Abstract
Evolutionary optimisation algorithm is employed to design networks of phaserepulsive oscillators that achieve an antiphase synchronised state. By introducing the link frustration, the evolutionary process is implemented by rewiring the links with probability proportional to their frustration, until the final network displaying a unique nonfrustrated dynamical state is reached. Resulting networks are bipartite and with zero clustering. In addition, the designed nonfrustrated antiphase synchronised networks display a clear topological scale. This contrasts usually studied cases of networks with phaseattractive dynamics, whose performance towards full synchronisation is typically enhanced by the presence of a topological hierarchy.
Introduction
Study of collective phenomena in complex systems has been massively enhanced in the last decade by introduction of the complex networks framework^{1}. This revolutionised our understanding of the real complex systems, ranging from gene and neuron networks, to air traffic, Internet and society^{2}. This paradigm allowed for novel theoretical and computational investigations, such as stability of gene regulatory networks examined through the networkinhibition of chaos^{3}, or detection of communities in social networks^{4}.
The phenomenon of synchronisation occupies a prominent place in the networks science, as a par excellence example of emergent collective behaviour^{5}. Synchronisation of various kinds of oscillating dynamical units have been extensively studied both theoretically^{6} and experimentally^{7}. Kuramoto oscillators^{8,9} allowed for the emergence of synchronisation to be examined in great detail^{5,6}, due to being conceptually simple and easy to implement numerically. This model is often used in theoretical and experimental studies, such as network reconstruction^{10}, or modelling of neural phaseresetting curves^{11}.
In parallel with synchronisation, research attention was devoted to various forms of antisynchronisation^{12}. This particularly refers to antiphase synchronisation^{13}, where the interacting oscillator pairs evolve towards having the opposite phases, i.e. towards being synchronised with a fixed phase difference π. In this context, networks of units with the repressive and/or repulsive interactions were investigated. Most famously, the system of repressively interacting genes – repressilator – received much attention^{14,15,16}. Repulsive interactions can generate dynamical frustration and multistability, since interacting units cannot always relax to a unique equilibrium state^{17,18}. Neurons are often coupled repulsively, which was examined theoretically^{19,20} and experimentally^{21,22}. In addition, repulsive Kuramoto oscillators can be used to model social interactions and social equilibria^{23,24}, or suppression of synchronisation^{25}.
Recently, complex networks of Kuramoto oscillators will solely repulsive coupling were examined^{26}, in opposition to studies involving both attractive and repulsive interactions^{27}. By borrowing the terminology from disordered systems^{28}, the dynamical states in these networks were characterised by defining the link frustration f. In general, complex networks were found to have positive total frustration, due to complexity of their topologies. In addition, depending on the initial conditions, networks typically display multiple final dynamical states. The network's emergent dynamics is thus strongly dependent on the topology, and for a general network does not reach “minimum energy” state with zero frustration.
In a different context, researchers have been studying the problem of network design, motivated by the technological need to engineer networks that perform prescribed functions. Models are typically based on a network whose topology is evolving towards the optimal target topology. Most used among such models is the evolutionary optimisation, a version of Monte Carlo simulated annealing algorithm^{29}. A myriad of results were obtained through this formalism, such as emphasising peculiarities of the scalefree topologies^{30}, design of biological transduction networks^{31,32}, design of robust genetic clocks^{33}, or design of networks that optimally perform a given dynamical behaviour^{34}. In the context of oscillators, design of easily synchronisable network was recently examined^{35}, and a method of achieving the prescribed synchronisation state by configuring a repulsive subnetwork was exposed^{36}.
