Inferred PPIs within the human endosome complex. Interaction network of the human endosome complexes as inferred by CAPPI-Pred. Nodes represent gene products and node colors represent protein families identified by sequence clustering. 49 predicted interactions which have been previously detected experimentally are denoted by green edges. 49 other PPI predictions are denoted by gray edges. The p-value of the predicted network is 3.977 × 10^-9.

Inferred PPIs within the human mRNA decay complexes. Interaction network of the human mRNA decay complexes as inferred by CAPPI-Pred. Nodes represent gene products and node colors represent protein families identified by sequence clustering. 53 of the predicted interactions which have been previously detected experimentally are denoted by green edges. 76 other PPI predictions are denoted by gray edges. The p-value of the predicted network is 1.868 × 10^-15.

Inferred PPIs within the A. thaliana SWI/SNF chromatin remodeling complex. Interaction network of the putative SWI/SNF complex in Arabidopsis as inferred by CAPPI-Pred. Nodes represent gene products and node colors represent protein families identified by sequence clustering. 13 of the predicted interactions which have been previously detected experimentally are denoted by green edges. 83 other PPI predictions are denoted by gray edges. The p-value of the predicted network is 6.381 × 10^-10.

Inferred PPIs within the yeast proteasome complex. Interaction network of the yeast 26S proteasome complex as inferred by CAPPI-Pred. Nodes represent gene products and node colors represent protein families identified by sequence clustering. 177 of the predicted interactions which have been previously detected experimentally are denoted by green edges. 66 other PPI predictions are denoted by gray edges. The p-value of the predicted network is 4.348 × 10^-16. The networks are visualized using the Cytoscape software [73].

Assessment of CAPPI-Pred predictions. Assessment of CAPPI-Pred predictions. (A) and (B): Histograms of BP scores for the predicted yeast (A) and human (B) PPI datasets of the same size (1576 yeast PPIs and 17105 human PPIs) from the Interlog method, CAPPI-Pred and CAPPI-Integ. (C) and (D): The ratio of true positives and false positives as a function of the number of interactions in yeast (C) and human (D). An interaction is deemed true positive if it is found in the reference dataset of either co-complex interactions (Complex) or binary PPIs (PPI), or in any available reference set (ALL) and false positive, if the proteins are assigned different cellular localization (see text). Plots labeled as "Filtered Complex" and "Filtered PPI" show the results of selected CAPPI predictions which are part of dense clusters tested against either the co-complex reference (Complex) or binary reference (PPI). The gray dashed line marks the level at which the number of true positive predictions is equal to the number of false positive predictions.

A family-oriented overview of the input PPIs. A 4-way Venn diagram illustrating the overlap of PPI evidence between four of the considered seven species: human, yeast, fly and worm. Each cell in the diagram is labeled with the number of pairs of protein families for which members interact in the corresponding species. For example, there are 742 pairs of protein families such that in both yeast and human there exists at least one interaction between members of the two families and no such interactions exist for fly and worm. Only about 0.5% (42514 of 8280415) of possible family pairs we consider have any evidence for interaction in any of the four species. Of these only 0.1% (45 of 42514) have evidence in all four species, which seems small, given that all considered families are evolutionarily conserved. However, the size of the overlap presumably corresponds to the fraction of the interactomes sampled experimentally, rather than to the actual level of conservation. For example, while there is a significant size difference between the overlap of the relatively best sampled yeast and human interactomes (742+175+45+42 = 1004 family pairs) and the overlap between yeast and worm interactomes (23+45+42+78 = 188 family pairs), the fraction of family pairs with PPI evidence from human and worm overlapping with such pairs in yeast is of the same magnitude (8% and 9%, respectively). It is highly probable that many of the homologous interactions in yeast and human have, yet unidentified, counterparts in worm and similarly in the other species. CAPPI uses phylogenetic information and probabilistic modeling to identify the most probable interactions in each species given the joint evidence from all input datasets and considering their reliability.

Assessment of CAPPI-Integ predictions. Assessment of CAPPI-Integ predictions. (A) Histogram of BP scores for the fly input datasets (combined) and corresponding inferred datasets of the same size (4049 PPIs in case of Input-3sp and CAPPI-Integ-3sp and 12107 PPIs in case of Input-7sp and CAPPI-Integ). Both CAPPI-Integ and CAPPI-Integ-3sp provide higher-scoring interactomes compared to their input datasets demonstrating the method's ability to use the interactions from distant species to make high quality predictions in other species. (B) Assessment of predicted yeast interactions using the three GO scores. The similarity of GO annotations of each pair of interacting proteins is measured in each ontology: biological process (BP), molecular function (MF) and cellular component (CC). CAPPI and Domain-ML predictions are ranked by their probabilities and the average GO score for the top n predictions is shown. CAPPI-Integ-3sp outperforms the domain based approach trained on the same experimental data. CAPPI-Integ integrates all available data from the seven species and further improves the predictions for yeast. (C) The ratio of true positives (TP) and false positives (FP) as a function of the number of yeast interactions. An interaction is deemed true positive if it is found in the reference dataset comprising co-complex and binary PPIs, and false positive, if the two proteins are assigned different localizations in the MIPS sub-cellular localization catalog (see text). The TP/FP ratios for the CAPPI-Integ, CAPPI-Integ-3sp and Domain-ML predictions are compared with the scores of the input experimental datasets. The gray dashed line marks the level at which the number of true positive predictions is equal to the number of false positive predictions.

Inferring protein interactions via message passing. A toy example of the Bayesian tree model of evolution of interactions between members of two protein families for three species: blue, yellow and red. For each species a certain number of experimental datasets is given: two for blue and red and one for yellow. Part (A) shows two reconciled trees for the considered families together with putative protein interactions at each level of evolution. The proteins in the trees are represented by ellipses (with color corresponding to their species). The speciation events are marked by horizontal lines and the duplication events are marked by filled squares. The evolution of the ancestral interaction between the root proteins (purple) can be traced down the trees to the extant interactions. Evidence for the extant interactions can be found in the experimental datasets. In (B) a random variable is associated with each putative interaction. A solid arrow indicates a dependence between two random variables which comes from the speciation event. Similarly, a dashed arrow indicates a dependence for the duplication event. Finally, dotted arrows represent an interface between the true interactions in extant species and the observed experimental evidence. The parameters p_s, δ_s, p_d and δ_d determine the probability of retaining or gaining an interaction during evolution, while the reliability of each dataset (Rel()) determines the probability of identifying a true interaction or a false positive one. Arrows colored blue, yellow, red and green represent messages corresponding to interaction evidence coming from each of the species. These messages are passed up the tree in the first phase of the MP algorithm. In the second phase, messages containing aggregated evidence from one side of the tree are passed down to the other side (orange arrows).