摘要
[目的]探索一种有效的组合预测方法,用于定量构效关系(QSAR)的研究分析。[方法]提出一种基于支持向量机回归(SVR)与K-最近邻法(KNN)的组合预测方法:以均方误差(MSE)最小为择优准则,对SVR实施核函数寻优;基于最优核函数以SVR进行描述符筛选并得到保留描述符;以"多轮末尾强制淘汰法"阐述各保留描述符对预测精度影响的程度;基于保留描述符,以不同KNN预测值反映样本集异质性并构建子模型,最后基于SVR以留一法实施组合预测。运用该组合预测方法研究磺酰脲和三唑并嘧啶磺酰胺类除草剂QSAR建模。[结果]建模结果表明,基于SVR与KNN的组合预测方法在参比模型中预测精度最高,具有结构风险最小、非线性、能有效克服过拟合、泛化推广能力优异等优点。[结论]基于SVR与KNN的组合预测具有许多优点,在QSAR研究中应用前景广泛。
[ Objective ] An effective combination forecasting method for QSAR （QSAR） analysis was explored. [ Method ] The combination forecasting method of SVR and KNN was presented. The optimal selection was based on the mean square error（MSE） and the SVR was optimally screened based on the kernel function. Based on the most optimal kernel function, the selection was done according to SVR descriptor and the descriptor was reserved. The impacting extent of the reservation descriptors on the accuracy of forecasting was explained with ＂The enforcement law of the last one being out＂. The quality of sample was reflected based on the different KNN forecast values and a sub-model was built, and finally, the combination forecasting was carried out based on the SVR. The QSAR model of the kinds of herbicide such as sulfonylurea, triazolo and pyrimidine sulfonamide was established with combination forecasting method. [ Resuh ] The modeling results showed that the forecasting method based on the combination of SVR and KNN was the most accurate in all of models, with smallest risk in the structure and non-linear, and also, it can effectively overcome the over-fit, and with the advantage such as good generalization ability etc. [ Conclusion] The forecasting method based on the combination of SVR and KNN had many advantages, which can be widely applied in study on the QSAR.
出处
《安徽农业科学》
CAS
北大核心
2008年第35期15284-15286,共3页
Journal of Anhui Agricultural Sciences
基金
湖南省科学计划一般项目(2008SK3056)
