文章摘要

文章信息/Info

Title:: MiRNA-disease association prediction model based on graph convolution neural network and topological features

作者:: 畅豫霄1; 2; 董亚东4; 徐永涛1; 2; 3△; （1.新乡医学院医学工程学院，新乡 453003；2.河南省临床与生物医学大数据融合技术工程实验室，新乡 453003；3.河南省神经信息分析与药物智能设计国际联合实验室，新乡 453003；4.郑州大学第一附属医院互联网医疗系统与应用国家工程实验室，郑州 450052）

Author(s):: CHANG Yuxiao1; 2; DONG Yadong 4; XU Yongtao 1; 2; 3; （1.School of Medical Engineering, Xinxiang Medical University, Xinxiang 453003, China; 2.Henan Engineering Laboratory of Combinatorial Technique for Clinical & Biomedical Big Data, Xinxiang 453003; 3. Henan International Joint Laboratory of Neural Information analysis and Drug Intelligent Design, Xinxiang 453003；4.The First Affiliated Hospital of Zhengzhou University,China National Engineering Laboratory of Internet Medical System and Application, Zhenzhou 450052,China）

Keywords:: Link prediction; Gaussian interaction profile kernel similarity; Convolution operation; Classification features; Graph iteration; Undirected graph features; 10-fold cross-validation

摘要:: 针对在微核糖核酸（micro ribonucleic acid，miRNA）-疾病关联性研究中信息使用不充分，且过度依赖网络节点的相似度信息，预测准确率较低的问题，本研究提出了基于图卷积神经网络（graph convolutional neural networks，GCN）及拓扑特征的miRNA和疾病的关联预测计算模型GTMDA。该模型综合miRNA相似度矩阵、疾病语义相似度网络和miRNA-疾病关联关系矩阵，首先使用GCN和随机游走算法，分别获取miRNA和疾病的子图顶点嵌入特征及miRNA与疾病相似性网络的拓扑结构特征;然后将其输入多层感知器（multilayer perceptron，MLP）,预测潜在关联性。结果表明，该模型的AUC值达到0.947 5±0.010 6，优于其他方法。此外，预测的前50个miRNA中，84%的乳腺恶性肿瘤、90%的食管癌和94%的肺癌得到了独立数据库的验证。因此，本研究模型可作为预测miRNA-疾病潜在关联的可靠模型。

Abstract:: In view of the problem of inadequate use of information, excessive dependence on similarity information of nodes in the network and low prediction accuracy in micro ribonucleic acid（miRNA） association studies, we proposed a computational model for miRNA-disease association prediction based on graph convolutional neural networks (GCN) and topological features (GTMDA). This model integrated miRNA similarity matrix, disease semantic similarity network, and miRNA disease association matrix. Firstly, GCN and random walk algorithm were used to obtain the vertex embedding characteristics of miRNA, disease subgraphs, the topological structure characteristics of miRNA and disease similarity network; then it was input into multilayer perceptron(MLP) to predict the potential relevance.The experimental results showed that the AUC value of this model reached 0.947 5±0.010 6, which was better than the other methods. Among the top 50 predicted miRNAs, 84% of breast cancer, 90% of esophageal cancer, and 94% of lung cancer were validated by independent databases. Therefore, this model can be a useful and reliable model for predict potential miRNA-disease associations.