【作者】 王文静；

【导师】 聂茹；

【作者基本信息】 中国矿业大学 ， 计算机技术（专业学位）， 2024， 硕士

【摘要】 环状RNA(circular RNA,circRNA)是一种具有共价闭环结构的内源性单链非编码RNA,稳定存在于生物体细胞核、细胞质和体液中。研究表明,circRNA与复杂疾病之间存在密切联系,如诱导血管生成、维持增殖信号和阻断凋亡途径等,可以将其作为癌症预后的高级生物标志物和治疗靶点,因此,识别潜在的circRNA与疾病之间的关联对生物医学的发展具有重要意义。然而,传统的生物学研究方法往往存在实验规模小、成本高昂且耗时较长等问题,因此,亟需建立更加高效合理的计算方法为该研究增添助力。本文基于图神经网络,分别从图分类和节点分类两个角度,对circRNA与疾病关联预测算法展开研究。针对传统计算模型难以充分捕捉节点集依赖关系的不足,本文从图分类的角度提出一种基于双半径节点标记和图神经网络的circRNA与疾病关联预测模型LMGATCDA。该模型首先整合circRNA与疾病的多源相似性信息,构建circRNA-疾病异构图,并提取目标关联的三跳封闭子图,从而将链路预测问题转化为图分类问题。然后,采用双半径节点标记方法获取子图拓扑特征,并将其与节点显性特征结合。最后,将整合特征馈入Graph SAGE和多跳注意力图神经网络以获取节点邻域中直接相连和非直接相连的邻居信息,进一步增强特征聚合。经过五折交叉验证,该方法在circR2Disease和Circ Atlas 2.0数据集上获得的平均AUC分别为94.25%和93.97%。通过案例研究分析,与结直肠癌和乳腺癌相关的前30个circRNA中,分别有23个和25个被Circ Atlas 2.0、Circ2Disease或circRNADisease数据库证实。针对LMGATCDA模型在表征子图拓扑结构特征时,其图池化过程存在信息丢失问题,本文从节点分类的角度提出一种基于线图神经网络的circRNA与疾病关联预测模型LGLPCDA。该模型首先将多源生物信息和局部拓扑信息整合为子图节点特征,并将链路两端节点特征处理为边缘特征。随后,将子图转化为线图形式,将原始链路特征转化为线图的节点特征,从而将链路预测问题转化为节点分类问题。最后,将线图中节点特征馈入简单谱图卷积网络中,以聚集k跳邻域的特征信息,获得更高阶的节点嵌入,从而完成对circRNA-disease未知关联的预测。在五折交叉验证实验中,该方法在circR2Disease和Circ Atlas 2.0数据集上平均AUC分别达到97.56%和97.49%。通过案例研究分析,与食管鳞状细胞癌和胃癌相关的前30个circRNA中,分别有24个和28个被Circ Atlas 2.0、Circ2Disease或circRNADisease数据库证实。更多还原

【Abstract】 Circular RNA(circRNA)is an endogenous single-stranded non-coding RNA with a covalent closed-loop structure that stably exists in the nucleus,cytoplasm,and bodily fluids of organisms.Studies have shown a close association between circRNA and complex diseases,such as inducing angiogenesis,maintaining proliferative signals,and blocking apoptotic pathways,making it a potential biomarker for cancer prognosis and a therapeutic target.Therefore,identifying the potential association between circRNA and diseases is of great significance to the development of biomedicine.However,traditional biological research methods often suffer from small experimental scales,high costs,and time-consuming processes.Therefore,it is urgently needed to establish more efficient and reasonable computational methods to provide assistance for this research.Based on graph neural networks,this thesis conducts a study on the prediction algorithm of the association between circRNA and diseases from two perspectives:graph classification and node classification.Given the limitations of traditional computational models in adequately capturing node set dependencies,this thesis proposes a novel prediction model for circRNAdisease associations,named LMGATCDA,from the perspective of graph classification,based on double-radius node labeling and graph neural networks.Initially,the model integrates multi-source similarity information of circRNAs and diseases to construct a circRNA-disease heterogeneous graph.It then extracts three-hop closed subgraphs of target associations,thus transforming the link prediction problem into a graph classification problem.The model employs the double-radius node labeling method to capture subgraph topological features,which are combined with explicit node features.These integrated features are subsequently fed into Graph SAGE and multi-hop attention graph neural networks to gather information from both directly and indirectly connected neighbors,enhancing feature aggregation.The model achieves average AUC scores of 94.25% and 93.97% on the circR2 Disease and Circ Atlas 2.0 datasets,respectively,as validated by five-fold cross-validation.Case studies indicate that among the top 30 circRNAs associated with colorectal cancer and breast cancer,23 and 25 were confirmed by the Circ Atlas 2.0,Circ2 Disease,or circRNADisease databases,respectively.Addressing the information loss problem during the graph pooling process in the LMGATCDA when representing subgraph topological features,this thesis also presents another prediction model for circRNA-disease associations,named LGLPCDA,from the perspective of node classification,based on line graph neural networks.This model integrates multi-source biological information and local topological information as subgraph node features,processing the features of the nodes at both ends of the link as edge features.Subsequently,the subgraph is transformed into a line graph,converting the original link features into node features of the line graph,thereby transforming the link prediction problem into a node classification problem.Finally,the node features in the line graph are fed into a simple spectral graph convolutional network to aggregate features from k-hop neighborhoods,resulting in higher-order node embeddings and completing the prediction of unknown circRNA-disease associations.In five-fold cross-validation experiments,this method achieves average AUC scores of 97.56% and97.49% on the circR2 Disease and Circ Atlas 2.0 datasets,respectively.Case studies reveal that among the top 30 circRNAs associated with esophageal squamous cell carcinoma and gastric cancer,24 and 28 were confirmed by the Circ Atlas 2.0,Circ2 Disease,or circRNADisease databases,respectively.更多还原