A recent research article by the National Center for Computational Toxicology (NCCT) (Kleinstreuer et al., 2013), indicated that high throughput screening (HTS) data from assays linked to hallmarks and presumed pathways of carcinogenesis could be used to predict classification of pesticides as either (a) possible, probable or likely rodent carcinogens; or (b) not likely carcinogens or evidence of non-carcinogenicity. Using independently developed software to validate the computational results, we replicated the majority of the results reported. We also found that the prediction model correlating cancer pathway bioactivity scores with in vivo carcinogenic effects in rodents was not robust. A change of classification of a single chemical in the test set was capable of changing the overall study conclusion about the statistical significance of the correlation. Furthermore, in the subset of pesticide compounds used in model validation, the accuracy of prediction was no better than chance for about three quarters of the chemicals (those with fewer than 7 positive outcomes in HTS assays representing the 11 histopathological endpoints used in model development), suggesting that the prediction model was not adequate to predict cancer hazard for most of these chemicals. Although the utility of the model for humans is also unclear because a number of the rodent responses modeled (e.g., mouse liver tumors, rat thyroid tumors, rat testicular tumors, etc.) are not considered biologically relevant to human responses, the data examined imply the need for further research with HTS assays and improved models, which might help to predict classifications of in vivo carcinogenic responses in rodents for the pesticide considered, and thus reduce the need for testing in laboratory animals.
Keywords: Carcinogenesis; High throughput screening; Prediction modeling.
Copyright © 2016 The Authors. Published by Elsevier Inc. All rights reserved.