【作者】 赵明慧；

【导师】 李丹；

【作者基本信息】 大连理工大学 ， 控制科学与工程， 2022， 硕士

【摘要】 基因表达数据蕴含着丰富的基因活动信息,分析基因表达数据中隐含的模式对生物基因功能理解及推断、基因调控机制研究等具有重要意义。随着DNA微阵列技术的发展,产生了数量庞大的基因表达数据,如何对海量的基因表达数据进行有效分析及理解已成为生物信息领域的一项重要挑战。聚类是一种重要的无监督数据挖掘方法,通过基因聚类有助于发现共表达基因,进而推断未知基因的功能。在聚类过程中融入先验信息能够有效提升聚类效果,相比于类标签信息,先验信息中的成对约束更易获取,因此应用更为广泛。现有的基于成对约束的半监督算法或直接使用已知标签信息生成成对约束,或根据数据特性等挖掘成对约束。实际应用中,基因表达数据通常为无标签数据集,自动挖掘获得的成对约束不可避免地存在噪声约束,即与真实簇信息不符的成对约束,严重影响基因表达数据半监督聚类的性能。针对这一问题,本文提出了聚类和约束选择一体化求解的基因表达数据多目标半监督聚类算法。(1)本文从剔除噪声约束,选择有效的成对约束作用于半监督聚类过程的角度出发,提出了基于约束选择的基因表达数据多目标半监督聚类算法。算法首先根据密度跟踪法获取初始成对约束集,其次将其引入到具有约束违反惩罚项的目标函数中实现半监督聚类。为实现NSGA-II框架下聚类结果与约束选择的协同优化,提出了一种约束选取与聚类中心的混合编码,在多目标进化过程中选取适用于聚类的成对约束,实现了监督信息与聚类结果的联合优化,进而有效提升了基因表达数据的聚类效果。(2)将基因生物学知识融入到基因表达数据的半监督聚类过程中,有助于进一步剔除噪声约束。鉴于此,本文提出了融合基因本体的基因表达数据多目标半监督聚类算法。算法首先从基因本体知识库得到基因的功能相似度,并生成基因本体成对约束集,其次综合考虑基于基因表达数据和基因本体的成对约束信息,改进半监督聚类算法中的约束违反惩罚项权重,最后通过所提混合编码实现生物学知识指导下的约束选取与聚类中心的协同优化。多个基因表达数据集上的实验结果表明,所提算法能够通过融合生物学信息从初始约束集中进一步优选成对约束,得到了更为准确且生物学显著的聚类结果。更多还原

【Abstract】 Gene expression data contain abundant gene activity information,and it is of great significance to analyze the implied patterns in gene expression data for the understanding and inference of biological gene function and the study of gene regulation mechanism.With the development of DNA microarray technology,a large amount of gene expression data has been generated.How to effectively analyze and understand massive gene expression data has become an important challenge in the field of bioinformatics.Clustering is an important unsupervised data mining method that can help discover co-expressed genes and infer the function of unknown genes.Integrating prior information in clustering process can effectively improve clustering performance.Compared with class label information,pairwise constraints in prior information are easier to obtain,so they are more widely used.The existing semi-supervised algorithms based on pairwise constraints either directly use the known label information to generate pairwise constraints,or mine pairwise constraints according to data characteristics.In practice,gene expression data are usually unlabeled data sets,and pairwise constraints obtained by automatic mining inevitably have noisy pairwise constraints,that is,pairwise constraints that do not match the real cluster information,which seriously affects the performance of semisupervised clustering on gene expression data.To solve this problem,two multi-objective semisupervised clustering algorithms for gene expression data are proposed in this paper.(1)From the perspective of eliminating noisy constraints and selecting effective pairwise constraints to act on semi-supervised clustering,this paper proposes a multi-objective semisupervised clustering algorithm for gene expression data based on constraint selection.The algorithm firstly obtains the initial pairwise constraint set based on the density tracking method,and then introduces it into an objective function with constraint violation penalty term to achieve semi-supervised clustering.To achieve the collaborative optimization of clustering solution and constraint selection under NSGA-II framework,this paper proposes a hybrid coding method of constraints selection and clustering center.In the process of multi-objective evolution,pairwise constraints suitable for clustering are selected,which realizes the joint optimization of supervision information and clustering results,and then effectively improves clustering performance of gene expression data.(2)Integrating knowledge of gene biology into semi-supervised clustering for gene expression data is helpful to further eliminate noisy constraints.In view of this,this paper proposes a multi-objective semi-supervised clustering algorithm for gene expression data integrating gene ontology.Firstly,the algorithm obtains the functional similarity of genes from gene ontology and generates the constraint set of gene ontology.Secondly,considering pairwise constraints information based on gene expression data and gene ontology,the weight of constraint violation penalty term in semi-supervised clustering algorithm is improved.Finally,the proposed hybrid coding is used to realize the joint optimization of constraint selection and clustering center.Experimental results on multiple gene expression data sets show that the proposed algorithm can further optimize pairwise constraints from the initial constraint set by integrating biological information,and obtain more accurate and biologically significant clustering results.更多还原