Machine learning approaches for toxicology and risk assessment

Abstract

Computational toxicology is a new approach methodology (NAM) for chemical toxicity testing and risk assessment in light of the replacement, reduction, and refinement (3Rs) of animal use. This thesis focuses on machine learning which involves the development of predictive models that extract patterns from existing datasets and generate new predictions. Four themes – awareness, predictivity, accessibility, and interpretability – are explored to reinforce the machine learning protocol for use in toxicology. Firstly, awareness of computational toxicology as a field is promoted, introducing how in silico approaches are applied in science then reviewing current machine learning research in toxicology. A novel machine learning framework, called grouped multitask learning, was designed to increase the predictivity of quantitative structure – activity relationship (QSAR) models predicting strain-specific Ames mutagenicity. The resulting machine learning QSAR models were then engineered to meet international regulatory standards to ensure its accessibility as a mutagenicity hazard screening tool. Finally, an interpretation protocol was formulated to confirm machine learning QSAR models extract toxicologically relevant patterns from data, which was applied to a model of retinoic acid receptor antagonists to support novel endocrine disruption hypotheses. The work of this thesis demonstrates machine learning to be a reliable non-animal NAM available to toxicologists for chemical toxicity testing and risk assessment.