【作者】 李辉；

【导师】 陈奎生； 朱真峰；

【作者基本信息】 郑州大学 ， 软件工程， 2023， 博士

【摘要】 生物分子(即实体)在复杂生物系统中起着至关重要的作用,各种生命过程离不开它们之间的协调配合。从复杂网络的角度研究生物分子的性质和功能是当前网络科学和生物信息学的重要研究方向,特别是识别生物实体关联网络中不同生物实体之间的关联关系是一个极具价值的研究热点。它不仅能增进研究者对生命系统运作方式的理解,还有助于发现药物靶点和治疗策略,促进精准医学的发展。图神经网络因其同时使用图中的拓扑信息和节点的属性信息,能有效建模图结构数据,在生物实体关联预测问题中愈发流行。然而,生物实体关联数据具有数据稀疏性、协作互补性、网络结构性和交互复杂性的特点,为基于图神经网络的方法带来了问题和挑战:如何有效利用实体节点的多视角属性和增强的拓扑信息,以缓解数据稀疏性问题;如何充分使用生物实体之间的互作和协作信息、生物网络中的领域先验知识、细粒度的潜在交互因素,以通过增强节点的表达能力来改进预测性能。针对上述挑战,本文提出相应的解决方案,并将它们应用到四类重要的生物实体关联预测问题中,主要研究成果如下:1)针对现有方法难以有效利用节点的多视角属性和增强的拓扑信息问题,提出注意力级联的多视角图神经网络。首先,提取节点的多视角特征并构建多属性的相互作用图。然后,设计一种级联注意力机制,包括视角级注意力、节点级注意力和层级注意力。其中,视角级注意力自适应融合节点的多视角属性信息,节点级注意力迭代聚合节点的邻域信息,而层级注意力组合不同阶的邻域信息。最后,学到的节点表示融合了节点的多视角属性信息和增强的拓扑信息。实验表明,在lncRNA-miRNA相互作用预测问题的四个基准数据集上,与主流的方法相比,该方法取得较好的预测效果,可以有效缓解数据稀疏性问题。2)针对现有方法不能同时建模实体之间的互作行为和协作行为的问题,提出跨域对比的图神经网络。首先,分别基于已知的关联和节点特征构建互作图和协作图。其次,设计一种多域编码器,由一个互作域编码器和两个协作域编码器组成。它们分别从互作域和协作域中捕获互作信息和协作信息。然后,使用域自适应融合获得节点的最终表示。同时利用跨域对比学习使得每个域学习到与预测任务相关的节点表示,并确保不同域之间的节点表示可以相互补充。最后,学到的节点表示融合了同类型节点之间的协作信息和不同类型节点之间的互作信息。实验表明,在lncRNA-蛋白质相互作用预测问题的四个基准数据集上,该方法优于其他主流方法。3)针对现有方法难以充分整合领域先验知识的问题,提出元知识增强的轻量图神经网络。首先,构建知识图谱并提取节点表示,其提供全局视角的知识。然后,设计轻量图神经网络捕获相互作用图中的拓扑信息,其提供局部视角的知识。最后,引入个性化知识迁移,从全局视角的知识中提炼出元知识,并将其迁移到局部视角的知识中,以获得高质量的节点表示。在药物-蛋白质相互作用预测问题上的实验结果表明,该方法可以充分整合领域先验知识,提高节点的表示能力,从而获得更好的预测性能。4)针对现有方法不能细粒度建模隐藏在关联形成背后的复杂机制问题,提出多因素解纠缠的图神经网络。首先,随机初始化节点的多因素表示。然后,基于多因素表示来构建因素感知图,并独立地迭代聚合每个因素感知图上的邻域信息,以学习因素感知的节点表示。最后,使用独立性约束增强不同因素的节点表示之间的独立性。在ncRNA-药物耐药性关联预测问题上的实验结果表明,该方法明显优于主流的对比方法。同时,这也说明了显式建模导致关联形成的潜在多因素,有助于增强节点表示,并有益于下游预测任务。综上所述,本文以生物实体关联网络为研究对象,提出的四个图神经网络具有一定的理论价值,并在不同的生物实体关联预测问题上取得很好的应用效果。此外,泛化性实验表明,它们具有一定的泛化能力,同样适用于其他相关的研究领域。更多还原

