Application Of QSAR Model Based On Artificial Neural Network In The Joint Toxicity Of Mixtures

Quantitative structure activity relationship(QSAR)is a method to study the relationship between the molecular structure of compounds and their physicochemical properties or biological activities.It has been just 80 years since QSAR was put forward,and the method has been already applied in the fields of chemistry,toxicology,environmental science and ecological science.Considering that in the actual ecological environment,componds are mostly in the form of a mixture,in recent years,researchers have begun to focus on research of QSAR model of mixtures.Predicting the joint toxicity of mixture with QSAR model is a research topic of great significance.Chapter 1 of the dissertation introducesd the significance of this study,the research status of QSAR at home and abroad,the content,method,innovation and technical route of this study.Chapter 2 introduced the development history of QSAR,several classical QSAR models,the modeling steps of QSAR model,the application of QSAR method,the current research situation of QSAR and the principle of artificial neural network.In the third chapter and the fourth chapter,the QSAR modeling of different mixtures was studied.First of all,12 kinds of benzene was chosen,including non-polar narcotic compounds and polar narcotic compounds.Mixtures of different classification in these compounds was generated,and then QSAR models for different mixtures was set up.Secondly,The three kinds of mixed molecular descriptors were calculated.The topological parameters needed to be calculated using the descriptor of single compound of the mixture.The molecular structure descriptors were divided into three categories: physical chemistry parameters,topological parameters and quantum chemical parameters,which were obtained respectively by consulting Chemistry Manual,software E-dragon and software Gaussian09.In the end,33 commonly used classical descriptors were selected,and the number of the three kinds of parameters were 10,12 and 11,respectively.Thirdly,the method of forward,backward,stepwise regression and principal component analysis in software SPSS19.0 were used to screen the descriptors,and the variables with serious collinearity were excluded.Because of the difference in structure and property between non-polar and polar materials,the descriptors selected by the two kinds of mixtures weren't the same,and the number of them was 6 and 7 respectively.The modeles of the single descriptor and toxicity data were set up,and the results showed that different descriptors made different contribution to mixture toxicity,and the same descriptors in different mixtures also made different contribution to mixture toxicity.Finnaly,R language was used to establish and validate the model with four methods,they were the multivariate linear regression,partial least-squares regression,support vector machine method and back-propagation neural network.And internal and external methods were used to verify the model,calculating three statistics-the correlation coefficient QCV2,root mean square error of validation RMSEV and QF12.The results showed that the 4 kinds of models of the mixtures built had good goodness of fit,robustness and predictability.But in contrast,whether in nonpolar mixtures or polar mixtures,the artificial neural network model was superior to the other 3 models.Thus,ANN method is suitable for the establishment of QSAR model of mixture.The innovation of this paper is that different kinds of molecular structure descriptors in QSAR model are used,which is more comprehensive than single descriptor or single class descriptor.In order to obtain the highest correlation descriptor of mixture toxicity,several methods are used to select variables.Artificial neural network modeling method belongs to supervised learning,it replaces simple linear modeling methods and can deal well with the nonlinear relationship.And results also show that the model established with ANN method has good predictive ability and robustness.