Predictive models of hepatotoxicity using gene expression data from primary rat hepatocytes

With the aim of evaluating the usefulness of an in vitro system for assessing the potential hepatotoxicity of compounds, the paper describes several methods of obtaining mathematical models for the prediction of compound-induced toxicity in vivo. These models are based on data derived from treating rat primary hepatocytes with various compounds, and thereafter using microarrays to obtain gene expression 'profiles' for each compound. Predictive models were constructed so as to reduce the number of 'probesets' (genes) required, and subjected to rigorous cross-validation. Since there are a number of possible approaches to derive predictive models, several distinct modelling strategies were applied to the same data set, and the outcomes were compared and contrasted. While all the strategies tested showed significant predictive capability, it was interesting to note that the different approaches generated models based on widely disparate probesets. This implies that while these models may be useful in ascribing relative potential toxicity to compounds, they are unlikely to provide significant information on underlying toxicity mechanisms. Improved predictivity will be obtained through the generation of more comprehensive gene expression databases, covering more 'toxicity space', and by the development of models that maximize the observation, and combination, of individual differences between compounds.