Optimization of the drug-induced cholestasis index based on advanced modeling for predicting liver toxicity

Cholestatic drug-induced liver injury (cDILI) is a frequent reason for drug failure and withdrawal during premarketing and postmarketing stages of drug development. Strategies for reliable detection of cDILI in early drug development are therefore urgently needed. The drug-induced cholestasis index (DICI) concept was previously introduced as a tool for assessing the cholestatic potential of drug candidates. DICI is calculated as the ratio between the viability values obtained in drug-treated liver cells in the presence and absence of bile acids. The present in vitro study was set up to investigate the applicability of DICI in a novel high-throughput and large sample setting. Furthermore, the improvement of the predictivity of the DICI by introduction of advanced modeling was explored. Fifty-eight well-documented drugs were selected and categorized as drugs inducing cDILI, non-cholestatic DILI (ncDILI), and not inducing DILI (non-DILI). Cultures of human hepatoma HepaRG cells in 3D spheroid configuration were exposed to 9 half-log concentrations of each drug for 1, 3 and 7 days in the absence or presence of a concentrated mixture of human bile acids. The highest concentration of each drug was based on solubility and the maximum concentrations in human plasma (total Cmax). DICI values were computed for all drugs and time points. In addition, the area under the curve ratio and the occurrence of a trend in the cytotoxicity profiles were included as modeling descriptors. As such, 3 time-related scenarios were considered upon modeling, while categories were modeled on a nominal or an ordinal scale. Applying DICI with a cut-off value of 0.8 resulted in a high sensitivity for the cDILI class, but in turn, a low sensitivity for the non- DILI class. From the 28 predictive models generated, the best performing models integrated all descriptors and the ordinal scale for either the 7-day time point from a 3-time-point model or the 3-day time point. While these models were unable to accurately identify ncDILI drugs, the 7-day time point identified 84 % of the cDILI drugs and the 3-day time point correctly identified 94 % of non-DILI drugs. Based on the obtained results, it can be concluded that the reported DICI modeling provides an optimized approach that could be applied in an integrated DILI testing strategy