One hub-one process: a tool based view on regulatory network topology
© Axelsen et al; licensee BioMed Central Ltd. 2008
Received: 17 August 2007
Accepted: 04 March 2008
Published: 04 March 2008
The relationship between the regulatory design and the functionality of molecular networks is a key issue in biology. Modules and motifs have been associated to various cellular processes, thereby providing anecdotal evidence for performance based localization on molecular networks.
To quantify structure-function relationship we investigate similarities of proteins which are close in the regulatory network of the yeast Saccharomyces Cerevisiae. We find that the topology of the regulatory network only show weak remnants of its history of network reorganizations, but strong features of co-regulated proteins associated to similar tasks. These functional correlations decreases strongly when one consider proteins separated by more than two steps in the regulatory network. The network topology primarily reflects the processes that is orchestrated by each individual hub, whereas there is nearly no remnants of the history of protein duplications.
Our results suggests that local topological features of regulatory networks, including broad degree distributions, emerge as an implicit result of matching a number of needed processes to a finite toolbox of proteins.
Contemporary systems biology have provided us with a large amount of data on topology of molecular networks, thereby giving us glimpses into computation and signaling in living cells. It have been found that 1) regulatory networks have broad out-degree distributions [1, 2], 2) transcriptional regulatory networks contains many feed forward motifs , and 3) highly connected hubs are often found on the periphery of the network . These findings are elements in understanding the topology of existing molecular networks as the result of an interplay between evolution and the processes they orchestrate in the cell.
In this paper we consider properties of proteins in the perspective of how they are positioned relative to each other in the network. This is in part motivated by the existence of highly connected proteins (hubs) and their relation to soft modularity [4, 5] in regulatory networks. In particular one may envision broad degree distributions and possible isolation of hubs as a reflection of a local "information horizon"  with partial isolation between different biological processes. We here address this problem by considering the yeast regulatory network  with regards to protein properties. Using the Gene Ontology (GO) Consortium annotations we will show that locality in the regulatory network primarily is associated to locality in biological process, and only weakly related to functional abilities of a protein.
More precisely, a GO-graph is an acyclic directed graph which organize proteins according to a predefined categorization. A lower ranking protein in a GO-graph share large scale properties with higher ranking proteins, but are more specialized. In the GO-database, proteins are categorized into three networks according to different annotations, ranking known gene products after respectively: P) biological process, F) functional ability/design of the protein and C) cellular components where the protein is physically located. For each of these three ways of categorization we examined two distinct ways to measure GO annotation difference (see box in Fig. 1).
In particular Fig. 2(a) shows that proteins separated by one or two links are involved in similar processes. Here distance l = 1 mostly count proteins on the periphery of a hub and their directly upstream and highly connected regulator. Distance l = 2 count proteins regulated by the same highly connected regulator. Note that we are averaging over all pairs in the whole regulatory network including connections to less well-connected regulators. In this way the highly connected nodes are counted for each of their downstream targets and therefore the larger hubs will make the dominant contributions to this calculation.
Figure 2(b) investigate the differences in GO-annotations, but with the hierarchical distance that emphasize differences close to the root of the GO-graph for processes(P). The fact that this measure correlate to larger distances in the regulatory network implies that proteins in a larger neighborhood of the regulatory network tends to be on the same larger subbranches on the GO(P)-hierarchy.
In all the panels in Fig. 2 we also compare to a null model, generated by keeping the regulatory network, but randomly reassigning which proteins from the GO-graph that are assigned to which positions on the network. This randomization maintain the positions of all nodes in the regulatory network exactly. By doing this randomization one loose any P, F or C correlation between a regulator and its downstream targets. Any conceivable GO-distance therefore becomes independent on the regulatory distance.
From Fig. 3(a) we see that in order to reproduce the observed local correlations of GO(P) in a random sample of networks, these need to be generated with maximal bias. That is, the network generated with ε = 0 reproduce observed correlations between processes of proteins which are downstream of the same regulator i.e. at distance l = 2 in the regulatory network. At distances l > 2 there are no detectable correlations, which in turn is reproduced by allowing small imperfections (ε ~ 0.15) in the rewiring.
In Fig. 3(b) we repeat the investigation from a), but with respect to the hierarchical GO(P) distance. In this case we see that ε ~ 0.15 → 0.30 reproduce the observed correlations between protein processes out to larger regulatory distances (l ~ 3). Figure 3(c)–(f), on the other hand, show that function or cellular localization are only moderately related within the same hub (l ~ 2), and unrelated at all larger distances.
Protein regulatory networks are highly functional information processing systems, evolved to perform a diverse sets of tasks in a close to optimal way. It is of no surprise that they are not random, also in ways that can be detected without knowing much about what actually goes on in the living system they regulate. However we do not, a priori, know much about the relative importance of function versus history: Is the topology of a network primarily governed by the processes it direct, or is its topology influenced by random gene duplications [9, 10] and "link" rewirings ?
