Computational integration of in vivo single cell and in vitro bulk transcriptomics across 236 human and mouse datasets differentiates physiological versus non-physiological hepatic cell lines for hepatotoxicity screening.

New approach methods (NAMs), including in vitro paradigms, are needed to increase throughput, sustainability, and ethicality in toxicity research. However, selecting optimal cell culture models that mimic in vivo physiological conditions is challenging. To identify cell lines that best recapitulate physiological cells, we compared gene expression signatures of cell lines and in vivo tissues. We curated 214 transcriptomics datasets from 17 human and mouse hepatic cell lines representing hepatocytes, hepatic stellate cells, and cholangiocytes and determined basal gene expression profiles for each. We also collected 7 in vivo single-cell RNA sequencing (scRNAseq) datasets from human and mouse livers, which provide physiologically relevant transcriptome profiles for hepatic cell types. We compared cell line transcriptome profiles to liver scRNAseq data to determine which cell lines best represent in vivo physiology for each cell type and compared genes, regulatory networks, and biological pathways between cell lines and hepatic cell types. We further analyzed 15 cell lines, in vivo, and primary hepatocyte datasets from hepatotoxicity studies to relate baseline patterns to toxicological responses. We identified HepaRG as optimal to model hepatocytes both at baseline and in hepatotoxicity application studies of diverse toxicants, and further provided biological insights into the key differences of some of the widely used hepatic cell lines from in vivo biology. Overall, we present a new in silico approach that leverages existing big data to guide the selection of cell lines with better functional relevance, which can be applied to in vitro modeling of other tissues and broad biomedical applications.