Initial experiments to adapt supervised machine-learning techniques (decision tree and rule induction) to the problem of predicting the carcinogenic potential of untested, noncongeneric chemicals provided very promising results. Each of the 301 chemicals in the training set was represented by biological attributes that encoded values for results from short-term mutagenesis assays, chemical-substructure alerts, the estimated minimally toxic dose (MTD) for chronic studies, and specific-organ toxicity determined from subchronic studies. The accuracy of the decision tree and rule induction for predictive toxicology (TRIPT) system in predicting the potential carcinogenicity for a standard set of 44 NTP chemicals compared favorably with the most accurate predictions, made by a human-expert-team using heuristic rules, and exceed the accuracy of 7 other computer-assisted methods. In addition, TRIPT provided predictions for all 44 chemicals, including 5 inorganic compounds and 1 polymer, and generated rule sets that 'rediscovered' some human-expert heuristics and suggested others that represent new possibilities. This work confirms the concept that the inductive approach can be used to address the noncongeneric-chemical prediction problem and indicates TRIPT is one of the most promising predictive methods under development. An International Workshop on Predicting Chemical Carcinogenesis in Rodents was conducted at the NIEHS in May, 1993. It concluded the first round of a collaborative experiment, where the standardized testing performed by the National Toxicology Program on 44 chemicals was employed as an experimental technique for validating predictive toxicology research methods under development. The lessons learned will be put to use in a second prediction experiment that utilizes ongoing NTP bioassays for about 30 chemicals and a special issue of Environmental Health Perspectives to accommodate publication of prospective sets of predictions submitted by any groups who wish to participate. The compilation of specific organ toxicity data from carcinogenesis bioassays begun in FY92 was extended from 20 to 111 chemicals. Four Phase 1, SBIR Program awards were made to study relationships in these data.
