QSAR Studies With The Descriptors Derived From Local Electrostatic Potentials On Molecular Surface

Quantitative structure-property/activity relationships (QSPRs/QSARs) have been widely used in the fields of chemistry, medicinal and environment sciences. One of the core works in QSAR is to find suitable structural descriptors. It is well known that physicochemical properties and biological activities depend on intermolecular non-covalent interactions. As these interactions are mainly electrostatic, the distribution of electrostatic potentials (ESPs) on molecular surface is therefore of particular importance. On the basis of the work of Politzer et al, we have developed a series of new ESP-derived descriptors, and applied successfully in many QSPR studies relevant to solute-solvent interactions. In this dissertation, we proposed a type of ESP-derived descriptors based on molecular fragments (or local molecular structure), and tried to apply them to the QSAR study. The full text is divided into six chapters, which can be described as follows:In the first chapter, we firstly summarized the development history and workflow of the QSAR/QSPR, highlighting the molecular descriptors and modeling methods. Subsequently, ESP, molecular surface and ESP-derived structural descriptors as well as the progress were introduced. Lastly, the research idea of this dissertation was put forward.The second chapter is related tothe QSAR study of the neonicotinoid insecticides based on local ESP-derived desriptors. Firstly, the molecular structure is divided into three parts, X, Y, and Core, based on their ligand-receptor binding mode. Then geometrical optimization and ESP calculations have been performed for a group of 173 neonicotinoid compounds. The global and local ESP-derived desriptorswere extracted, and the quantitative relationships between insecticidal activity and surface ESP descriptors have been established with multiple linear regression (MLR) and some nonlinear methods, SVM, GP, LSSVM. The results showed that, the local ESP-derived descriptors,Vmin(Core)、Vs,max(Core)、Vind,s0(Y)、∏(Core)、σtot 2t(Y)、∏ind- (X) �'� ∑s,ind V+(X), together with molecules volume V, the lowest unoccupied molecular orbital energy level (εLUMO) can be well used to build a quantitative structure-activity relationship of these compounds. This information may help us to better understand the binding mode of neonicotinoid compounds with the corresponding receptor (nAChRs). At the same time, the models have strong predictive power, it is expected to provide us a new mean to predict other similar neonicotinoid compounds.The third chapter is committed to QSAR study of human urate transporter (hURAT1) inhibitors. Firstly, the molecular structure is divided into three parts, (1) carbonyl group, (2) functional group 1 and (3) functional groups 2 or 3, according to the structural feature of 38 hURATl inhibitors.Then geometrical optimization and ESP calculations have been performed for these series compounds. The global and local ESP-derived descriptors were extracted, and the QSAR models were established with MLR and some nonlinear modeling methods. Among all models, the GP model exhibits good robustness and best predictability.3D-QSAR study was also performed for comparison. It is shown that both the CoMFA and CoMSIA modelsare inferior to the 2D-QSAR model based on local ESP descriptors.In the fourth chapter, QSAR studyofa group of 65 stibeneTTR inhibitors was conducted. The results show that:the ESP-derived descriptors,∑ind V+、∑ind V+、Vs-、 combined with the lowest unoccupied molecular orbital energy level (εLUMO) can be used to build quantitative structure-activity relationship of these compounds. Stability and predictive ability of the model is not as satisfactory as expected. Subsequently, the nonlinear model has been established with Gaussian process. It has been revealed that GP model exhibits better robustness and predictability as compared to the linear model. Finally, considering the fact that activity data of these compoundsis not of high accuracy (inhibitory percentage), we performed a discriminant analysis for this data set with similar descriptors. The results of Fisher discriminant analysis showed that the accuracy of this series of compounds itself to verify is 90.4%. cross-validation accuracy is 88.5%, and predictive accuracy is 92.3%.The fifth chapter was devoted to QSPR study of adsorption of aromatic pollutant compounds on the carbon nanotubes. It appears that the quantities derived from electrostatic potentials, Vmin、σ+2, and ∑ind V+, together with the molecular surface area (S) and the energy level of lowest occupied molecular orbital (εLUMO) can be well used to express the quantitative structure-property relationship.For the data set investigated, linear model seems to be enough. Nonlinear SVM and LSSVM models exhibit strong fitting abilities, their predictive powers are inferior to other models. The GP model yields the best fitting and predictive abilities among all models, its advantage over the linear model, however, is not so remarkable as expected.In the last chapter, we summarized the full text and proposed future research direction.