The current gold-standard method for cancer safety assessment of drugs is a rodent two-year bioassay, which is associated with significant costs and requires testing a high number of animals over lifetime. Due to the absence of a comprehensive set of short-term assays predicting carcinogenicity, new approaches are currently being evaluated. One promising approach is toxicogenomics, which by virtue of genome-wide molecular profiling after compound treatment can lead to an increased mechanistic understanding, and potentially allow for the prediction of a carcinogenic potential via mathematical modeling. The latter typically involves the extraction of informative genes from omics datasets, which can be used to construct generalizable models allowing for the early classification of compounds with unknown carcinogenic potential. Here we formally describe and compare two novel methodologies for the reproducible extraction of characteristic mRNA signatures, which were employed to capture specific gene expression changes observed for nongenotoxic carcinogens. While the first method integrates multiple gene rankings, generated by diverse algorithms applied to data from different subsamplings of the training compounds, the second approach employs a statistical ratio for the identification of informative genes. Both methods were evaluated on a dataset obtained from the toxicogenomics database TG-GATEs to predict the outcome of a two-year bioassay based on profiles from 14-day treatments. Additionally, we applied our methods to datasets from previous studies and showed that the derived prediction models are on average more accurate than those built from the original signatures. The selected genes were mostly related to p53 signaling and to specific changes in anabolic processes or energy metabolism, which are typically observed in tumor cells. Among the genes most frequently incorporated into prediction models were Phlda3, Cdkn1a, Akr7a3, Ccng1 and Abcb4.