Prediction of the acute toxicity of chemical compounds to the fathead minnow by machine learning approaches

Support vector machines (SVM) and artificial neural networks (ANN) are applied for prediction of the acute toxicity of compounds to fathead minnow from molecular structure. A diverse set of 611 compounds, including 442 fathead minnow toxicity (FMT) agents and 169 non-FMT agents, are adopted to develop the classification models. A hybrid feature selection method, which combines Fischer's score and Monte Carlo simulated annealing embedded in the SVM approach, is used to select the relevant descriptors from 1559 molecular descriptors. Five-fold cross-validation method is used to optimize the model parameters and select the relevant descriptors. Using the 60 selected descriptors, SVM model gives an averaged prediction accuracy of 95.5% for FMT, 79.3% for non-FMT and 91.0% for all samples, while the corresponding values of the ANN model are 92.5%, 75.2% and 87.7%, respectively. The study indicates that the hybrid feature selection method is very efficient and the selected descriptors from the SVM approach have also a good performance for the ANN approach. A hold-out method is used to build the final classification models by using the selected descriptors and optimized model parameters from the 5-fold cross-validation. The SVM model gives an excellent prediction accuracy of 96.6% for FMT, 93.0% for non-FMT and 95.1% for all samples, while the corresponding values of the ANN model are 91.4%, 90.7% and 91.1%, respectively.