The current strategy for identifying the carcinogenicity of drugs involves the 2-year bioassay in male and female rats and mice. As this assay is cost-intensive and time-consuming there is a high interest in developing approaches for the screening and prioritization of drug candidates in preclinical safety evaluations. Predictive models based on toxicogenomics investigations after short-term exposure have shown their potential for assessing the carcinogenic risk. In this study, we investigated a novel method for the evaluation of toxicogenomics data based on ensemble feature selection in conjunction with bootstrapping for the purpose to derive reproducible and characteristic multi-gene signatures. This method was evaluated on a microarray dataset containing global gene expression data from liver samples of both male and female mice. The dataset was generated by the IMI MARCAR consortium and included gene expression profiles of genotoxic and nongenotoxic hepatocarcinogens obtained after treatment of CD-1 mice for 3 or 14 days. We developed predictive models based on gene expression data of both sexes and the models were employed for predicting the carcinogenic class of diverse compounds. Comparing the predictivity of our multi-gene signatures against signatures from literature, we demonstrated that by incorporating our gene sets as features slightly higher accuracy is on average achieved by a representative set of state-of-the art supervised learning methods. The constructed models were also used for the classification of Cyproterone acetate (CPA), Wy-14643 (WY) and Thioacetamid (TAA), whose primary mechanism of carcinogenicity is controversially discussed. Based on the extracted mouse liver gene expression patterns, CPA would be predicted as a nongenotoxic compound. In contrast, both WY and TAA would be classified as genotoxic mouse hepatocarcinogens.