生物的生长发育、形态、对内外部环境的响应均受到生物内在的复杂生物网络控制。在基因层面上，基因调控网络反映了生物系统中各类元件在基因层面的相互作用关系。基因调控网络重构的研究对人们解析复杂性状的遗传构成和调控机制起到重要的作用，是系统生物学、生物信息学中极具挑战性的前沿课题。近年来，面对日益增长的粮食需求，使得从生物信息学角度出发，研究发现控制优质农艺性状的关键基因和关键性状的多基因遗传调控网络的计算方法成为该领域重要的研究课题。为此，本文以模式生物拟南芥、水稻为对象，以基因调控网络构建算法为主线，主要对候选基因识别、网络重构算法、数据方法等问题进行了深入研究。主要研究内容包括以下四个方面： 第一，研究了单一时序表达谱数据中差异表达基因的识别及排序方法。 当前生物数据的获取与筛选成本高昂，使用计算方法完成差异表达基因的识别与候选基因排序能有效地推动基因调控网络构建的研究。本文提出了一种基于平滑基因过滤器和样条曲线拟合的差异表达基因识别算法，并提出了一种基于伙伴评价原则的基因优先级排序策略用于候选基因的排序。首先，算法充分考虑了数据的时序特性，提出了基于Ljung-box检验的平滑基因过滤器滤除平滑基因；然后，提出了基于B样条拟合的检测器来识别统计上差异表达的候选基因；最后，提出了基于伙伴评价原则的优先级排序指标，使用共表达信息作为伙伴评价，对差异表达基因进行重新排序。新的排序能够反映特定过程或者条件下候选基因的生物学重要性。实验结果表明，本文提出的差异表达基因识别以及排序算法能够有效地识别单一时序表达数据中的差异表达基因，并按照基因的生物学重要性排序，为关键基因的发现以及调控网络的构建奠定基础。 第二，研究了基于贝叶斯网络模型构建基因调控网络的候选基因自动选择算法，并进一步对其改进提出了泛洪剪枝爬山法。 随着生物技术的快速发展，积累了大量的转录组数据，利用这些数据进行基因调控网络重构算法的研究成为生物信息学领域的热点与难点。基于贝叶斯网络模型构建基因调控网络因其内在的概率特性受到关注，但是目前的方法面临数据稀缺以及搜索空间复杂的问题，无法高效准确的学习网络结构。因此，本文提出了基于互信息和断点检测思想的候选基因自动选择算法CAS。CAS算法首先使用互信息度量节点之间的相关性，然后利用断点检测思想自动完成关联节点的识别，达到缩减搜索空间的目的。考虑到关联节点并不都是邻居节点，本文根据数据处理不等式原理提出了DPILevel的概念，通过对关联节点排序来区分邻居节点与非邻居节点，进一步缩减结构学习的搜索空间、降低邻居节点识别的假阳性。然后，基于DPILevel的概念提出了用于基因调控网络构建的泛洪剪枝爬山算法FPHC，加快了网络结构的学习速度。实验验证了上述算法的有效性。本文提出的算法能够高效的重构基因调控网络，为下一步进行网络结构分析识别关键基因、解析遗传构成提供生物信息学依据。 第三，提出了层次聚类引导的图格兰杰因果算法用于较大规模基因调控网络构建。 通过计算基因之间的因果关系快速构建较大规模的基因调控网络，对于理解生物网络结构特性以及识别候选关键节点具有重要意义。在基因间因果关系的发现上，基于格兰杰因果模型的方法受到了广泛关注。现有的基于格兰杰因果概念的方法假阳性率较高，而改进的图格兰杰因果方法不能有效的处理关联特征问题，基于生物先验知识进行分组的方法受到生物数据稀缺的限制应用场景十分有限。针对这些问题，本文提出了一种使用层次聚类改进图格兰杰因果模型的算法。首先，利用皮尔森相关系数和层次聚类原理将基因进行分组。然后采用分治的策略学习其调控因子。最后，将各分组结果进行二次预测，合并构建基因调控网络。与同类方法相比较，本文提出的方法不需要生物先验知识，获得了更准确的计算结果，能够为网络拓扑特性分析提供依据。 第四，在水稻多组学数据融合方面，研究了注释稀缺物种中组织特异蛋白质相互作用网络构建方法。 受限于注释数据较少，当前积累的大量水稻相关的组学数据不能很好地融合形成可利用先验知识。因此，亟需一种合理的融合多组学数据的生物信息学方法，为基因调控网络的构建提供可利用先验知识。组织特异的基因表达和蛋白质相互作用对研究基因调控、蛋白质功能、细胞过程有重要意义。本文提出了一种融合多组学数据构建组织特异蛋白质相互作用网络的方法。该方法首先确立了统一的评价标准和多表达数据集成方法进行组织特异基因识别；然后，提出了一种新的同源映射方法来构建目标物种蛋白质相互作用网络；最后，融合数据构建不同组织的蛋白质相互作用子网，并筛选高可靠的蛋白质相互作用。利用上述框架，本文构建了水稻组织特异的蛋白质相互作用网络，并对构建的网络进行了详细分析，验证了框架的有效性。预测的组织组织特异的蛋白质相互作用网络，作为预测的先验知识，能够为水稻高产性状相关关键基因发现以及多基因互作调控网络的构建提供帮助。 关键词：基因调控网络；差异表达基因；贝叶斯网络；图格兰杰因果；候选自动选择；组织特异蛋白质相互作用；

 The growth and development of organisms, morphology and response to internal and external environment are controlled by intrinsic complex biological networks. At the gene level, gene regulatory networks(GRNs) depict this complex mechanism of regulatory elements at the level of genes. Research on the reconstruction of GRNs plays an important role in the analysis of genetic structure and regulatory mechanism of complex traits, and is a very challenging frontier subject in system biology and bioinformatics. In recent years, in the face of growing demand for food, it has been an important research topic to develop methods for locating the key genes and for constructing GRNs of important agronomic traits related to yield in the field of bioinformatics. Therefore, this dissertation, taking the model organism Arabidopsis and Rice as the research object, making GRNs reconstruction algorithm as the main line, mainly focuses on the the calculation problems of GRNs, including candidate gene identification, reconstruction algorithms and data fusion problems. The main contributions of this dissertation mainly include the following four aspects: Firstly, a method for identification and ranking of differentially expressed genes in single time-series expression data is proposed. Due to the high cost of biological data acquisition and screening, identification and ranking of candidate genes can promote the research on the reconstruction of GRNs. Therefore, a differentially expressed gene identification algorithm based on flat gene filter and spline curve fitting is proposed, and then a new gene prioritization strategy is proposed for candidate gene ranking. Firstly, we design a flat gene filter based on the Ljung-box test, which utilizes the temporal properties of data, to filter out flat genes. Then, we put forward a B-spline model detector to identify differentially expressed genes in statistical sense. Finally, we propose a new gene prioritization strategy based on the principle of partner evaluation, which uses the coexpression information as the partner evaluation index to rerank the differentially expressed genes. The new rank can reflect the biological importance of genes in specific processes or conditions. Experimental results show that the proposed algorithm in this paper can effectively identify the differentially expressed genes in single time-series expression data and rank them according to biological significance, and then help to find key genes and to construct the regulation network in the following research. Secondly, a candidate gene automatic selection algorithm and an improved version using data processing inequality named flooding-pruning hill-climbing to construct the GRNs based on