Drug-induced liver injury (DILI) is a major health and economic problem and the leading cause of hepatic dysfunction, drug failure during clinical testing and post-market withdrawal of approved drugs. Pre-clinical testing should be able to detect potential hepatotoxins early in the drug development process in order to minimize health risks and financial losses. Several liver-derived in vitro models have been developed to be used in pharmacology and toxicology research to understand the mechanism of DILI and to evaluate potential hepatotoxicity of new chemical entities. Although they fail to reproduce the complexity of a whole organ, their low cost, high reproducibility, and the possibility of a human origin make them a good complement to traditional in vivo tests. Monoparametric strategies used in in vitro toxicity testing have been proved insufficient to predict human DILI. The application of the new 'omics' technologies allows the simultaneous determination of multiple parameters in a single biological sample and represents a more sensitive, comprehensive and powerful tool for the study of hepatotoxic events. Among them, metabolomics measures the downstream products of the 'omics cascade', thus representing a closer approximation to phenotype than the study of genes, transcripts or proteins. Based on the previous evidences we decided to evaluate whether metabolomics, in combination with in vitro cellular models and in vivo animal models, could be a useful tool for the disvovery of characteristic patterns associated to specific mechanisms of DILI. First, we defined a suitable framework that, thanks to a careful design of sample analysis and the incorporation of different internal standards and quality controls, allowed us to perform metabolomic analysis within a quality assurance environment. Then, uni- and multivariate statistical tools were selected in order to be able to identify mechanism-specific alterations and to develop predictive/classificatory models. Finally, we optimized a sample processing and analysis strategy that allowed the differential extraction and detection of a broad range of metabolites ranging from highly polar to highly apolar ones thus maximizing metabolome coverage. The application of the developed tools to HepG2 cells exposed to subcytotoxic concentrations of model hepatotoxins acting through different mechanisms of toxicity allowed us to identify specific metabolomic patterns associated to each of the mechanisms of interest. Moreover, the application of multivariate data analysis techniques allowed the development of predictive/classificatory models able not only to distinguish between toxic and non-toxic compounds, but also to specifically classify drugs according to their mechanism of hepatotoxicity. The proposed strategy could be of interest for the identification of early markers of hepatotoxicy and for the prediction of mechanism of hepatotoxicity of new drug entities. The usefulness of the analytical strategy was also confirmed by its application with in vivo models using both medaka (Oryzias latipes) and rat. Studies in medaka revealed common liver altered pathways with HepG2 cells, suggesting medaka as a useful model for human hepatotoxicity prediction. Toxicity studies in rats allowed us to identify common serum markers of hepatotoxicity which could be used as biomarkers in pre-clinical studies or even extrapolated to humans.