A kinetic model of the evolution of a protein interaction network
 Piotr H Pawlowski^{1}Email author,
 Szymon Kaczanowski^{1} and
 Piotr Zielenkiewicz^{1, 2}
DOI: 10.1186/1471216414172
© Pawlowski et al.; licensee BioMed Central Ltd. 2013
Received: 22 October 2012
Accepted: 8 March 2013
Published: 14 March 2013
Abstract
Background
Known protein interaction networks have very particular properties. Old proteins tend to have more interactions than new ones. One of the best statistical representatives of this property is the node degree distribution (distribution of proteins having a given number of interactions). It has previously been shown that this distribution is very close to the sum of two distinct exponential components. In this paper, we asked: What are the possible mechanisms of evolution for such types of networks? To answer this question, we tested a kinetic model for simplified evolution of a protein interactome. Our proposed model considers the emergence of new genes and interactions and the loss of old ones. We assumed that there are generally two coexisting classes of proteins. Proteins constituting the first class are essential only for ecological adaptations and are easily lost when ecological conditions change. Proteins of the second class are essential for basic life processes and, hence, are always effectively protected against deletion. All proteins can transit between the above classes in both directions. We also assumed that the phenomenon of gene duplication is always related to ecological adaptation and that a new copy of a duplicated gene is not essential. According to this model, all proteins gain new interactions with a rate that preferentially increases with the number of interactions (the rich get richer). Proteins can also gain interactions because of duplication. Proteins lose their interactions both with and without the loss of partner genes.
Results
The proposed model reproduces the main properties of proteinprotein interaction networks very well. The connectivity of the oldest part of the interaction network is densest, and the node degree distribution follows the sum of two shifted powerlaw functions, which is a theoretical generalization of the previous finding. The above distribution covers the wide range of values of node degrees very well, much better than a power law or generalized power law supplemented with an exponential cutoff. The presented model also relates the total number of interactome links to the total number of interacting proteins. The theoretical results were for the interactomes of A. thaliana, B. taurus, C. elegans, D. melanogaster, E. coli, H. pylori, H. sapiens, M. musculus, R. norvegicus and S. cerevisiae.
Conclusions
Using these approaches, the kinetic parameters could be estimated. Finally, the model revealed the evolutionary kinetics of proteome formation, the phenomenon of protein differentiation and the process of gaining new interactions.
Background
Although an evolutionary viewpoint in network studies is not a new concept [1], it still gains new followers [2], especially in the field of the evolution of protein interactions [3, 4] and in regulatory [5] and metabolic [6] networks. Investigators of protein–protein interaction (PPI) networks indicate that functional evolution [7], modular organization [8], evolutionary pressures [9] and genome duplications [10, 11] are crucial factors in shaping network architecture, and several of these researchers negatively correlate the connectivity of wellconserved proteins in the network with their individual rate of evolution [12, 13]. Numerous studies indicate that local network growth rules, such as gene duplication and gene diversification, can give rise to scalefree connectivity distributions and an effective linear preferential attachment [14]. Original approaches, such as the evolutionary excess retention method [15] or modeling of protein evolution using a lattice representation of their structures, were proposed to determine the effect of explicit selection on PPI [16]. The most popular ideas for the main mechanisms for generating the scalefree, older core and hierarchically modular topology of protein interaction networks are the BarabasiAlbert “the rich get richer” model of preferential attachment and the gene duplication and divergence model of Ispolatov et al. [17]. They were recently criticized on the basis of the KimMarcotte stochastic crystal growth model [18], which captures the agedependency of the interaction density along with the hierarchical modularity. Nevertheless, some of the elements of the previous models can still be beneficial in modeling the overall kinetics of interactome evolution.
Consequently, one may expect that the current network architecture may provide quantitative information about the network history. Comparing the presented kinetic model for the evolution of the protein interaction network with the data for S. cerevisiae and nine other species allows us to estimate the rates of the basic processes of interactome evolution, i.e., the emergence of new genes and the loss of the old ones, the duplication phenomenon, the differentiation of functional significance, the obtaining of new interactions and the deactivation of active ones.