Concerning gene duplications [9, 12–19], we detected 581 paralogous pairs among the 848 gene products in YPD, see methods. Of these 581 pairs, only ~15% significantly retained their common regulator, and only ~0.6% of the proteins pairs at distance l = 2 are detectable paralogs. Therefore the contribution from duplication events to any GO-similarity within hubs can be ignored.
Our analysis in Figs. 2, 3 emphasize the strong correlations between network localization and process, in particular very strong (maximally possible) correlation between process annotation of proteins in the same hub. In addition, we see some functional similarities between proteins in the same hub, in particular when considering the hierarchical GO distances at l = 2 in Fig. 3(d). However we also find that the functional diversity within hubs are large in terms of the direct GO distance (l = 2 in Fig. 3(c)). Combined Fig. 3(c,d) therefore show that proteins in the same hub have quite large direct function-GO distances, but rarely belong to entirely different function-GO categories.
In any case we emphasize that we primarily find GO-processes localized on hubs, and only weak correlations of the functional abilities between proteins involved in the same process.
The idea that process similarity are associated to network localization is not new, and implicitly behind attempts to infer gene networks from similarity in gene expression . In the supplement we use gene expression from micro-arrays to re-investigate the correlation between process and locality in the regulatory network. Thereby, we provide a broader support for our findings, and present a quantitative illustration of the extent to which gene-expression studies can be used to deduce co-regulation.
Support for the ubiquity of the "one hub-one process" association is also found from the fact that the likelihood that a regulatory protein is essential is nearly independent on how many proteins it regulate . That is, the question of whether a null mutant of a certain protein is viable is keyed to the essentiality of the regulated process, and not to whether the process needs many or few different "tools" to be performed.
Overall we suggest that the topology of the yeast regulatory network is governed by processes located on hubs, each consisting of a number of tools in the form of proteins with quite different functional abilities. This is consistent with a network evolution where gene duplication occur, but where rewiring of regulatory links plays a bigger role [14, 19, 21–23]. The regulatory network is designed to co-regulate processes, and its evolutionary history must include a bias towards hub-regulation of individual processes. Degree distributions are not broad because of duplication events, but because a given biological task sometimes needs many, but typically require few tools.
Finally our analysis have consequences for development of null models for network topologies, and thereby for identifying functionally important network motifs . While the previous null model  maintain in- and out- degrees of each protein, it ignore correlations associated to cellular process. When nearby proteins are associated to the same processes one statistically expect an increased probability for cliques [24, 25]. We therefore expect that some of the many feed-forward loops in transcription networks  will be explained by a new type of null model: A null model where proteins contributing to a given process are forced to remain close in the randomized network.
The GO-annotations are used without any filtering. This does not preclude bias introduced from using inferred annotations. Of the 848 genes in the YPD, 52 are not annotated and were thus not included in the analysis. 142 genes has more than one molecular function, 314 genes takes part in more than one cellular component and 463 genes participates in more than one biological process. To accommodate this the analysis was carried out by choosing the annotations which minimized the mutual distance for each pair of proteins. This choice maximally resolves significant signals, since we minimize the effect of the finite size of the GO-tree, and in the case of no signal this choice introduces no bias.
Of the 848 gene products in YPD, we found 581 paralogous pairs using BLASTP with E-value cutoff of 10-10 [14, 26]. For the YPD network 132 of these paralogous pairs are at distance l = 2. This should be compared to a null expectation of 50 ± 6 paralogous pairs at l = 2 found by randomizing the YPD network while keeping in- and out-degrees . Therefore at max 132-50 = 82 of the paralogous pairs are in the same hub due to their history of common origin. This correspond to 82/581 ~15% of duplicated proteins in YPD. The excess of 82 paralogous pairs at distance 2 should also be compared to the total of 13554 protein pairs that the YPD network have at distance l = 2. Thus only ~0.6% of all proteins pairs at l = 2 are detectable paralogs.
As seen in our Additional file 1, we reach the same basic conclusion of hubs being functionally isolated using a completely different approach based on gene expression data. Analyzing micro-array data from 482 stress experiments from Saccharomyces Genome Database  and managing the false discovery rate as in  we indeed find localization of perturbations on our regulatory network. Thus the appendix supports the robustness of our results to an independent categorization of protein processes.
We acknowledge the support from the Danish National Research Foundation through "Center for Models of Life" at the Niels Bohr Institute. KS and JBA wishes to thank the Lundbeck Foundation. JBA wish to thank The Eva and Henry Frænkel Memorial Foundation.
