Geometric Interaction Graph Neural Network for Predicting Protein-Ligand Binding Affinities from 3D Structures (GIGN)

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Predicting protein-ligand binding affinities (PLAs) is crucial for drug discovery.
• Current machine learning (ML) methods often neglect 3D structures and physical interactions, limiting their predictive power.
• Understanding binding mechanisms requires incorporating structural and interactional data.

Purpose of the Study:

• To develop a novel ML model for accurate PLA prediction.
• To integrate 3D protein-ligand complex structures and physical interactions into the prediction framework.
• To improve the biological interpretability of PLA predictions.

Main Methods:

• Proposed a geometric interaction graph neural network (GIGN).
• Designed a heterogeneous interaction layer to unify covalent and noncovalent interactions.
• Ensured model invariance to translations and rotations, reducing data augmentation needs.

Main Results:

• GIGN achieved state-of-the-art performance on three external test sets.
• The model effectively learned node representations by incorporating diverse interactions.
• Visualizations confirmed the biological meaningfulness of GIGN's predictions.

Conclusions:

• GIGN offers a powerful new approach for predicting protein-ligand binding affinities.
• The model's ability to integrate structural and interactional data enhances prediction accuracy.
• This method holds significant promise for accelerating drug discovery processes.