【作者】 张琦；

【导师】 李瑛；

【作者基本信息】 吉林大学 ， 计算机应用技术， 2021， 硕士

【摘要】 自人类基因组测序完成以来,学者们发现除少量编码蛋白的基因序列外,基因组中大部分不参与蛋白质编码工作,仅转录成RNA,这些转录产物被称为非编码RNA(Non-coding RNAs,nc RNAs)。物种进化等级越高,非编码RNA占该基因组的比例越高(人类基因组高达98%)。越来越多的研究揭示非编码RNA在众多非常重要的生理和病理过程中起着关键的调控作用,与多种疾病密切相关,可作为疾病诊断和治疗的分子靶标。非编码RNA种类繁多,常见的家族种类包括mi RNA,pi RNA,sno RNA,sn RNA,si RNA和lnc RNA等。随着新一代测序技术的广泛使用,涌现了数量庞大的非编码RNA,其中绝大多数非编码RNA的功能和分子机理未知,从而在海量非编码RNA与匮乏的注释信息之间产生了巨大的数据鸿沟。基于生物实验的方法往往需要投入巨大的人力财力成本,周期长且难以应用于大规模数据分析中。因此亟需使用机器学习和信息技术来构建高效率、高准确度的非编码RNA功能分析模型。判定不同非编码RNA在序列特别是结构上的关系是理解和推断非编码RNA功能的最主要途径之一。已有的非编码RNA关系判定的算法主要都是基于非监督学习的,并且由于绝大多数的非编码RNA的序列保守性很低,为此如何提取和更好利用结构信息是非编码RNA的研究的挑战之一。在本论文中,我们提出了一种基于卷积神经网络和深度森林算法(多粒度级联森林,Gc Forest)的深度融合学习框架,整合多种序列结构比对特征的非编码RNA分类识别模型,称为GCFM(Gcforest fusion method)。相比于非监督学习算法,基于监督学习框架的GCFM能够更好的利用已知的非编码RNA家族信息,进而有助于深入挖掘非编码RNA间复杂抽象的内在关系。GCFM包含多视图结构表示模块和深度集成模块两部分:(i)多视图结构表示模块:提出了三类多视图表示方法,序列-结构比对编码表示、结构图像表示和局部结构形状比对编码表示。这些不同角度和层次的比对和结构特征表示,使得非编码RNA之间潜在的特异性能被GCFM捕获到。(ii)深度集成模块:提出了基于卷积神经网络和深度森林算法的深度集成模型,卷积模块用来学习更高级的特征表示。深度森林算法中的级联森林模块训练得到最终分类模型,每个级联层都是由XGboosting、Random Forest和Extra Trees三种算法构成。相较于其他深度学习架构,深度森林算法不需要调整大量的参数,具有更好的分类预测精度。与现有基于比对的非编码RNA分类方法相比,GCFM方法的F值提高了6%。并且通过全面系统的实验探究了多视图结构特征表示和深度集成架构的有效性,分析了GCFM方法相较于仅包含卷积模块方法的时间消耗情况。此外,为了进一步评估GCFM模型的有效性和可用性,我们设计了GCFM模型在三种非编码RNA任务中的应用实例:基于GCFM的非编码RNA聚类任务,GCFM推断非编码RNA系统进化树和GCFM预测RNA相互作用。在非编码RNA家族的聚类实验中,基于GCFM生成的分类矩阵,以行向量为特征应用多种聚类方法得到最终聚类结果。与现有的非编码RNA聚类方法(RNAclust,Ensembleclust和CNNclust)相比,基于GCFM的研究方法在包含未知非编码RNA家族的聚类研究中,准确率提高了20%。在应用GCFM构建了非编码RNA的系统进化树构建中,大部分非编码RNA在系统进化树中定位正确。在RNA相互作用的研究中,基于GCFM方法的预测准确率为90.63%。最后,为了最大化本研究方法的可用性,开发了GCFM的在线服务平台(http://bmbl.sdstate.edu/gcfm/),源代码和相关数据也可以在平台中获得。更多还原

