基于支持向量回归的非线性变量筛选用于酚类化合物QSAR建模的研究

首先基于支持向量回归（SVR）依均方根误差最小原则确定最优核函数,再以最优核函数为基础,进行SVR非线性自变量筛选,最后以所选自变量进行建模预测.将该方法应用于酚类化合物的QSAR研究,最优核函数确定为径向基核,最终保留自变量为疏水性参数（lgp）与拓扑指数（Am3）.结果表明：基于SVR进行变量筛选能有效地剔除无关自变量,进一步改进SVR对小样本数据的建模预测能力.该方法在农业环境毒性污染物的QSAR研究领域有较广泛的应用前景.

Nonlinear Variable Screening Based on Support Vector Regression and Its Application on Phenol Compounds QSAR Modeling

Firstly,the optimal kernel function was determined in accordance with the minimum root mean square error based on support vector regression （SVR）,and then on the basis of the optimal kernel function,the independent variables were screened nonlinearly using SVR.Finally,modeling was conducted on the training set and prediction was performed on the test set using the selected independent variables.The method was applied to QSAR study of phenolic compounds,the optimal kernel function was determined as RBF kernel,the retained independent variables as hydrophobic parameter （lgP） and topological index Am3.The results show that irrelevant variables can be effectively eliminate using SVR to screen variables and prediction ability was further improved for SVR modeling on small sample data,this method has a potential application prospect in the QSAR research field of environmental toxic pollutants of agriculture.