Explainable Graph Neural Networks in Chemistry: Combining Attribution and Uncertainty Quantification.

Area of Science:

• Computational Chemistry
• Machine Learning for Chemistry

Background:

• Graph Neural Networks (GNNs) are effective for chemical property prediction.
• The black-box nature of GNNs hinders trust and practical application in chemistry.
• Explainability and uncertainty quantification are crucial for reliable GNN deployment.

Purpose of the Study:

• To systematically evaluate post-hoc feature attribution methods for GNNs in chemistry.
• To investigate the integration of uncertainty quantification with feature attribution in chemical GNNs.
• To enhance the interpretability and actionable insights of GNN models for chemical research.

Main Methods:

• Evaluation of various post-hoc feature attribution techniques (e.g., Feature Ablation, Shapley Value Sampling).
• Integration of uncertainty quantification strategies with GNN architectures.
• Application to prediction tasks such as aqueous solubility and molecular weight.

Main Results:

• A strong synergy was observed between feature attribution and uncertainty quantification.
• Attributing uncertainty to specific input features (atoms, substructures) provides granular model confidence insights.
• Feature Ablation and Shapley Value Sampling effectively identified molecular substructures driving predictions and their uncertainty.

Conclusions:

• Combining feature attribution and uncertainty quantification significantly enhances the interpretability of chemical GNNs.
• This approach highlights potential data gaps and model limitations.
• Facilitates the development of more reliable and actionable GNN models for chemical research and development.