Unveiling molecular moieties through hierarchical Grad-CAM graph explainability.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Virtual screening (VS) is crucial for identifying potential drug candidates efficiently.
• Graph Neural Networks (GNNs) excel at modeling molecular structures but lack interpretability.
• Explaining GNN predictions for drug discovery remains a challenge, hindering rational therapeutic design.

Purpose of the Study:

• To develop an explainable AI method for GNNs in drug discovery.
• To identify specific molecular substructures responsible for biological activity.
• To enhance the interpretability of GNN models in virtual screening.

Main Methods:

• Trained 20 GNN models for predicting small molecule activity against 20 Kinase targets.
• Implemented the Hierarchical Grad-CAM graph Explainer (HGE) framework.
• Utilized Grad-CAM at atom, ring, and whole-molecule levels to analyze binding stabilization.

Main Results:

• GNN models achieved state-of-the-art performance in virtual screening.
• HGE successfully identified key molecular moieties driving protein-ligand binding.
• Explainer validated against literature data, correctly annotating drug-target interactions.

Conclusions:

• The developed approach can accelerate drug screening and hit discovery.
• Understanding substructure contributions aids structure optimization and drug repurposing.
• Provides valuable insights for computational chemists in therapeutic development.