Drugs that modulate mitochondrial function can cause severe adverse effects. Unfortunately, mitochondrial toxicity is often not detected in animal models, which stresses the need for predictive in silico approaches. In this study we present a model for predicting mitochondrial toxicity focusing on human mitochondrial respiratory complex I (CI) inhibition by combining structure-based methods with machine learning. The structure-based studies are based on CI inhibition by the pesticide rotenone, which is known to induce parkinsonian motor deficits, and its analogue deguelin. After predicting a common binding mode for these two compounds using induced-fit docking, two structure-based pharmacophore models were created and used for virtual screening of DrugBank and the Chemspace library. The hit list was further refined by three different machine learning models, and the top ranked compounds were selected for experimental testing. Using a tiered approach, the compounds were tested in three distinct in vitro assays, which led to the identification of three specific CI inhibitors. These results demonstrate that risk assessment and hazard analysis can benefit from combining structure-based methods and machine learning.
Troger, F., Delp, J., Funke, M., van der Stel, W., Colas, C., Leist, M., … Ecker, G. F. (2020). Identification of mitochondrial toxicants by combined in silico and in vitro studies – A structure-based view on the adverse outcome pathway. Computational Toxicology, 14. https://doi.org/10.1016/j.comtox.2020.100123