Two sources for new interactions were considered: one, newly emerging proteins and two, proteins within the currently existing interactome. The overall preference for gaining new interactions was assumed to be related to the node degree. In addition, two methods for losing new interactions were considered; the first was related to protein deactivation, and the second one was spontaneous. Because there is evidence that more important proteins evolve similarly to others [21], the kinetic parameters for the evolution of the number of interaction partners were assumed to be independent of protein class.
The described model predicts a doubleshifted powerlaw distribution for the node degree. Therefore, it confirms the earlier proposal of a double exponential distribution for the node degree [22] in the range of small degrees. The model also reveals parabolic relationships between the total number of interactions and the total number of interacting proteins. The parameters of the derived mathematical formulas were estimated by fitting the theoretical predictions of the model to the existing data for the interactomes of 10 different species. This model enabled us to reveal the evolutionary kinetics of proteome formation, the differentiation process and the process of gaining new interactions.
Results
Kinetic model of the evolution of a protein interaction network
Proteome formation
All parameters of the model describing the rates are treated as fixed.
Interactome formation
where τ is the protein age.
For simplicity, only the proteins that emerged in the steady state of proteome formation (dX/dt = 0, dY/dt = 0) were analyzed in the following.
where ξ_{ r } is presented as it is defined in the Methods.
Node degree distribution
The amplitudes A_{ i } and the powers β_{ i } (i = 1,2) are defined in the Methods.
Total number of links
The probabilities p_{ i } (i = 1,2) are defined in the Methods.
The cited quantities ξ_{ r }, A_{ i }, β_{ i }, p_{ i } can be related to reality by fitting eqs. 5 and 6 to experimental data. However, they are dependent on the kinetic parameters of the processes considered in the model (see the Methods). This approach may lead to the quantitative estimation of these parameters.
Computer simulations
Experimental data
Experimental data for the studied interactomes
Interactome  N _{ ∞ }  L  Database 

A. thaliana  487  959  BIND 
B. taurus  129  107  DIP 
C. elegans  3227  5026  BIND 
D. melanogaster  7910  23128  BIND 
E. coli  399  312  BIND 
H. pylori  724  1403  COSIN 
H. sapiens  2529  3376  DIP 
M. musculus  1003  994  DIP 
R. norvegicus  349  304  DIP 
S. cerevisiae  4135  7839  COSIN 
Fitting the model of the node degree distribution to the experimental data
The results of fitting the model to the experimental data
Quantity  Estimate  SE (%) 

A _{1}  3184.82  32 
A _{2}  49.8628  77 
β _{1}  4.80485  29 
β _{2}  2.1242  20 
ξ _{ r }  6.30779  51 
p _{1}  0.73526  13 
p _{2}  0.000552383  5 
Fitting the model of the dependence of N_{∞}and L to the experimental data
The Mathematica 4.1 standard procedure (NonlinearRegress), from the package Statistics`NonlinearFit`, was applied to fit the proposed model (eq. 6) to the set of (N_{ ∞ }, L) pairs for 10 different interactomes (Table 1). The results of the fitting, i.e., the values of the quantities p_{ i }, are listed in Table 2. The corresponding experimental points and fitted plot are presented in Figure 3d.
Finding the values of the kinetic parameters of the model
The general parameters of both the node degree distribution (A_{ i }, β_{ i } and ξ_{ r }) and the total number of links (p_{ i }) can be related to the parameters of the kinetic model. Using both sets of parameters increases the universality and the credibility of the final estimated parameters of model.
with the assumption that N_{∞} = 4135, as for S. cerevisiae.
