Gene expression as governed by the interplay of the components of regulatory networks is indeed one of the most complex fundamental processes in biological systems. Although several methods have been published to unravel the hierarchical structure of regulatory networks, weaknesses such as the incorrect or inconsistent assignment of elements to their hierarchical levels, the incapability to cope with cyclic dependencies within the networks or the need for a manual curation to retrieve non-overlapping levels remain unsolved.
We developed HiNO as a significant improvement of the so-called breadth-first-search (BFS) method. While BFS is capable of determining the overall hierarchical structures from gene regulatory networks, it especially has problems solving feed-forward type of loops leading to conflicts within the level assignments. We resolved these problems by adding a recursive correction approach consisting of two steps. First each vertex is placed on the lowest level that this vertex and its regulating vertices are assigned to (downgrade procedure). Second, vertices are assigned to the next higher level (upgrade procedure) if they have successors with the same level assignment and have themselves no regulators. We evaluated HiNO by comparing it with the BFS method by applying them to the regulatory networks from Saccharomyces cerevisiae and Escherichia coli, respectively. The comparison shows clearly how conflicts in level assignment are resolved in HiNO in order to produce correct hierarchical structures even on the local levels in an automated fashion.
We showed that the resolution of conflicting assignments clearly improves the BFS-method. While we restricted our analysis to gene regulatory networks, our approach is suitable to deal with any directed hierarchical networks structure such as the interaction of microRNAs or the action of non-coding RNAs in general. Furthermore we provide a user-friendly web-interface for HiNO that enables the extraction of the hierarchical structure of any directed regulatory network.
HiNO is freely accessible at http://mips.helmholtz-muenchen.de/hino/.
Citation: Hartsperger ML, Strache R, Stümpflen V (2010) HiNO: An Approach for Inferring Hierarchical Organization from Regulatory Networks. PLoS ONE 5(11): e13698. https://doi.org/10.1371/journal.pone.0013698
Editor: Francesco Falciani, University of Birmingham, United Kingdom
Received: July 6, 2010; Accepted: September 27, 2010; Published: November 4, 2010
Copyright: © 2010 Hartsperger et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This work was supported by the Helmholtz Alliance on Systems Biology (project ‘CoReNe’), the Federal Ministry of Education and Research (project SysMBo, FKZ: 0315494A) and the Graduate School for Information Science in Health (GSISH). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
One of the fundamental problems in systems biology is the understanding of regulatory networks and the interdependencies between their components. Of particular importance is how transcription factors (TFs) mutually coordinate the expression of thousands of genes and noncoding RNAs (ncRNAs)  like microRNAs (miRNAs) or long ncRNAs in response to various stimuli. The qualitative interaction between TFs, miRNAs and their targets can be modeled in terms of directed regulatory networks.
The determination of hierarchical interdependencies within gene regulatory networks (GRNs) is important for the understanding of the potential impact of perturbations on underlying cellular processes and their correlated diseases , . Yu and Gerstein  pointed out that in contrast to well investigated undirected biological networks the superior feature of GRNs is their directedness and hence their orientation towards control rather than communication. The direct implication is a hierarchical organization, where a higher-level structure results from multiple instances of lower-level structure of different types . As a consequence small groups of nodes organize in a hierarchical manner to increasingly large groups on many different scales –.
It is evident that the correct determination of the overall topology of the network as well as the unambiguous prediction of the position of individual regulatory elements are essential for the comprehensive understanding of regulatory effects on a systems level (see Figure 1). For example, master regulators like HOXA10 display significant downstream effects in postnatal hematopoietic development depending on different concentrations of the key regulator . This can lead to complete blocking of erythroid and megakaryocyte development for high concentrations of HOXA10, whereas intermediate concentrations result in increased stem cell proliferation. In contrast to master regulators the so called mid-level regulators can not only be mediators of different incoming regulatory signals but also influence multiple downstream components or pathways. In the case of TP53 as an important factor in cancer development it has long been known to be influenced by multiple upstream signals such as stress leading to different downstream effects such as apoptosis or development of cancer , .
An abstract example of a directed regulatory network is shown (A) unstructured and (B) hierarchically structured. We schematically show in (C) and (D) that for master regulators like HOXA10 significant downstream effects in postnatal hematopoietic development depending on different concentrations can be observed. In contrast to master regulators so called mid-level regulators can not only be mediators of different incoming regulatory signals but also influence multiple downstream components or pathways. We schematically display that TP53 is influenced by various upstream signals such as (E) DNA damage or (F) oncogenes leading to different downstream effects such as apoptosis or development of cancer. The detailed understanding of the hierarchical topology is necessary to comprehend the dynamic behavior and the possible malfunction of regulatory networks.
