Predicting toxicity of complex mixtures by artificial neural networks

Abstract

Industrial and municipal wastewaters constitute major sources of contamination of the aquatic compartment and represent a threat to the aquatic life. Artificial neural networks, based on three different learning paradigms, were studied to predict the acute toxicity of trouts (5 days exposure to wastewaters) with the inputs of two simple microbiotests which requires only 5 and 15 min incubation time. These microbiotests were 1) the chemoluminescent peroxidase (CI-Per) assay which can detect radical scavengers and enzyme-inhibiting substances, and 2) the bacterial luminescent toxicity test (Microtox^TM) which is responsive to toxic substances affecting the vital function of bacteria. The responses obtained with the the trout bioassay, the C1-Per and the Microtox^TM microbiotests were analyzed for statistical correlations (Pearson-moment correlation), unsupervised learning by self-organizing network, and assisted learning by the backpropagation and the Hopfield (probabilistic) paradigms. The results showed that no significant correlation (p<0.05) was obtained between either the responses obtained with Cl-Per (p = 0.121) or the Microtox^TM (p = 0.061) microbiotests with the ones obtained with the trout bioassay. The self-organizing network was able to identify by itself a maximum number of 5 classes that were more or less related to fish toxicity: class 1 contained 2 samples that were toxic to fish, class 2 contained 2/3 samples that were toxic to fish, class 3 showed 6/8 samples that were not toxic, class 4 contained 5/6 samples that were non-toxic and class 5 identified one sample that was toxic. Supervised learning with backpropagation analysis yielded 2 kinds of networks that proved promising. The first one, was able to predict the actual toxic wastewater concentration with an overall performance of 65 % with unseen data while the second one, which was designed to identify toxic effluents from those that were not, had a much better performance (90 %). However, the probabilistic network proved also a very good prediction model to predict toxicity to fish with an overall performance of 90 %. Although more data is needed, the network based on the backpropagation paradigm seems a better predictor or classifier of trout toxicity with the Cl-Per and the Microtox^TM microbiotests.