Prediction of acute toxicity of organic chemicals on rare minnow using machine learning

Acute toxicity data of organic chemicals were collected for rare minnows (Gobiocypris rarus). A machine learning method was developed to predict the acute toxicity of organic chemicals specially for rare minnow, using existing acute toxicity data for zebrafish (Danio rerio) and fathead minnow (Pimephales promelas). A binary model was used to determine whether the organic chemical have acute toxicity to rare minnow; if yes, a regression model was then utilized to predict the median lethal concentration LC50. Different machine learning models were compared, and it was found that the support vector machine performed the best in the binary classification model, with the accuracies of 92.4% in the training set and 88.6% in the test set, respectively. The elastic net regression method demonstrated the best performance in the regression model. The adjusted R2 of the training set was 0.87, while the adjusted R2 of the test set was 0.75. The cross-validation coefficient Q2 LOO of the left-one-out method was 0.52, and the external validation coefficient Q2EXT was 0.71. The two models exhibited commendable accuracy, robustness, and predictive capability. The first ionization potential and n-octanol water partition coefficient had a greater effect on the classification, and the regression prediction results were more heavily influenced by the topological charge. The above results offer a precise and efficient prediction method for assessing the acute toxicity of the rare minnow, an endemic model organism in China, significantly expediting the environmental risk assessment of organic chemicals. © 2024 Chinese Society for Environmental Sciences. All rights reserved.