北京大学学报自然科学版 ›› 2022, Vol. 58 ›› Issue (1): 37-44.DOI: 10.13209/j.0479-8023.2021.105

上一篇    下一篇

基于异质信息网络元路径的药物‒靶标相互作用预测模型

廖懿鸣, 欧阳纯萍†, 刘永彬, 胡富裕   

  1. 南华大学计算机学院, 衡阳 421001
  • 收稿日期:2021-05-08 修回日期:2021-08-09 出版日期:2022-01-20 发布日期:2022-01-20
  • 通讯作者: 欧阳纯萍, E-mail: ouyangcp(at)126.com
  • 基金资助:
    国家自然科学基金(61402220)、湖南省自然科学基金(2020JJ4525)、湖南省教育厅重点科研项目(19A439)和南华大学研究生科研创新项目(213YXC007)资助

Drug-Target Interactions Prediction Based on Meta-path of Heterogeneous Information Network

LIAO Yiming, OUYANG Chunping†, LIU Yongbin, HU Fuyu   

  1. Computer College, University of South China, Hengyang 421001
  • Received:2021-05-08 Revised:2021-08-09 Online:2022-01-20 Published:2022-01-20
  • Contact: OUYANG Chunping, E-mail: ouyangcp(at)126.com

PDF

摘要：

提出一种融合元路径信息的图神经网络模型, 用于预测药物-靶标相互作用(GMDTI)。首先根据8个数据集中的药物、靶标、疾病和副作用数据以及它们之间的8种作用关系, 构建药物-靶标异质信息网络(HIN); 然后定义两条元路径来捕获药物-靶标HIN 中的不同子结构信息和不同节点间隐藏的语义信息, 并应用图神经网络的方法聚合节点的一阶邻居信息和元路径中节点间的语义信息; 最后利用端到端的学习方法完成DTIs预测。该方法同时考虑药物-靶标HIN的结构特性和元路径语义信息, 有助于学习到更多潜在的药物-靶标作用关系。实验结果表明, GMDTI的预测准确率高于所有基线模型, AUC达到98.6%, AUPR达到94.5%。同时通过调整数据的稀疏度和降噪实验, 证明GMDTI具备优于所有基线模型的鲁棒性。

关键词: 药物?靶标相互作用预测, 图神经网络, 异质信息网络, 元路径, 特征表示

Abstract:

The paper proposes a graph neural network model based on meta-path to predict drug target interactions (GMDTI). Firstly, based on drugs, targets, diseases and side effects in eight datasets, and the eight different types of action relationships between them, the authors construct a drug-target heterogeneous information network (HIN). Then, two different meta-paths are defined to capture the different sub-topology information of HIN and the latent semantic information between different nodes. Especially, the graph neural network method is applied to represent the node by aggregating the information of the first-order neighbor nodes and the nodes of the meta-path. Finally, DTIs prediction is completed effectively by end-to-end learning method. This method takes the first-order topology and the semantic information of meta-path of the drug-target HIN into account, which is helpful to learn more potential drug target relationships. The experiment results show that the proposed method achieves 98.6% in AUC and 94.5% in AUPR, which are higher than all baseline models. At the same time, GMDTI has better robustness than all baseline models by sparsity experiments of datas and reduction experiments of noise.

Key words: drug-target interaction prediction, graph neural network, heterogeneous information network, metapath; feature representation

版权所有 © 《北京大学学报（自然科学版）》编辑部
技术支持：北京玛格泰克科技发展有限公司