Four groups of HRSS values with different confidences for positive and negative protein-protein interactions in BP and CC ontologies.

An ROC analysis showed that the HRSS method using the maximum strategy on GO annotations including IEA code performed best among all methods on distinguishing true protein interactions from false interactions. Since no one ‘best’ threshold could be obtained directly from the ROC analysis, five indexes, TPR, FPR, PPV, NPV and the Youden index, were used to determine the reasonable HRSS cutoffs (in bold) for selecting positive and negative interactions. Accordingly, the HRSS values were divided into four groups with low (L), medium-low (ML), medium-high (MH) and high (H) confidences.