Quantitative Structure-property(Activity) Relationships Of Carbon-based Nanomaterials And Carbon Cluster Derivatives

Carbon nanomaterials are the most potential functional materials in the 21st century.They have unique electrical,mechanical,optical,antioxidant and anti-bacterial properties,widely applied in semiconductors,biomedicine,new materials and other fields.Quantitative structure-activity(property)relationships(QSAR/QSPR)is a research method based on the quantitative relationship between the molecular structure characteristics and the biological activity(physical and chemical properties)of the compound,which builds a mathematical model by using statistical methods.QSAR/QSPR researches on carbon nanomaterials or carbon cluster derivatives are of great significance for deeply understanding such special materials and better utilizing their functions.This thesis is divided into four parts.In the first part,basic concepts,basic principles,research processes,development history of QSAR/QSPR and so on were introduced.We focused on introducing structural descriptors,modeling methods,and especially the molecular surface electrostatic potential parameters developed by our research group.The research idea of this thesis was put forward.The second part is related to QSPR studies on the dispersibility of graphene and single-walled carbon nanotubes in different solvents.The results show that the surface electrostatic potential parameters combined with conventional quantum chemical descriptors can be well used to describe the quantitative relationship between the molecular structure of solvents and the dispersion of graphene in various solvents.The correlation coefficient and standard deviation of the linear model obtained by using multiple linear regression analysis(MLR)are R2=0.951,RMSEE=0.103,respectively.The correlation coefficient and standard deviation of Leave-one-out cross-validation are RPred 2=0.937,RMSEP=0.086,respectively.Eight structural parameters were introduced in the model.These parameters have clear physical meanings and can better explain the rationality of the model from the perspective of solvent-solute molecular interaction.Using these parameters and up-to-date machine learning algorithm,we have constructed a variety of nonlinear models.Among these nonlinear models,the results of Partial Least Squares Support Vector Machine(LSSVM)and Gaussian Process(GP)are more ideal,which are better than the linear model.Monte Carlo Cross-Validation(MCCV)analysis shows that the peak and median qext2 values of the linear model and the GP model are 0.935 and 0.915,respectively.Moerover,The integral qext2 value of the two models are all above 0.85.The results of MCCV show that predictive ability of our model is pretty good.This result is better than the existing literature reports(Yousefinejad S et al.used the Dragon descriptors to build a model,which R2 and RMSEE values are respectively 0.850,0.380,and Rpred2 and RMSEP values are respectively 0.900 and 0.320).However,the modeling results for the dispersion of single-walled carbon nanotubes in different solvents are not satisfactory.The correlation coefficients for the training set and the predicted set are R-=0.654 and Rpred2=0.483,respectively.By contrast,the fitting ability and predictive ability of nonlinear LSSVM and GP models have been significantly improved(The correlation coefficient of them are 0.738 and 0.743,respectively.The correlation coefficient of their predicted set are respectively Rpred2=0.662 and 0.748).But compared with the literature results,they are still slightly inferior.Carbon nanotubes can be regarded as graphene molecules curled,in which the surface electrostatic potential parameters may have certain limitations in reflecting the changes of electronic structure properties.In the third part,QSPR of the solubility of C60 in different solvents was conducted.The linear model of the surface electrostatic potential descriptor and solubility established by the MLR method yield R2=0.819.MCCV analysis shows that the peak,median,and iitegral qex2t are all greater than 0.770,indicating that the model's stability and predictive ability are good.Relative to linear models,the stability and predictive ability of nonlinear models such as LSSVM,GP and Random Forest(RF)have improved to a large extent.Among them,the GP model performed best,with R2 reaching 0.988.The fourth part is related to QSAR of the C60 derivatives as HIV-1 protease inhibitors.For the linear model of the training set,with the value of R2 is 0.805,indicating that the model's the fitting ability is acceptable,but the model has poor predictive ability for the prediction set.The HIV-1 protease inhibitory activity reflects the complicated interaction between the active site of HIV-1 protease and the C60 derivative.However,the structural information reflected by the surface electrostatic potential descriptor is mostly related to the intermolecular electrostatic interaction,while the influence of steric factors is not considered.Therefore,there are still limits for the surface electrostatic potential parameters in QSAR research.