QSAR Of Chemical Pesticides Based On Support Vector Machine

Chemical pesticide plays a very importance role in plant protection. Quantitative Structure-activity Relationship (QSAR) researches always connect with chemical and biology. The chemical pesticide modeling based on QSAR could effectively increase the activity of chemical pesticide, and saving the time and costs of new pesticide exploited. Traditional, the QSAR modeling is based on multiple linear regression (MLR), stepwise linear regression (SLR), principal component regression (PCR), partial least square regression (PLR), artificial neural networks (ANN), and so on. But, these methods always have some flaws itself. Support vector machine (SVM) based on statistical learning theory is the machine learning method with fastest improving speed. SVM is based on structural risk minization, and it was generated form solved small sample, capable of dealing with issues under over-fitting, non-linear, circumstance of high dimensional and small sample. SVM is divided into support vector classification (SVC) and support vector regression (SVR), and it could effectively fit QSAR modeling. But SVM also has some flaws itself, such as kernel selecting and lack of epexegetically.Because of the advantage of SVM, we apply it to quantitative structure-activity relationship (QSAR) research of chemical pesticide. Based on SVM, we develop some algorithms to achieve it. Firstly, based on mean square error (MSE) minimization principle we choose the best kernel. Secondly, develop multi-round optimization method to step-to-step select descriptors. Finally, using single-factor's importance and effect analysis, we get the retained descriptors'influence to final model, and SVM model has some explanation.Based on above improvement, we apply SVM to QSAR research of neonicotinoid insecticides and sulfonylurea herbicides, and both of it were get good results. First, the neonicotinoid insecticides model after filtering retained six descriptors. The six descriptors were:Hydrophobic parameter, Molecular van der Waals volume, Molecular size volume, Dihedral angle2, Net charge of the pyridine ring, Orbital energy difference. At the same time, we obtained the importance and changing trends of the activity of the insecticide against the values of different descriptors through the analysis of single descriptor's importance and effect. Finally, the six retained descriptors were used to establish the QSAR model, of which the fitting MSE is 0.0803, the fitting R2 0.9584, the leave-one-out MSE 0.3004, the leave-one-out R2 0.8479. The precision of SVR-QSAR model is better than PCR, PLS methods. Second, the sulfonylurea herbicides model after filtering retained seven descriptors. The seven descriptors were:molecular polarizability, the 11th,12th,14th,15th,16th atomics' Milliken Electric charge, the Hydrogen atoms Milliken Electric charge of 10th nitrogen atomics connected, and we obtained the importance and changing trends of the activity of the herbicides against the values of different descriptors through the analysis of single descriptor's importance and effect. Finally, the seven retained descriptors were used to establish the QSAR model, of which the fitting MSE is 0.0442, the fitting R2 0.9577, the leave-one-out MSE 0.2579, the leave-one-out R2 0.7500. The precision of SVR-QSAR model is better than PCR, PLS methods.The improved SVR have extensive application prospect in non-linear chemical pesticide QSAR modeling.