Low-Rank Matrix Completion for Inference of Protein-Protein Interaction Networks

We propose a new model for protein-protein interaction networks that is based on the assumption that interaction affinities approximately satisfy sets of linear constraints, and that there exist relatively few factors that influence the affinity levels. This model allows for inferring unknown protein-protein interactions using emerging algorithmic solutions from the area of low-rank matrix completion. Low-rank matrix completion algorithms predict interactions using only a small number of known affinity values, and in addition, they are robust to measurement noise. We illustrate the use of the new modeling approach to protein interaction prediction for Saccharomyces cerevisiae, based on data from the well known STRING repository. For 1200 proteins, a rank 25 model recovers more than 84% of test interactions reported in STRING.