The detailed understanding of the hierarchical topology is essential to understanding the dynamic behavior and the possible malfunction of regulatory networks. This can be achieved only by accurate determination of the hierarchical organization.
Several methods exist that identify the hierarchical structure within directed regulatory networks like the leaf removal algorithm , , or the breadth-first-search (BFS) method . Although these methods are capable of extracting hierarchical structures from GRNs, they have several shortcomings as discussed by Jothi et al. : They can either extract the hierarchical structure incorrectly or they are not scalable. Another important drawback is that they are only partly applicable to networks containing loops. However, loops are an important feature of regulatory networks . Network motifs represent the building blocks of complex networks and their properties determine the local and global organization , . Common networks motifs are feed-forward-loops (FFLs) or feedback loops (FBLs). The leaf-removal algorithm that was used for network decomposition and to infer hierarchical structures in biological networks ,  is therefore not applicable to GRNs containing loops. The BFS approach presented by  can be applied to cyclic networks, but the inferred hierarchical structure may contain conflicts in the level assignment. Jothi et al.  presented a method capable of dealing with these shortcomings. However, it does not determine a clear hierarchical position for each vertex in the network rather vertices are assigned to an interval of possible positions. The algorithm does not count or enumerate all feasible topological orderings of nodes in a network because this is a NP-hard problem. It only outputs a linear ordering of nodes containing all feasible solutions rather than reporting just a single solution . Subsequently the final transformation into a graph has to be done manually which is a quite cumbersome procedure for large regulatory networks.
Here, we present significant improvements of the BFS approach that directly reveal the hierarchical structure from GRNs by considering the occurrence of network motifs. This idea is implemented by expanding the BFS method with two correction steps, a “downgrade” and an “upgrade” procedure. Therefore we can determine the accurate hierarchy of a GRN that we define as its hierarchical structure without having any inconsistencies in the level assignments. To evaluate our method we applied both algorithms to the regulatory networks from Saccharomyces cerevisiae (YRN) and Escherichia coli (ERN), respectively. Our results demonstrate that the introduction of the correction steps clearly overcomes the limitations of the BFS-method. We provide a user-friendly web-interface to HiNO that enables the extraction of the hierarchical structure of any directed regulatory network.
Materials and Methods
GRNs were reconstructed using publicly available data sources. We downloaded and reconstructed the GRNs for S. cerevisiae and E. coli from http://www.gersteinlab.org/proj/nethierarchy/.
The E. coli gene regulatory network (ERN) consisted of regulatory interactions between TFs and targets. The extracted TF-subgraph consisted of interactions between TFs with a total of self-loops (auto-regulatory edges). The S. cerevisiae gene regulatory network (YRN) contained interactions involving TFs and targets. The extracted TF-subgraph consisted of interactions involving TFs with a total of self-loops. For detailed numbers see Tables 1 and 2.
Design and implementation
HiNO – an algorithm for inferring hierarchical organization from regulatory networks.
A directed network is a graph of directed edges between a set of vertices . Given a directed network, our algorithm first identifies all vertices at the bottom representing the lowest level in the hierarchy. A vertex is assigned to the bottom level if and only if it does not regulate any other vertices or if it is only regulating itself (i.e. auto-regulation). In the next step, we traverse the network into a “shortest-path-tree” . Here, we define the level in the hierarchy for any non-bottom vertex as its shortest distance to a bottom vertex. Now each vertex is assigned to a certain level within the hierarchy. However, in the case of occurring loops it can be observed that regulators have a lower level annotation than their targets . To account for the occurrence of loops in the network (such as feed-forward loops) and to extract the underlying hierarchical structure accurately we added a recursive correction approach consisting of two steps. First, vertices are assigned to the lowest level that this and its regulating vertices are assigned to. This “downgrade” step resolves conflicts in level assignments where regulators have a lower level annotation than their targets. Second, vertices with no predecessors are placed to the next higher level if some of their successors are located on the same level. In this “upgrade” step vertices that are regulators only get the next higher level annotation than their targets. Finally, the hierarchical structure of the directed regulatory network is determined without conflicts in the vertices' level assignments. The hierarchy extraction procedure is illustrated in Figure 2. A pseudocode representation of HiNO is shown in Figure 3.
