A priority control list for LCMs in freshwater food chain by deep learning

Liquid crystal monomers (LCMs) are widely distributed globally, and their persistence, bioaccumulation, and toxicity (PBT) effects pose potential threats to environmental homeostasis and human health. This study employed molecular docking method to investigate the 12879 values for the PBT effects of 1431 commercial LCMs in the freshwater food chain (i.e., Daphnia pulex - Danio rerio - Pelecanus crispus), forming a matrix of 1431 LCMs × 3 trophic levels × 3 PBT effects. A priority control list for 1431 LCMs affecting the freshwater food chains was developed and optimized by machine learning methods alongside Residual Neural Network (ResNet) deep learning model. The results of model evaluation demonstrated that the ResNet deep learning model achieved high accuracy of 0.84 and 0.85 on the test and validation sets, respectively. According to the priority control list optimized by the ResNet deep learning model, 509 LCMs were identified as high-risk category. Besides, the factors affecting the PBT effects were analyzed through the SHapley Additive exPlanations visualization, finding that introducing functional groups with lower electronegativity may reduce the PBT effects of LCMs. This is the first priority control list targeting the PBT effects of commercial LCMs affecting the freshwater food chain. The constructed ResNet deep learning model not only enhances the accuracy of the list, but can also enables prediction for the PBT risks of other LCMs in the freshwater food chain.