Prediction of Chemical Carcinogenicity in Discovery Challenges

Abstract

The Predictive Toxicology Challenge (PTC) workshop was held in ECML/PKDD conferences this summer. The Challenge was to obtain models that predict the outcome of biological tests for the carcinogenicity of chemicals using information related to chemical structure only. The learning data comes from NTP results, containing 417 compounds for rodent carcinogenicity. The test data consists of 185 compounds provided by FDA. The 7 descriptor sets are submitted for these compounds. 14 groups of researchers proposed more than 26 models to predict the carcinogenicity of test compounds. Their submissions were judged by ROC analysis and by how "toxicologist-friendly" each model was presented. In this paper, we first introduce the PTC workshop outline and gives the overview of the submitted models. The results show the usefulness of various data mining methodology for the analysis of structure activity relationships. The author was one of the submitters and our model was rated as the best. In the latter half of this paper, we show the results derived by the cascade model in our lab. It uses linear fragments and 7 physicochemical descriptors. The derived rules have been able to give us useful insights to the SAR of rodent carcinogenicity.