A question that remains open in the context of phaserepulsive networks, regards the topologies that allow for the zero frustration state to be attained. These are the networks whose dynamics leads to antiphase synchronisation for any choice of initial conditions. While essentially any oscillator network with phaseattractive coupling generates full synchrony, obtaining antiphase synchronisation is more demanding. As shown in^{26}, despite that phaserepulsive interactions naturally tend towards antiphase synchrony, a generic network typically can not attain fully antiphase synchronised state due to the topological constraints. In this paper we employ the evolutionary design algorithm^{29,31,33}, to obtain phaserepulsive oscillator networks that always achieve nonfrustrated, i.e. antiphase synchronised state. By choosing the links to rewire according to their frustration, the initial network topology evolves towards minimising global frustration, which inevitably leads to nonfrustrated topology. As we show, in opposition to the well studied synchronisable networks^{35}, whose synchronisation performance is typically enhanced by hierarchical or scalefree structures^{37}, nonfrustrated networks are topologically very uniform, and exhibit a precise scale^{38,39} rather than being scalefree. This problem is similar to the problem of 2color vertex colouring encountered in graph theory^{40,41}. The difference is that the process of vertex colouring is here automatically done by the network phaserepulsive dynamics.
We develop our model by considering a network that consists of N oscillators (nodes), connected via L nondirected links. Dynamical state of the oscillator i is given by the phase variable ϕ_{i} ∈ [0, 2π), and its dynamics is defined by:where k_{i} is the node's degree (Σ_{i}k_{i} = 2L), ε is the coupling strength, and ω_{i} is the oscillator's natural frequency. Network's topology is expressed via symmetric adjacency matrix A_{ij}, defined as A_{ij} = 1 if nodes i and j are connected, and A_{ij} = 0 otherwise. We consider identical oscillators ω_{i} = ω, and set g = sin, reducing our system to the simple Kuramoto model^{8}. Following the results exposed in^{26}, we examine the network model with phaserepulsive coupling. To that end we fix ε = –1, and for simplicity take ω = 0. The equation for our model reads:Timeevolution starts from a random set of initial phases (IP), selected independently for each oscillator from ϕ_{i}(0) ∈ [0, 2π). The interacting oscillator pairs evolve towards maximising the phase difference between them, i.e. their phase values “stretch” apart from each other as much as possible. Preferably, the maximal phase difference is π. However, for a general network, due to its complex topology, the phase difference along certain links is often less than π, or even zero. The final (stationary) dynamical state is quantified by introducing frustration f_{ij} for each link as^{26,27}:Frustration measures how “squeezed” is a link: it can be pictured as the elastic potential energy contained in it: a link stretched to the phase difference π has zero frustration, whereas a link forced to synchronise carries the maximal frustration that equals 2. We characterise the network dynamical states by specifying the link frustration values f_{ij}. To measure the global frustration we define F as the network average of f:which quantifies how much does the network topology allow for links to stretch. Note that F can be seen as the nonequilibrium potential, since Eq.2 can be written as^{27}:As concluded in^{26}, phaserepulsive dynamics for a general network leads to a positive total frustration, whose value strongly depends on the details of network topology. In widely studied phaseattractive case, full synchronisation is essentially the only final dynamical state, independently of the topology. In contrast, topology in our model crucially determines the dynamics. In addition, a general phaserepulsive network displays multiple final frustration states, each occurring for a certain fraction of IP.
Interestingly, some networks exhibit a unique dynamical state with zero total frustration F = 0. These networks are completely antiphase synchronised – each coupled oscillator pair has the opposite phases for any IP. This comes from peculiar topologies of such networks that allow for all links to stretch, thus completely avoiding frustration. While in the phaseattractive case, basically any network achieves the fully synchronised state, in the phaserepulsive case only a limited set of networks attain the antiphase synchronised state. In addition, such state is also stable, as it is the only possible dynamical equilibrium state. How can we construct nonfrustrated networks? What are their topological properties? The rest of this paper is devoted to these two questions.
Results
We employ the evolutionary algorithm based on simulated annealing in order to design nonfrustrated networks. The evolutionary process involves rewiring links with the probability proportional to their frustration f_{ij}, until a network with zero frustration along all links for all IP is obtained. Further details on our network design method are provided in Methods section below.
