Abstract
<ns3:p>Neurotoxicity is undesirable property of substances, especially those reaching the central nervous system. In this study we investigated neurotoxicity by modelling two distinct endpoints: neuronal metabolic activity and electrical activity. Based on data from in vitro experiments on human cell lines, we built predictive models using automated machine learning (AutoML) and deep symbolic optimization (DSO), trained on six molecular representations. The mathematical equations generated using DSO outperformed AutoML models, offering superior predictive accuracy and an advantage in interpretability. The best results for electrical activity prediction were achieved with RDKit descriptors (R2 of 0.863 on the train set and 0.861 on the test set). For metabolic activity, the best model relied on MACCS fingerprints (R2 of 0.645 on the train set, 0.627 for the test set). These findings highlight the potential of data-driven approaches to anticipate neurotoxicity and demonstrate that DSO can create accurate and transparent models to support the design of safer chemicals.</ns3:p>