Abstract
Background
Data from highthroughput experiments of proteinprotein interactions are commonly used to probe the nature of biological organization and extract functional relationships between sets of proteins. What has not been appreciated is that the underlying mechanisms involved in assembling these networks may exhibit considerable probabilistic behaviour.
Results
We find that the probability of an interaction between two proteins is generally proportional to the numerical product of their individual interacting partners, or degrees. The degreeweighted behaviour is manifested throughout the proteinprotein interaction networks studied here, except for the highdegree, or hub, interaction areas. However, we find that the probabilities of interaction between the hubs are still high. Further evidence is provided by path length analyses, which show that these hubs are separated by very few links.
Conclusion
The results suggest that proteinprotein interaction networks incorporate probabilistic elements that lead to scalerich hierarchical architectures. These observations seem to be at odds with a biologicallyguided organization. One interpretation of the findings is that we are witnessing the ability of proteins to indiscriminately bind rather than the proteinprotein interactions that are actually utilized by the cell in biological processes. Therefore, the topological study of a degreeweighted network requires a more refined methodology to extract biological information about pathways, modules, or other inferred relationships among proteins.
Background
Experimental proteinprotein interaction (PPI) data and related networks, obtained from highthroughput methodology as well as handcuration, are being widely used to probe the nature of biological organization and extract functional relationships among sets of proteins [1,2]. What has not been appreciated is that the guiding principles involved in assembling these networks may exhibit considerable probabilistic behaviour. Here, we show that the probability of an interaction between two proteins is generally proportional to the product of their individual numbers of interacting partners (or degrees) and discuss the consequences of this for probing PPI networks. Understanding the underlying organizational principles in assembling PPI networks holds the key for interpreting and analyzing the observed interactions.
Highthroughput methodologies [36] to determine PPI networks have been used to probe the interactome of a range of organisms. The organization of these interaction networks has been studied using graphtheoretical techniques [79] to find global characteristics that can be mapped back to biological phenomena, such as evolutionary conserved interactions, pathway or module organization, and localization of essential proteins in the network, to mention a few. Since we know that outcomes of cellular actions are biologically "deterministic" in the sense that cells use energy, synthesize proteins, duplicate DNA, etc., the analysis of PPI networks is aimed at finding and extracting causative components. If this information is to be mined from a global dataset, it is vital to have an accurate model of the architecture of the determined PPI networks. The incorporation of the underlying determining principle of PPI organization into graphtheoretical topological studies will provide a baseline from which biologicallyrelevant insights could be extracted. For example, a guided biological framework implies that cellular processes consist of precise and unique proteinprotein interactions, whereas a probabilistic model is suggestive of an underlying principle that is more chemical than biological, describing the ability of proteins to bind.
We show here that currently available PPI data support the latter interpretation and demonstrate that the probability of an interaction between two proteins is proportional to their numbers of interacting partners. The observations suggest that PPI networks are almost completely probabilistic and, therefore, in a proteome context, PPI interactions for specific biological processes are generally not distinct. From a purely biological point of view, the knowledge of any potential interactions between proteins is useful. However, by identifying common themes in large PPI networks, the underlying principles responsible for the discovered interactions may become more apparent.
