AI snake oil? A risk/benefit analysis for toxicology

Area of Science:

• Toxicology
• Drug Development
• Artificial Intelligence

Background:

• Artificial intelligence (AI) is increasingly utilized for large-scale predictive, mechanistic, and human-relevant toxicology.
• Integration of AI into regulatory science, especially drug development, is inconsistent due to a gap between technical performance and regulatory trust.
• Current AI toxicology studies cover vast datasets but show variable performance, from modest in screening to strong on specific endpoints.

Purpose of the Study:

• To critically analyze the dual role of AI in toxicology.
• To review AI applications in areas like Green Toxicology, the Human Exposome, and safety assessment for biologics and peptides.
• To address the challenges hindering regulatory acceptance of AI, including interpretability, data bias, validation, and complex endpoints.

Main Methods:

• Review of state-of-the-art AI-enabled toxicology applications.
• Analysis of the gap between AI model development and regulatory acceptance.
• Discussion of the proposed 'e-validation' framework and the TREAT principle (Trustworthiness, Reproducibility, Explainability, Applicability, Transparency).

Main Results:

• AI applications in toxicology span diverse areas, facing challenges in regulatory trust.
• Key barriers to regulatory acceptance include model interpretability, dataset bias, and insufficient external validation.
• The TREAT principle and continuous 'e-validation' are proposed as foundational for building regulatory trust in AI toxicology.

Conclusions:

• AI-based toxicology methods can gain regulatory acceptance by adhering to the TREAT criteria and undergoing continuous 'e-validation'.
• A robust framework is needed to distinguish credible AI applications from unreliable ones, focusing on measurable criteria for trust.
• AI serves as a powerful evidence engine for advancing predictive and ethical toxicological science when rigorously validated.