A deep learning algorithm for predicting protein-protein interactions with nonnegative latent factorization

Abstract

Protein-protein interaction (PPI) networks play an essential role in the study of proteomics. Given the fact that known PPI data are extremely incomplete, high-throughput technologies have been developed to significantly increase the amount of PPI data, but they are prone to generate false positive PPIs and accordingly affect the performance of computational prediction algorithms. To overcome this problem, we propose a novel deep learning algorithm for predicting PPIs with symmetric nonnegative latent factorization (SNLF). In particular, we first improve the quality of PPI data by applying an established SNLF model. Quasi-Sequence-Order is then used to encode proteins based on the modality of their sequence information. Principal component analysis is utilized to generate the features of proteins in a more compact manner. After that, we adopt graph variational autoencoder to learn the embedding of each protein by considering protein features and network topology. Finally, the embeddings of proteins are concatenated in pairs as input to train a simple feedforward neural network for prediction. Experiments have been performed on five different PPI datasets by comparing the performance of our algorithm with the state-of-the-art prediction algorithms, and the results demonstrate that the proposed model is promising in predicting PPIs.