Advancing Computational Toxicology by Interpretable Machine Learning.

Area of Science:

• Toxicology
• Computer Science
• Data Science

Background:

• Traditional animal models for chemical toxicity are costly, slow, and often inaccurate for human health.
• Computational toxicology using machine learning (ML) and deep learning (DL) shows promise but often lacks interpretability.
• The
• black box
• nature of many ML/DL models hinders their use in chemical risk assessment.

Purpose of the Study:

• To review the applications of interpretable machine learning (IML) in computational toxicology.
• To highlight how IML can unveil toxicity mechanisms and domain knowledge.
• To encourage the development of interpretable models for chemical safety assessments.

Main Methods:

• Focus on IML applications in toxicology, covering toxicity feature data and interpretation methods.
• Exploration of knowledge base frameworks for IML development.
• Review of recent applications and challenges in IML for toxicology.

Main Results:

• IML addresses the need for transparency in computational toxicology models.
• IML facilitates the understanding of toxicity mechanisms and aids risk assessment.
• The review synthesizes current IML approaches and their utility in predicting chemical toxicity.

Conclusions:

• IML is crucial for advancing computational toxicology and improving chemical safety.
• Developing interpretable models with novel IML algorithms is essential for future chemical assessments.
• IML can assist toxicologists by illustrating human toxicity mechanisms, enhancing risk evaluation.