Toxicogenomics strategies for predicting drug toxicity

Introduction: The failure of pharmaceutical drug candidates due to toxicity, especially hepatotoxicity, is an important and continuing problem for drug development. The current manuscript explores newtoxicogenomics approaches to better understand the hepatotoxic potential of human pharmaceutical compounds and to assess their toxicity earlier in the drug development process by means of a toxicity screen. Resources: Data consisted of two commercial knowledgebases that employed a hybrid experimental design in which human drug toxicity information was extracted from the literature, dichotomized and merged with rat-based gene expression measures. One knowledgebase used gene expression from rat primary hepatocytes while the other employed whole rats. Approximately 100 compounds were used in each. Methods: Toxicity classification rules were built using a stochastic gradient boosting machine learner, with classification error estimated using a modified bootstrap estimate of true error. Several types of clustering methods were also applied, some based on sets of compounds and others based on sets of genes. Results: Robust classification rules were constructed for both in vitro (hepatocytes) and in vivo (liver) data, based on a high dose, 24-hour design. There appeared to be little overlap between the two classifiers, at least in terms of their gene lists. Robust classifiers could not be fitted when earlier timepoints and/or low dose data were included, indicating that experimental design is important for these systems. Conclusions: In light of these findings, a working compound screen based on these toxicity classifiers appears feasible, with classifier operating characteristics used to tune a screen for a specific implementation. To ensure robust and optimal performance, issues such as site variability of microarrays and generalizability of findings should be addressed as indicated.