MOA-based linear and nonlinear QSAR models for predicting the toxicity of organic chemicals to Vibrio fischeri

Risk assessment of pollutants to humans and ecosystems requires much toxicological data. However, experimental testing of compounds expends a large number of animals and is criticized for ethical reasons. The in silico method is playing an important role in filling the data gap. In this paper, the acute toxicity data of 1221 chemicals to Vibrio fischeri were collected. The global models obtained showed that there was a poor relationship between the toxicity data and the descriptors calculated based on linear and nonlinear regression analysis. This is due to the fact that the studied compounds contain not only non-reactive compounds but also reactive and specifically acting compounds with different modes of action (MOAs). MOAs are fundamental for the development of mechanistically based QSAR models and toxicity prediction. To investigate MOAs and develop MOA-based prediction models, the compounds were classified into baseline, less inert, reactive, and specifically acting compounds based on the modified Verhaar’s classification scheme. Satisfactory models were established by multivariate linear regression (MLR) and support vector machine (SVM) analysis not only for baseline and less inert chemicals, but also for reactive and specifically acting compounds. Compared with linear models obtained by the MLR method, the nonlinear models obtained by the SVM method had better performance. The cross validation proved that all of the models were robust except for those for reactive chemicals with nN (number of nitrogen atoms) = 0 and n(C=O) (number of carbonyl groups) > 0 (Q²ₑₓₜ < 0.5). The application domains and outliers are discussed for those MOA-based models. The models developed in this paper are significantly helpful not only because the application domains for baseline and less inert compounds have been expended, but also the toxicity of reactive and specifically acting compounds can be successfully predicted. This work will promote understanding of toxic mechanisms and toxicity prediction for the chemicals with structural diversity, especially for reactive and specifically acting compounds.