Enhancing property and activity prediction and interpretation using multiple molecular graph representations with...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Graph Neural Networks (GNNs) are powerful for predicting compound properties and activities.
• The choice of molecular graph representation critically impacts GNN model learning and interpretability.
• Existing representations like atom-level graphs may miss crucial substructures, while reduced graphs integrate higher-level chemical information.

Purpose of the Study:

• To investigate the impact of multiple molecular graph representations on GNN model learning and interpretation.
• To introduce MMGX (Multiple Molecular Graph eXplainable discovery) for evaluating diverse graph types.
• To assess how different graph perspectives enhance understanding of model decisions.

Main Methods:

• Utilized multiple molecular graph representations: Atom, Pharmacophore, JunctionTree, and FunctionalGroup.
• Implemented the MMGX framework to analyze model performance and interpretation across these graphs.
• Evaluated the impact of combining different graph views on learning and explainability.

Main Results:

• Employing multiple molecular graphs generally enhances GNN model performance, with improvements varying by dataset.
• Interpreting models using multiple graph views provides more comprehensive features and identifies potential substructures aligned with chemical knowledge.
• Diverse graph perspectives offer richer insights compared to single representation approaches.

Conclusions:

• Multiple molecular graph representations are beneficial for improving GNN performance and interpretability in cheminformatics.
• MMGX facilitates a deeper understanding of model behavior and aids in identifying key chemical features.
• The approach of using multiple graphs and interpretation perspectives has broad applicability in drug discovery and related fields.