An example of a simple evolutionarily designed nonfrustrated network with N = 20 nodes and L = 30 links is shown in Fig. 1. It was obtained after 15 evolution steps. The network appears to be uniform in structure, without prominent hubs. It is also bipartite, as illustrated by different shades of adjacent nodes. Large presence of 4node rings, which are small nonfrustrated networks, is clearly visible. There are also 6node rings, which are unable to exhibit the higher F state due to being topologically intertwined with the rest of the network, which is forcing them into F = 0 state. Clustering coefficient is zero for all nodes. Due to a large number of 4node rings, the network topology seems more regular than expected from an evolution governed by random rewirings.
To quantify more precisely the topological properties of nonfrustrated networks, we now consider larger networks with N = 100 nodes. We run evolution simulations with various numbers of links, ranging from L = 200 to L = 2000: all evolution processes yield a final nonfrustrated network. In particular, we here focus on networks with L = 200, 400, 600 and 800, which are shown in Fig. 2, and indicated via the corresponding L (details on this evolution process are included in the Methods section below). As observed earlier with Fig. 1, the nonfrustrated networks seem not to posses any hierarchical structure, which is even more apparent here. Of course, all the obtained networks are bipartite and have zero clustering. We run many evolution simulations with the same N and L: the same global topological features were always observed. As we show more quantitatively in what follows, the structures of these networks are indeed surprisingly regular.
We first compute the histograms of node degrees. To this end, we average the degree histograms over 5 realisations of the nonfrustrated networks with given N and L, and report the results in Fig. 3. All histograms are very localised in range and display a well pronounced peak in the centre, whose value coincides with the average degree . As expected, for the increasing network size, the peak becomes gradually less pronounced. This confirms that our networks are not hierarchical. In order to nullify the frustration, network evolution leads towards evenly spreading the remaining link frustration over the network. This indicates that rather than being scalefree, nonfrustrated networks are “scalerich”, i.e. posses a well defined scale^{38} (this holds for small L, i.e. sparse networks). The scale is determined by the average node degree.
The node degree histograms suggest that nonfrustrated networks posses structures that are similar to ErdősRényi random networks, that the evolutionary process started from. However, two types of structure also have profound differences. Nonfrustrated networks, by being bipartite, always have zero clustering. In contrast, the average expected clustering coefficient for random networks is ^{1}. This difference is a clearly pronounced in dense nonfrustrated networks with high 〈k_{i}〉. In addition, we found that nonfrustrated networks systematically display somewhat longer average shortest paths than random networks with the same N and L. This can again be attributed to zero clustering: since no three nodes can be mutually interconnected, shortest paths are on average somewhat longer.
There is an upper bound on L for any given N, beyond which it is not possible to design nonfrustrated networks, since avoiding connected node triangles becomes impossible. Close to this limit the network evolution process gradually becomes slower. The network structure however becomes even more regular, since avoiding frustration requires more and more ordered organisation of links, and more uniform node degrees. In the supplementary figure to this paper, we show the additional six networks with N = 100, and L ranging from 1000 to 2000. The regularity and bipartitness of these dense networks are clearly visible.
Finally, we quantity the topological regularity of nonfrustrated networks. We achieve this by measuring the concentration of motifs in our networks, using the mfinder software^{42}. We first introduce Z_{score}, as a measure of how over or under represented is a given motif w within some network. It is defined as^{43}: where N_{actual}(w) is the number of motifs w found in the examined network, N_{random}(w) is the average expected number of those motifs in a network with the same N and L, while SD(w) is the standard deviation of N_{random}(w). Thus, any motif whose Z_{score} is considerably bigger (smaller) than 1 (−1) is over (under) represented. Since the studied networks differ in their numbers of links, rather them employing Z_{score}, we instead rely on normalised Z_{score}, which is defined as^{32}:We calculate Z_{normalised} for all 4node nondirected motifs (shown in Fig. 5c), for four evolutionarily designed networks shown in Fig. 2. The results are reported in Fig. 4. In all networks, nonfrustrated motifs are overwhelmingly overrepresented, while the frustrated motifs are drastically underrepresented (cf. Fig. 5). Both over and under representation are invariant to varying realisations of nonfrustrated networks, and to the number of links L. The same result is found for both 3node motifs: the chain is systematically overrepresented, while the ring is strongly underrepresented. This suggests that, regardless of their size, nonfrustrated networks have an inherent motif pattern, rooted in their dynamical properties. This result confirms that large nonfrustrated networks are essentially composed of small nonfrustrated motifs. By locally forming nonfrustrated motifs, mechanism of network evolution evolves towards global zero frustration. Interestingly, this means that the emergence of global F = 0 is achieved locally, by each network building block evolving towards F = 0. This implies that any large nonfrustrated network can be decomposed into smaller nonfrustrated subnetworks (an alternative scenario can be imagined, where the desired global property is achieved only globally, i.e. only at the level of the entire network, with no subnetworks exhibiting the same property).
