【作者】 刘涛；

【导师】 甘杨兰； 芮欣凯；

【作者基本信息】 东华大学 ， 电子信息（专业学位）， 2025， 硕士

【摘要】 基因表达调控是对生物体内的基因表达过程的调节,它是调节细胞代谢、控制细胞分化及实现细胞环境适应性应答等生物学过程的核心机制。通过推断转录因子与靶基因间的调控关系并重构基因表达调控网络,可以深入了解生物体的分子机制,揭示细胞生长分化、发现疾病的关键致病基因和潜在的临床药物靶点。随着单细胞测序技术的快速发展,研究者可用更高的分辨率观察细胞内的基因表达水平,为解析细胞动态变化和重构基因调控网络等提供了新的机遇。近年来,各种不同的方法被陆续提出用于从单细胞RNA测序数据集中推断基因调控网络。然而,现有的方法仍然存在一些局限性。首先,大多数基因调控推断方法仅关注基因之间的成对调控关系,未能充分关注基因调控网络的长程调控关系;同时,现有方法普遍忽视了基因对之间的因果关联分析,而这对于探索转录因子与其靶基因之间的复杂调控相互作用至关重要。为解决上述问题,本文的主要研究工作概括如下:首先,针对现有算法对基因调控网络拓扑信息利用不足和基因对复杂调控模式捕获能力欠缺的问题,本文提出了一种新的有监督学习框架GTGRNS,通过联合图神经网络和Transformer模型从sc RNA-seq数据推断GRN。GTGRNS算法通过Graph SAGE算法捕获网络局部和全局结构信息,将调控网络的拓扑信息与基因表达特征相结合,从而学习基因的上下文信息并生成初始基因嵌入;然后利用Transformer模型学习基因间的深层内在相互作用,引入边位置编码以丰富节点的语义特征,进一步提取基因对的复杂依赖关系从而构建基因特征矩阵;最后由多层感知器模型预测基因对之间存在调控关系的概率,从而构建基因调控网络。在七个基准数据集上的对比实验表明,GTGRNS在AUROC和AUPRC指标上分别平均提升了3.9%和3.6%,充分展示了GRN推断任务中具有显著的有效性和竞争力。并且在真实的数据集上GTGRNS也能有效预测潜在的基因相互作用,尤其是某些关键枢纽基因或调控因子。其次,针对现有基于图的GRN推断算法在学习长距离调控关系和利用基因对因果关系的局限性,本文提出了一种融合因果关系和图神经网络的基因调控网络推断算法GCGRNS。首先利用自注意力机制来学习基因调控的上下文依赖关系以生成基因节点嵌入,通过引入全局空间编码机制来进一步捕获基因间的长距离调控关系;接着利用结构因果模型显式地学习基因间潜在的因果关系并将其融合进节点嵌入中;使用解码器重构初始的基因表达值以优化基因节点嵌入的质量。最后通过全连接层来预测基因对的调控概率,构建最终调控网络。本文将GCGRNS应用于七种细胞类型特异性网络的sc RNA-seq测序数据集,并与七种先进的有监督GRN推断算法进行了比较,与性能次优的方法相比,GCGRNS在AUROC和AUPRC指标上分别实现了平均1.37%和1.15%的提升。实验结果表明GCGRNS能够有效推断出基因对的作用关系,同时,消融实验进一步验证了其关键子模块在模型中的重要作用。更多还原

【Abstract】 Gene expression regulation refers to the control of gene expression processes within organisms.It serves as a fundamental mechanism underlying various biological processes,including the regulation of cellular metabolism,the control of cell differentiation,and the adaptive responses to environmental stimuli.By inferring the regulatory relationships between transcription factors and their target genes and reconstructing gene regulatory networks(GRNs),researchers can gain deeper insights into the molecular mechanisms of living organisms.Such efforts are crucial for understanding cellular growth and differentiation,identifying key pathogenic genes associated with diseases,and discovering potential clinical drug targets.With the rapid development of high-throughput sequencing technologies,single-cell RNA sequencing(sc RNA-seq)data enables researchers to observe gene expression levels within cells at a finer resolution,providing new opportunities for interpreting dynamic cell differentiation and reconstructing gene regulatory networks.In recent years,various methods have been proposed for inferring gene regulatory networks from single-cell RNA sequencing datasets.Although most of these models have achieved promising results,the inherent high-dimensionality and noise characteristics of sc RNA-seq data can affect the performance of unsupervised learning models on large-scale datasets.As for existing supervised learning models,the majority primarily focus on pairwise regulatory relationships between genes without fully leveraging the global regulatory structure of GRNs,moreover,they often overlook the causal relationships between gene pairs,which are crucial for uncovering the complex regulatory interactions between transcription factors and their target genes.To address these issues,the main research contributions of this paper are summarized as follows:Firstly,to address the limitations of existing algorithms in effectively mining gene regulatory structures and capturing complex regulatory patterns between gene pairs,this paper proposes a novel supervised learning framework,GTGRNS,which infers GRNs from sc RNA-seq data by jointly integrating graph neural networks and the Transformer model.The GTGRNS algorithm primarily consists of three modules:the gene node representation learning module,the gene pair interaction learning module,and the regulatory relationship prediction module.Initially,the Graph SAGE algorithm is employed to capture both local and global structural information of the network,integrating the topological information of the regulatory network with gene expression features to embed contextual information of genes and generate initial gene embeddings.Subsequently,the Transformer model is utilized to learn deep intrinsic interactions between genes,while edge positional encoding is introduced to enrich the semantic features of nodes,further extracting complex dependencies between gene pairs to construct the gene feature matrix.Finally,a multilayer perceptron(MLP)model is used to predict the probabilities of regulatory relationships between gene pairs,thereby constructing the gene regulatory network.Comparative experiments conducted on seven benchmark datasets show that GTGRNS achieves average improvements of 3.9%in AUROC and 3.6%in AUPRC,demonstrating its notable effectiveness and strong competitiveness in gene regulatory network(GRN)inference tasks.Moreover,GTGRNS can effectively predict potential gene interactions on real datasets,particularly for certain key hub genes or regulatory factors.Furthermore,to address the limitations of existing graph-based gene regulatory network(GRN)inference algorithms—particularly their challenges in capturing long-distance regulatory dependencies and incorporating causal relationships between gene pairs—we propose a novel GRN inference framework named GCGRNS,which integrates causal inference with graph neural networks.This framework comprises three main components:an encoder module,a decoder module,and a gene interaction prediction module.First,a self-attention mechanism is employed to capture the contextual dependencies in gene regulation,thereby generating gene node embeddings.Additionally,a global spatial encoding mechanism is introduced to further capture long-range interactions and global regulatory patterns among genes.Then,structural causal models are leveraged to explicitly learn the underlying causal relationships between genes,which are integrated into the node embeddings.The decoder subsequently reconstructs the original gene expression values,aiming to enhance the quality of the learned embeddings.Finally,a fully connected layer is used to predict the regulatory probability between gene pairs,thereby constructing the final regulatory network.In this study,GCGRNS was applied to sc RNA-seq datasets derived from seven cell type-specific gene regulatory networks and compared against seven state-of-the-art supervised GRN inference algorithms.Compared to the second-best method,GCGRNS achieved average improvements of 1.37%in AUROC and 1.15%in AUPRC.The experimental results demonstrate that GCGRNS can effectively infer regulatory relationships between gene pairs,and ablation studies further confirm the critical contributions of its key submodules to overall performance.更多还原