Douglas W. Bristol, Ph.D.
ToxicoInformatics Predicts Chemical Carcinogenesis and Indicates Mechanistic Pathways.
ToxicoInformatics Unit, Laboratory of Environmental Carcinogenesis and Mutagenesis, Division of Intramural Research, National Institute of Environmental Health Sciences, PO Box 12233, Research Triangle Park, NC 27709, USA.
Classifying the activity expected from exposure of whole-animal systems to noncongeneric-chemicals is a formidable challenge, because the number and diversity of pure chemicals, mixtures, and substances that must be dealt with is immense and the mechanistic pathways that govern most toxicity endpoints are multifactorial. Toxicoinformatics is a novel, flexible approach that can accomodate the complex nature of the noncongeneric-chemical-bioactivity-prediction problem. We begin by factoring aspects of the complex problem into components, each of which addresses a related part of the overall risk-assessment process. This makes it possible to develop a continuum of models, where each model addresses special aspects of the overall problem fittingly and generates information that is suitable as input for the next in the series. The first model must deal with special challenges posed by noncongeneric-chemical-bioactivity prediction, i.e., it must classify the activity, or potential hazard, expected from exposure to noncongeneric pure chemicals, physical substances, or mixtures. If no activity is expected, then the need for further assessment is low. If activity is expected, then assessment using SAR and PB-PK models, or evaluation in higher animal test systems is indicated. Thus, activity classification models can support of the first stage of risk assessment, hazard identification
The toxicoinformatics, or database-mining, approach to the development of activity-classification models is based on three, interrelated concepts that deal with: representation of the nature of the problem, compilation of the learning set or database, and pattern-recognition analysis accomplished by machine-learning techniques. These integral parts of toxicoinformatics will be illustrated by models where pattern-recognition analysis was performed using decision-tree-induction and neural-network techniques. The examples will also be used to illustrate how the attributes, relationships, and empirical rules identified by the toxicoinformatics approach have heuristic value and contribute to the confirmation of existing domain knowledge or support the formation of new, data-based, mechanistic hypotheses.Back to Program Page