Empirical Assessment of Graph Embedding Techniques for Predicting Missing Links in Biological Networks

Network science tries to shed light into the complex relationships among entities of a system. For instance, biological networks represent relations between macro molecules such as genes, proteins or other small chemicals. Often potential links are guessed computationally due to expensive nature of wet lab experiments. Conventional link prediction techniques consider local network wiring structure, which may not able to infer true relationships. The recent approaches of graph embedding (or representation learning) aims to capture the complete network structure that may be utilized for link prediction.

In this work, we assess the performance of ten (10) state-of-the-art embedding techniques for their effectiveness of link prediction in homogeneous and heterogeneous biological networks. Majority of the graph embedding methods, in its original form, not in a position to predict links. We use the latent representation of the network produce by the embedding methods and recreate the network using various similarity and kernal functions. We evaluate nine (09) such functions in combination with candidate embedding methods. We even compare the performance of five (05) traditional, local structure based link prediction methods to show the superiority.

Experimental results clearly reveal that Graph Neural Network (GNN) and Attention based encoders with dot product based decoder are the best performers in predicting missing links for both homogeneous and heterogeneous biological network.