We illustrate the approach of hierarchy extraction from directed networks. In the first step, the original network (grey) is traversed into a shortest path-tree (step1) with an initial hierarchical organization consisting of three layers. As this alone does not capture all feasible solutions we introduce two correction steps to resolve conflicts in the level assignments (e.g. node 1 is located in level 2, but it is regulating node 4 that is located in level 3). First, in the “downgrade” step each vertex is placed on the lowest level that this vertex and its regulating vertices are assigned to (step2). Thereby it is assured that each regulator has at least the same level assignment as its targets. Here, node 4 and node 5 are downgraded from level 3 to level 2 as its regulators - node 1 and node 3- are assigned to level 2. Second, vertices are “upgraded” if they have successors with the same level assignment and have no regulators themselves. Here, node 1 has no predecessor and targets on the same level (level 2). Consequently it is upgraded from level 2 to level 3 (step3). In the end HiNO determined a three layered hierarchical structure of the directed network without conflicts in the level assignment.
We implemented a user-friendly web application providing the functionality of HiNO, for illustration see Figure 4. Directed regulatory networks can be uploaded as text files. Users can select the node and edge types that should be used for the analysis. Additionally, TF annotation can be done for human or mouse regulatory networks using the transcription factor atlas presented by . The deduced subgraph then consists of selected node and edge types only and from this its hierarchical structure is extracted. Finally, statistics on level distribution and a graphical representation of the hierarchical structure can be downloaded as text file or GraphML (an XML-based file format for graphs) file. The web interface is freely available online at http://mips.helmholtz-muenchen.de/hino/.
An illustration of the user-friendly web-interface for HiNO is shown. The user can upload (A) a text file containing the representation of any directed network. Afterwards (B) the user retrieves information about node and edge types within the uploaded network and can decide which types should be used for the hierarchy extraction. Additionally, TF annotation can be added for human or mouse networks. (C) Information about the extracted hierarchical structure can be downloaded either as text file or in a graphical representation.
The source code of HiNO is available at https://sourceforge.net/projects/hino.
We applied HiNO to the ERN and the YRN to deduce their hierarchical structure. For this, we reconstructed the GRNs provided by  and subsequently extracted sub-graphs containing TFs and regulatory interactions between TFs only. The hierarchical topology is deduced only from the regulatory interdependencies of transcription factors. In Figure 5 the different analysis steps are shown, precise numbers are shown in Tables 1,2. To evaluate our method we compared the hierarchical structures extracted with the results when applying the BFS method of  (see Tables 3,4). The complete hierarchical structure of the ERN and the YRN can be found at http://mips.helmholtz-muenchen.de/hino/help.jsp.
Illustration of (A) the GRN of E. coli with 1095 nodes and 2044 edges and (D) the GRN of S. cerevisiae with 3458 nodes and 8371 edges. Red nodes indicate transcription factors, green nodes genes. For extracting the hierarchical structure we deduced transcriptional regulation only. The transcriptional regulatory networks (TRNs) are shown in (B) for E. coli and (E) S. cerevisiae only. A snapshot of the hierarchical structure of the TRNs is shown for E. coli in (C) and for S. cerevisiae in (F). The TRN of E. coli has a four-layered pyramidal-shaped hierarchical structure, whereas the TRN of S. cerevisiae consists only of three layers. The different colors represent the distinct hierarchical levels: level 4– yellow; level 3– red; level 2– green; level 1– blue.
The E. coli regulatory network (ERN)
Applying HiNO to the ERN reveals a pyramidal-shaped hierarchical structure consisting of four levels with most TFs on the bottom level (here level 1) and only a few regulators on the top (level 4). The global findings are consistent with the results when applying the method of  where also four levels could be identified. However, the BFS-method incorrectly assigns four TFs to the top level, whereas our method only assigns two TFs to this layer. The misassigned two elements, gntR and yhiW are placed by HiNO on the next lower level due to the correction step procedure. Both TFs regulate yhiE assigned to level 3 by . This TF is regulated by several TFs assigned to level 2 and therefore the hierarchical structure is incorrect as a target has a higher level annotation than its regulator. In contrast, HiNO directly assigns yhiE to level 2 and subsequently downgrades the level annotation for its regulators gntR and yhiW to level 3. In the “upgrade” step TFs are assigned to the next higher level if they are not regulated at all and if they have targets on the same hierarchical layer. Exemplarily, this is the case for yieH which is upgraded from level 2 to level 3. For illustration see Figure 6, details on the level distribution are shown in Table 3.