【Abstract】 Biomolecules(i.e.,entities)are essential in complex biological systems,where life processes rely on their coordinated actions.Examining the properties and roles of biomolecules through the complex network analysis represents a significant research avenue in network science and bioinformatics.In particular,uncovering the relationships among different biological entities within biological association networks is a valuable area of focus.It not only deepens our comprehension of life system operations,but also aids in identifying drug targets and therapeutic strategies,thereby advancing precision medicine.Graph neural networks are adept at modeling graph-structured data,leveraging both node attributes and topological information.This capability renders them particularly well-suited for predicting associations among biological entities.However,biological entity association data is characterized by data sparsity,collaborative complementarity,network structure and interaction complexity,which brings problems and challenges for graph neural network-based methods:How can multi-view attributes and enhanced topological information be effectively harnessed to mitigate data sparsity?How can interactive and collaborative information among biological entities,domain prior knowledge in biological networks,and fine-grained latent interaction factors,be leveraged to improve prediction performance by enhancing node expressiveness?To address the above challenges,this thesis proposes corresponding solutions and applies them to four important classes of biological entity association prediction problems.The main contributions are summarized as follows:1)To address the problem that existing methods fail to effectively utilize the multi-view attributes of nodes and enhanced topological information,we propose a multi-view graph neural network with cascaded attention.First,we extract the multi-view features of nodes and construct a multi-attributed interaction graph.Then,we design a cascaded attention mechanism,which includes view-level attention,node-level attention,and layer-level attention.Among them,view-level attention adaptively integrates multi-view attribute information of nodes,node-level attention iteratively aggregates neighborhood information of nodes,and layer-level attention combines neighborhood information of different orders.Finally,the learned node representations fuses both multi-view attribute information and enhanced topological information of nodes.Experiments show that on the four benchmark datasets of lncRNA-miRNA interaction prediction problems,this method achieves the best prediction performance compared to the state-of-the-art methods and can effectively alleviate the data sparsity problem.2)To address the problem that existing methods are unable to simultaneously model both interactive and collaborative behaviors among entities,we propose a crossdomain contrastive graph neural network.First,interactive and collaborative graphs are constructed based on known associations and node features,respectively.Next,a multi-domain encoder is designed,comprising one interactive domain encoder and two collaborative domain encoders,which capture interactive and collaborative information from their respective domains.Then,domain-adaptive fusion is employed to obtain the final representations of nodes.Cross-domain contrastive learning is also utilized to enable each domain to learn node representations relevant to the prediction task while ensuring that node representations from different domains can complement each other.Finally,the learned node representations integrate collaborative information among nodes of the same type and interactive information between nodes of different types.Experiments demonstrate that the method outperforms other state-of-theart methods on four benchmark datasets for the lncRNA-protein interaction prediction problem.3)To address the problem of existing methods in fully integrating domain prior knowledge,we propose meta-knowledge-enhanced light graph neural network.First,a knowledge graph is constructed,and node representations are extracted to provide knowledge from a global perspective.Then,a light graph neural network is designed to capture topological information from the interaction graph,thereby providing local perspective knowledge.Finally,personalized knowledge transfer is introduced to extract meta-knowledge from the global perspective knowledge and transfer it to the local perspective knowledge,resulting in high-quality node representations.Experimental results on the drug-protein interaction prediction problem show that the method can fully integrate domain prior knowledge to enhance the representation power of nodes,thereby achieving better prediction performance.4)To address the problem that existing methods are unable to model the complex mechanisms underlying association formation at a fine-grained level,we propose a disentangled multi-factor graph neural network.First,the multi-factor representations of nodes are randomly initialized.Then,factor-aware graphs are constructed based on these multi-factor representations,and neighbor information on each factor-aware graph is aggregated independently and iteratively to learn factor-aware node representations.Finally,an independence constraint is imposed to enhance the independence between node representations of different factors.Experimental results on the ncRNAdrug resistance association prediction problem show that this method is significantly superior to the state-of-the-art methods.Moreover,these results indicate that explicitly modeling the underlying multiple factors contributing to association formation enhances the representation power of nodes and benefits the downstream prediction task.In summary,this thesis focuses on biological entity association networks as the research object.Our proposed four graph neural networks possess the significant theoretical value and have achieved favorable application outcomes in various biological entity association prediction problems.Furthermore,generalization analysis indicates that they exhibit a certain degree of generalizability and are applicable to other related research fields.更多还原