在QSAR中应用SVR的非线性组合预测

Application of a non-lineal combinatorial forecasting method in QSAR modeIing based on support vector regression

为提高定量构效关系(quantitative structure-activity relationship,QSAR)模型预测的精度,以支持向量回归(support vector regression,SVR)全局与局部核函数,发展出1种非线性组合方法GK-LK-SVR,其基本思路为:依均方误差(MSE)最小原则,分别基于SVR的全局与局部核函数筛选描述符后预测,实测值与不同核函数的预测值组合成混合样本,然后再依MSE最小原则基于SVR对混合样本实施核函数寻优及子模型筛选,最后以留一法完成预测。对2种化合物QSAR建模结果表明:GK-LK-SVR方法预测精度高,有望在QSAR研究中得到广泛应用。

To improve the performance of QSAR model, a novel non-lineal combinatorial method named GK-LK-SVR based on global kernel and local kernel was proposed in this paper. Firstly, descriptors were screened by global kernel and local kernel, and then, SVR based prediction was carried out. The prediction results of different kernel and the experiment results assembled to be a mixed sample set. Kernels and sub-models screening based on SVR was evaluated by the rule of minimum MSE value. The forecast was took by the leave-one-out method based on optimized kernel and remained sub-model. Results of QSAR model for two compounds showed that GK-LK-SVR method has precise predicting ability and will be Widely used in QSAR study.