Estimation of carcinogenicity using molecular fragments tree

Carcinogenicity is an important toxicological endpoint that poses high concern to drug discovery. In this study, we developed a method to extract structural alerts (SAs) and modulating factors of carcinogens on the basis of statistical analyses. First, the Gaston algorithm, a frequent subgraph mining method, was used to detect substructures that occurred at least six times. Then, a molecular fragments tree was built and pruned to select high-quality SAs. The p-value of the parent node in the tree and that of its children nodes were compared, and the nodes that had a higher statistical significance in binomial tests were retained. Finally, modulating factors that suppressed the toxic effects of SAs were extracted by three self-defining rules. The accuracy of the 77 SAs plus four SA/modulating factor pairs model for the training set, and the test set was 0.70 and 0.65, respectively. Our model has higher predictive ability than Benigni's model, especially in the test set. The results highlight that this method is preferable in terms of prediction accuracy, and the selected SAs are useful for prediction as well as interpretation. Moreover, our method is convenient to users in that it can extract SAs from a database using an automated and unbiased manner that does not rely on a priori knowledge of mechanism of action.