TIRESIA: An eXplainable Artificial Intelligence Platform for Predicting Developmental Toxicity.

Area of Science:

• Toxicology
• Computational Chemistry
• Artificial Intelligence

Background:

• Developmental toxicity assessment is crucial for human health and heavily relies on animal testing.
• There is a growing need for reliable alternative methods in regulatory toxicology.
• Explainable Artificial Intelligence (XAI) offers a promising avenue for transparent and interpretable predictive toxicology.

Purpose of the Study:

• To develop and validate a robust and reproducible XAI approach for predicting developmental toxicity.
• To provide a reliable support system for developmental toxicity assessment, ensuring informativeness, uncertainty estimation, generalization, and transparency.
• To offer an alternative to animal testing in a critical area of regulatory toxicology.

Main Methods:

• Utilized the CAESAR training set (234 chemicals) and two test sets (585 chemicals) for model development and validation.
• Employed the eXtreme Gradient Boosting (XGB) algorithm for predictive modeling.
• Incorporated SHAP (SHapley Additive exPlanations) for model interpretability and identified key molecular descriptors and structural alerts.

Main Results:

• The proposed XAI framework demonstrated high accuracy, sensitivity, and specificity, outperforming state-of-the-art methods.
• The model successfully identified toxic and non-toxic chemicals based on specific molecular features.
• Results were summarized in a standard report format, including applicability domain and SHAP values, adhering to OECD principles.

Conclusions:

• The developed XAI approach provides a reliable, transparent, and accurate method for predicting developmental toxicity.
• This computational tool can significantly reduce the reliance on animal testing in regulatory toxicology.
• The model is accessible via the free web platform TIRESIA (Toxicology Intelligence and Regulatory Evaluations for Scientific and Industry Applications).