计算机科学与探索 ›› 2023, Vol. 17 ›› Issue (3): 701-708.DOI: 10.3778/j.issn.1673-9418.2105041

• 人工智能·模式识别 • 上一篇    下一篇

Protein-HVGAE:一种双曲空间中的蛋白质编码方法

王皓白,沈昕,黄尉健,陈可佳   

  1. 1. 南京邮电大学 计算机学院,南京 210023
    2. 南京邮电大学 理学院,南京 210023
    3. 江苏省大数据安全与智能处理重点实验室(南京邮电大学),南京 210023
  • 出版日期:2023-03-01 发布日期:2023-03-01

Protein-HVGAE: Protein Encoding Method in Hyperbolic Space

WANG Haobai, SHEN Xin, HUANG Weijian, CHEN Kejia   

  1. 1. School of Computer Science, Nanjing University of Posts and Telecommunications, Nanjing 210023, China
    2. School of Science, Nanjing University of Posts and Telecommunications, Nanjing 210023, China
    3. Jiangsu Key Laboratory of Big Data Security & Intelligent Processing (Nanjing University of Posts and Telecommunications), Nanjing 210023, China
  • Online:2023-03-01 Published:2023-03-01

摘要: 蛋白质相互作用(PPI)网络中的蛋白质功能预测、蛋白质交互预测和复合物识别是生物信息学的重要任务,非常依赖于对蛋白质的编码。由于PPI网络是由少量中枢节点主导的无标度网络,传统欧氏空间嵌入方法难以捕捉网络中的层次结构,导致蛋白质编码效果并不理想。提出一种基于双曲空间图嵌入的蛋白质自编码器Protein-HVGAE,该模型采用两个双曲图卷积网络作为编码器,计算隐藏层的均值和方差,并在不同曲率的双曲空间中捕捉网络的层次结构,以区分各节点的低维表示;采用Fermi-Dirac函数做解码器,在双曲空间上通过内积运算重构网络。实验结果表明,该模型在3个PPI数据集中的两个下游任务(PPI预测和蛋白质功能预测)上的表现优于以往在欧氏空间中的编码方法(在PPI预测中AUC值高于VGAE模型0.07左右,在蛋白质功能预测中Macro-F1值高于VGAE模型0.02左右)。

关键词: 蛋白质交互网络, 双曲空间, 图卷积, 变分图自编码器(VGAE), 蛋白质功能预测

Abstract: Protein function prediction, protein interaction prediction and complex identification in protein-protein interaction (PPI) networks are important tasks in the field of bioinformatics, which rely heavily on the protein expression. Since the PPI network is a scale-free network dominated by a small number of hub nodes, it is difficult for the embedding method in traditional Euclidean space to capture the hierarchical structure in the network, resulting in unsatisfactory protein embeddings. This paper proposes a protein auto-encoder in hyperbolic space, Protein-HVGAE (hyperbolic graph auto-encoder for protein interaction networks). This paper uses two hyperbolic graph convolutional networks as encoders, calculates the mean and variance of the hidden layer and captures the hierarchical structure of the PPI network in hyperbolic spaces with different curvatures to distinguish the low-dimensional representation of each node; it uses the Fermi-Dirac function as the decoder, and reconstructs the network through the inner product operation on the hyperbolic space. Experimental results in three PPI networks show that the performance of this model in two downstream tasks (i.e., PPI prediction and protein function prediction) is superior to the previous methods in Euclidean space (around 0.07 improvement of AUC in PPI prediction and 0.02 improvement of Macro-F1 in protein function prediction compared with VGAE model).

Key words: protein-protein interaction network, hyperbolic space, graph convolution, variational graph auto-encoder (VGAE), protein function prediction