Networks can be constructed directly from probabilistic procedures where the interactions, or edges, between two nodes is determined from an a priori probability distribution of edges, the simplest being the ErdösRényi random model [10,11]. However, biological networks, including PPI typically show powerlaw scaling in their degree distributions, in that the probability of any node having a given number of interactions follows a power law [1214]. As such, the ErdösRényi model, which generates Poissonian degree distributions, is an unsuitable archetype for PPI networks. Networks with powerlaw degree distributions can be constructed using a number of techniques, including those based on preferential attachment [15,16], duplication [1719], and hierarchical [20,21] approaches. Alternatively, the geometric random model generates networks that nearly follow a powerlaw distribution [22]. While each of these models may have qualitatively simulated biological networks, none have consistently and accurately reproduced properties of individual PPI networks.
Here, we describe insights into the topologies of PPI networks that should serve to enhance the development of future models. A degreeweighted network is one in which the probability of an interaction between two nodes is proportional to the product of their degrees, i.e., P_{ij }∝ k_{i}k_{j}, where k_{i }and k_{j }are the degrees, or number of interactions, associated with nodes i and j, respectively [23]. A type of degreeweighted network denoted "STICKY" [24] has been proposed as a model for PPI networks on the basis of similarities in derived global, or average, network properties, e.g., graphlet frequencies and average clustering coefficients. However, this model generates far too many nodes of zero degree and is therefore an unsuitable prototype for PPI networks. It is thus of importance to both qualitatively and quantitatively ascertain the extent of degreeweighted behaviour in biological networks. Here, we explore the nature of the proteinprotein connectivities more directly and conclusively demonstrate that PPI networks indeed contain degreeweighted elements.
Results
Probabilistic behaviour in protein interaction networks
A total of nine PPI networks from six unique organisms were studied. Full details of these networks are provided in Additional file 1. Their sizes are given in Table 1. For each network we calculated the probabilities P(k_{1}, k_{2}) of interaction between two proteins of degrees k_{1 }and k_{2}. A probability of interaction P(k_{1}, k_{2}) is calculated by counting the total number of interactions occurring between all proteins of degree k_{1 }and all proteins of degree k_{2}, and dividing this by the total number of all pairs of combinations that can be made. Degreeweighted behaviour was then established by comparing the probabilities P(k_{1}, k_{2}) of interaction with the products k_{1}k_{2 }of the degrees (Figure 1). We find that each PPI network exhibits perfect degreeweighted behaviour up to a characteristic value of k_{1}k_{2}, or cutoff, which depends on the network studied. Cutoffs have been estimated for each network and these are shown as dashed lines in the graphs (Figure 1). For Plasmodium falciparum, degreeweighted behaviour is exemplary throughout, thus no such value could be determined. Cutoff estimates range from 200 (WormCORE) to 4000 (Escherichia coli). It is unclear why cutoff values vary greatly between networks but this is presumably related to the differences in their degree distributions in that the actual degrees of the hub proteins vary from network to network. As the number of hub proteins of a particular degree is consistently very small (one or two), one might expect more noise in the hubhub interaction regions (largest values of k_{1}k_{2}). Correlation coefficients between P(k_{1}, k_{2}) and k_{1}k_{2 }determined using data with product degrees less than the cutoff are 0.97 or higher (Table 1), indicating an unmistakable degreeweighted signature in the PPI networks.
Additional file 1. Descriptions and sources of the proteinprotein interaction networks used in this work.
Format: PDF Size: 44KB Download file
This file can be viewed with: Adobe Acrobat Reader
Figure 1. Evidence of degreeweighted connectivity in nine PPI networks. a, Homo sapiens (human); b, Drosophila melanogaster (fruit fly); ce, Saccharomyces cerevisiae (yeast): YeastDIP, YeastCORE, YeastY2H; f, Escherichia coli (bacterium); gh, Caenorhabditis elegans (nematode): WormY2H, WormCORE; i, Plasmodium falciparum (malariacausing parasite). For k_{1}k_{2 }> 10, probabilities of interaction P(k_{1}, k_{2}) were ordered by k_{1}k_{2 }and averaged in groups of 10.
Table 1. Properties of PPI networks
