Using explainable machine learning to predict the irritation and corrosivity of chemicals on eyes and skin.

Summary

This study assembled >6,000 experimental labels to train explainable ML models that predict eye and skin irritation with balanced accuracies of 73–75%. It identifies structural alert fragments, provides multi-level interpretability, and offers a user-friendly interface, positioning it as a practical alternative-to-animal screening tool for cosmetics and related chemicals.

Key Findings

• Best models achieved balanced accuracies of 73.0% (eye) and 75.1% (skin) on external validation.
• Dataset-, molecule-, and atom-level interpretability identified structural alert fragments linked to irritation.
• A visualization interface enables non-specialists to predict irritation potential.
• Integrated 3316 eye and 3080 skin irritation data points across chemicals relevant to cosmetics and pharmaceuticals.

Clinical Implications

Early screening of ingredient irritation risk could inform formulation decisions, reduce late-stage failures, and support regulatory submissions aligned with alternative-to-animal testing paradigms.

Limitations

• Balanced accuracy indicates moderate performance; false positives/negatives may persist.
• Potential dataset bias and limited coverage of rare chemotypes.
• Regulatory acceptance requires further prospective validation.

Future Directions

Expand datasets to underrepresented chemotypes, calibrate thresholds to specific use-cases, conduct prospective validation against in vitro alternatives, and integrate with regulatory frameworks (e.g., OECD QSAR principles).