A multi-geometric graph fusion network for protein-ligand affinity prediction

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Protein-ligand binding affinity prediction is vital for drug discovery.
• Existing 2D graph neural networks do not fully utilize 3D structural information.
• Limited exploration of 3D spatial conformations in current models.

Purpose of the Study:

• To develop a model that effectively incorporates 3D geometric information for improved protein-ligand binding affinity prediction.
• To address the limitations of existing methods by exploring 3D spatial conformations.

Main Methods:

• Introduced MGGNet, a multi-geometric graph fusion model utilizing geometric graph neural networks.
• Leveraged 3D structural data of protein-ligand complexes.
• Employed heterogeneous and homogeneous networks with 3D message passing to capture interactions and conformations.
• Incorporated geometric features from multiple coordinate systems for spatial transformation invariance.

Main Results:

• MGGNet demonstrated superior performance compared to state-of-the-art methods on benchmark datasets.
• The model effectively captures atomic interactions and 3D spatial conformations.
• Visualizations confirmed the biological relevance of MGGNet's predictions.

Conclusions:

• MGGNet advances protein-ligand binding affinity prediction by integrating 3D geometric data.
• The model's ability to learn spatial conformations enhances its accuracy and biological relevance.
• MGGNet shows promise for accelerating drug discovery pipelines.