WOMBAT: Whitebox Oracle for Molecular Benchmarking and Attribution Testing

Explainers can fail in ways that are hard to detect: When AI systems explain their decisions about molecules, it's unclear whether unexpected explanations mean the explainer tool is broken or the underlying model learned something wrong. WOMBAT solves this by creating 14 test models with known decision rules, allowing researchers to definitively identify explainer failures.

Current explanation tools have significant blind spots: The study found that popular explainers like Integrated Gradients can spread credit across an entire molecular structure even when the model is actually focusing on a specific chemical pattern, demonstrating that these tools can provide misleading explanations despite the model working correctly.

Transparency requires intentional design: By hand-crafting neural networks with known internal logic, the researchers showed that understanding why AI makes decisions requires building systems with transparency from the ground up—not just adding explanation tools after the fact.

Benchmarking with ground truth is essential for responsible AI: The release of WOMBAT's dataset, models, and code provides the research community with a standardized way to test and improve explanation tools, similar to how medical imaging benchmarks help validate diagnostic AI systems.

Molecular AI needs trustworthy explanations: In high-stakes applications like drug discovery, scientists need to understand which parts of a molecule triggered an AI's prediction; this work highlights that current explanation methods may give false confidence in AI recommendations.