【作者】 韩旭东；

【导师】 郭雪江；

【作者基本信息】 东南大学 ， 基础医学， 2024， 博士

【摘要】 背景与目的:单细胞测序技术的迅速发展,为在单细胞分辨率上研究细胞异质性和动态变化以及揭示生物学机制提供了前所未有的机遇。生物通路分析在理解生物学机制、疾病研究、药物开发、精准医学、数据整合和生物标志物发现等方面具有重要意义。然而,由于单个细胞的信使核糖核酸(Messenger ribonucleic acid,m RNA)拷贝数较低以及测序技术的限制,单细胞转录组测序数据通常存在“drop-out”现象。这种现象会导致单细胞测序数据具有高稀疏和高噪声的特性,为数据处理和分析带来了巨大挑战。传统的基因集富集分析方法(如GSEA、ss GSEA和GSVA)在处理单细胞数据时,它们的准确性和鲁棒性方面存在显著不足。为此,本研究开发了一个系统性的基于图神经网络的单细胞生物通路分析框架scapGNN,以填补当前在单细胞数据处理和生物通路分析方面的空白。研究方法:(1)为了确保单细胞测序数据分析的准确性和可靠性,本研究制定了一套标准化的数据预处理程序。首先,通过严格的质量控制和标准化步骤,去除低质量的细胞和低表达的基因,以减少噪声并消除技术变异。随后,构建细胞-细胞相关性网络和基因-基因相关性网络,以揭示细胞状态和基因表达模式间的复杂关系。最后,通过计算基于排秩的经验显著性水平,对关联网络进行修剪,提升数据的信噪比,保留具有生物学意义的显著关联。(2)为了将稀疏和不稳定的单细胞谱数据转换成稳定的基因-细胞关联网络,本研究提出了一个图神经网络(Graph neural network,GNN)框架。首先,基于转录调控状态(Transcriptional regulation states,TRS)正则化的深度神经网络自编码器(Deep neural network autoencoder,DNNAE)来学习细胞和基因的信息并提取具有代表性的低维嵌入特征。随后,基于细胞和基因节点特征及其对应的相关网络,图自编码器(graph autoencoder,GAE)对数据进行整合和重构,推断出稳定的有边权重的基因-细胞关联网络。(3)基于稳定的基因-细胞关联网络,本研究利用重启动随机游走(RWR)算法,来捕捉网络中的高阶和间接关系,揭示深层次的生物学关联信息。以先验的生物通路或基因集合的信息作为种子节点,RWR来量化每个节点与种子之间的空间距离。随后,扰动分析进一步推断单细胞通路活性分数和评估细胞表型关联的基因模块显著性。(4)对于单细胞多组学数据,本研究开发了一个基于边权重的网络融合方法。各组学数据通过GNN框架将谱数据转换成基因-细胞关联网络。对于这些网络,基于费舍尔联合概率检验的布朗扩展方法被用于计算边的联合权重。基于联合边权重的基因-细胞关联网络,多组学支持的单细胞通路活性分数和基因模块被推断来解析其在不同组学角度的生物学变异。(5)本研究收集了16个单细胞转录组测序(Single-cell RNA sequencing,sc RNA-seq)数据集、2个单细胞转座酶可及染色质测序(Single-cell assay for transposase-accessible chromatin sequencing,sc ATAC-seq)数据集以及2个单细胞多组学数据集,系统评估了scapGNN在细胞聚类、通路和基因模块识别精度指标以及鲁棒性等方面的性能。此外,我们将scapGNN应用于具体的生物学问题,包括精子发生、胚胎发育和2019冠状病毒(COVID-19)感染,以展示其在解决实际生物学问题和可视化方面的能力。成果和结论:本研究开发了一个基于图神经网络的单细胞数据分析平台scapGNN,创造性地将稀疏的单细胞谱数据转换为稳定的基因-细胞关联网络,用于推断单细胞通路活性分数,识别细胞表型关联的基因模块以及整合单细胞多组学数据。采用真实和模拟的单细胞多组学数据集来对scapGNN的性能进行系统的基准实验测试。结果表明scapGNN在各种下游单细胞分析中,如单细胞数据去噪声、消除批次效应、细胞聚类、细胞轨迹推断以及活性通路或基因模块识别等方面,比现有的最先进的方法更准确且具有更好的鲁棒性和扩展性。此外,scapGNN被应用于真实的生物学问题中,发现其计算的通路活性分数能够更好地重建精子发生和胚胎发育的细胞轨迹,精确地识别出细胞分化或发育过程相关的活性通路。细胞表型关联的基因模块能够揭示胞表型特异或与细胞状态转变相关的关键基因子网络。在COVID-19数据集中,scapGNN揭示了与COVID-19发生和发展密切相关的生物通路,并确定了COVID-19与IAV相比的特异性特征和潜在的治疗靶点。scapGNN被封装为一个综合的R软件包,能够高效、灵活地对单细胞多组学数据进行处理、分析和展示。总之,scapGNN不仅为单细胞多组学数据的分析提供了强有力的工具,还为生物学和临床医学的研究带来了新的视角和方法。通过深入应用scapGNN,研究人员能够更全面地解析细胞功能和疾病机制,从而推动生物医学领域的不断进步和发展。更多还原

