【作者】 张燕；

【导师】 戴道清；

【作者基本信息】 中山大学 ， 信息计算科学， 2010， 硕士

【摘要】 随着人类基因组序列草图的完成,有关功能基因组的研究在生命科学领域中占据越来越重要的地位。阐明基因选择性表达所依赖的调控信息及其相互作用的分子机制,成为揭示生命现象本质的核心问题,是功能组研究的重要内容。随着基因组学研究的深入展开,基因的表达调控研究已经从单个基因、线性的调控拓展到立体层面上多基因、基因簇乃至整个基因组的调控网络。如何有效地利用现已有的基因组学数据,充分整合多学科的思路,建立新的试验系统和技术体系,阐明基因组表达的调控网络,分析基因之间的相互制约关系,已经成为功能基因组学领域内国际竞争的焦点。贝叶斯网络方法将概率理论知识与图论结合,其有图形化表示、因果关系清晰以及不确定性推理等优点,本文将贝叶斯网络引入到微阵列数据中并进行分析,从概率角度描述了各基因间的依赖关系,从而阐明了整个基因组之间的调控网络。本文首先阐述了贝叶斯网络的基本概念、发展历史和分类情况及其特点,以及贝叶斯网络构造的一些基本方法,详细阐述了贝叶斯网络的结构学习和参数学习的原理,然后分几种情况将结构学习与参数学习相结合来建立表示基因之间相互影响的贝叶斯网络模型。在实例分析中,本文详细描述了微阵列数据的贝叶斯网络模型构建的整个流程。本文将基因表达谱数据进行了三值离散化后构建贝叶斯网络,并分析了具有较多子结点的几个基因在网络中的作用与影响。本文主要进行了以下几个方面的研究:(1)在完整数据集中,使用贝叶斯网络结构学习中的K2算法进行结构学习研究,由于K2需要事先确定各结点间的的排序问题,此时采用决策树算法完成排序问题,从而提高了学习效率。并在网络学习过程中讨论了设置最大父结点的个数问题。通过结构学习最终获得了可反映基因之间调控关系的贝叶斯网络模型,在此基础上,使用极大似然方法进行参数估计,从而掌握基因间的后验概率。最后求出贝叶斯等价类,找到在原来的网络中可以反转的有向弧,从而可以在通过实验等手段获得先验知识后进行网络结构的相应调整,这样做不影响网络的结构,更具有实用性。(2)在有先验知识即已知网络结构的基础上,通过对数据随机分别设置1/3,1/4,1/5的缺失值,使用期望最大化算法进行参数学习,获得期望最大化算法对含缺失值数据的处理能力。(3)在无先验知识又含有缺失数据的的基础上,使用结构期望最大化算法进行了完整的学习得到了网络结构和相应的参数结构,并获得了结构期望最大化算法的处理能力。贝叶斯网络有着很好的理论知识和清楚的知识表达形式,是不确定性研究的一种重要方法,在数据挖掘中有着重要作用。将其引入微阵列实验数据的分析中,能较好的构建网络模型,分析各基因间的相互作用与影响,可广泛应用于生物学和肿瘤学的研究,观察疾病所引起的基因表达变化,并找出作用重要的变量基因。更多还原

【Abstract】 With the completion of human genome sequence sketches, functional genomics re-search plays more and more important part in the field of life science. Clarifying geneexpression regulation of selective on information and the molecular mechanism of in-teraction, and revealing the essence of life is the core problem and an important partof the study of functional groups. With the deepening of the research on genomics,gene expression studies in a single gene regulation has been extended to the regulation,linear three-dimensional level multiple genes and gene regulation of the whole genomeclusters. How to effectively use the existing data, fully integrate genomics of multi-disciplinary new ideas, establish a new test system and technical system, expound thegenome of the regulation network analysis and express the relationship between genesand functional genomics has become the focus of international competition within the?eld.Bayesian network method goes with the graph theory knowledge to probability,combining with graphical representation, causal relationship and uncertainty reasoningclearly, and it has good features, this paper will take bayesian network into microar-ray data analysis, it describes the relationship between genes with probability, whichillustrates the entire genome control network.This paper discusses the basic concept of bayesian network, development his-tory, classiffcation and its characteristics, and the bayesian network structure of somebasic methods, this paper expounds the bayesian network structure and parameters oflearning principle, and then puts forward the bayesian score study method, the struc- ture of the K2 algorithm and parameter study method to establish the mutual influencebetween genes in the bayesian network model. In the example, this paper carefully de-scribes a microarray data as a bayesian network model and the whole construction ofbayesian network. This gene expression data gets discretization after three values of thebayesian network construction. The paper constructs a bayesian network model andanalyses the function and influence between several nodes which have more children inthe network.This paper made several aspects of the research as follows:(1) For complete datas, we use the K2 algorithm in the bayesian network struc-ture learning. Because each node shoule be predetermined when using K2, we use De-cision Tree classic algorithm to complete sequencing problem and improve the learninge?ciency. The numbers of father nodes are discussed in the network learning process.At last by structure learning, we get a bayesian network model that can express therelationship between genes. Based on this, we use the Maximum Likelihood Estima-tion to estimate parameters, thereby mastering posteriori probabilities between genes.Finally we find the bayesian equivalence class, and we can reverse the arcs in the orig-inal network with a priori knowledge which can be obtained by means of experiments.Reversing the arcs does not affect the structure of networks, which is more practical.(2) If we have known a priori knowledge such as the network structure, we setrespectively random 1/3, 1/4, 1/5 loss value of the data, then use Expectation Maxi-mum algorithm for learning parameters to acquire the processing ability of ExpectationMaximum algorithm in datas containing missing values.(3) If we have no prior knowledge and the data contains loss values, we useStructural Expectation Maximum algorithm to complete the network structure learningand get the corresponding parameters, and then we obtain the handling ability ofStructural Expectation Maximum.Bayesian network has very good theoretical knowledge and clear knowledgeexpression of uncertainty, it is an important method in data mining, and plays an important role in datamine. Taking bayesian network into the microarray experimentdata analysis can better construct network models, analyze the interaction of in?uence,and it can be widely used in the study of biological and oncology to observe the diseasecaused by the change of gene expression.更多还原