Modeling the toxicity of ionic liquids based on deep learning method

This work proposed a hybrid molecular descriptor combined with deep learning method to model and evaluate the rational application of ionic liquids (ILs) through deep convolutional neural networks (DCNN). A total toxicity dataset of ILs against the leukemia rat cell line (ICP-81) was collected from the literature. The MACCS fingerprint and sigma profiles of the ILs were calculated using the RDKit packet and the COSMO-SAC model, respectively. The hyperparameters of the DCNN model were optimized by combining Bayesian optimization and local search algorithm. The importance of the feature descriptors was determined based on their influence on the DCNN model. The obtained results showed that the proposed model had a satisfactory prediction accuracy, and the coefficient of determination (R2) for the train set and test set were 0.972 and 0.965. This work provides guidance for the screening of ILs and their rational application in the industry.