QSAR Studies On Real Biochemical Systems Using Novel Chemometric Algorithms

Quantitative structure-activity relationship (QSAR) is an important branch in chemometrics. Based on several novel algorithms proposed by our laboratory, four applications have been performed in the presented Master D. thesis.The first step in building a QSAR model is to calculate a great number of molecular descriptors, then the second is to select the descriptors most correlated to the studied object, that is to say, variable selection. Using particle swarm optimization (PSO) algorithm to select variables and multiple linear regression (MLR) method to build models, the activity of cyclooxygenase (COX) and the toxicity of aromatic compounds have been predicted. The conclusions show that PSO can get satisfactory results comparable to genetic algorithm.Artificial Neural Network (ANN), a powerful modeling method for the nonlinear statistical studies, has some drawbacks like the possible premature convergene to local optima and overfitting to samples in the training set. In order to improve the ability of ANN, chaos theory was introduced to chemometrics and chaotic genetic artificial neural network (CGANN) was proposed. Using this method, QSAR model of distamycin derivatives was constructed. Compared with PLS, CGANN has superiority in both the training ability and the predicting ability. In addition, hybrid PSO was also introduced to train ANN (HPSONN). Based on this novel method, QSAR analysis of inhibitors of platelet-derived growth factor receptor (PDGFR) phosphorylation was performed. The results have been improved comparing to MLR.Through the analysis of the used descriptors in building models some information about the interaction between compounds and biomacromolecules can be achieved, which will guide the subsequent drug design and synthesis.