支持向量机回归方法用于黄酮类化合物醛糖还原酶抑制活性的QSAR研究

Using support vector regression for QSAR study on the aldose reductase inhibitory activity of some flavones

用支持向量机回归(SVR)方法研究了48个黄酮类化合物醛糖还原酶抑制活性的定量构效关系。建模过程中利用留一法交叉验证(LOOCV)优化了核函数的类型、惩罚系数C和不敏感函数ε.所建模型最终采用了227个变量中的7个:dChivps9, ESHaaCH,EsssCH2,n2pag[1,2],degree2,I'3和I'4。所得SVR模型的预测相对误差为0.0622,小于多元线性回归(MLR)和偏最小二乘法(PLS),以及文献报道模型的预测相对误差。

Support vector regression （SVR）, a powerful machine learning technology based on statistical learning theory （SLT） was applied to QSAR on the aldose reductase inhibitory activity of 48 flavones. Leaving-one-out cross-validation （LOOCV） of SVR was introduced to optimize the parameters （the type of kernel function, the regularization parameter C, and ε-insensitive loss function） used in SVR model. The optimal model was built based on 7 descriptors： dChivps9, ESHaaCH, EsssCH2, n2Pag[1,2], degree2, I＇3, I＇4, which come from 227 descriptors available. The average of relative error （ARE） of SVR model was 0.0622 that was less than those of multiple linear regression （MLR） and partial least squares （PLS） methods. It was also found that the performance of SVR model was better than that reported.