Advances in Artificial Intelligence and Machine Learning for Toxicity Prediction in Computational Toxicology: A...

Area of Science:

• Computational toxicology
• Environmental health
• Biomedical informatics

Background:

• AI and ML have significantly advanced computational toxicology over three decades.
• This review synthesizes 30 years of AI/ML in toxicology, integrating historical, data, and regulatory aspects.

Purpose of the Study:

• To provide a comprehensive 30-year review of AI and ML in toxicology (1995-2025).
• To highlight advancements, challenges, and future directions in AI/ML for chemical safety assessment.

Main Methods:

• Review of historical evolution, multi-omics data fusion, nanotoxicity, regulatory frameworks, and AI models.
• Analysis of deep learning (CNNs, GNNs) and traditional ML (RF, SVM) performance on datasets like ToxCast and Tox21.
• Examination of regulatory acceptance and stakeholder perspectives.

Main Results:

• AI/ML models, particularly deep learning, show >85% accuracy in retrospective toxicity predictions.
• Traditional ML methods effectively handle high-throughput screening data and support multi-omics integration.
• A critical gap exists in prospective validation for novel compounds, despite regulatory progress.

Conclusions:

• AI/ML significantly reduces animal testing and accelerates safety evaluations but faces challenges with bias and opacity.
• Hybrid mechanistic-AI models, federated learning, and collaboration are key for future progress.
• Enhanced validation and transparency are crucial for broader regulatory acceptance and application.