基于支持向量机回归的抗癌药物活性研究

Predicting bis[(acridine-4-carboxamides)propyl]methylamines analog compounds by using Support Vector Regression

采用支持向量机回归(SVR)方法研究了40个抗癌化合物-二取代[(吖啶-4-酰胺基)丙基]甲胺类衍生物的定量构效关系,基于留一法交叉验证的结果,其平均相对误差是6.56%。结果表明,所建SVR模型的精度高于逆传播人工神经网络(BPANN)、多元线性回归(MLR)和偏最小二乘法(PLS)所得的结果。

In this work,support vector regression（SVR）,an effective machine learning method,proposed by Vapnik was applied to establish QSAR model for a series of novel anticancer agents-Ais[（acridine-4-carboxamides）propyl]methylainines.Six descriptors（including HOMO＂,LUMO~＋, Surface Area Grid,RMS Gradient,Polarizability and LogP） were selected for constructing the SVR mode by using floating searching feature selection method.The kernel function（including the linear kernel function,the polynomial kernel function,and the RBF kernel function） and parameters（e,C,and g） were adjusted by leave-one-out cross validation（LOOCV） method which was used to judge the predictive power of different models.After optimization,one optimal SVR-QSAR model was attained,and the mean relative errors（MREs） of LOOCV by using SVR is 6.56%. Based on the LOOCV test,the performance of SVR model is also compared with back-propagation neural networks（BP-ANN）,multiple linear regression（MLR） and partial least squares（PLS） for this real world data set.The results show that the performance of SVR model outperforms those of MLR,PLS and BP-ANN for this case study.Finally,sensitivity analysis was employed to study how the six descriptors affect the activity.As a result,HOMO,Polarizability,Surface Area Grid negatively affected the activity,LogP positively affected the activity.