【作者】 刘坤；

【导师】 潘毅； 王建新；

【作者基本信息】 中南大学 ， 计算机应用技术， 2022， 硕士

【摘要】 拓扑关联域是一种自相互作用强的局部染色质区域,具有复杂的三维空间结构。研究表明拓扑关联域边界与一些功能性调控元件密切相关,还与遗传疾病和癌症有一定的联系,因此拓扑关联域检测对于理解染色质空间结构、基因表达、基因调控具有十分重要的意义。尽管许多计算方法被广泛地应用到拓扑关联域检测领域,然而它们在高维、高噪声、稀疏的Hi-C数据中检测拓扑关联域依然有一定的不足,比如丢失远距离的信息、对拓扑关联域的数量和尺寸进行不同的假设、未对Hi-C矩阵进行降噪处理等。因此,本文基于拓扑关联域及其边界的特征提出了不同的计算方法在Hi-C数据上检测拓扑关联域。本文的主要贡献如下:第一,针对目前拓扑关联域检测方法的存在的调控元件富集水平不够高的问题,本文提出了一种基于深度学习的拓扑关联域检测模型YOLOTAD。YOLOTAD模型将拓扑关联域检测任务转化成目标检测任务,使用在模拟数据上完成训练的模型识别拓扑关联域。该模型先使用CSPDarknet53网络和空间金字塔池化网络等网络学习不同层次的拓扑关联域特征和语义信息,再使用特征金字塔网络和路径聚合网络将浅层网络提取的特征和深层网络提取的语义信息进行不同层次的多尺度特征融合,提高预测拓扑关联域的准确性。在Hi-C实验数据上,YOLOTAD模型识别的拓扑关联域边界附近调控元件的富集水平整体要高于其它比较的模型,具有一定的适用性。以拓扑关联域整体作为分析目标时,YOLOTAD模型的拓扑关联域再现性、具有显著DCC指标的拓扑关联域比例、模拟数据上拓扑关联域的准确性超过了大部分方法。结果表明YOLOTAD模型能准确地学习到拓扑关联域特征并应用于检测拓扑关联域。第二,针对Hi-C矩阵高维、稀疏、高噪声等特点影响拓扑关联域检测准确性问题,本文提出了一种基于对称非负矩阵分解的拓扑关联域检测模型SNMFTAD。该模型先使用网络增强技术对Hi-C数据进行降噪处理,结合对称非负矩阵分解有效地从Hi-C数据提取到低维图嵌入,能够学习到准确的节点相似特征并应用于拓扑关联域的检测。在Hi-C实验数据上,SNMFTAD模型识别的拓扑关联域边界附近调控元件的富集水平整体要高于YOLOTAD模型和其它比较的模型,且适用性高于YOLOTAD模型。在模拟数据上,噪声水平较低时,SNMFTAD模型的性能整体上能够超越其它模型,噪声水平较高时,SNMFTAD模型的性能处于领先地位。实验结果表明SNMFTAD模型的适用性较强,能够有效地应用于拓扑关联域的检测。图30幅,表22个,参考文献86篇更多还原

【Abstract】 Topologically associating domains are local chromatin regions with strong self-interaction and complex spatial structure.Studies have shown that topologically associating domain boundaries are closely related to some functional regulatory elements,and also have links to genetic diseases and cancers.Therefore,detecting topologically associating domains is very important for understanding chromatin spatial structure,gene expression,and gene regulation.Although many computational approaches are widely used to detect topologically associating domains,they still have shortcomings when detecting topologically associating domains on highdimensional,sparse,and noisy Hi-C data.The shortcomings of these methods include the loss of information between long distances,different assumptions about the number and size of topologically associating domains,and failure to perform noise reduction on Hi-C matrices.Therefore,according to the features of topologically associating domains and their boundaries,different computational methods are proposed to detect topologically associating domains on Hi-C data in this thesis.The main contributions of this thesis are as follows.First,to address the problem that enrichment of regulatory elements near the topologically associated domain boundaries identified by current methods is not high enough,a deep learning-based model called YOLOTAD is proposed.Formulating the task of topologically associating domain detection as the task of object detection,YOLOTAD detects topologically associating domains using a model trained on simulated data.YOLOTAD uses networks like CSPDarknet53 network and spatial pyramid pooling network to learn features and semantic information of topologically associating domains at different levels.To improve the accuracy in detecting topologically associating domains,YOLOTAD then uses feature pyramid networks and path aggregation networks to fuse multi-scale features and semantic information in different levels from shallow and deep networks.On the Hi-C experimental data,the enrichment of regulatory elements around topologically associating domain boundaries identified by YOLOTAD is higher than that of other methods overall.It shows YOLOTAD has a degree of applicability.When analyzing topologically associating domains as a whole,YOLOTAD exceeds most methods in terms of the reproducibility of topologically associating domains,the proportion of topologically associating domains with significant DCC,and the accuracy of topologically associating domains on the simulated data.The results show that YOLOTAD can accurately learn topologically associating domain features and apply them to detect topologically associating domains.Second,in terms of high-dimensional,sparse and noisy Hi-C matrices affect the accuracy of topologically associating domains,a model called SNMFTAD is proposed.SNMFTAD detects topologically associating domains based on symmetric non-negative matrix decomposition.SNMFTAD first uses network enhancement technique to reduce noise for Hi-C data.SNMFTAD then combines symmetric non-negative matrix decomposition to effectively extract low-dimensional graph embeddings from Hi-C data.SNMFTAD can learn accurate similarity features between nodes,and apply them to detect topologically associating domains.On the Hi-C experimental data,the enrichment of regulatory elements around topologically associating domain boundaries identified by SNMFTAD is higher than that of YOLOTAD and other methods overall.It demonstrates SNMFTAD has a higher applicability than YOLOTAD.On the simulated data,SNMFTAD outperforms other methods overall when the noise levels are low,and is in the lead when the noise levels are high.The results show that SNMFTAD is more applicable and can effectively detect topologically associating domains.更多还原