ABSTRACT This paper presents the results from an evaluation of the noncommercial, (quantitative) structure-activity relationship ([Q]SAR) models for the prediction of mutagenicity and carcinogenicity, carried out in a collaboration between the European Chemicals Bureau Group on Computational Toxicology and the Italian Istituto Superiore di Sanita'. Local QSAR models for congeneric chemical classes and structure alert (SA) models were investigated. The studied models can be interpreted mechanistically, agree with, and/or support the available scientific knowledge, and exhibit good statistics. These models were subjected to external prediction tests with chemicals not considered by the authors of the models. Local QSARs that estimated the potency of congeneric chemicals were 30% to 70% correct, whereas the models that discriminated active and inactive chemicals had considerably higher accuracy (70% to 100%). In addition, the commonly used statistical internal cross-validation procedures were poorly correlated with external validation statistics. The genotoxic-based SA models had an accuracy of about 65% for rodent carcinogens, and about 75% for Salmonella mutagens. However, the SA models did not discriminate well active and inactive chemicals within individual chemical classes, and are more suited for preliminary or large-scale screenings. Overall, the (Q)SAR-based predictions are able to significantly enrich the target of safer chemicals, contribute to the organization and rationalization of data, elucidate mechanisms of action, and complement data from other sources.