Predicting Chemical Carcinogenesis: Problem Representation Governs Model Performance

Douglas W. Bristol1, NIH-NIEHS

Title: Predicting Chemical Carcinogenesis: Problem Representation Governs Model Performance Author information: Douglas W. Bristol, Toxicology Operations Branch, Environmental Toxicology Program, Division of Intramural Research, National Institute of Environmental Health Sciences, Research Triangle Park, NC 27709, USA Long Abstract: Predictive models for chemical carcinogenesis are needed to help manage the safe use of over 100,000 chemical substances worldwide, as well as to guide decisions made during development of new chemicals. Three Predictive-Toxicology Challenges (PTC) have stimulated the development of such models over the past decade. Each was an open, collaborative experiment that attracted cross-disciplinary participants and resources. Learning sets were compiled from standardized chemical bioassays for carcinogenesis conducted in rodents by the U.S. National Toxicology Program. ROC convex-hull analysis was applied to evaluate the predictive performance of 54 individual models, within and across the three PTC experiments. Evaluation of model comprehensibility or the coherence of individual models with domain knowledge requires further development effort. Additionally, the three PTC experiments and individual models utilized three different paradigms for representing the nature of the problem. When classified according to whether a model was developed using only chemical-structure attributes, only biological-system attributes, or a mix of chemical and biological attributes, it is clear that mixed-attribute models, which reflect interactions between chemical and biological features, clearly outperform those based on only biological or chemical attributes. Thus, the interaction paradigm offers opportunities to develop predictive models with high accuracy and broad coverage, but, more importantly, to discover features and relationships that govern mechanistic pathways for chemical carcinogenicis and provide insight into confounding effects, such as those that transcend species, gender, and route of exposure.