If we express the probability of interaction as P(k_{1}, k_{2}) = γ(k_{1}k_{2})^{θ}, then the power, θ, and proportionality constant, γ, can be determined for each network by linear regression on data with product degrees less than the cutoff (Table 1). We find that all powers, θ, are very close to one, which is consistent with a probability function that is linear in each degree [23,24]. The proportionality constants γ determined from the regressions can also be calculated from normalizations via γ(cal) = E/Σ_{i<j}(k_{i}k_{j}), where E is the total number of interactions in the network and the summation is over all pairs of proteins. We find that the fitted and calculated proportionality constants are in good agreement (Table 1). Therefore, not only is degreeweighted behaviour evident in the networks but this property can straightforwardly be extracted, and modelled by P_{ij }= γk_{i}k_{j}, where the proportionality constant γ is determined from the degrees of the proteins.
Having demonstrated that PPI networks exhibit degreeweighted behaviour up to a certain value of the degree product k_{1}k_{2}, we turn our attention to these nonconforming regions of the networks. Of the networks analyzed here, only that of P. falciparum (Figure 1i) shows a degreeweighted tendency throughout. In terms of the number of interactions, this network is the second smallest only to that of the highconfidence network of Caenorhabditis elegans (Figure 1h), which shows more consistent behaviour in the highdegree product range than the other networks. However, there does not seem to be any association between levels of consistency and the sizes of the PPI networks. The nature of the deviations from degreeweighted behaviour is similar in all networks (Figure 1) and consists of a levelling off in values of P(k_{1}, k_{2}) together with increased variability. An important observation is that the probabilities of interaction in these highdegree areas are still quite high when compared to the wellbehaved, lowerdegree interaction regions. Thus, even though the highdegree nodes (or hub proteins) do not seem to obey degreeweighted behaviour, they still prefer to interact with each other rather than with lowerdegree proteins. These findings are similar to that reported previously, in that the hub proteins act somewhat differently to the remainder of the proteins [25]. However, in contrast, we find that interactions between hub proteins have high probability compared to an interaction between lowdegree nodes. It has been commonly accepted that hubs in a network avoid each other [25], however, we do not find this to be so.
Impact of degreeweighted behaviour upon network topology
Further insight regarding the hubhub connectivity, as well as the overall topology of the PPI networks, can be gained by determining the average path lengths L(k_{1}, k_{2}) between proteins of degrees k_{1 }and k_{2}. We investigate such maps for each of the PPI networks studied here. Figure 2 illustrates maps for the networks of Homo sapiens and Drosophila melanogaster, while maps for the remaining networks are provided in Additional file 2. As expected, the lowestdegree proteins are typically separated by the largest number of links, and the distance between proteins is decreased as either, or both, of their degrees are increased. For H. sapiens, this trend extends through to the highdegree interacting proteins, as most of the hubs are separated by only one or two links, indicating that they do not avoid each other. Therefore, this network incorporates a scalerich element [26] as well as a hierarchical nature in that the hub proteins are somewhat interconnected and generally closer to higherdegree proteins. For D. melanogaster, the hubs appear slightly less clustered than in H. sapiens, with separations of mostly one, two, and three links. If most of the shortest paths between proteins traverse the interconnected hub areas, then this explains why the overall average path length for the H. sapiens network (4.28) is smaller than for the D. melanogaster network (4.41) even though the former is much larger. Maps for all PPI networks studied here show similar features. The only real differences are in the connectivities of the highdegree proteins, which can be enhanced (H. sapiens, E. coli, C. elegans, P. falciparum) or slightly diminished (D. melanogaster, Saccharomyces cerevisiae).
Additional file 2. Supplementary Figure – Distance profiles in proteinprotein interaction networks. ac, Saccharomyces cerevisiae (yeast): YeastDIP, YeastCORE, YeastY2H; d, Escherichia coli (bacterium); ef, Caenorhabditis elegans (nematode): WormY2H, WormCORE; g, Plasmodium falciparum (malariacausing parasite). Distances shown as average shortest path lengths L(k_{1}, k_{2}) between proteins of degrees k_{1 }and k_{2}.
Format: PDF Size: 507KB Download file
This file can be viewed with: Adobe Acrobat Reader