A snapshot of the results (the comparison of the BFS-method and HiNO) using the TRN of E. coli is shown. Node colors indicate the level assignment in HiNO: blue - level 1; green - level 2; red - level 3; yellow - level 4. Dashed edges represent additional regulatory interdependencies (further up- and/or down-stream factors) that are not shown. The results of the BFS method are shown in (A). The node coloring indicates conflicts in level assignments. In (B) the result of the “downgrade” step of HiNO is shown. In (C) the result of the “upgrade” step is shown. This is also the final hierarchical assignment of the elements by HiNO.
The S. cerevisiae regulatory network (YRN)
Applying HiNO to the YRN we extracted a pyramidal-shaped hierarchical structure consisting of three levels. In comparison to this the method of  reveals a four layered topology. The BFS method placed ADA2 on the top level (level 4) although it is regulated by RAP1 that is assigned to the second lowest level (level 2). This hierarchical assignment is inconsistent to the fact that a regulator has to be at least on the same hierarchical layer as its target. Similar conflicts can also be observed for SPT23, NGG1, GAT1 or MOT3 having all predecessors with lower level annotations (see Figure 7). In contrast to this HiNO assigns the correct level in the hierarchy to these TFs. In the “downgrade” step, ADA2 is placed on the same level as its regulators RAP1 and FHL1 (level 2). Interdependencies are resolved subsequently by downgrading other factors such as RTG3 or MSN4. Afterwards, elements are “upgraded” to the next higher level if they have no upstream elements and if they are positioned at the same layer as one of their targets. This step is illustrated in Figure 7 for HAL9 which is upgraded from level 2 to level 3. Details on the level distribution are shown in Table 4.
A snapshot of the results (the comparison of the BFS-approach and HiNO) using the TRN of S. cerevisiae is shown. Node colors indicate the level assignment in HiNO: blue - level 1; green - level 2; red - level 3; yellow - level 4. Dashed edges represent additional regulatory interdependencies (further up- and/or down-stream factors) that are not shown. The results of the BFS method are shown in (A). The node coloring indicates conflicts in level assignment. In (B) the result of the “downgrade” step of HiNO is shown. In (C) the result of the “upgrade” step is shown. This is also the final hierarchical assignment of the elements by HiNO.
We presented HiNO a method that significantly improves the BFS approach presented by Yu and Gerstein  that directly extracts the hierarchical structure from GRNs considering the occurrence of network loops. While the BFS method returns the same hierarchical organization of the network, inconsistencies on the individual levels in local neighborhoods can be resolved by HiNO. Small, local motifs and modules define the basic dynamic properties of regulatory processes . Their correct identification is necessary for the deduction of experimentally testable hypotheses deduced from large data sets. Our approach expands the BFS method with two correction steps, a “downgrade” and an “upgrade” procedure. We compared HiNO and the BFS-method using the YRN and the ERN data sets. Results demonstrate that our correction procedure clearly improves the BFS-method as it is able to reveal a hierarchical structure without inconsistencies. We provide a user-friendly web-interface that enables the extraction of the hierarchical structure of any directed network.
Conceived and designed the experiments: MLH VS. Performed the experiments: MLH. Analyzed the data: MLH. Contributed reagents/materials/analysis tools: MLH RS. Wrote the paper: MLH VS.
