UNDERSTANDING CONDITIONAL MODES OF ACTIONS IN CHEMICAL-INDUCED TOXICITY USING RULE MODELS
It is estimated that 115 million animals are used in experimental testing each year. Hence, shifting efforts toward alternative methods for toxicity assessment is essential. However, slow regulatory acceptance of new approaches is governed by knowledge gaps in toxicity modes of action. In this thesis, I describe these challenges and the use of in vitro screening as an alternative of animal testing. I also discuss common data-based methods to derive hypotheses about toxicity modes of actions, and the associated limitations in capturing multiple biological perturbations. I applied novel data-based workflows, using rule models, to prioritize in vitro assays predictive of toxicity as well as to detect significant polypharmacology profiles. I explain how constraints were applied to rule-based models to inform meaningful mechanistic interpretation for two toxicity endpoints: rat hepatotoxicity and acute toxicity. I compared assays selected, by rules, for predicting hepatotoxicity with endpoints used in in vitro models from commercial sources. An overlap was observed including cytochrome activity, mitochondrial toxicity and immunological responses. However, nuclear receptor activity, identified in rules, is not currently covered in commercial setups. I also demonstrate that endocrine disruption endpoints extrapolate better into in vivo toxicity when a set of specific conditions are met, such as physicochemical properties associated with good bioavailability. Next, I examined synergistic interactions between conditions in rules describing acute toxicity. I gained novel insights into how specific stressors potentiate the perturbation by known key events, such as acetylcholinesterase inhibition and neuro-signalling disruption. I show that examining polypharmacology profiles is particularly important at low bioactive potencies. Further, the overall predictive performance of rules describing acute toxicity was tested against a benchmark Random Forest model in a conformal prediction framework. Irrespective to the data type used in the training, the models were prone to bias over compounds promiscuity, by which high promiscuous compounds were more likely to be predicted as toxic. Overall, the studies conducted in this thesis provide novel insights into molecular mechanisms of toxicity, namely hepatotoxicity and acute toxicity, and with regards to chemical properties and polypharmacology. This knowledge can be used to improve the utility and design of alternative methods for toxicity, and hence, accelerate the regulatory acceptance.