流形学习的四嗪衍生物抗癌活性预测模型研究

Study on prediction model of tetrazine derivatives anticancer activity with manifold learning

由于四嗪衍生物抗癌活性与其结构之间可能存在非线性关系,本文引入非线性的流形学习方法对计算出的四嗪衍生物分子描述符进行特征提取,以提高其预测模型的准确性。便于分析,分别采用特征选择的逐步回归法、线性特征提取的主成份分析法以及非线性特征提取的流形学习方法对四嗪衍生物分子描述符进行筛选,然后基于偏最小二乘和支持向量回归机构建其定量构效关系模型。计算结果表明,本文中四嗪衍生物的描述符数据为非线性流形,并且它们的结构与活性之间呈非线性关系,基于支持向量回归机模型的最优预测结果达到了97.4%。所以,利用非线性特征提取的流形学习预处理的QSAR模型可以为此类化合物抗癌活性的预测提供指导。

Because of the probably nonlinear relationship between the antitumor activity and structure of tetrazine derivatives, in order to improve the efficiency of prediction model, the nonlinear manifold learning method is introduced to extract feature for the descriptors of tetrazine derivatives. For convenient for analysis, the stepwise regress method for feature selection, PCA for linear feature extraction and manifold learning for nonlinear feature extraction are used to select descriptors of tetrazine derivatives, respectively. After that, building the quantitative structure-activity relationship （QSAR） models for the selected descriptors based on Partial Least Squares （PLS） and Support Vector Regress （SVR）, respectively. The calculation results show that, the descriptors dataset of tetrazine derivatives in this paper is nonlinear manifold, and the relationship between structure and activity of these is nonlinear, the optimal prediction result based on SVR reaches 97.4 %. So, after the pretreatment of manifold learning for nonlinear feature extraction, the QSAR model can provide guidance for predicting such compounds anticancer activity.