Figure 2. Distance profiles in two proteinprotein interaction networks. a, Homo sapiens; b, Drosophila melanogaster. Distances shown as average shortest path lengths L(k_{1}, k_{2}) between proteins of degrees k_{1 }and k_{2}.
The path length maps clearly demonstrate that almost every highdegree protein has another highdegree protein located within one or two steps. This implies that any existing modules that incorporate one or more highdegree proteins are very likely to overlap or neighbour each other. As such, it is doubtful that isolated modules, or dense clusters, will contain highdegree proteins. Rather, they might contain proteins of more modest degree. However, the maps also indicate that any protein is, on the average, within three steps of a hub. Therefore, isolated complexes, if they exist, are likely to be few steps away from a highdegree protein. The observed trends in degreeweighted behaviour and shortest path lengths suggests that the PPI networks are extremely dense in their core, or interconnected hub region, and become somewhat sparser as the number of steps from the core is increased. Therefore, if any concentrated clusters are identified by some graph theoretical criterion, then there are probably many other complexes satisfying, or very nearly satisfying, this criterion. Thus, the concept of an isolated module becomes indistinct.
Analogy between degreeweighted connectivity and randomness
To illustrate the concept of inherent randomness in networks displaying degreeweighted behaviour, we show how ErdösRényi (ER) random graphs [10,11] also exhibit a degreeweighted characteristic. ER random networks have degree distributions that are Poissonian about the average degree and are, therefore, different from those of PPI networks, which show powerlaw scaling. Nonetheless, examination of the connectivity profile in ER networks will shed light on the interpretation of degreeweighted behaviour. The model we studied here is an ER random graph equivalent to the PPI network of P. falciparum, i.e., the probability of any edge is determined from the number of nodes and edges in the network of P. falciparum. We analyzed the extent of degreeweighted behaviour in this ER model by computing the probabilities P(k_{1}, k_{2}) of interaction between two proteins of degrees k_{1 }and k_{2 }over 10^{4 }realizations of the network. However, each probability P(k_{1}, k_{2}) was only averaged over the number of generated networks that contain nodes of degree k_{1 }and k_{2}. The reason for this is that nodes of higher degree may not occur in every realization. The resulting relationship between the probability of interaction P(k_{1}, k_{2}) and the degree product k_{1}k_{2 }is shown in Figure 3. As expected, this plot clearly suggests that ER random networks are inherently degree weighted, i.e., P(k_{1}, k_{2}) ∝ k_{1}k_{2}. The results further suggest that degreeweighted behaviour may be an indication, or property, of randomness. The ramifications of these findings are not immediately obvious and further analysis is required to comprehensively assess whether PPI networks incorporate random elements. However, our preliminary analysis indicates that randomness may play a significant role in these biological networks.
Figure 3. Degree weighted connectivity in the ErdösRényi random graph model equivalent to the PPI network of P. falciparum (1304 nodes, 2745 edges). Probabilities of interaction P(k_{1}, k_{2}) are calculated for 10^{4 }realizations, which are then averaged over the number of simulated networks that contain nodes of degree k_{1 }and k_{2}.
Discussion and Conclusion
The degreeweighted nature of PPI networks as well as the hub grouping present a quandary in that it implies that the assembly of these networks may be less biologically guided and more probabilistic in nature. One reason for this may be that highthroughput methods make little assumption about a protein's locality in the cell and therefore allow for more interactions than might be observed in vivo. In fact, only 40–50% of the identified interactions from highthroughput yeast twohybrid (Y2H) analyses of S. cerevisiae were between proteins occurring in the same cellular compartment [27,28]. However, the YeastCORE PPI network, which is considered to be high confidence and has a high conservation of interactions between proteins of the same compartment [29], exhibits a high level of degreeweighted behaviour (Figure 1d). Another consideration is that the various approaches to identify proteinprotein interactions unintentionally bias their collation from the different functional and cell component categories [28]. However, all the PPI networks studied here show similar degreeweighted connectivity even though five of them (Figures 1b (D. melanogaster), 1e (S. cerevisiae), 1g–h (C. elegans), and 1i (P. falciparum)) are almost completely determined from Y2H screens, while the remaining four are compiled from a variety of experimental sources (see Additional file 1).
