ABSTRACT
The application and reliance on structure-activity relationship (SAR) approaches to predict and understand mechanisms of toxicity is increasing. The absolute predictivity of SAR models of rodent carcinogenicity cannot be determined due to lack of information regarding experimental reproducibility of the cancer bioassays. By classification of compounds with probabilities greater than 0.50 as carcinogens, we obtain a concordance of 0.735 for predictions based on the battery of carcinogenic SAR models. An “independent” test set of compounds not included in any of our carcinogenicity data bases with known carcinogenic activities was selected to compare the predictive ability of individual SAR models as well as batteries thereof. Consideration of only the 26 compounds for which there are experimentally derived data on carcinogenicity and Salmonella mutagenicity data further substantiates the problem of relying solely on mutagenicity as an indicator of carcinogenicity.
