Multi-species Protein Association Prediction Using Residual Graph Convolutional Network

Abstract

Graph Convolutional Networks have recently received a lot of attention for their capability of representation learning on non-Euclidian feature spaces. Graph convolutional networks aggregate the neighbouring node features and attribute to learn graph representations. Like traditional deep learning models, the representation power of graph convolutional networks also increases with the increasing number of layers. However, it also increases the difficulty associated with training such models. Deep graph convolutional networks suffer from issues like vanishing gradient or over-fitting. In this paper, we explored skip connections in graph convolutional networks and proposed a deep residual graph convolutional neural network for predicting node properties in a protein-protein interaction network. The proposed model is an improvement over traditional deep learning models and present state-of-the-art graph learning algorithms. The implementation of the algorithm, as well as the saved model, is available online for reproducibility at https://github.com/rangan2510/R-GCN.