【作者】 徐栋；

【导师】 朱其立（Kenny Qili Zhu）；

【作者基本信息】 上海交通大学 ， 计算机科学与技术， 2016， 硕士

【摘要】 随着人类在生物医学领域的不断探索,生物医学领域的文献数量也在呈指数级的快速增长着。大量的文献使得研究人员从中获取相关知识变得越来越困难。本项目中,利用文本挖掘和自然语言处理的相关技术,我们提出一个半自动的系统框架,用于从海量的生物医学文献中提取致病基因。我们致病基因提取框架包含三个部分:命名实体识别、关系抽取和基因排序。对于命名实体识别,我们通过扩展现有的基因库和疾病库,实现一个基于词典的基因、疾病匹配程序,去识别Medline的摘要中的基因和疾病。同时,我们还设计一些启发式的规则,过滤掉一些常见的误识别的基因和疾病。该混合方法的F-score可以达到0.822。在关系抽取阶段,所有出现在同一个句子中的基因、疾病二元组都被视为候选对象,我们训练了一个基于支持向量基的二分类器,用来判断这些候选对象中的基因和疾病是否相关。我们考虑了两类不同的特征,一是基因和疾病的上下文,即基因和疾病周围的单词,二是依赖树上的信息,1)依赖树上基因和疾病之间的最短路径上的一元组、二元组和三元组,以及2)基因和疾病的最小公共祖先到根节点上的一元组、二元组和三元组。在包含了 1000个正例和1000个负例的测试集上,10-交叉验证得到的F-score为 0.863。对于每一个疾病,我们使用三种不同的方法对与它相关的基因进行排序。最简单的是利用疾病和基因共同出现的文章个数作为分数进行排序。第二种方法考虑到不同文章的权重不同,我们利用Page Rank算法及文章之间相互引用图来计算文章的权重。第三种方法则考虑到同一个作者会在不同文章提及同一基因、疾病二元组,我们对这样的情况进行抑制。我们在10个疾病上进行了测试,三种排序方法的MRR分别为0.249,0.281和0.293。此外,我们取50个与疾病最为相关的基因,得到的F-score为0.259,这比已知的所有提取基因、疾病关系的系统都要高。更多还原

【Abstract】 Alone with the rapid growth of biomedical literatures,it becomes increasingly difficult to find out useful knowledge from them.In this paper,we proposed a simi-automatic framework to extract gene-disease associations from a huge amount of biomedical literatures based on the technology of text mining and natural language processing.Our association mining framework consists of three phases,name entity recognition,as-sociation detection and ranking.In the name entity recognition(NER)phase,we extended the exsiting gene and disease databases.Then,we designed a MEDRA-based longest match strat-egy to recongnize gene and disease term in Medline abstract and article titles.Also,A number of heuristic rules are applied to filter out some error terms which are recognized by the dictio-naries.Such hybrid technique achieves 0.84 Fl-score in recognizing genes and diseases from the MEDLINE abstracts.In the association detection phase,all recognized gene-disease pairs that co-occur within the same sentence are considered as candidate evidences.A binary SVM classifier is used to determine the plausibility of the candidate pair.Two types of features are extracted by this classifier.The local lexical features are words surrounding the gene or the disease terms in the original text.The global syntactic features are unigrams,bigrams and trigrams drawn from 1)the shortest path between the gene and the disease terms in the dependency tree,and 2)the path between the least common ancestor of the two terms and the root of the dependency tree.Ten-fold cross validation of the model with 1000 positive and 1000 negative samples shows an Fl-score of 0.934.In the ranking phase,each of the positive pairs can be ranked by three methods.The basic method is by the co-occurrence frequency.The second method is to weigh each co-occurrence by the page rank of the paper from which the evidence was extracted,in a paper citation net-work constructed from PubMed.The last and most advanced method considers the duplicated evidence published by the same author,and thus suppresses the contribution of such evidence.Our evaluation of the 10 diseases shows that the MRR scores of the above three rankings are 0.249,0.281 and 0.293,respectively.In addition,if we consider a disease to gene association problem as an information retrieval problem,the Fl-score for the top 50 genes associated to a disease by the third ranking methods reaches 0.259,which are significantly higher than existing systems on similar tasks.更多还原