Proximity Graph Networks: Predicting Ligand Affinity with Message Passing Neural Networks

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Structural bioinformatics

Background:

• Message Passing Neural Networks (MPNNs) excel at encoding molecular graphs for tasks like protein-ligand complex scoring.
• Existing methods may not fully leverage structural information between interacting molecules.

Purpose of the Study:

• Introduce the Proximity Graph Network (PGN) package, an open-source toolkit for applying MPNNs to ligand-receptor interactions.
• Establish benchmarks for affinity and docking score prediction using PGN.
• Evaluate the performance of MPNNs with proximity graphs.

Main Methods:

• Constructing ligand-receptor graphs based on atom proximity.
• Applying MPNN architectures to these graphs.
• Developing benchmarks for affinity and docking score prediction tasks.

Main Results:

• MPNNs utilizing proximity graph data structures demonstrate strong performance, particularly for docking score prediction.
• Graph networks show better generalization capabilities compared to traditional fingerprint-based models.
• The PGN package facilitates rapid application and evaluation of MPNNs.

Conclusions:

• MPNNs combined with proximity graph representations significantly enhance the prediction of ligand-receptor complex properties.
• The PGN toolkit provides a valuable resource for researchers in computational chemistry and drug discovery.
• Proximity graph data structures are effective for augmenting predictive models when interaction data is available.