For assessing carcinogenicity in animals, it is difficult and costly, an alternative strategy has been desired. We explored the possibility of applying a toxicogenomics approach by using comprehensive gene expression data in rat liver treated with various compounds. As prototypic non-genotoxic hepatocarcinogens, thioacetamide (TAA) and methapyrilene (MP) were selected and 349 commonly changed genes were extracted by statistical analysis. Taking both compounds as positive with six compounds, acetaminophen, aspirin, phenylbutazone, rifampicin, alpha-naphthylisothiocyanate, and amiodarone as negative, prediction analysis of microarray (PAM) was performed. By training and 10-fold cross validation, a classifier containing 112 probe sets that gave an overall success rate of 95% was obtained. The validity of the present discriminator was checked for 30 chemicals. The PAM score showed characteristic time-dependent increases by treatment with several non-genotoxic hepatocarcinogens, including TAA, MP, coumarin, ethionine and WY-14643, while almost all of the non-carcinogenic samples were correctly predicted. Measurement of hepatic glutathione content suggested that MP and TAA cause glutathione depletion followed by a protective increase, but the protective response is exhausted during repeated administration. Therefore, the presently obtained PAM classifier could predict potential non-genotoxic hepatocarcinogenesis within 24 h after single dose and the inevitable pseudo-positives could be eliminated by checking data of repeated administrations up to 28 days. Tests for carcinogenicity using rats takes at least 2 years, while the present work suggests the possibility of lowering the time to 28 days with high precision, at least for a category of non-genotoxic hepatocarcinogens causing oxidative stress.