Protein-Protein Interaction Prediction via Structured Matrix Completion

This paper considers how to computationally predict unknown protein-protein interactions (PPIs) given the experimentally verified PPIs. Matrix completion, a very popular machine learning technique that can be used to to infer the missing part of a matrix, has been introduced to recover the missing interactions of an incomplete PPI network. The benefit of Matrix completion is that it does not rely on unavailable negative samples, which are crucial for existing supervised classification methods. However, current matrix completion solutions for recovering missing PPIs fail to capture the important topology information of the underlying network. That is, the underlying network is a sparse network with skewed degree distribution. In this paper, we design a structured matrix completion method that is suitable for capturing the skewed degree distribution of the underlying true PPI network. Theoretical analysis and extensive experimental results on known PPIs of three species (Plasmodium falciparum, Rattus norvegicus and Caenorhabditis elegans) show that our method outperforms related state-of-the-art protein-protein prediction approaches. We also tested the predicted networks of Plasmodium falciparum in terms of GO similarities. In turns out that our predicted network is with the highest GO score. To further demonstrate the power of our algorithm in predicting new PPIs, we compare the predicted Rattus norvegicus PPI networks using relatively old release from BioGrid with the much newer releases. Comparing with other methods, our predicted networks are much more similar with the newer releases. Our code is available upon request and will be public available in our website after the paper published.