Figure 2

Assessing the performance of each dataset in predicting the human protein-protein interactions. A. Large-scale protein-protein interaction (PPI) datasets from model organisms and human. The datasets SC5 and DM1 are binned by their confidence level given by the original studies: LC, low confidence; HC, high confidence. B-D. Phenotypic datasets from model organisms. The fly genome-wide RNAi dataset is evaluated by the arithmetic difference in phenotypic values between a pair of genes (D). The phenotypic similarity of yeast genes upon knock-out is evaluated by cosine distance for the discrete values (C) and PCC for the continuous values (D). E. Yeast genetic interaction datasets. Yeast genetic interactions are grouped by the number of shared neighbors between a pair of genetically interacting genes. F. Large-scale human gene expression datasets. Gene pairs are binned by their Pearson Correlation Coefficient (PCC) between the expression profiles of the pair. The purple, yellow and blue curves are derived from three different expression datasets [57-59]. G. Domain-domain interaction (DDI) score. The DDI score of a domain pair is assigned to a pair of proteins containing the domains. If different scores exist between a pair of proteins arising from different interacting domain pairs, the maximum of the scores is assigned to the pair. The protein pairs are grouped according to their DDI scores. H. Smallest number of shared biological processes (SSBP) of yeast (SC), worm (CE), fruitfly (DM), mouse (MM) and human GO annotations. Gene pairs are binned by the smallest number of shared GO annotations between a pair of genes. Then the LR of being GSP versus GSN is calculated and plotted for gene pairs within each bin for each organism. I. Gene context analysis to predict PPIs. Three types of in silico prediction results are evaluated (gene fusion, gene co-occurrence and gene neighborhood).