A potential of transcriptomic data is to derive a single or small number of highly sensitive genes or transcriptional markers (e.g. single nucleotide changes or SNPs within a gene) that have high specificity for a particular adverse outcome. An example of this is a relatively small number of genes, isolated by predictive modeling of whole transcriptomic changes, which in aggregate produce changes in gene expression that are highly predictive of a genotoxic response to exposure. This concept has been applied to a gene panel where human TK6 cells are exposed to a chemical or environmental perturbagen, and expression of 65 genes is rapidly and inexpensively measured by qPCR, and mathematical models of the resolved expression patterns indicates a chemical probability of inducing genotoxicity.
Development of a molecular fingerprint for predicting drug-induced cholestasis Cholestasis is a hepatic disease resulting from impaired bile acid metabolism and transport. Select drugs are well-established causal agents of cholestasis. Predicting the potential for a small molecule to cause cholestasis is a priority in the drug safety space due to the potential of these effects to derail development. Transcriptomic and statistical modeling approaches to identify molecular signatures of cholestasis have not yet been implemented on a wide scale. Here, we explain a meta-analysis strategy used to develop a preliminary classifier of hepatic cholestasis through public transcriptomic data sets. Briefly, 199 experimental transcriptome samples, corresponding to either control, cholestatic, or non-cholestatic hepatotoxic labels, were extracted from 8 published studies in human, rat, and mouse and from both RNA-seq and array-based platforms. Experiments included genetically, pharmacologically, surgically, and dietary induced models of disease. Using a uniform computational pipeline, we calculated the relative expression for each gene, across the transcriptome, as well as differential gene expression between experimentally controlled cholestasis and control samples. The resulting standardized transcriptomic data were used to construct a model-based sample classifier of cholestasis through multiple feature selection and machine learning statistical approaches. Our final logistic regression model predicted cholestatic status from expression of 9 genes with 88% accuracy by Monte Carlo cross-validation. These studies support experimental meta-analysis as biological informative for defining cholestasis gene regulatory networks, even with diverse organisms and model systems. Ongoing work seeks to validate performance of the predictive model by independent drug treatment experiments in the HepaRG® human hepatocyte model.