MEG: Generating Molecular Counterfactual Explanations for Deep Graph Networks

Explainability is critical for high-stakes chemistry applications: The paper emphasizes that when AI models predict molecular properties like toxicity for drug development, understanding why the model made a prediction becomes essential for human health and safety—not just a nice-to-have feature.

Counterfactual explanations make AI decisions understandable to non-experts: MEG generates alternative molecules that are structurally similar to an input molecule but have different predicted properties. This approach helps chemists and non-technical stakeholders understand what features the AI model considers important, without requiring deep machine learning knowledge.

The system maintains scientific validity while explaining decisions: A key innovation is that MEG restricts its explanations to chemically valid compounds, ensuring that the counterfactual explanations are actually realistic and could theoretically exist—rather than generating nonsensical molecular structures that wouldn't be useful to domain experts.

Reinforcement learning can be guided toward trustworthy explanations: By using an RL agent that receives feedback from the trained model at each step, MEG efficiently explores the space of similar molecules to find meaningful counterfactuals, demonstrating a practical method for generating explanations rather than simply extracting them post-hoc.