【Abstract】 Background and Purpose:The rapid advancement of single-cell sequencing technology offers unprecedented opportunities to study cellular heterogeneity and dynamic changes at single-cell resolution,thereby revealing underlying biological mechanisms.Pathway analysis plays a crucial role in understanding biological mechanisms,disease research,drug development,precision medicine,data integration,and biomarker discovery.However,due to the low copy number of messenger ribonucleic acid(m RNA)in individual cells and the limitations of sequencing technologies,single-cell transcriptome sequencing data often exhibit a"drop-out"phenomenon.This phenomenon results in high sparsity and high noise characteristics in single-cell sequencing data,posing significant challenges for data processing and analysis.Traditional gene set enrichment analysis methods(such as GSEA,ss GSEA,and GSVA)show notable deficiencies in accuracy and robustness when dealing with single-cell data.To address these issues,this study has developed a systematic graph neural network-based framework for single-cell pathway analysis,named scapGNN,to fill the current gaps in single-cell data processing and pathway analysis.Research Methods:(1)To ensure the accuracy and reliability of single-cell sequencing data analysis,we have established a standardized data preprocessing protocol.Initially,we perform stringent quality control and normalization steps to remove low-quality cells and lowly expressed genes,thereby reducing noise and eliminating technical variations.Subsequently,we construct cell-cell and gene-gene correlation networks to uncover the complex relationships between cell states and gene expression patterns.Finally,we prune the correlation networks by calculating rank-based empirical significance levels,enhancing the signal-to-noise ratio and preserving biologically significant associations.(2)To transform sparse and unstable single-cell profile data into stable gene-cell association networks,we propose a graph neural network(GNN)framework.First,a deep neural network autoencoder(DNNAE)regularized by transcriptional regulation states(TRS)is used to learn and extract representative low-dimensional embedding features of cells and genes.Then,based on the features of cell and gene nodes and their corresponding correlation networks,a graph autoencoder(GAE)integrates and reconstructs the data,inferring a stable gene-cell association network with weighted edges.(3)Utilizing the stable gene-cell association network,we apply a restart random walk(RWR)algorithm to capture high-order and indirect relationships within the network,revealing deeper biological associations.Using prior information from biological pathways or gene sets as seed nodes,RWR quantifies the spatial distance between each node and the seeds.Perturbation analysis then further infers single-cell pathway activity scores and assesses the significance of gene modules associated with cell phenotypes.(4)For single-cell multi-omics data,we develop an edge-weight-based network fusion method.Each omics dataset is converted into gene-cell association networks using the GNN framework.For these networks,combined edge weights are calculated using Brown’s extension of Fisher’s combined probability test.Based on the combined edge weights of the gene-cell association network,multi-omics-supported single-cell pathway activity scores and gene modules are inferred to analyze biological variations from different omics perspectives.(5)We collected 16 single-cell RNA sequencing(sc RNA-seq)datasets,2 single-cell assay for transposase-accessible chromatin sequencing(sc ATAC-seq)datasets,and2 single-cell multi-omics datasets to systematically evaluate the performance of scapGNN in terms of cell clustering,pathway and gene module identification accuracy,and robustness.Additionally,we applied scapGNN to specific biological problems,including spermatogenesis,embryonic development,and COVID-19 infection,demonstrating its capability in addressing real biological questions and visualization.Results and Conclusions:This study developed a graph neural network-based platform,scapGNN,for single-cell data analysis,which innovatively converts sparse single-cell profile data into stable gene-cell association networks.These networks are used to infer single-cell pathway activity scores,identify gene modules associated with cell phenotypes,and integrate single-cell multi-omics data.We systematically benchmarked scapGNN’s performance using both real and simulated single-cell multi-omics datasets.The results demonstrate that scapGNN outperforms existing state-of-the-art methods in various downstream single-cell analyses,such as data denoising,batch effect elimination,cell clustering,trajectory inference,and pathway or gene module identification,exhibiting superior accuracy,robustness,and scalability.Moreover,scapGNN was applied to real biological problems,showing that its computed pathway activity scores can more accurately reconstruct cell trajectories in spermatogenesis and embryonic development,identifying active pathways related to cell differentiation or developmental processes.Gene modules associated with cell phenotypes revealed key gene sub-networks specific to cell phenotypes or related to cell state transitions.In COVID-19 datasets,scapGNN identified biological pathways closely related to the onset and progression of COVID-19,distinguishing specific features and potential therapeutic targets unique to COVID-19 compared to IAV.scapGNN is encapsulated in a comprehensive R package that efficiently and flexibly processes,analyzes,and visualizes single-cell multi-omics data.In summary,scapGNN not only provides a powerful tool for the analysis of single-cell multi-omics data but also introduces new perspectives and methods for biological and clinical research.By leveraging scapGNN,researchers can gain a more comprehensive understanding of cellular functions and disease mechanisms,thereby advancing the field of biomedical research.更多还原