- 1. Häsler J, Strub K (2006) Alu elements as regulators of gene expression. Nucleic Acids Res 34: 5491–5497.J. HäslerK. Strub2006Alu elements as regulators of gene expression.Nucleic Acids Res3454915497
- 2. He L, Hannon GJ (2004) Micrornas: small rnas with a big role in gene regulation. Nat Rev Genet 5: 522–531.L. HeGJ Hannon2004Micrornas: small rnas with a big role in gene regulation.Nat Rev Genet5522531
- 3. Hobert O (2008) Gene regulation by transcription factors and micrornas. Science 319: 1785–1786.O. Hobert2008Gene regulation by transcription factors and micrornas.Science31917851786
- 4. Yu H, Gerstein M (2006) Genomic analysis of the hierarchical structure of regulatory networks. Proc Natl Acad Sci U S A 103: 14724–14731.H. YuM. Gerstein2006Genomic analysis of the hierarchical structure of regulatory networks.Proc Natl Acad Sci U S A1031472414731
- 5. Clauset A, Moore C, Newman MEJ (2008) Hierarchical structure and the prediction of missing links in networks. Nature 453: 98–101.A. ClausetC. MooreMEJ Newman2008Hierarchical structure and the prediction of missing links in networks.Nature45398101
- 6. Ravasz E, Barabási AL (2003) Hierarchical organization in complex networks. Phys Rev E Stat Nonlin Soft Matter Phys 67: 026112.E. RavaszAL Barabási2003Hierarchical organization in complex networks.Phys Rev E Stat Nonlin Soft Matter Phys67026112
- 7. Barabási AL, Oltvai ZN (2004) Network biology: understanding the cell's functional organization. Nat Rev Genet 5: 101–113.AL BarabásiZN Oltvai2004Network biology: understanding the cell's functional organization.Nat Rev Genet5101113
- 8. Balázsi G, Barabási AL, Oltvai ZN (2005) Topological units of environmental signal processing in the transcriptional regulatory network of escherichia coli. Proc Natl Acad Sci U S A 102: 7841–7846.G. BalázsiAL BarabásiZN Oltvai2005Topological units of environmental signal processing in the transcriptional regulatory network of escherichia coli.Proc Natl Acad Sci U S A10278417846
- 9. Magnusson M, Brun ACM, Miyake N, Larsson J, Ehinger M, et al. (2007) Hoxa10 is a critical regulator for hematopoietic stem cells and erythroid/megakaryocyte development. Blood 109: 3687–3696.M. MagnussonACM BrunN. MiyakeJ. LarssonM. Ehinger2007Hoxa10 is a critical regulator for hematopoietic stem cells and erythroid/megakaryocyte development.Blood10936873696
- 10. Levine AJ (1997) p53, the cellular gatekeeper for growth and division. Cell 88: 323–331.AJ Levine1997p53, the cellular gatekeeper for growth and division.Cell88323331
- 11. Vogelstein B, Lane D, Levine AJ (2000) Surfing the p53 network. Nature 408: 307–310.B. VogelsteinD. LaneAJ Levine2000Surfing the p53 network.Nature408307310
- 12. Ma HW, Buer J, Zeng AP (2004) Hierarchical structure and modules in the escherichia coli transcriptional regulatory network revealed by a new top-down approach. BMC Bioinformatics 5: 199.HW MaJ. BuerAP Zeng2004Hierarchical structure and modules in the escherichia coli transcriptional regulatory network revealed by a new top-down approach.BMC Bioinformatics5199
- 13. Ma HW, Kumar B, Ditges U, Gunzer F, Buer J, et al. (2004) An extended transcriptional regulatory network of escherichia coli and analysis of its hierarchical structure and network motifs. Nucleic Acids Res 32: 6643–6649.HW MaB. KumarU. DitgesF. GunzerJ. Buer2004An extended transcriptional regulatory network of escherichia coli and analysis of its hierarchical structure and network motifs.Nucleic Acids Res3266436649
- 14. Jothi R, Balaji S, Wuster A, Grochow JA, Gsponer J, et al. (2009) Genomic analysis reveals a tight link between transcription factor dynamics and regulatory network architecture. Mol Syst Biol 5: 294.R. JothiS. BalajiA. WusterJA GrochowJ. Gsponer2009Genomic analysis reveals a tight link between transcription factor dynamics and regulatory network architecture.Mol Syst Biol5294
- 15. Shen-Orr SS, Milo R, Mangan S, Alon U (2002) Network motifs in the transcriptional regulation network of escherichia coli. Nat Genet 31: 64–68.SS Shen-OrrR. MiloS. ManganU. Alon2002Network motifs in the transcriptional regulation network of escherichia coli.Nat Genet316468
- 16. Oltvai ZN, Barabási AL (2002) Systems biology. life's complexity pyramid. Science 298: 763–764.ZN OltvaiAL Barabási2002Systems biology. life's complexity pyramid.Science298763764
- 17. Alon U (2007) Network motifs: theory and experimental approaches. Nat Rev Genet 8: 450–461.U. Alon2007Network motifs: theory and experimental approaches.Nat Rev Genet8450461
- 18. Cormen TH, Leiserson CE, Rivest RL, Stein C (2009) Introduction to Algorithms. MIT Press, third edition, 1312: TH CormenCE LeisersonRL RivestC. Stein2009Introduction to Algorithms.MIT Press, third edition,1312
- 19. Ravasi T, Suzuki H, Cannistraci CV, Katayama S, Bajic VB, et al. (2010) An atlas of combinatorial transcriptional regulation in mouse and man. Cell 140: 744–752.T. RavasiH. SuzukiCV CannistraciS. KatayamaVB Bajic2010An atlas of combinatorial transcriptional regulation in mouse and man.Cell140744752