Flowchart of experimental method.

<p>There are 2 main steps in the method. The first is to rank protein pairs according to their probabilities or confidence scores of being true protein-protein interaction (PPI) pairs by defining a machine learning task. The second is to apply the state-of-the-art protein complex detection algorithm to a reconstructed PPI network. In the first part, we represented protein pairs based on 6 sources, then used learning from positive and unlabeled data (LPU) to predict PPIs. We also applied a five fold cross-validation for evaluating the LPU model. In the second part, we applied some popular protein complexes detection methods (COACH, CMC, CFinder, MCODE, IPCA, MCL, Clusterone) on a reconstructed PPI network to identify protein complexes.