【Abstract】 Since the completion of human genome sequencing,scholars have found that,except for a small number of protein-coding gene sequences,most of the genome does not participate in protein-coding work and is only transcribed into RNA.These transcripts are called non-coding RNAs(nc RNAs).Also found that the higher the evolutionary level of the species,the higher the proportion of non-coding RNA in the genome(up to 98% in the human genome).More and more studies have revealed that non-coding RNA plays a key regulatory role in many very important physiological and pathological processes,is closely related to a variety of diseases,and can be used as molecular targets for disease diagnosis and treatment.There are many types of non-coding RNAs,and common family types include mi RNA,pi RNA,sno RNA,sn RNA,si RNA,and lnc RNA.With the widespread use of the new generation of sequencing technology,a large number of non-coding RNAs have emerged,most of which have unknown functions and molecular mechanisms,resulting in a huge data gap between the vast amount of non-coding RNAs and the lack of annotated information.The methods based on biological experiments often require huge human and financial costs,and the cycle is long and difficult to be applied in large-scale data analysis.Therefore,it is urgent to use machine learning and information technology to build an efficient and accurate functional analysis model of non-coding RNA.It is one of the most important ways to understand and infer the function of non-coding RNAs to determine the relationship between different non-coding RNAs in sequence,especially in structure.Existing algorithms for determining the relationship between non-coding RNAs are mainly based on unsupervised learning,and since the sequence conserved of most non-coding RNAs is very low,how to extract and make better use of structural information of non-coding RNAs is one of the challenges in research.In this paper,based on the deep fusion learning framework of convolutional neural network and deep forest algorithm(GCForest),we propose a non-coding RNA classification and recognition model,which integrates multiple sequence-structure alignment features,called GCFM(GCforest Fusion Method).Compared with the unsupervised learning algorithm,GCFM based on the supervised learning framework can make better use of the known information of the non-coding RNA family,thus helping to dig deeper into the complex and abstract internal relationships among non-coding RNAs.GCFM consists of a multi-view structure representation module and deep integration module:(i)Multi-view structure representation module:three types of multi-view representation methods are proposed,including sequence-structure alignment coding representation,structure image representation,and local structure shape alignment coding representation.These different angles and levels of alignment and representation of structural characteristics enable potential specific properties between non-coding RNAs to be captured by GCFM.(ii)Deep integration module: a deep integration model based on convolutional neural network and deep forest algorithm is proposed.The convolution module is used to learn more advanced feature representations.The cascading forest module in the deep forest algorithm is used to train the final classification model,and each cascading layer is composed of XGBoosting,Random Forest,and Extratrees algorithms.Compared with other deep learning architectures,the deep forest algorithm does not need to adjust a large number of parameters and has better classification and prediction accuracy.Compared with the existing non-coding RNA classification methods based on the comparison,the F value of the GCFM method is improved by 6%.Besides,the effectiveness of multi-view structure feature representation and deep integration architecture is explored through comprehensive and systematic experiments,and the time consumption of the GCFM method compared to the method containing only convolution module is analyzed.Also,to further evaluate the validity and availability of the GCFM model,we designed examples of the GCFM model in three non-coding RNA tasks: GCFM-based non-coding RNA clustering task,GCFM inferred non-coding RNA phylogenetic tree,and GCFM predicted RNA interactions.In the clustering experiment of the non-coding RNA family,based on the classification matrix generated by GCFM,a variety of clustering methods were used to obtain the final clustering results with row vectors as the characteristics.Compared with the existing non-coding RNA clustering methods(RNAclust,Ensembleclust,and CNNclust),the accuracy of the GCFMbased method is improved by 20% in the clustering studies involving unknown non-coding RNA families.In the phylogenetic tree construction of non-coding RNAs constructed by GCFM,most of the non-coding RNAs were located correctly in the phylogenetic tree.In the study of RNA interaction,the prediction accuracy based on the GCFM method is 90.63%.Finally,to maximize the availability method of research,we developed GCFM online server(http://bmbl.sdstate.edu/gcfm/),source code,and related data is available in the server.更多还原