Prediction Of Persistence And Accumulation Of PFAS Based On Machine Learning

Perfluoroalkyl substances are a class of compounds containing alkylfluoroalkyl fragments,mainly used in fire protection,electroplating,textile,medicine,kitchenware coating and food packaging.Due to the high energy C-F bond contained in its molecules,it has extremely stable physical and chemical properties,and is difficult to be hydrolyzed and biodegradable.It is easy to combine proteins and accumulate in organisms.In recent years,a large variety of new PFAS with strong structural diversity have been detected in a variety of environmental media,and their environmental persistence,bioaccumulation and toxicity have attracted wide attention.Evaluation of persistence and bioaccumulation by experimental methods is often limited by the difficulty of obtaining standard substances,the high cost of testing,and the inability to meet the chemical safety screening of numerous novel PFAS.Based on theoretical simulation methods such as machine learning,molecular dynamics and quantum chemistry,high-throughput virtual screening and mechanism analysis can be realized for the persistence and bioaccumulation of novel PFAS.Therefore,this study conducted computational simulation studies on defluorination ability and plasma protein binding ability,two typical PFAS with persistence and blood accumulation characteristics,and constructed prediction models based on machine learning,quantum chemistry and molecular dynamics simulation,to provide theoretical basis and technical support for environmental and health risk assessment of novel PFAS.Specific research results are as follows:（1）Hydration electrons produced by UV play a key role in the defluorination of PFAS.However,existing experimental data limit the in-depth understanding of the defluorination ability of novel PFAS.Therefore,a quantitative constitutive activity relationship model based on machine learning algorithm was used in this study to establish a predictive model of PFAS defluorination capability.The hyperparameter optimization of the model was performed by using half-fold cross validation.The results showed that the gradient lifting algorithm using Pa DEL descriptor was the best model,and the statistical evaluation indexes were R_test²=0.944and RMSE_test=0.114.The evaluation of the importance of the descriptors showed that the electrostatic properties and topological structures of the compounds had significant effects on the defluorination ability of PFAS.For novel PFAS,such as potential alternatives to PFOS,the defluorination capacity is weak,while perfluoroalkyl ether carboxylic acids have a stronger defluorination capacity than perfluorooctanoic acid.Quantum chemical calculations show that the addition of additional electrons to the PFAS structure leads to molecular deconstruction,changes in the dihedral Angle of the fluorocarbon chain,and cleavage of the C-F and ether C-O bonds.（2）In this study,we developed a machine learn-based predictive model for PFAS plasma protein-binding ratio by collecting a large number of high-quality fluorochemicals data sets from relevant environmental and pharmacochemical databases.By comparing the performance of the traditional machine learning model and the graph neural network model,it is found that the accuracy of the best prediction model is 0.901.Predictions of more than7,000 PFAS show that most（～92%）are highly bound in plasma.The model interpretation suggests that the introduction of an alkaline amino group into the PFAS structure increases the binding affinity of the molecule to plasma proteins.Molecular dynamics simulations have shown that highly bound PFAS can bind at different binding sites of serum albumin,thereby increasing its accumulation in the blood.The above workflow integrating machine learning and molecular dynamics can be used to predict the bio-accumulative properties of novel PFAS and facilitate the molecular design and structural modification of environmentally friendly PFAS.