Explaining compound activity predictions with a substructure-aware loss for graph neural networks.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• Explainable machine learning (ML) is crucial for understanding compound property predictions in drug discovery.
• Feature attribution methods help identify molecular substructures influencing predicted properties.
• Existing methods show limitations with deep learning models like graph neural networks (GNNs).

Purpose of the Study:

• To enhance the explainability of graph neural networks (GNNs) for drug discovery applications.
• To address the low performance of current feature attribution techniques with GNNs.
• To develop a more accurate method for rationalizing compound property predictions.

Main Methods:

• A modified regression objective for GNNs was developed.
• The approach specifically accounts for common core structures between molecule pairs.
• Performance was evaluated on a recent explainability benchmark.

Main Results:

• The proposed method demonstrated higher accuracy compared to existing techniques.
• Improved performance was observed for GNNs in molecular feature attribution.
• The approach showed superior results over simpler modeling alternatives.

Conclusions:

• The novel methodology significantly improves GNN explainability in drug discovery.
• This approach can aid in rationalizing compound property predictions and lead optimization.
• It offers a promising tool for investigating specific chemical series in drug development pipelines.