Predicting direct hepatocyte toxicity in humans by combining high-throughput imaging of HepaRG cells and machine learning-based phenotypic profiling.

Accurate prediction of drug- and chemical-induced hepatotoxicity remains to be a problem for pharmaceutical companies as well as other industries and regulators. The goal of the current study was to develop an in vitro/in silico method for the rapid and accurate prediction of drug- and chemical-induced hepatocyte injury in humans. HepaRG cells were employed for high-throughput imaging in combination with phenotypic profiling. A reference set of 69 drugs and chemicals was screened at a range of 7 concentrations, and the cellular response values were used for training a supervised classifier and for determining assay performance by using tenfold cross-validation. The results showed that the best performing phenotypic features were related to nuclear translocation of RELA (RELA proto-oncogene, NF-kB subunit; also known as NF-kappa B p65), DNA organization, and the F-actin cytoskeleton. Using a subset of 30 phenotypic features, direct hepatocyte toxicity in humans could be predicted with a test sensitivity, specificity and balanced accuracy of 73%, 92%, and 83%, respectively. The method was applied to another set of 26 drugs and chemicals with unclear annotation and their hepatocyte toxicity in humans was predicted. The results also revealed that the identified discriminative phenotypic changes were related to cell death and cellular senescence. Whereas cell death-related endpoints are widely applied in in vitro toxicology, cellular senescence-related endpoints are not, although cellular senescence can be induced by various drugs and other small molecule compounds and plays an important role in liver injury and disease. These findings show how phenotypic profiling can reveal unexpected chemical-induced mechanisms in toxicology.