【作者】 陈智超；

【导师】 李杰；

【作者基本信息】 哈尔滨工业大学 ， 计算机技术（专业学位）， 2023， 硕士

【摘要】 传统的新药研发过程存在着投入成本高、耗时长的缺点。为了缩短研发周期、降低研发成本,研究人员尝试利用计算的方法发现已有药物的新的适应症。药物-疾病关系预测作为一种发现已有药物新的适应症的计算方法,受到广泛关注。随着计算机技术与人工智能技术的快速发展,研究人员开始利用深度神经网络来预测药物-疾病的关系,来发现已有药物的新的适应症。已有研究表明,基于图神经网络的药物-疾病关系预测方法在药物-疾病关系预测中取得了较好的预测效果,但是现有的基于图卷积神经网络的预测方法在提取异构图的特征表示时存在特征损失问题,另外现有方法的预测精度仍需进一步提高,因此需要发展新的药物-疾病关系预测方法来解决以上问题。为了解决以上问题,本文提出了基于图神经网络的药物-疾病关系预测方法(Graph Neural Network Drug-Disease Association prediction,GNDDA)。在GNDDA方法中,本文针对图卷积神经网络在提取药物-疾病异构图的节点特征表示时出现的特征损失问题,提出了异构特征表示提取方法(Heterogeneous Feature Representation Extraction,HFRE),提升了预测性能。其次,为了进一步提升预测的准确率,本文提出了基于注意力机制的多种特征表示聚合方法。随后本文进行了消融实验,实验结果表明两种方法均能提升预测性能。GNDDA方法的具体流程为:1)融合多种生物组学数据构建异构网络并计算药物的分子结构相似性与疾病语义相似性,得到药物与疾病的初始特征表示;2)使用异构特征表示提取方法完成药物节点、疾病节点和蛋白质节点的属性特征与结构特征的融合,同时令每个节点完成对不同类型的邻居特征学习,得到药物与疾病的异构特征表示;3)使用图卷积神经网络提取药物-蛋白质-疾病异构图特征,得到基于图卷积神经网络的药物与疾病的特征表示;4)将异构网络进行分解后使用异构特征提取方法提取子网特征,得到基于子网信息的药物与疾病的特征表示;5)对药物及疾病的四类特征表示进行线性转换,并使用多种特征表示聚合方法进行聚合得到药物与疾病的最终特征表示;6)将药物与疾病的特征表示输入线性内积解码器重构药物-疾病关系预测矩阵,实现药物-疾病的关系预测。最后本文通过对比实验来验证GNDDA的预测性能,对比实验结果表明GNDDA在AUC、AUPR指标上优于现有的药物-疾病预测方法。为了帮助研究人员快速发现潜在的药物-疾病关系,降低发现已有药物的新的适应症的成本,提高发现的效率,本文开发了一种在线的药物-疾病关系查询与预测系统。为了开发该系统,我们首先对药物与疾病的数据进行了整理,构建了包含多种药物信息的数据库,然后实现了在线药物-疾病关系预测和检索的算法及药物-疾病关系图形化展示等功能,方便研究人员开展后续研究工作。更多还原

【Abstract】 The traditional process of developing new drugs has the drawbacks of high investment cost and long time consumption.In order to shorten the research and development cycle and reduce research and development costs,researchers have attempted to use computational methods to discover new indications for existing drugs.Drug disease relationship prediction,as a computational method for discovering new indications for existing drugs,has received widespread attention.With the rapid development of computer technology and artificial intelligence technology,researchers have begun to use deep neural networks to predict drug disease relationships and discover new indications for existing drugs.The existing research shows that the drug disease relationship prediction method based on graph neural network has achieved good prediction effect in the drug disease relationship prediction,but the existing prediction method based on graph Convolutional neural network has the problem of feature loss when extracting the feature representation of heterogeneous graphs.In addition,the prediction accuracy of the existing methods still needs to be further improved,so it is necessary to develop new drug disease relationship prediction methods to solve the above problems.To address the above issues,this article proposes a Graph Neural Network Drug-Disease Association prediction method(GNDDA)based on graph neural networks.In the GNDDA method,this paper proposes the Heterogeneous Feature Representation Extraction(HFRE)to solve the problem of feature loss when graph Convolutional neural network extracts the node feature representation of drug disease heterogeneous graph.Secondly,in order to further improve the accuracy of prediction,this paper proposes multiple feature representation aggregation methods based on attention mechanism.Subsequently,this article conducted ablation experiments,and the experimental results showed that both methods can improve prediction performance.The specific process of GNDDA method is as follows: 1)building heterogeneous networks by fusing multiple biomics data and calculating the similarity of drug molecular structure and disease semantics to obtain the initial feature representation of drugs and diseases;2)Using heterogeneous feature representation extraction methods to fuse the attribute and structural features of drug nodes,disease nodes,and protein nodes,while enabling each node to learn different types of neighbor features and obtain heterogeneous feature representations of drugs and diseases;3)The graph Convolutional neural network is used to extract the features of drug protein disease isomerism graph,and the feature representation of drugs and diseases based on graph Convolutional neural network is obtained;4)Decomposing heterogeneous networks and using heterogeneous feature extraction methods to extract subnet features,obtaining feature representations of drugs and diseases based on subnet information;5)Perform linear transformation on the four types of feature representations of drugs and diseases,and use multiple feature representation aggregation methods to aggregate to obtain the final feature representations of drugs and diseases;6)Input the feature representations of drugs and diseases into the linear inner product decoder to reconstruct the drug disease relationship prediction matrix,achieving drug disease relationship prediction.Finally,this article verifies the predictive performance of GNDDA through comparative experiments,which show that GNDDA outperforms existing drug disease prediction methods in AUC and AUPR indicators.In order to help researchers quickly identify potential drug disease relationships,reduce the cost of discovering new indications for existing drugs,and improve the efficiency of discovery,this article develops an online drug disease relationship query and prediction system.In order to develop this system,we first organized the data of drugs and diseases,constructed a database containing multiple drug information,and implemented algorithms for online prediction and retrieval of drug disease relationships,as well as graphical display of drug disease relationships,to facilitate researchers’ subsequent research work.更多还原