Relative evidence contribution to FunCoup networks viewed by input data type (A) or species (B). For each evidence category in each species and FC class, we calculated the sum of its contributions to the final Bayesian score (FBS) over the whole predicted network, i.e., the set of links with FBS > 3. Then, these sums were normalized by dividing with the sum of FBS scores over the same set. Note that no input data from Ciona was used. Data-type legends as in Figure 1.

Scenarios for interaction inheritance. (A) Interacting genes A and B in an ancestral species are duplicated into two and three genes in species 1, while B is duplicated into two genes in species 2. These duplicated genes are clusters of inparalogs in relation to the other species, meaning that they are co-orthologous to the corresponding A and B genes in that species. When transferring functional coupling (arrow marked FC) between orthologs, one can consider the interaction either to be valid for all inparalogs in a cluster, or to be specific for a particular inparalog pair (e.g., the seed orthologs, i.e., the two most similar ones, dotted arrows). The transfer of interaction information thus can either be done from an arbitrary inparalog in a cluster, which maximizes the coverage, or only be transferred between, e.g., seed orthologs. Benchmarking showed that transferring the best interaction in a set of inparalogs to all inparalogs in the other species yields the best results (Supplemental Material). (B) Real example of this situation: An interaction in fly ptc–babo (Shyamala and Bhat 2002) can either be considered to be valid for all six pairs of mouse inparalogs {Ptch1, Ptch2} vs. {Tgfbr1, Acvr1b, Acvr1c} or to be specific for a particular inparalog pair. Seed orthologs in this case are ptc/Ptch1 and babo/Tgfbr1. The two fly genes are connected with solid arrows; the others are mouse genes. The arrow denotes the known protein–protein interaction between ptc and babo. The leaves of the gene trees present protein domain architectures.

Outline of the FunCoup network reconstruction process. Amounts of input data and sizes of training sets are shown for each species in FunCoup version 1.0. Input data are as follows: MEX, mRNA coexpression; PHP, phylogenetic profile similarity; PPI, protein–protein interactions; SCL, subcellular colocalization; TFB, shared transcription factor binding; PEX, protein coexpression: MIR, miRNA targeting of transcripts; and DOM, domain associations. Training sets are as follows: ML, links between proteins from the same metabolic pathways; SL, links between proteins from the same signaling pathways; PI, experimentally observed protein–protein interactions; and CM, pairs of protein-members of the same complex. The Bayesian framework processes the input data using the training sets. The input datapoints are converted into raw interaction scores, which are grouped into discrete regions. Each such bin is assigned an FC score using the training sets. The “cards” illustrate the results of this process, showing the raw interaction score along the horizontal axis. For each training set, or functional class, the resulting bins are shown as colored rectangles: (green) positive evidence of FC; (white) either close to neutral or insignificant; (red) negative evidence of FC. Finally, the FC scores are calculated for all possible gene pairs in each species. For brevity, the predicted links of different functional classes have been combined into one network per species.

Example of predicting functional coupling (FC) by FunCoup's naïve Bayesian network. To evaluate the cumulative likelihood of the link CDC2–KPNB1, evidence (only nine shown) from a number of species were available. The metrics computed from each evidence were scored by the four models of FC and summed up to a total likelihood value, the final Bayesian score (FBS). In this example, a physical protein interaction (FC-PI) is more likely than membership in the same protein complex (FC-CM), a metabolic (FC-ML), or a signaling (FC-SL) link. The web interface to the database of predictions (http://FunCoup.sbc.su.se) displays likelihoods of the FC classes as four colored lines with FBS values transformed into the pfc confidence scores. This coupling was supported by many evidences from several species (top). The strongest came from the IntAct database that reported two experiments (Thelemann et al. 2005; Koch et al. 2007) where CDC2 and KPNB1 were mentioned as preys interacting with the same bait (in parallel with 219 and 68 other proteins, respectively). On its own, this is rather weak and would only yield pfc = 0.35 to prove a physical interaction (FBS = 6.3). However, combined with other evidence such as coexpression in human and mouse, the score was substantially strengthened (FBS = 11.2; pfc = 0.98). Note that the website can also display the coupling in terms of the individual evidence or species contributions. (MEX) mRNA expression; (SLC) subcellular localization; (PPI) known protein–protein interaction; (PLC) Pearson's linear correlation coefficient; (WMI) weighted mutual information score; and PPI score. See Supplemental Methods.

Comparative network analysis in FunCoup. (A) Subnetworks in human (middle), mouse (top), and fly (left) were generated by submitting human presenilin 1 and 2 (PSEN1 and PSEN2) to FunCoup, asking for one step of network expansion keeping the 20 strongest links with P > 0.5, and inclusion of orthologous subnetworks in mouse and fly. At the lower right are two newly predicted interactors of the gamma secretase complex, BET1 and LFNG. On the right, the color legend for the links is shown in terms of evidence type. Species are indicated by gene symbol prefixes: has, human; mmu, mouse; and dme, fly. Supplemental Figure 8 presents alternative views by species source or predicted class. The jSquid XML source to this figure is available as supplementary file psen_lfng_bet.xml. (B) Using FunCoup to identify novel candidate genes in Parkinson's disease. We here employ orthologous networks in human (left) and yeast (right, squares). Novel candidates were extracted by looking for human genes that were coupled both to known PD genes and to orthologs of yeast alpha-synuclein toxicity modifiers. This resulted in 12 novel candidate PD genes (red circles). PD and alpha-synuclein toxicity modifier genes not connected to these 12 genes were omitted from the network for clarity. Edge categories are shown in the figure. Node categories are as follows: (yellow) known human PD genes from the pathway KEGG05020 (triangles) and their yeast orthologs (squares); (blue) yeast modifiers of a-synuclein toxicity (squares) and their human orthologs (circles); (red) 12 novel human PD candidate genes (circles) and their yeast orthologs (squares); and larger shapes labeled “collapsed,” grouped genes—inparalogs, except for UBE2 that represents four ubiquitin-conjugating enzymes with similar molecular function. For details on how the subnetwork was generated, see Supplemental Methods; for a large scale prospective of the presented subinteractome, see Supplemental Figure 10.