An attributed network embedding method to predict missing links in protein-protein interaction networks

Abstract

Predicting missing links in noisy protein-protein interaction networks is an essential~computational method. Recently, attributed network embedding methods have been shown to be significantly effective in generating low-dimensional representations of nodes to predict links; in these representations, both the nodes'features and the network's topological information are preserved. Recent research suggests that models based on paths of length 3 between two nodes are more accurate than models based on paths of length 2 for predicting missing links in a protein-protein interaction network. In the present study, an attributed network embedding method termed ANE-SITI is recommended to combine protein sequence information and network topological information. In addition, to improve accuracy, network topological information also considers paths of length 3 between two proteins. The results of this experiment demonstrate that ANE-SITI outperforms the compared methods on various~protein-protein interaction (PPI) networks.