基于多核学习的肿瘤—药物—基因语义关系提取

发布时间：2018-02-25 17:18

本文关键词： 文献挖掘语义关系提取多核学习个体化用药肿瘤-药物-基因关系　出处：《北京协和医学院》2015年硕士论文　论文类型：学位论文

【摘要】：在科学研究发展的推动下,肿瘤类疾病的药物治疗进入针对个体基因特征的个性化治疗阶段。生物医学文献作为科学研究成果重要呈现方式之一,记录了大量有关肿瘤、药物和基因的数据,为深入研究和发现肿瘤个体化用药治疗方案提供重要的数据支持。然而,面对海量生物医学文献信息资源,从这类非结构化数据中自动地挖掘有价值的信息,尤其是肿瘤、基因和药物三者相互作用关系,需要借助于信息处理技术与方法实现。本研究采用多核机器学习方法从呈指数增长的海量生物医学文献中自动提取肿瘤—药物—基因三者之间的语义关联关系。多核机器学习的核函数选取方面,针对文献信息数据结构特点,分别从词形、语法和语义三方面选择向量空间核(Vector Space Kernel)和字符串核(String Kernel)作为词汇核(Lexical Kernel, LK),卷积树核(Tree Kernel)作为句法核(Syntax Kernel, SyK),以及基于WordNet的语义核(Semantic Kernel, SeK)。实验语料主要是从CTD数据库(The Comparative Toxicogenomics Database, http://ctdbase.org/)中的采集经过人工审编的肿瘤-药物、肿瘤-基因和药物-基因等相互作用关联关系数据及其来源于PubMed数据库文献信息。利用SVM训练多种方式组合核函数语义分类器并测试其分类性能。利用词汇核、句法核和语义核三者构成的线性组合核方法提取语义关系的性能优于其他核方法。构建基于多核机器学习的肿瘤—药物—基因语义关系提取模型,实现了自动提取肿瘤—药物、药物—基因和肿瘤—基因三种关系的具体语义关系类型。实验结果表明,肿瘤-药物关系的F-值为88.41%,肿瘤-基因关系的F-值为85.68%,以及药物-基因关系的F-值为71.31%,本研究的方法优于其他方法。在此基础上,本研究结合共现关联关系方法和多核学习语义关系提取模型,设计并实现了面向肿瘤个体化用药的语义关系提取原型系统。该系统可以从生物医学文献中自动识别和提取肿瘤—药物一基因三者间语义关系,并且支持用户对多个命名实体的多种组合查询和查询结果批量下载。
[Abstract]:With the development of scientific research, the drug therapy for tumor diseases has entered the stage of individualized treatment for individual genetic characteristics. As one of the important presentation methods of scientific research achievements, biomedical literature has recorded a large number of tumors. The data of drugs and genes provide important data support for the further study and discovery of individualized drug therapy for cancer. However, in the face of vast amounts of biomedical literature and information resources, The automatic mining of valuable information from such unstructured data, in particular the interaction of tumors, genes and drugs, The multi-core machine learning method is used to automatically extract the semantic relationship between tumor, drug and gene from a large number of biomedical literature with exponential growth. The kernel function selection aspect of multicore machine learning, According to the characteristics of the data structure of literature information, respectively from the word form, In syntax and semantics, vector space kernel Space Kernel and string kernel string Kernelare selected as lexical Kernel, LKO, convolutional tree Kernel. as syntactic cores Syntax Kernel, Sykneland semantic kernels based on WordNet, semantic Kernel. Sekel. the experimental corpus is mainly from CTD data. In the Comparative Toxicogenomics Database, http: r / ctdbase.org / r / ctdbase.org / r / c / ctdbase.org / r / ctdbase.org. Tumor-gene and drug-gene interaction correlation data and their sources from PubMed database literature information. Using SVM to train a variety of ways to combine kernel function semantic classifier and test its classification performance. The linear combination kernel composed of syntactic kernels and semantic kernels is superior to other kernel methods in extracting semantic relations. A model of extracting tumor-drug-gene semantic relationship based on multicore machine learning is constructed to extract tumor-drug automatically. The specific semantic relationship types of drug-gene and tumor-gene relationships. The F- value of the tumor-drug relationship is 88.41, the F- value of the tumor-gene relationship is 85.68, and the F- value of the drug-gene relationship is 71.31. The method in this study is superior to other methods. In this study, the co-occurrence correlation method and the multi-core learning semantic relation extraction model are combined. A prototype system of semantic relation extraction for tumor individualized drug use is designed and implemented, which can automatically identify and extract the semantic relationship between tumor and drug gene from biomedical literature. And support users to multiple named entities of multiple combinations of queries and query results download batch.
【学位授予单位】：北京协和医学院
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：R-05;G254

【相似文献】