An ensemble of neural predictors is used to develop a set of QSAR models for the prediction of the carcinogenicity TD50 index. The proposed approach makes use of the self organizing feature map algorithm to select diverse subsets of molecular descriptors that are used afterwards to train an ensemble of predictive fuzzy ARTMAP networks. The diversity introduced by the predictors trained using different subsets of descriptors produces better generalization results than single models. Comparison of the developed models with published models is to be used to assess the quality of the prediction system.