Discussion
Natural evolutionary processes in real complex networks often lead to scalefree structures in the emerging networks^{30}. However, the ubiquity of scalefree topologies has been recently challenged. New findings include examples of real complex networks, that despite emerging from biological selforganisation, posses a clear scale in their structure, contrasting widely accepted universality of scalefree topologies^{38,39}. In this paper, we showed a model of network topology that develops under a simple evolutionary constraint, whose topology evolves away from hierarchical structure, and towards a “scalerich” structure^{38}. As shown above, this can be attributed to the nature of phaserepulsive dynamics, which requires certain topological properties in order to attain fully antiphase synchronised state (i.e., nonfrustrated state). Hence, in opposition to the phaseattractive dynamics and resulting full synchronisation, that are typically improved by the presence of a topological hierarchy, antiphase synchronisation appears to be enhanced by the lack of hierarchy.
Bipartite networks are usually considered in relation to two different types of nodes, such as online comments and the corresponding webpage users (persons)^{44}. Here we showed that bipartitness can spontaneously emerge from the evolution of a network with a single type of nodes (identical oscillators). Besides these findings, this work confirmed again that evolutionary design through random link rewirings is a viable method of constructing complex networks that exhibit prescribed dynamical patterns. Our method mimics the natural evolutionary processes, by implementing random mutations (rewirings). To make mutations completely random, one could modify the algorithm to choose links for rewiring regardless of their frustration. This would however only slow down the process, without changing the final result. On the other hand, it is possible that the nonfrustrated networks can be obtained by means of other design methods. Such methods are however still poorly explored, even in the context of simple models such as oscillator networks.
We conclude this section by examining the limitations and extensions of our results. As noted earlier, zero frustration is not achievable beyond a certain limit of L. Our algorithm can in this case be applied to minimising the frustration, and studying the ways in which the resulting networks organise the remaining frustrated links. Another question refers to maximising instead of minimising the total network frustration. By accepting the mutations that increase rather than decrease F, our algorithm can be used to design networks with maximal obtainable frustration. Intriguing problem here is the difference between the frustrated (squeezed) and nonfrustrated (relaxed) network topologies. This is related to elastic network models, where mechanical properties of networks of elastic springs are studied^{45}. It would be also interesting to apply the same algorithm to design nonfrustrated networks using other models of repulsive interactions. This particularly refers to neural and gene interactions, where networks with repressive interactions are widely studied^{17,18,20}. This could aid engineering genetic or neural networks with a given prescribed functionality. Finally, we note that our networks were designed to satisfy a simple dynamical rule, formulated as F = 0. By varying the dynamical rules, one could use this or similar algorithms to design networks whose topological properties reflect a predetermined dynamical rule. This could lead to the discovery of new complex network topologies, that do not fit into any of the currently studied topological categories.
Methods
In this section we describe our method of evolutionarily network design and explain its implementation.
