An improved model of predicting hepatocarcinogenic potential in rats by using gene expression data

Carcinogenicity studies using animals are expensive and time consuming. Therefore, the development of a highly accurate carcinogenicity prediction system to interpret short-term test results would be beneficial. The Ames test is popular for mutagens; however, it cannot detect non-genotoxic carcinogens. Previously, we reported a prediction system using gene expression data obtained from a short-term (28-day) study that screened candidate compounds for testing in long-term carcinogenicity studies. In this study, our system was improved by adding more gene expression data. To establish our new system, we used the data of 93 test compounds (41 hepatocarcinogens and 52 non-hepatocarcinogens). Analysis of liver gene expression data by dividing compounds into 'for training' and 'for test' categories (20 cases assigned randomly) using Support Vector Machine (SVM) identified a set of marker probe sets that could be used to predict hepatocarcinogenicity. The assigned 42 probe sets have included the cancer- or c-Myc-related genes such as Hsp90, Pink1, Hspc111, Fbx29, Hepsin, Syndecan2 and Synbindin. Compared with the older version, the improved system had a higher concordance rate with the training data and a good performance with the external test data.