It could also be that PPI networks determined from highthroughput methods contain nonspecific interactions. Such variability is not unexpected considering the large amount of irreproducibility of onceidentified interactions [30]. In such a case, we might expect to see similar probabilistic behaviour as that observed here. Contrary to this, though, the highconfidence network of C. elegans [30], which contains interactions found in three independent repeated experiments, exhibits clear degreeweighted characteristics (Figure 1h).
Obviously, proteinprotein interactions are necessary for a myriad of biological processes, however, if the event is "controlled" by other time and locationdependent processes, the actual binding or interaction could be of secondary importance. If degreeweighted behaviour is observed in a network, i.e., if protein interactions appear probabilistic, an analysis of expected binding events will determine whether the observed binding events are guided by their interactions or just by their ability to bind. This will greatly enhance the capability of interpreting and extracting biological information from proteinprotein interaction networks. The findings presented here provide a cautionary note on the biological interpretation of large PPI networks. One interpretation of the observed degreeweighted networks is that we are witnessing the ability to bind, and not necessarily what connections/interactions are actually present in the cell. The true biological connections that are used in a pathway or biological process cannot be backengineered from this type of data without taking into account a degreeweighted model, and hence the topological study of a degreeweighted network requires a more refined methodology to extract biological information about pathways, modules, or other inferred relationships among proteins. A priori knowledge of a protein's degree or connectivity is not available, however, algorithms to predict this [31,32], as well as their interactions [3234], are being developed. Whether application of these predictive algorithms on genomic scales yield degreeweighted networks remains to be seen, and may even serve as a test for the verity of the resultant network topologies.
Further insight into the degreeweighted nature of PPI networks may be obtained from analyses of the interacting protein pairs at more elementary levels. An avenue for this dissection has been to characterize the structural and functional domains present in each protein [35,36] and subsequently identify consistent signatures, i.e., pairs of domains that are more likely to be involved in binding [37,38]. In this way, domaindomain interaction (DDI) networks can be derived and then compared against PPI networks to see if they have similar topological properties such as degreeweighted behaviour. If, for example, degreeweighted behaviour is not observed in DDI networks, then one would anticipate consistent precepts for the allowed interactions, thereby allowing for alternative, and more insightful, analyses of PPI networks.
One utility of knowing that a network is degreeweighted is to use the probabilistic interpretation to find nodes that deviate from degreeweighted probability. Such nodes would represent a potential network that is biologically deterministic by its proteinprotein interactions alone. For example, clusters of lowdegree proteins might imply selective complex formation, and hubs found to be isolated from other highdegree proteins may represent important bottlenecks.
Authors' contributions
All authors contributed to the design and coordination of the study. JI performed the computational implementations and prepared the original draft, which was revised by AW and JR. All authors read and approved the final manuscript.
Acknowledgements
We thank the referees for valuable feedback which helped improve the paper. The authors were supported, in part, by the Military Operational Medicine research program of the U.S. Army Medical Research and Materiel Command, Ft. Detrick, Maryland. This effort was supported by the U.S. Army's Network Science initiative. The opinions and assertions contained herein are the private views of the authors and are not to be construed as official or as reflecting the views of the U.S. Army or the U.S. Department of Defense. This paper has been approved for public release with unlimited distribution.
References

