Development of Methods to Ascertain the Predictivity and Consistency of SAR Models: Application to the U.S. National Toxicology Program Rodent Carcinogenicity Bioassays

Publisher Website
Google Scholar
Abstract
Models investigating relationships between chemical structures and biological activities are receiving increased recognition for the identification of chemicals with the potential for inducing adverse health effects. The relationships can be either qualitative (noted as SAR) or quantitative (noted as QSAR). The objective of the present study was to define an effective process for evaluating such models. The predictivity of SAR/QSAR models derived from the U.S. National Toxicology Program Rodent Carcinogenicity Bioassay endeavor by CASE/MultiCASE was evaluated by several different approaches: leave‐one‐out tests, 10‐fold cross‐validations and by the use of an independent test set. The goodness‐of‐fit for the data used in the model building, the predictivity for the chemicals not contained in the model, and the consistency of the predictions for a group of chemicals by different SAR/QSAR sub‐models were examined systematically. Individual prediction indices generated by CASE/MultiCASE, arbitrary combinations thereof, as well as weighted combinations using Bayes' theorem, were utilized to derive predictions of Carcinogenicity. Combinations derived using Bayes' theorem provided the most predictive model. The closeness between sub‐models based on the leave‐one‐out procedure and the full model (all chemicals used for model building) makes it the most reliable process for the estimation of a model's predictivity. However, the similarity between the predictions of the leave‐one‐out models and the 10‐fold cross‐validation models indicates that the latter process provides an acceptable approach.