Usage of model combination in computational toxicology.

Area of Science:

• Toxicology and computational chemistry
• Development of New Approach Methodologies (NAMs) for chemical safety assessment

Background:

• New Approach Methodologies (NAMs) aim to replace traditional animal testing in toxicology.
• NAMs still encounter significant complexities in study endpoints, including mechanistic, chemical space, and methodological challenges.
• Classical Quantitative Structure-Activity Relationship (QSAR) models struggle with these complexities, necessitating advanced approaches.

Approach:

• This review focuses on predictive QSAR metamodels, which combine multiple low-level models (LLMs) to address toxicological complexities.
• Mechanistic complexity is tackled by integrating multiple Molecular Initiating Events (MIEs) into mechanistic-based metamodels.
• Chemical space complexity is addressed using fragment-based metamodels, with or without structure sharing (e.g., federated learning).
• Methodological complexity is managed by combining diverse algorithms, descriptors, and balanced datasets to create methodological-based metamodels.

Key Points:

• Metamodels effectively integrate diverse information to model complex biological systems.
• Fragment-based metamodels with and without structure sharing offer solutions for chemical space disparities.
• Federated learning and prediction sharing are viable for proprietary data in the pharmaceutical industry.
• Algorithmic and descriptor limitations, along with class imbalance, are mitigated through ensemble methods.

Conclusions:

• Metamodels consistently demonstrate superior performance compared to classical QSAR models across various toxicological challenges.
• The strategic combination of predictive models (metamodels) provides a robust framework for navigating the complexities inherent in NAMs.
• This work highlights the critical role of metamodels as advanced alternatives to classical QSAR for reliable chemical safety assessments.