Bayesian network model are proposed. With the rapid development of biotechnology, a large number of transcriptome data has been accumulated. The reconstruction of GRNs using these data has become a hotspot in the field of bioinformatics. The reconstruction of GRNs based on Bayesian network model is concerned with its intrinsic probabilistic characteristics, but current methods still faces the problem of scarcity of data and complex search space, which cannot efficiently and accurately learn the network structure. Therefore, this paper proposes a candidate gene automatic selection algorithm (CAS). The CAS algorithm uses mutual information to measure the correlation between nodes, and then automatically identify the associated nodes by using the idea of breakpoint detection to reduce the search space. Considering that the correlated nodes are not all neighbor nodes, the concept of DPILevel is put forward according to the principle of data processing inequality. The algorithm sorts all the related nodes using DPILevel to distinguish the neighbor nodes from the indirectly related nodes. This will make further efforts to reduce the search space and false positive rate. Based on the concept of DPILevel, we propose an improved flood pruning algorithm (FPHC) to speed up the learning of the network structure. Experimental results prove the validity of the proposed algorithms. The proposed algorithm can effectively reconstruct the GRNs, which can provide bioinformatics basis for the identification of key genes and the analysis of genetic components. Thirdly, a hierarchical clustering guided graphical Granger causality algorithm for the construction of large-scale GRNs is proposed. It is of great significance to construct a large-scale GRNs by calculating the causal relationship between genes for understanding the structure characteristics of the biological network and the identification of the candidate key nodes. Reconstruction of GRNs based on Granger causality has been paid much attention in the aspect of discovering causality between genes. The existing method based on Granger causality has higher false positive rate, while the improved graphical Granger method cannot effectively deal with the problem of correlated feature, and grouping genes based on biological priori knowledge is limited due to scarcity of biological prior knowledge. To mitigate these problems, this paper proposes an improved graphical Granger causality method based on hierarchical clustering, which can be used on larger scale data. First, genes are grouped using hierarchical clustering method with Pearson correlation coefficient. Then the regulators are identified by the divide and conquer strategy. Finally, results are merged to construct the final GRN. Compared with concerned methods, the method proposed in this paper does not require biological prior knowledge, obtains more accurate calculation results, and can provide basis for analyzing network structure and further refining and controlling relation analysis. Finally, in the field of rice data fusion, the construction method of tissue specific protein interaction network for species with less biological annotation is studied. Limited by the lack of annotation data, the large amount of rice related omics data cannot be well integrated to form available prior knowledge. A reasonable bioinformatics method is urgently needed to integrate these multisource omics data, which can provides prior knowledge for the construction of GRNs. Tissue-specific gene expression and protein interaction are important for the study of gene regulation, protein function and cell process. Therefore, in this paper, a computing framework to predict tissue-specific protein interaction networks for the species with less annotation data by combining multiple omics data is proposed. Firstly, a unified evaluation standard and data integration method are established to identify tissue-specific genes. Then, a new Interolog mapping method is proposed to construct the target species protein interaction network. Finally, the protein interaction subnets of different tissues are constructed, and highly reliable protein interactions are extracted. Based on the above framework, we construct the first integrated network for tissue-specific protein interactions of rice. Then, we analyse this network in detail to validate the effectiveness of the framework. The proposed framework and the rice tissue-specific protein interaction network , as predicted prior knowledge, are helpful for the key gene discovery and multi-gene regulatory network construction of rice high-yield traits. Keywords: gene regulatory networks, differential expressed genes, Bayesian networks, graphical Granger causality, candidate auto-selection, tissue-specific protein interaction