Variational Autoencoder Based Network Embedding Algorithm For Protein Function Prediction

The development of high-throughput technology has produced a large number of protein-protein interaction datasets, which provide an effective way to infer the functional annotation of proteins. However, how to make proper use of these datasets to extract effective low-dimensional feature representation of proteins for functional prediction is a challenge. Most existing network integration methods for protein function prediction have some limitations to capture complex and highly non-linear network structure information due to their design architecture. Therefore, we propose a novel multi-network embedding method deepVAE based on deep variational autoencoder (VAE), which uses the variational autoencoder to extract low-dimensional features of proteins from multiple various interactive network datasets and then trains a SVM classifier to predict protein function. Particularly, we denoise the original networks before network embedding, thus the new proposed method is called deepVAE-NE. The experiments are conducted on the yeast and human protein-protein interaction datasets and the experimental performance shows that our methods perform better than the other four compared advanced approaches, which greatly improves the accuracy of functional prediction.