Fig 1.
Contact map and protein-protein interaction network of small and large ribosomal subunits.
The contact map and the protein-protein interaction network for A the small ribosomal subunit and B the large ribosomal subunit (proteins only), using a distance cutoff of 8Å between heavy atoms. The upper diagonal part shows the contact map, with red dots indicating intra-protein contacts, and blue dots inter-protein contacts. The lower triangular part shows the coarse graining into the corresponding protein-protein interaction networks, with the color levels indicating the number of intra- resp. inter-protein contacts, cf. the scales. The sparse character of both the contact network and the interaction network is clearly visible.
Fig 2.
Concatenating two multiple sequence alignments.
Sketch of the matching procedure that allows us to concatenate two different MSAs, here MSA1,MSA2. π represents the optimal permutation of the sequences on the second MSA computed using a standard linear programming routine.
Fig 3.
Residue-residue structure of both artificial and ribosomal complex.
A Architecture of the artificial protein complex. Arcs width are proportional to the number of inter-protein residue contacts. B Inferred PPI network for both paired and combined strategy for different number M of sequences generated from the artificial model. Green arcs are true positives, red false positives, gray low-ranking predictions. Arc widths are proportional to the inter-protein interaction score. C SRU architecture (same color code as A). D Inferred PPI network (same color code as B). E Same as C for LRU. F Same as D for LRU. Arc width in panels C-F is provided by the number of inter-protein contacts, as a measure of interface size. It becomes obvious that mainly large interfaces are recognized by our approach.
Fig 4.
Architecture and inferred protein-protein interaction network of the artificial protein complex.
A Residue-residue interaction structure of the generating model for the artificial data. Colored arcs represent the protein chain. Non-zero couplings in the coupling matrix of the generating model are represented as curves between the nodes. The width of the curves is proportional to the interaction score. Only the 10 strongest intra/inter-protein scores are shown. B Same as A, but based on the inferred couplings. Green arcs are true positives, red false positives. Note that not all green arcs have a corresponding arc in A due to our choice to display only the 10 strongest couplings, which not always correspond to the strongest score. C Same as A for SRU. All links represent a contact in the PDB structure and have equal width. D Same as B for SRU. E Same as C for LRU. All links represent a contact in the PDB structure and have equal width. F Same as D for LRU.
Fig 5.
The plots illustrate the performance in predicting protein interaction partners. The left panels show the fraction of true positives among the first n PPI predictions, with n being the number indicated on the horizontal axis (solid lines). The dashed lines show the best possible (upper dashed line) and the mean of a random prediction (lower dashed line). The right panels show ROC-curves, which indicate the dependence of the true-positive predictions (TP/P) from the false positive predictions (FP/N). The area under the curve (AUC) is a global global measure for the prediction quality; it is 1/2 for a random, and 1 for a perfect prediction. A protein pair is identified as an interacting (true positive) pair, if at least one PDB structure with at least one inter-protein contacts exists.
Fig 6.
Tryptophan biosynthesis pathway.
A Architecture of the known protein-protein interaction among the 7 enzymes which are coded in the Trp operon. The widths of the arcs are proportional to the number of inter-protein residues (which in this case is almost equal for the two interacting pairs). B Inferred PPI network, here the width of the arcs is proportional to the interaction score. Green arc correspond to the protein pairs for which a known structure exist.
Fig 7.
Efficacy of the different matching procedures.
True-positive rates for inter-protein residue contact prediction for different matching procedures. Shown are means for all protein pairs that have at least 100 residue pairs in contact. The ribosomal and the trp proteins were tested independently. The red curves correspond to a matching including only protein sequences without paralogs inside the same species (“matching by uniqueness in genome”). The low performance of this approach on Trp proteins is due to a very low number of species without homologs, which leads to very small matched alignments. The blue curves show the results for our matching procedure as described in the text. The green curves correspond to alignments that have been obtained by first applying our matching procedure and then randomizing the matching within individual species. The definition of “contact” was the same as used above (a distance of less than 8.0Åbetween two heavy atom in the residues).