Predict Toxicity Effects of Endocrine Disruptor Chemicals on Aquatic Organisms Using Machine Learning

Key words: 
• endocrine disruptor chemicals /
• Pimephales promelas /
• chronic toxicity /
• machine learning /
• QSAR

Abstract: The time-consuming and high costs of reproductive toxicity test of endocrine disruptor chemicals (EDCs) lead to a relatively lack of reproductive toxicity data for aquatic species, which restrict the ecological risk assessment and management of EDCs. The prediction of toxicity data is one of the important methods to solve the above problems, and it is also one of the hotspots and difficulties in the field of ecotoxicology. Based on the review of related research using machine learning to predict chemicals’ toxicity effects on aquatic organisms, a support vector machine (SVM) and a linear neural network (LNN) coupled with quantitative structure-activity relationship (QSAR) were used respectively, to build binary classification models to predict reproduction toxicity for Pimephales promelas, and the models were validated and evaluated using the reproduction toxicity dataset. The results of review showed that SVM was the most widely used model to predict the toxicity effects of compounds on aquatic organisms, followed by linear regression and neural network. Acute toxicity has been studied more than chronic toxicity in application of the machine learning. There was no clear theoretical guidance for the selection of molecular descriptors subset in the field of QSAR. Generally, the combination of experiences and algorithms was applied to filtrate molecular descriptors. The descriptors related to octanol-water partition coefficient were considered to be of high importance. The experimental results are as follows: four descriptors that related to atomic mass, polarizability, ionization potential and bond order were obtained as input variables. The prediction accuracies of SVM for the training set and the test set are 0.91 and 0.88 respectively, and the area under the curve (AUC) of the training set and the test set obtained from the receiver operating characteristic (ROC) curve are 0.93 and 0.88 respectively. The accuracies of LNN for the training set and the test set are both 0.82, and the AUC are 0.87 and 0.88, respectively, indicating that LNN and SVM have good generalization and prediction ability. The results of SVM are better than that of LNN, which means that SVM is more suitable for small dataset. The results can provide an important supplement for the ecotoxicological studies of EDCs and enrich the toxicity data, as well as provide a scientific reference for the ecological risk management of EDCs.