The Classification And Regression Model For HPLC Solvents Based On Support Vector Machine

Chemical Pattern Recognition and QuantitativeStructure-Activity/Property Relationship (QSAR/QSPR) are veryimportant research branches of Chemometrics. In this dissertation,the main work is to build the QSPR models based on support vectormachine for classification and regression researches of HPLCsolvents.Either in chemical pattern recognition or QSPR researches, thefirst work you have to do is to calculate the molecular descriptors.However, some of these descriptors correlate each other stronglyand may be useless to the QSPR researches. So, it’s very importantto select the important descriptors first. In this dissertation, wecalculated molecular descriptors by computer softwareADMEWORKS ModelBuilder, which is calculated specially, and did acorrelation test with it. Then, according to the method of supportvector machine (SVM), the HPLC solvents were classified and themodels of their physicochemical properties based on descriptors selected by genetic algorithm were built. The models were used inQSPR researches of the physicochemical properties such asKamlet-Taft polar parameter, dipole moment, Hildebradt solubilityparameter.Support Vector Machine (SVM),which being developed since mid1990s[1],was introduced in chemometrics as a method from 2000,includes support vector classification (SVC) and support vectorregression (SVR). SVM can be used to solve not only linear problemsbut also the nonlinear ones efficiently by introducing suitablekernel functions. Moreover, SVM doesn’t exist the local minimumpoint in theory. SVM show strong stabilities and generalizationabilities in dealing with the small number of sample and nonlinearproblems. In this dissertation, we used multiple SVM classifiers toclassify HPLC solvents and built the regression models by SVR.Main work included in the dissertation:(1) First, we introduced the chemometric methods used in thedissertation and discussed the importance of molecular descriptorsin QSRP/QSAR(Quantitative Structure-Property Relationship/Quantitative Structure-Activity Relationship) researches. Then, awide set of molecular descriptors were calculated by computersoftware.(2) The basical theory of SVM was introduced simply. A classification model was built by multiple SVM classifiers based onmolecular descriptors. The model tested by independent sampleswith a good predictive power was got.(3) Regression models for solvents’polar, diplor moment andsolubility were built, and good predictive powers were got.