Modeling the toxicity of chemicals to Tetrahymena pyriformis using heuristic multilinear regression and heuristic back-propagation neural networks.

During the last years, considerable effort has been devoted to model the toxicity of chemicals to Tetrahymena pyriformis for medium and large sized data sets using various artificial neural network (ANN) techniques. Motivation behind this has been to model highly complex relationships with nonlinear character making it possible to describe wide structural diversity within one model. The current work compares the performance of two heuristic methods in developing quantitative structure-activity relationship (QSAR) models: the best multilinear regression (BMLR) approach and the heuristic back-propagation neural networks (hBNN). The modeling is based on a diverse data set of 1371 organic chemicals with toxicity data (log(1/IGC50)) collected from the literature. The toxicity values correspond to the static 40-h Tetrahymena pyriformis population growth impairment assay. The comparison of the two methods showed that the BMLR approach produces acceptable QSAR models (R2 = 0.726), whereas the hBNN method produced a statistically more significant model (R2 = 0.826) for the given endpoint. The hBNN method was able to relate different descriptors to the toxicity than the BMLR method. Both models were validated with an external prediction set. The descriptors in the models were analyzed and discussed.