The intent of this project is to develop a method of automatically programming neural-networks to extract needed information from databases of biological and chemical information. The specific application example we will pursue in Phase I will be the creation of a framework for training a neural network to predict the toxic and carcinogenic properties of chemicals using data stored in the Organ Toxicity Database. The neural network will be trained to recognize patterns and correlations between chemical structure/substructures, physiochemical properties, and toxicologic/carcinogenic findings stored in the electronic compilation of a large subset of NTP TR and NTP TOX documents. A system which automatically selects the pertinent data fields for accurately predicting toxicity and then trains a neural network will be developed. The ability of the resulting neural network to predict the carcinogenic and toxic properties of known liver and kidney toxicants and carcinogens will be evaluated during Phase I. If accurate prediction is achieved, we will build a generalized framework for extending our techniques to other types of biological and chemical databases during Phase II. The value of many databases will be greatly increased by the availability of methods to automatically create neural networks which extract needed information for various applications. For example, our toxicity prediction system would reduce the exposure of the general public to possibly hazardous chemicals, would allow chemical manufacturers to quickly determine the hazards of new and existing products, and would same enormous amounts of time and money currently spent on animal testing. Each application developed using our framework could offer similar benefits.
