I‑GAT: Interpretable Graph Attention Networks for Ligand Optimization.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Medicinal chemistry

Background:

• Designing selective and potent ligands is crucial but challenging in drug discovery.
• Machine learning (ML) offers advanced computational solutions for ligand design.
• Existing ML models often lack interpretability, hindering optimization efforts.

Purpose of the Study:

• To develop a composite ML model for predicting ligand selectivity and potency with high accuracy.
• To enhance ML model interpretability for guiding ligand optimization.
• To create a framework for efficient and targeted drug discovery.

Main Methods:

• Compiled a dataset of 757 ligands for metabotropic glutamate receptor subtype 2 (mGlu2) and subtype 3 (mGlu3) negative allosteric modulators (NAMs).
• Developed a composite ML model integrating graph architecture and transfer learning (Phase 1).
• Employed attention mechanisms and attention gradients for model interpretability and ligand modification (Phases 2 & 3).

Main Results:

• Achieved over 97% accuracy in predicting ligand NAM selectivity and over 78% accuracy in potency.
• The model provided atom-level interpretability, illuminating ML decision-making processes.
• Successfully designed a novel ligand with predicted superior properties through intelligent modification.

Conclusions:

• The developed Interpretable Graph Attention (I-GAT) model offers both high predictive accuracy and atom-level interpretability for ligand design.
• This approach can accelerate drug discovery by providing a powerful framework for optimizing ligands.
• The model's performance matches or exceeds state-of-the-art ML models and is adaptable to other targets.