A comparative study of molecular similarity, statistical, and neural methods for predicting toxic modes of action

Quantitative structure-activity relationship (QSAR) models are routinely used in predicting toxicologic and ecotoxicologic effects of untested chemicals. One critical factor in QSAR-based risk assessment is the proper assignment of a chemical to a mode of action and associated QSAR. In this paper, we used molecular similarity, neural networks, and discriminant analysis methods to predict acute toxic modes of action for a set of 283 chemicals. The majority of these molecules had been previously determined through toxicodynamic studies in fish to be narcotics (two classes), electrophiles/proelectrophiles, uncouplers of oxidative phosphorylation, acetylcholinesterase inhibitors, and neurotoxicants. Nonempirical parameters, such as topological indices and atom pairs, were used as structural descriptors for the development of similarity-based, statistical, and neural network models. Rates of correct classification ranged from 65 to 95% for these 283 chemicals.