Kitano H: Systems biology: a brief overview.
Science 2002, 295(5560):16621664. PubMed Abstract  Publisher Full Text

Aloy P, Russell RB: Structural systems biology: modelling protein interactions.
Nat Rev Mol Cell Biol 2006, 7(3):188197. PubMed Abstract  Publisher Full Text

Fields S: Highthroughput twohybrid analysis. The promise and the peril.
The FEBS journal 2005, 272(21):53915399. PubMed Abstract  Publisher Full Text

Gavin AC, Bosche M, Krause R, Grandi P, Marzioch M, Bauer A, Schultz J, Rick JM, Michon AM, Cruciat CM, Remor M, Hofert C, Schelder M, Brajenovic M, Ruffner H, Merino A, Klein K, Hudak M, Dickson D, Rudi T, Gnau V, Bauch A, Bastuck S, Huhse B, Leutwein C, Heurtier MA, Copley RR, Edelmann A, Querfurth E, Rybin V, Drewes G, Raida M, Bouwmeester T, Bork P, Seraphin B, Kuster B, Neubauer G, SupertiFurga G: Functional organization of the yeast proteome by systematic analysis of protein complexes.
Nature 2002, 415(6868):141147. PubMed Abstract  Publisher Full Text

Ho Y, Gruhler A, Heilbut A, Bader GD, Moore L, Adams SL, Millar A, Taylor P, Bennett K, Boutilier K, Yang L, Wolting C, Donaldson I, Schandorff S, Shewnarane J, Vo M, Taggart J, Goudreault M, Muskat B, Alfarano C, Dewar D, Lin Z, Michalickova K, Willems AR, Sassi H, Nielsen PA, Rasmussen KJ, Andersen JR, Johansen LE, Hansen LH, Jespersen H, Podtelejnikov A, Nielsen E, Crawford J, Poulsen V, Sorensen BD, Matthiesen J, Hendrickson RC, Gleeson F, Pawson T, Moran MF, Durocher D, Mann M, Hogue CW, Figeys D, Tyers M: Systematic identification of protein complexes in Saccharomyces cerevisiae by mass spectrometry.
Nature 2002, 415(6868):180183. PubMed Abstract  Publisher Full Text

Zhu H, Bilgin M, Bangham R, Hall D, Casamayor A, Bertone P, Lan N, Jansen R, Bidlingmaier S, Houfek T, Mitchell T, Miller P, Dean RA, Gerstein M, Snyder M: Global analysis of protein activities using proteome chips.
Science 2001, 293(5537):21012105. PubMed Abstract  Publisher Full Text

Barabasi AL, Oltvai ZN: Network biology: understanding the cell's functional organization.
Nature reviews 2004, 5(2):101113. PubMed Abstract  Publisher Full Text

Przulj N: Graph Theory Analysis of ProteinProtein Interactions. In Knowledge Discovery in Proteomics. Edited by Jurisica I, Wigle DA. CRC Press; 2005.

Zhu X, Gerstein M, Snyder M: Getting connected: analysis and principles of biological networks.
Genes & development 2007, 21(9):10101024. PubMed Abstract  Publisher Full Text

Jeong H, Tombor B, Albert R, Oltvai ZN, Barabasi AL: The largescale organization of metabolic networks.
Nature 2000, 407(6804):651654. PubMed Abstract  Publisher Full Text

Jeong H, Mason SP, Barabasi AL, Oltvai ZN: Lethality and centrality in protein networks.
Nature 2001, 411(6833):4142. PubMed Abstract  Publisher Full Text

Wagner A: The yeast protein interaction network evolves rapidly and contains few redundant duplicate genes.
Mol Biol Evol 2001, 18(7):12831292. PubMed Abstract  Publisher Full Text

Barabasi AL, Albert R: Emergence of scaling in random networks.
Science 1999, 286(5439):509512. PubMed Abstract  Publisher Full Text

Barabasi AL, Albert R, Jeong H: Meanfield theory for scalefree random networks.
Physica A 1999, 272(1–2):173187. Publisher Full Text

Vazquez A, Flammini A, Maritan A, Vespignani A: Modeling of protein interaction networks.
Complexus 2003, 1:3844. Publisher Full Text

PastorSatorras R, Smith E, Sole RV: Evolving protein interaction networks through gene duplication.
Journal of theoretical biology 2003, 222(2):199210. PubMed Abstract  Publisher Full Text

Chung F, Lu L, Dewey TG, Galas DJ: Duplication models for biological networks.
J Comput Biol 2003, 10(5):677687. PubMed Abstract  Publisher Full Text

Barabasi AL, Ravasz E, Vicsek T: Deterministic scalefree networks.
Physica aStatistical Mechanics and Its Applications 2001, 299(3–4):559564. Publisher Full Text

Ravasz E, Somera AL, Mongru DA, Oltvai ZN, Barabasi AL: Hierarchical organization of modularity in metabolic networks.
Science 2002, 297(5586):15511555. PubMed Abstract  Publisher Full Text

Przulj N, Corneil DG, Jurisica I: Modeling interactome: scalefree or geometric?
Bioinformatics (Oxford, England) 2004, 20(18):35083515. PubMed Abstract  Publisher Full Text