- Albert R, Barabasi AL: Rev Modern Phys. 2002, 74: 47-10.1103/RevModPhys.74.47.View ArticleGoogle Scholar
- Maslov S, Sneppen K: Phys Biol. 2005, 2: 94-10.1088/1478-3975/2/4/S03.View ArticleGoogle Scholar
- Milo R: Science. 2002, 298: 824- 10.1126/science.298.5594.824View ArticlePubMedGoogle Scholar
- Maslov S, Sneppen K: Science. 2002, 296: 910- 10.1126/science.1065103View ArticlePubMedGoogle Scholar
- Hartwell LH, Hopfield JJ, Leibler S, Murray AW: Nature. 1999, 402: C47- 10.1038/35011540View ArticlePubMedGoogle Scholar
- Trusina A, Rosvall M, Sneppen K: Phys Rev Lett. 2005, 94: 238701- 10.1103/PhysRevLett.94.238701View ArticlePubMedGoogle Scholar
- Costanzo MC: Nucleic Acids Res. 2001, 29: 75- 10.1093/nar/29.1.75PubMed CentralView ArticlePubMedGoogle Scholar
- Ashburner M: Nat Genet. 2000, 25: 25- 10.1038/75556PubMed CentralView ArticlePubMedGoogle Scholar
- Bhan A, Galas DJ, Dewey G: Bioinformatics. 2002, 18: 1486- 10.1093/bioinformatics/18.11.1486View ArticlePubMedGoogle Scholar
- Sole RV, Pastor-Satorras R, Smith E, Kepler TB: Adv Complex Syst. 2002, 5: 43-10.1142/S021952590200047X.View ArticleGoogle Scholar
- Bornholdt S, Sneppen K: Proc Roy Soc London B. 2000, 267: 2281-10.1098/rspb.2000.1280.View ArticleGoogle Scholar
- Wagner A: Mol Biol Evol. 2001, 18: 1283-View ArticlePubMedGoogle Scholar
- Gu Z: Nature. 2003, 421: 63- 10.1038/nature01198View ArticlePubMedGoogle Scholar
- Maslov S, Sneppen K, Eriksen KA, Yan KK: BMC Evolutionary Biology. 2004, 4: 9- 10.1186/1471-2148-4-9PubMed CentralView ArticlePubMedGoogle Scholar
- Teichmann SA, Babu MM: Nature Genetics. 2004, 36: 492- 10.1038/ng1340View ArticlePubMedGoogle Scholar
- Koonin EV, Wolf YI, Karev GP: "Power Laws, Scale-Free Networks and Genome Biology" Springer, ISBN: 0387258833. 2006View ArticleGoogle Scholar
- Rodriguez-Caso C, Medina MA, Sole RV: FEBS Journal. 2005, 272: 6423- 10.1111/j.1742-4658.2005.05041.xView ArticlePubMedGoogle Scholar
- Foster DV, Kauffman SA, Socolar JES: Phys Rev E. 2006, 73: 031912-10.1103/PhysRevE.73.031912.View ArticleGoogle Scholar
- Enemark J, Sneppen K: J Stat Mech. 2007, P11007-10.1088/1742-5468/2007/11/P11007.Google Scholar
- Haverty PM, Hansen U, Weng Z: Nucleic Acids Research. 2004, 32: 179- 10.1093/nar/gkh183PubMed CentralView ArticlePubMedGoogle Scholar
- Gu Z, Nicolae D, Lu H-S, Li W-H: Trends in genetics. 2002, 18: 609- 10.1016/S0168-9525(02)02837-8View ArticlePubMedGoogle Scholar
- Berg J, Lassig M, Wagner A: BMC Evolutionary Biology. 2004, 4: 51- 10.1186/1471-2148-4-51PubMed CentralView ArticlePubMedGoogle Scholar
- Ihmels J, Bergmann S, Gerami-Nejad M, Yanai I, McClellan M, Berman J, Barkai N: Science. 2005, 309: 938- 10.1126/science.1113833View ArticlePubMedGoogle Scholar
- Trusina A, Maslov S, Minnhagen P, Sneppen K: Phys Rev Lett. 2004, 92: 178702- 10.1103/PhysRevLett.92.178702View ArticlePubMedGoogle Scholar
- Axelsen JB, Bernhardsson S, Rosvall M, Sneppen K, Trusina A: Phys Rev E Stat Nonlin Soft Matter Phys. 2006, 74: 036119-View ArticlePubMedGoogle Scholar
- Axelsen JB, Yan KK, Maslov S: Biology Direct. 2007, 2: 32- 10.1186/1745-6150-2-32PubMed CentralView ArticlePubMedGoogle Scholar
- SGD project. "Saccharomyces Genome Database"., ftp://ftp.yeastgenome.org/yeast/
- Benjamini Y, Drai D, Elmer G, Kafkafi N, Golani I: Behav Brain Res. 2001, 125: 279- 10.1016/S0166-4328(01)00297-2View ArticlePubMedGoogle Scholar