.
• This limitation hinders accurate binding affinity prediction.

Purpose of the Study:

• To introduce a novel method, the Equivariant Line Graph Network (ELGN), for improved 3D protein-ligand binding affinity prediction.
• To overcome the limitations of 2D graph representations in capturing complex structural information.
• To enhance the accuracy of virtual drug screening and repositioning.

Main Methods:

• The proposed ELGN method constructs a line graph from the 3D protein-ligand complex.
• It incorporates a super node and utilizes an E(3)-equivariant network layer for message passing.
• This approach leverages the global coordinate system to learn node and edge features.

Main Results:

• ELGN effectively captures global information, including physical symmetry and bond topology, which are missed by 2D methods.
• Experiments on two real-world datasets show ELGN outperforms several state-of-the-art baseline methods.
• The model demonstrates superior performance in binding affinity prediction.

Conclusions:

• The Equivariant Line Graph Network (ELGN) offers a significant advancement in predicting binding affinity for 3D protein-ligand complexes.
• ELGN's ability to learn from global 3D structural information enhances prediction accuracy.
• This method holds promise for more effective virtual drug screening and repositioning strategies.