Preliminaries
For small networks it is easy to find nonfrustrated topologies, as shown in Fig. 5. A connected pair of nodes, 3node chain, 4node star, 4node chain and 4node ring (cycle), are networks that show a single frustration state F = 0. However, finding a large nonfrustrated network of size , is not straightforward. A typical network will in general have F > 0, for instance, due to having at least one 3node ring. On the other hand, it is easy to notice that for a network to be nonfrustrated, it needs to be bipartite, and hence have zero clustering. Indeed, for having two groups of nodes with opposite phases, the network needs to involve links that exclusively go from nodes of one group to the nodes of the other group. However, although necessary for antiphase synchronisation, bipartitness is not sufficient for it. In Fig. 6 we show the 6node ring, which is a bipartite network. However, besides displaying the expected nonfrustrated state, the network possesses another frustrated state with F = 0.5, that occurs for 14% of IP. Bipartitness alone guarantees the existence of a nonfrustrated state, but it does not guarantee its uniqueness. This indicates that we need a more elaborated mechanism for designing nonfrustrated networks with the unique F = 0 state.
Evolutionary algorithm
We employ the simulated annealing algorithm as follows. Starting from a random network, at each evolution step we rewire one of the frustrated links, and accept the mutation if the new network exhibits smaller total frustration F than the original one. The network topology evolves until the target F = 0 is reached. More precisely, our algorithm goes as follows:
0. Start from a nondirected, connected network , with N nodes and L links , and no selfloops. We use ErdősRényi random network. Compute link frustration values and total frustration , averaged over M random IP. We use M = 5.
1. Select a link ij in with probability proportional to (we use α = 0.01). Detach this link from one of its endnodes (randomly chosen), and rewire it to a different node chosen with uniform probability, making sure that the network stays connected and without selfloops. This is the mutated network .
2. For compute and . If the mutation is accepted, and the network is updated: . If the mutation is accepted with probability . Otherwise, the mutation is rejected and remains unchanged. We use σ = 0.02.
3. The process continues from step 1 by selecting a new link in to be rewired.
The evolution continues until the final network with a unique frustration state F = 0 is reached. Potential multiple frustration states are excluded by averaging over M random IP. The numerical constant α skews the probability distribution allowing for nonfrustrated links to be mutated as well, while ensuring that more frustrated links are still mutated more often. This drastically expedites the evolution process, while at the same time allowing for all links to be mutated. Parameter σ allows that, with a small probability, even mutations leading to a small increase of total frustration get accepted. By playing the role of temperature, σ prevents the evolution process from getting stuck into local minima.
Algorithm's performance
We tested the evolution algorithm exposed above on numerous network examples: the process invariably leads to a nonfrustrated network. Since the evolution involves randomness, the final network does not depend on the starting one. On the other hand, for any fixed N and L there are many different realisations of final nonfrustrated networks, all sharing the key topological properties. To illustrate the progress of the network evolution, we show in Fig. 7 the gradual decrease of for four examples of evolution simulations related to networks from Fig. 2. The evolution speed decreases with the increase of L. In a denser network it is harder to find a mutation that leads to a decrease of , and when found, that mutation typically decreases for a small amount. In addition, denser networks start with a larger initial , thus additionally increasing the needed number of evolution steps. We adjusted for the optimality of the process through the parameters M, α and σ. The “bumpiness” of the processes is due to Monte Carlo algorithm's temperature σ > 0. This parameter needs to be adequately set to optimise the evolution speed and effectiveness. Too small σ might result in process getting suck into a local minimum of , which may incorrectly suggest that smaller is not obtainable. Too large σ could induce the process to wander too excessively, without ever finding solution. Our choice of σ = 0.02 takes into account both issues. At each evolutionary iteration, the network usually has multiple frustration states with different Fvalues. To obtain a single frustration value corresponding to the current network, we average Fvalues over M = 5 IP. This is an additional reason for introducing temperature σ: due to a potentially wide range of multiple states, averaging might not give the best measure of the total frustration. Upon each completion of evolutionary simulation, we checked that the obtained network is indeed nonfrustrated.
References
 1.
Dorogovtsev, S. N. Lectures on Complex Networks. Oxford University Press, Oxford, (2010).
 2.