Estimated kinetic parameters of the model
Kinetic parameter  Best estimation χ^{2} = 0.14  Average (the 10 best)  SE (%) (the 10 best) 

k _{ i }  8.61692  12.683567  10.1 
k _{ 2 }  0.122669  0.052556819  27.7 
k _{ XY }  0.0611168  0.09274057  10.4 
k _{ YX }  2.8542  4.249004  10.1 
ξ _{ 0 }  0.426372  0.4256051  0.1 
μ  0.00237779  0.003526257  10.1 
Δ ε  10.6696  15.91212  9.9 
r  0.107993  0.204924912  45.9 
f _{ 0 }  34376.8  51108.95  10.1 
Simulations of the kinetics of the protein interactome evolution
Summary of the most important results
The proposed kinetic model (Figure 2a,b) of the evolution of a protein interaction network agrees very well with the experimental data. The node degree distribution of S. cerevisiae (Figure 3a) and the nonlinear dependence of the total number of links on the total number of interacting proteins (Figure 3d) can be successfully described with the derived theoretical formulas (eqs. 5 and 6). Thus, amplitudes, powers and probabilities (Table 2) were obtained according to the model in the Methods. In addition to providing a nontrivial explanation of the recently observed picture of the node degree distribution or the N_{ ∞ } and L dependence, these values led to the estimation of the kinetic parameters (Table 3) of the dynamic processes governing the evolution, differentiation and crosslinking of the protein interaction network. Finding these parameters enables numerical simulations of the evolution of the following: the total proteome (Figure 4a,b), the decrease and differentiation of a small sample of synchronized proteins (Figure 5) and the expansion of a single protein node degree (Figure 6). The estimated characteristic times of evolution are 1/γ_{1} = 0.12, 1/γ_{2} = 0.35 and 1/v = 0.45, indicating that the evolution of the node degree is slower than the evolution of the proteome. The estimated fraction of essential proteins κ = Y_{ ∞ }/N_{ ∞ } equals 0.02.
Discussion
The presented kinetic model of the evolution of a protein interactome is an extension of the previous twoclass model [22] describing a double exponential distribution of the node degree. The current version of the model additionally postulates asymmetry in the functional importance of the considered protein classes and takes into account a possible evolutionary transition between the classes. This model also considers gene doubling and preferential attachment.
From a cognitive point of view, the proposed model led to a satisfactory fit to the node degree histogram (Figure 3a) and to the picture of the nonlinear dependence of N_{∞} and L (Figure 3d). Moreover, the node degree fit according to the derived eq. 5 is 50% better than that of the power law [1] or 25% better than that of the generalized power law with exponential cutoff [23]. This fit is also much better than the fit from our previous double exponential model (Figure 3c), neglecting gene doubling, preferential attachment and interclass transitions. Moreover, the current model led to the estimation of unknown values of kinetic parameters (Tables 2 and 3). Thus, this model reveals the kinetics of evolution of the interactome (Figure 4a,b), the final result of which (approaching the steady state) does not depend on the initial state of the protein’s importance. Although the evolution of the total proteome stabilizes, individual proteins are eliminated (Figure 5). Gaining new interactions from a single protein (Figure 6) is much slower than the evolution of the proteome, but the increase of protein degree with protein age confirms the trend observed for proteins of eukaryotic and posteukaryotic origin [7].
The model and its estimated kinetic parameters allow a sketch of a hypothetical picture of proteome evolution, indicating that class Y of proteins that are functionally essential for basic processes of S. cerevisiae finally includes approximately 2% of protein population. One could expect that all the genes from this class and a portion of the genes belonging to the first class (proteins important for ecological adaptations) are strictly essential (their deletion is lethal). We compared this expectation with experimental results. Deutschbauer [24] and coworkers showed that the deletion of 19% of genes causes lethality. This finding is in agreement with our results. The second expectation is that some proteins belonging to the first class (proteins important for ecological adaptations) have a function only in particular conditions. This hypothesis was shown experimentally by Hillenmeyer [25] and coworkers, who performed 1144 chemical genomic assays on the yeast wholegenome heterozygous and homozygous deletion collections and quantified the growth fitness of each deletion strain in the presence of chemical or environmental stress conditions. In their first experiment, only approximately 40% of the gene deletion strains performed phenotype. However, 97% of the gene deletions also exhibited a measurable growth phenotype in one of the tested conditions. In conclusion, our results fit well to the experimental data.
