Better Link Prediction for Protein-Protein Interaction Networks

In functional genomics, experimentally obtained protein-protein interaction (PPI) data is often incomplete. To deal with this issue, computational approaches are used to infer missing data and to evaluate confidence scores. Link prediction is one such approach that uses the structure of the network of PPIs known so far to find good candidates for missing PPIs. In a recent study by Kovács et al., a novel PPI-specific link predictor was proposed. Their link predictor is biologically motivated by the so-called L3 principle and it was shown to be superior to other general link predictors when applied to PPI data. However, the L3 link predictor is only an approximate implementation of the L3 principle. As such, not only is the full potential of the L3 principle not realized, it may even penalize candidate PPIs that otherwise fit the L3 principle. In this paper, we formulate an L3-based link predictor without approximation, coined ExactL3. We show computationally that ExactL3 is better than the previously proposed methods on four major PPI datasets (STRING, BioGRID, IntAct/HuRI, and MINT). The predicted PPIs are also shown to be much more functionally relevant. This confirms that ExactL3 is a better link predictor for PPI networks, and demonstrates its ability to characterize PPIs by only the topological features of binary PPI networks.