Quantitative Evaluation of Explainable Graph Neural Networks for Molecular Property Prediction

Explainability methods vary significantly in quality: The research found that GradInput and Integrated Gradients (IG) consistently outperformed other explanation techniques for understanding how graph neural networks make predictions about molecular properties, suggesting that not all interpretability methods are equally reliable for drug discovery applications.

Model-method combinations matter: The study demonstrated that the effectiveness of explainability techniques depends on which specific GNN algorithm they're paired with, with GraphNet and CMPNN showing the best results when combined with top-performing XAI methods—indicating that practitioners cannot assume one explanation approach works universally.

Quantifying interpretability remains challenging: The authors highlight a fundamental problem in the field: because determining whether an AI explanation is actually correct ultimately requires human judgment rather than objective measures, evaluating the quality of model interpretations is inherently difficult, which has limited adoption of AI in drug discovery.

Benchmarking tools are now available: The researchers created benchmark datasets across three levels of complexity and released an open-source XAI package, providing drug discovery practitioners with concrete tools to test and compare how well different AI models can explain their predictions on new tasks.

Interpretability is critical for real-world adoption: The paper emphasizes that despite strong performance, graph neural networks have seen limited acceptance in actual drug discovery because scientists need to understand why the AI makes specific recommendations—not just that it works—making explainability a prerequisite for practical use.