Combining molecular informatics and computational modeling to predict chemical toxicity
The ability to predict chemical toxicity is important for many reasons. Designing processes and products that have minimal negative impact on health and environment is essential in the development of pharmaceuticals, foods and food additives, consumer products, cosmetics, and specialty chemicals. It is simply not possible to experimentally study every new or proposed chemical compound to determine its toxicity potential. Also, recent initiatives and regulations that limit the use of animal testing for certain types of products (e.g., cosmetics) have motivated the development of alternative ways to evaluate toxicity, especially in vitro experimental approaches (as replacements for in vivo assays) and in silico computational approaches.
We are developing computational modeling methods to meet this need. We are attempting to model extremely complex phenomena; for example, will a new drug candidate react with DNA to cause mutations? Will a proposed new cosmetics ingredient possibly be a skin sensitizer, eliciting an allergic response in some people? To answer such questions we combine cheminformatics methods with advanced computational modeling techniques, generally referred to as “machine learning”, to build predictive models. Our models are trained, tested and validated using the extensive and diverse databases available that summarize chemical property and biological endpoint data for many thousands of molecules for which experimental data from some of these endpoints are available.