Chung F, Lu L: The average distances in random graphs with given expected degrees.
Proceedings of the National Academy of Sciences of the United States of America 2002, 99(25):1587915882. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Przulj N, Higham DJ: Modelling proteinprotein interaction networks via a stickiness index.
Journal of the Royal Society, Interface/the Royal Society 2006, 3(10):711716. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Maslov S, Sneppen K: Specificity and stability in topology of protein networks.
Science 2002, 296(5569):910913. PubMed Abstract  Publisher Full Text

Tanaka R: Scalerich metabolic networks.
Physical review letters 2005, 94(16):168101. PubMed Abstract  Publisher Full Text

Sprinzak E, Sattath S, Margalit H: How reliable are experimental proteinprotein interaction data?
Journal of molecular biology 2003, 327(5):919923. PubMed Abstract  Publisher Full Text

von Mering C, Krause R, Snel B, Cornell M, Oliver SG, Fields S, Bork P: Comparative assessment of largescale data sets of proteinprotein interactions.
Nature 2002, 417(6887):399403. PubMed Abstract  Publisher Full Text

Deane CM, Salwinski L, Xenarios I, Eisenberg D: Protein interactions: two methods for assessment of the reliability of high throughput observations.
Mol Cell Proteomics 2002, 1(5):349356. PubMed Abstract  Publisher Full Text

Li S, Armstrong CM, Bertin N, Ge H, Milstein S, Boxem M, Vidalain PO, Han JD, Chesneau A, Hao T, Goldberg DS, Li N, Martinez M, Rual JF, Lamesch P, Xu L, Tewari M, Wong SL, Zhang LV, Berriz GF, Jacotot L, Vaglio P, Reboul J, HirozaneKishikawa T, Li Q, Gabel HW, Elewa A, Baumgartner B, Rose DJ, Yu H, Bosak S, Sequerra R, Fraser A, Mango SE, Saxton WM, Strome S, Van Den Heuvel S, Piano F, Vandenhaute J, Sardet C, Gerstein M, DoucetteStamm L, Gunsalus KC, Harper JW, Cusick ME, Roth FP, Hill DE, Vidal M: A map of the interactome network of the metazoan C. elegans.
Science 2004, 303(5657):540543. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Deeds EJ, Ashenberg O, Shakhnovich EI: A simple physical model for scaling in proteinprotein interaction networks.
Proceedings of the National Academy of Sciences of the United States of America 2006, 103(2):311316. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Thomas A, Cannings R, Monk NA, Cannings C: On the structure of proteinprotein interaction networks.
Biochemical Society transactions 2003, 31(Pt 6):14911496. PubMed Abstract  Publisher Full Text

Bowers PM, Pellegrini M, Thompson MJ, Fierro J, Yeates TO, Eisenberg D: Prolinks: a database of protein functional linkages derived from coevolution.
Genome biology 2004, 5(5):R35. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

von Mering C, Jensen LJ, Kuhn M, Chaffron S, Doerks T, Kruger B, Snel B, Bork P: STRING 7 – recent developments in the integration and prediction of protein interactions.
Nucleic acids research 2007, (35 Database):D358362. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Mistry J, Finn R: Pfam: a domaincentric method for analyzing proteins and proteomes.
Methods Mol Biol 2007, 396:4358. PubMed Abstract  Publisher Full Text

Mulder N, Apweiler R: InterPro and InterProScan: Tools for Protein Sequence Classification and Comparison.
Methods Mol Biol. 2007, 396:5970. PubMed Abstract  Publisher Full Text

Sprinzak E, Altuvia Y, Margalit H: Characterization and prediction of proteinprotein interactions within and between complexes.
Proceedings of the National Academy of Sciences of the United States of America 2006, 103(40):1471814723. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Sprinzak E, Margalit H: Correlated sequencesignatures as markers of proteinprotein interaction.
Journal of molecular biology 2001, 311(4):681692. PubMed Abstract  Publisher Full Text