Applying machine-learning approaches to identify key genes associated with drug-induced cholestasis

Background and Objectives: Drug-induced cholestasis (DIC) is one of the most severe manifestations of adverse drug reactions, constituting a major subgroup (up to 50%) of total drug-induced liver injury (DILI) cases 1. Due to its complex process, early detection of DIC during drug development remains challenging. Preclinical animal studies, a standard model in drug safety evaluation, often fail to detect DIC in humans mainly due to interspecies differences 2. Recently, toxicogenomics in vitro assays, especially based on human liver cells, have become a more convenient and practical approach for the prediction of human-relevant DILI. Over the past decade, the established large-scale databases, combined with machine-learning (ML) approaches, give us the opportunity to identify transcriptome signatures of DILI. In the present study, we leveraged the publicly available database, Open TG-GATEs3, for the identification of transcriptomic signatures of DIC.
Material and Methods: We retrieved toxicogenomics data derived from in vitro cultured primary human and rat hepatocytes following exposure to 18 compounds (9 cholestatic compounds and 9 non-cholestatic compounds). These transcriptome profiles were measured at two time points (8 and 24h) following a single exposure to a given compound at three dosages (control, middle and high) with two biological replicates. Due to the mechanistic complexity of DIC, the model cholestatic compounds were selected because of their potential to cause cholestatic hepatotoxicity through diverse toxic mechanisms. Several supervised ML approaches, including Random Forest, Support Vector Machine and Logistic Regression, were applied to the human liver TG-GATEs dataset to develop a prediction model.
Results: We identified a signature consisting of 20 genes that predicted cholestatic hepatic injury with high specificity and selectivity. The selected feature genes and model were validated using the in vitro rat TG-GATEs dataset.
Discussion and Conclusion: Our transcriptomic signature has yielded high accuracy in the identification of potential cholestasis-inducing compounds.