Prediction of aquatic toxicity of pesticides by using linear and nonlinear techniques

Abstract

Economical and environmental considerations for assessing toxicity of chemicals have led to a considerable amount of studies on the computational techniques. Pesticides allocate a significant part in these chemicals, mainly for their toxic effects on nontarget organisms. In the present study, the toxicities of 91 organic compounds including pesticides to freshwater algae, Chlorella vulgaris; and the toxicities of a set of 34 pesticides to Oncorhynchus mykiss were modeled employing Counter Propagation Neural Network (CPNN) and Multiple Linear Regression (MLR). The analyses were performed with about 1500 descriptors calculated using Dragon 5.4, Spartan 06, and Codessa 2.2 software. Additionally, we used the Characteristic Root Index (CRI) which was proved to be a significant descriptor in previous QSPR/QSTR studies. Descriptor selection was made by Heuristic Method. Kohonen network was used for splitting the data set into training and test sets. Linear and nonlinear 3, 4 and 5-descriptor models were compared according to their statistics such as squared correlation coefficient and Root Mean Squared Error (RMSE). All models were validated externally by using test sets. BLTD48 from Dragon, electrophilicity from Spartan, and the CRI appeared to be significant for the developed QSTR models of Chlorella vulgaris. Oncorhynchus mykiss model underscores the Dragon descriptors. The statistical quality of the models for Chlorella vulgaris is compared to those of the previously published models using the same experimental data and found to be superior to those models. Oncorhynchus mykiss models are compared to literature models in terms of chemical classes, mechanism of action, and statistical tools and fits. Linear and nonlinear methods were found to be comparable for both species. .