of Work:PREDICTIVE-TOXICOLOGY EVALUATION PROJECT: The second NIEHS Predictive-Toxicology Evaluation involves 30 NTP chemobioassays for carcinogenesis. It has generated 28 sets of predictions, submitted by 20 groups in 8 countries. About 25 manuscripts have been published thus far, 14 of them together in an EHP Supplement. A parallel Predictive-Toxicology Evaluation Challenge, based on this project and sponsored by the International Joint Conference on Artificial Intelligence, has generated 9 submissions that are based on computer machine- learning algorithms. Results will be evaluated during the 1999 conference. TOXICOINFORMATICS MODELS: Approaches are being developed that utilize human-heuristic expertise and several based on machine- learning algorithms, such as decision trees by induction, rule set from decision trees, back-propagation neural network, rule set from trained neural net, Bayesian-belief net, and inductive- logic programming. Each uses a fundamentally different approach to perform pattern- recognition analysis of learning sets, to identify specific biological and chemical features and relationships that may augment human-hypothesis formation about mechanistic pathways of chemotoxicity. The multiple-model/common learning-set approach creates an ideal opportunity to evaluate model differences, identify and combine unique aspects of many models to consensus models that may provide greater confidence and accuracy. DATABASE COMPILATION AND REPRESENTATION: This is an ongoing activity, because opportunities and success of the database-mining research approach are limited only by the availability of enough data of suitable quality. We expanded our main training set from 200 to 350 classified-chemical examples. Animal doses for all routes of exposure were converted to common units and to molar basis. New features incorporated include: clustered structural alerts, structural de-alert, SMILES code, 2- D structure, molecular weight, ClogP, highest-occupied and lowest-unoccupied molecular-orbital energies, and COMPACT ratio. Representations were developed that incorporate specific- morphology @ site information into our models and enable us to model individual sex-species toxicity.
