The impact of collapsing data on microarray analysis and DILI prediction

In this work, we focus on two fundamental problems of toxicogenomics using the data provided by the Japanese toxicogenomics project. First, we analyze to what extent animal studies can be replaced by in in vitro assays. We show that the probeset-level representation achieves poor agreement between in vivo and in vitro data. We present a data collapsing approach to resolve poor data agreement between in vivo and in vitro data, as measured by GSEA analysis and AUC scores. Second, we address the difficult problem of predicting DILI using available microarray data. Using a binary classification framework, our results suggest that rat in vivo data are more informative than human in vitro data to predict DILI.