Explainability Methods from Machine Learning Detect Important Drugs' Atoms in Drug-Target Interactions.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Bioinformatics

Background:

• Predicting drug-target interactions (DTI) is crucial for drug discovery.
• Graph neural networks (GNNs) show promise for DTI prediction but lack interpretability.
• Explainable AI (XAI) methods are needed to understand GNN decision-making.

Purpose of the Study:

• To benchmark explainable AI (XAI) attribution methods for GNNs in DTI prediction.
• To assess the consistency and biological relevance of XAI attributions.
• To enhance the interpretability of GNN models for drug discovery.

Main Methods:

• Benchmarking four XAI attribution methods on GNNs for kinase and G-protein-coupled receptors (GPCR) targets.
• Assessing method consistency using atom-level intersection over union (IoU).
• Validating biological relevance by mapping attributed atoms to 3D protein-ligand structures.

Main Results:

• Modest consistency was observed across different XAI methods.
• Consensus attributions were highly enriched for atoms contacting the protein binding pocket (up to 76% within 2 Å).
• Attributed atoms frequently contacted experimentally important residues, like those in the DFG motif.

Conclusions:

• XAI methods, despite disagreements, can identify chemically meaningful ligand features for DTI prediction.
• These findings provide a foundation for developing more interpretable GNNs in drug discovery.
• XAI enhances the understanding of GNNs, facilitating rational drug design.