Equivariant Interaction-Aware Graph Network for Predicting the Binding Affinity of Protein-Ligand.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Predicting protein-ligand binding affinity is crucial for drug discovery.
• Deep learning, particularly graph neural networks (GNNs), shows promise but often overlooks interaction details.
• Existing GNNs represent biomolecules well but lack rational modeling of complex interactions.

Purpose of the Study:

• To develop a novel deep learning model for accurate protein-ligand binding affinity prediction.
• To enhance interaction modeling within protein-ligand complexes.
• To improve the efficiency and reduce resource consumption in drug discovery.

Main Methods:

• Developed the Equivariant Interaction-aware Graph Network (EIGN).
• Incorporated a distance-inspired edge-gated attention layer for uniform inter-node interaction learning.
• Utilized equivariant convolutional layers to capture 3D geometric structure.
• Considered local structural information around nodes for precise interaction simulation.

Main Results:

• EIGN demonstrated exceptional performance on two benchmark datasets.
• The model showed strong generalization capabilities.
• Accurate interaction modeling was highlighted as key to improved predictions.

Conclusions:

• EIGN offers a significant advancement in predicting protein-ligand binding affinity.
• The model's ability to learn 3D geometric and interaction information is critical.
• This approach can accelerate drug discovery by improving prediction accuracy and efficiency.