摘要
在药物设计中,可以利用药物分子的构效关系模型进行药物活性的预报,从而降低药物开发的成本、缩短开发的周期。本文尝试结合Co-Training方法和嵌入式特征选择方法,提出了一种新的FESCOT(FeatureSelectionforCo-Training)算法。算法在药物活性数据集上进行了实验,结果显示结合了特征选择的Co-Training方法较之以前泛化能力有所提高。
The activity of drug molecule can be predicted by the QSAR （Quantitative Structure Activity Relationship） model, which overcomes the disadvantages of high cost and long cycle with the traditional experimental method only. With the fact that the number of drug molecule with known activity is less than those of unknown activity, it is important to predict molecular activities with the semi-supervised learning method. However, the numerous features of drug molecule affect the prediction accuracy of the QSAR model. Therefore, a novel algorithm named FESCOT （Feature Selection for Co-Training）is proposed in this paper, which combines Co-Training and an embedded feature selection method. Experiments are carried out on the data set of molecular activities, and the results show that generalization ability of FESCOT is better than that of Co-Training without feature selection.
出处
《计算机科学》
CSCD
北大核心
2006年第12期159-161,共3页
Computer Science
基金
国家自然科学基金(20503015)
上海市教委自然科学一般项目(05AZ67)
上海市教委E研究院-上海高校网格项目(20030301)的资助。
关键词
药物活性
半监督学习
特征选择
Molecular aetivity,Semi-supervised learning,Feature selection
