Exploiting the Japanese Toxicogenomics Project for Predictive Modelling of Drug Toxicity

Motivation In the last decade, surprisingly few drugs reached the market. Many promising drug candidates (approx. 80%) failed during or after Phase I, inter alia, due to issues with undetected toxicity [1]. The problem of undetected toxicity becomes even more apparent in the context of drug-induced illness which causes approximately 100,000 deaths per year solely in the USA [2]. Toxicogenomics tries to avoid such problems by prioritizing less toxic drugs over more toxic ones in early drug discovery. To this end, toxicogenomics employs high throughput molecular profiling technologies and predicts the toxicity of drug candidates. For this prediction, large-scale -omics studies of drug treated cell-lines and/or pharmacology model organisms are necessary. However, data exploitation of such large-scale studies requires a highly optimized analysis pipeline, that provides methods for correction of batch effects, noise reduction, dimensionality reduction, normalization, summarization, filtering and prediction. In this work, we present a novel pipeline for the analysis of large-scale data sets in particular for transcriptomics data. Our pipeline was tested on the Japanese Toxicogenomics Project (TGP) [3], where we evaluated to what degree in vitro bioassays can be used to predict in vivo responses.