Abstract

Thousands of new molecules are being created every year; ensuring that they are safe for human exposure is paramount. While classical toxicology has been focused on animal models, there is an increasing focus towards predictive computational models based on data, with machine learning models showing great promise. However, these models are viewed with suspicion due to lack of interpretability. The conception of the adverse outcome pathway (AOP) further provides a framework for understanding the mechanism of toxicity, with the molecular initiating event (MIE) being of particular interest to chemists. Identifying MIEs based on chemical structure can improve our understanding of why molecules are toxic.

Information regarding MIEs in toxicological data is rare, but a recent Seahorse assay ran by the US EPA contains multi-class labels (as opposed to binary labels). This allowed construction of structural alerts specific to each mechanism, but the small sample space limited the scope of this study.

Combined data for mitochondrial toxicity was extracted from ChEMBL, ToxCast and the EPA Seahorse assay. A neural network model was built from the data, and the resulting classifier was shown to outperform commercially available toxicity screening software such as DEREK Nexus. Analysis of the model uncovered certain structural features most significant to these predictions, and a visualization can be generated to aid chemists in decision-making, allowing some insight to be gleaned into the chemistry of MIEs related to mitochondrial toxicity.