Costa, L. F. et al. Analyzing and modeling realworld phenomena with complex networks: a survey of applications. Adv. Phys. 60, 329–412 (2011).
 3.
Levnajić, Z. & Tadić, B. Stability and chaos in coupled twodimensional maps on gene regulatory network of bacterium E. coli. Chaos 20, 033115 (2010).
 4.
Fortunato, S. Community detection in graphs. Phys. Rep. 486, 75–174 (2010).
 5.
Pikovsky, A., Rosenblum, M. & Kurths, J. Synchronization. A Universal Concept in Nonlinear Sciences. Cambridge University Press, Cambridge, (2003).
 6.
Arenas, A. et al. Synchronization in complex networks. Phys. Rep. 469, 93 (2008).
 7.
Ravoori, B. et al. Robustness of Optimal Synchronization in Real Networks. Phys. Rev. Lett. 107, 034102 (2011).
 8.
Kuramoto, Y. Chemical Oscillations, Waves, and Turbulence. Dover, New York, (2003).
 9.
Acebrón, J. A. et al. Kuramoto model: A simple paradigm for synchronization phenomena. Rev. Mod. Phys. 77, 137–185 (2005).
 10.
Levnajić, Z. & Pikovsky, A. Network Reconstruction from Random Phase Resetting. Phys. Rev. Lett. 107, 034101 (2011).
 11.
Perez Velazquez, J. L. et al. Phase response curves in the characterization of epileptiform activity. Phys. Rev. E 76, 061912 (2007).
 12.
Kim, C. M. et al. Antisynchronization of chaotic oscillators. Phys. Lett. A 320, 39–46 (2003).
 13.
Liu, W., Xiao, J., Qian, X. & Yang, J. Antiphase synchronization in coupled chaotic oscillators. Phys. Rev. E 73, 057203 (2006).
 14.
Elowitz, M. B. & Leibler, S. A synthetic oscillatory network of transcriptional regulators. Nature 403, 335 (2000).
 15.
Gardner, T. S., Cantor, C. R. & Collins, J. J. Construction of a genetic toggle switch in Escherichia coli. Nature 403, 339 (2000).
 16.
Hast, J., McMillen, D. & Collins, J. J. Engineered gene circuits. Nature 420, 2244 (2002).
 17.
Jensen, M. H., Krishna, S. & Pigolotti, S. Repressor Lattice: Feedback, Commensurability, and Dynamical Frustration. Phys. Rev. Lett. 103, 118101 (2009).
 18.
Koseska, A. et al. Inherent multistability in arrays of autoinducer coupled genetic oscillators. Phys. Rev. E 75, 031916 (2007).
 19.
Wang, Q., Perc, M., Duan, Z. & Chen, G. Synchronization transitions on scalefree neuronal networks due to finite information transmission delays. Phys. Rev. E 80, 026206 (2009).
 20.
Wang, Q., Chen, G. & Perc, M. Synchronous bursts on scalefree neuronal networks with attractive and repulsive coupling. PLoS ONE 6, e15851 (2011).
 21.
Balázsi, G. et al. Synchronization of hyperexcitable systems with phaserepulsive coupling. Phys. Rev. E 64, 041912 (2001).
 22.
Balázsi, G., CornellBell, A. H. & Moss, F. Increased phase synchronization of spontaneous calcium oscillations in epileptic human versus normal rat astrocyte cultures. Chaos 13, 515 (2003).
 23.
Marvel, S. A., Strogatz, S. H. & Kleinberg, J. M. Energy landscape of social balance. Phys. Rev. Lett. 103, 198701 (2009).
 24.
Hong, H. & Strogatz, S. H. Kuramoto Model of Coupled Oscillators with Positive and Negative Coupling Parameters: An Example of Conformist and Contrarian Oscillators. Phys. Rev. Lett. 106, 054102 (2011).
 25.
Louzada, V. H. P. et al. How to suppress undesired synchronization. Sci. Rep. 2, 658 (2012).
 26.
