The induction of rules for predicting chemical carcinogenesis in rodents

This paper presents results from an ongoing effort in applying a variety of induction-based methods to the problem of predicting the biological activity of noncongeneric (structurally dissimilar) chemicals. It describes initial experiments, the long-term goal of which is to assist toxicologists, cancer researchers, regulators, and others to predict the toxic effects of chemical compounds. We describe a series of experiments in tree and rule induction from a set of example chemicals whose carcinogenicity has been determined from long-term animal studies, and compare the resulting classification accuracy with eight published human and computer predictions for a common set of 44 test chemicals. The accuracy of our system is comparable to the most accurate human expert prediction yet published, and exceeds that of any of the computer-based predictions in the literature. The induced rules provide confirmation of current expert heuristic knowledge in this domain. These early results show that an inductive approach has excellent potential in predictive toxicology.