Levnajić, Z. Emergent multistability and frustration in phaserepulsive networks of oscillators. Phys. Rev. E 84, 016231 (2011).
 27.
Zanette, D. H. Synchronization and frustration in oscillator networks with attractive and repulsive interactions. Europhys. Lett. 72, 190 (2005).
 28.
Mezard, M., Parisi, G. & Virasoro, M. Spin Glass Theory and Beyond, World Scientific., Singapore, 1987.
 29.
Bornholdt, S. & Sneppen, K. Robustness as an Evolutionary Principle. Proc. R. Soc. Lond. B 267, 2281–2286 (2000).
 30.
Oikonomou, P. & Cluzel, P. Effects of topology on network evolution. Nat. Phys. 2, 532–536 (2006).
 31.
Kaluza, P. et al. Design and statistical properties of robust functional networks: A model study of biological signal transduction. Phys. Rev. E 75, 015101(R) (2007).
 32.
Kaluza, P., Vingron, M. & Mikhailov, A. S. Selfcorrecting networks: Function, robustness, and motif distributions in biological signal processing. Chaos 18, 026113 (2008).
 33.
Kobayashi, Y. et al. Robust network clocks: Design of genetic oscillators as a complex combinatorial optimization problem. Phys. Rev. E 83, 060901(R) (2011).
 34.
Karalus, S. & Porto, M. Network evolution towards optimal dynamical performance. Europhys. Lett. 99, 38002 (2012).
 35.
Yanagita, T. & Mikhailov, A. S. Design of easily synchronizable oscillator networks using the Monte Carlo optimization method. Phys. Rev. E 81, 065204 (2010).
 36.
Arizmendi, F. & Zanette, D. H. Adaptation to synchronization in phaseoscillator networks. Physica A 387, 5631 (2008).
 37.
Motter, A. E., Zhou, C. & Kurths, J. Network synchronization, diffusion, and the paradox of heterogeneity. Phys. Rev. E 71, 016116 (2005).
 38.
Tanaka, R. Scalerich metabolic networks. Phys. Rev. Lett. 94, 168101 (2005).
 39.
Pržulj, N. Proteinprotein interactions: making sense of networks via graphtheoretic modeling. Bioessays 33, 2 (2011).
 40.
Dvořák, Z., Král, D. & Škrekovski, R. Nonrainbow colorings of 3, 4 and 5connected plane graphs. J. Graph Theory 63, 129–145 (2010).
 41.
Diestel, R. Graph Theory., Springer, 2010.
 42.
Software package mfinder is available at http://www.weizmann.ac.il/mcb/UriAlon/ (Accessed Dec. 4, 2012).
 43.
Milo, R. et al. Network Motifs: Simple Building Blocks of Complex Networks. Science 298, 824–827 (2002).
 44.
Mitrović, M., Paltoglou, G. & Tadić, B. Networks and emotiondriven user communities at popular blogs. Eur. Phys. J. B 77, 597–609 (2010).
 45.
Flechsig, H. & Mikhailov, A. S. Tracing entire operation cycles of molecular motor hepatitis C virus helicase in structurally resolved dynamical simulations. Proc. Nat. Acad. Sci. USA 107, 20875 (2010).
Acknowledgements
Thanks to Alexander Mikhailov for useful suggestions, and Uri Alon for making mfinder publicly available.
Author information
Affiliations
Faculty of information studies in Novo mesto, Novo mesto, Slovenia
 Zoran Levnajić
Authors
Search for Zoran Levnajić in:
Competing interests
The author declares no competing financial interests.
Corresponding author
Correspondence to Zoran Levnajić.
Supplementary information
PDF files
 1.
Supplementary Information
Supplementary figure
Rights and permissions
This work is licensed under a Creative Commons AttributionNonCommercialNoDerivs 3.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/byncnd/3.0/
To obtain permission to reuse content from this article visit RightsLink.
About this article
Further reading

Untangling complex dynamical systems via derivativevariable correlations
Scientific Reports (2015)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.