In this picture, the origination of new species may be related to variations in the value of the parameters governing the kinetics of evolution (e.g., f_{ 0 }, which directly determines the value of N_{∞}), resulting in the origination of a new steady state of proteome organization. In addition, the results indicate that entering the important class Y is approximately 50fold slower than leaving it. This finding illustrates how difficult it is to become a member of a protein “gentlemen's club” and how easy it is to lose this position. Mechanisms of selection and adaptation certainly play an important role in this type of arrangement, ensuring stability in the composition of backbone biochemical reactions. The stability is one of the most important factors supporting organisms’ survival. During evolution, organisms investigate optimal paths of growth and replication, which is possible if and only if the organisms preserve certain optimal and stable biochemical machinery [26].
The obtained results also show how large dynamic changes involving new protein emergence and inactivation may occur in class X proteins without disturbing the steady state of the entire system. The results also revealed an essential preference for gaining new interactions. Within the interactome of S. cerevisiae, the first interaction of a given protein increases its rate of gaining a new one by approximately 100%.
To relate these findings to the timescale of real evolution, it is reasonable to arbitrarily assume that a unit of time in the model corresponds to 10^{9} years. Then, an f_{ 0 } of 34376.8 means approximately 30 new proteins per 10^{6} years. Consequently, the characteristic times of proteome evolution can be estimated to equal 1.2·10^{8} and 3.5·10^{8} years. The shorter time describes the timescale of entering the “higher” class, and the longer time describes the timescale of protein deactivation. The characteristic time of gaining a new interaction is 4.5·10^{8} years.
From the perspective of describing the current distribution of protein degree or the dependence of the total number of links on the size of the interactome, a steadystate approximation for proteome evolution appears to be a correct simplification. Most of the observed proteins most likely originated during the “steady state era”. For a more precise description of the connectivity of older proteins, e.g., those from the preeukaryotic radiation era, the model should also take into account the variations with time in both the proteome size and the values of kinetic parameters.
One of the main predictions of the proposed model (Figure 6) is consistent with the finding that, on average, evolutionarily older proteins have more interactions with other proteins than do their younger counterparts [27]. Because the discussed model only addresses the overall PPI network evolution, the more detailed features of this process, i.e., the fast asymmetric functional divergence of duplicated genes [28] or the modular preferential attachment [18] were disregarded, offering a large simplification with no loss of prediction ability. Nevertheless, some asymmetric divergence and modularity is still contained in our model, mainly from the assumption of two different classes of protein importance.
Finally, the proposed model relates the static observables, such as the node degree distribution, to many dynamic evolutionary processes. The discussed dynamics are not a trivial consequence of the birth and death of proteins. The dynamics also involve the transition of proteins between classes, which leads to a dynamic balance, in which a given protein may change its importance class several times depending on the environmental conditions. Thus, the amplitudes in the derived formula for node degree distribution describe an effective dynamic content of each protein class but not the number of specific proteins.
As previously shown, the presented kinetic model of the evolution of a protein interaction network offers a solid foundation for future development and provides a productive research approach to protein interaction networks.
In future studies, it would be nice to have a more definitive evaluation of how the model’s simplifications affect its accuracy. Standard errors of the estimation (Table 3) show that the spontaneous loss of interactions, r, is statistically insignificant and is, thus, not likely to be critical for the stability of the model. Furthermore, the duplication rate, k_{ 2 }, is of less statistical significance. Possibly, these parameters could be omitted in simplifications that neglect parameters of the second order without considerable loss in the accuracy of the model.
Despite good fits, we are aware of the fact that the cited experimental methods have enormous potential for false data. The PPI data are full of false positives and false negatives, which, when unquestioningly included, tend to generate false conclusions. Necessarily, the model was applied to the data that exist. Highthroughput data tend to be worse than lowthroughput data [29]. We expect that the errors in the set of interactions can mainly disturb the estimation of the general parameters of the extensive type (amplitudes A_{ i }, probabilities p_{ i }). Test simulations that were performed indicate that a 10% increase in the value of those parameters may result in a change of the final estimated kinetic parameters of the model reaching up to 70%. Thus, the results may change in the face of future data.
The presented and applied model of the evolution of the protein interactome by its nature contains some abstraction, which does not invalidate the results (see Hamilton [30]). For example, the central concept of “essentiality” is a significant binary simplification of a gene's ability to survive and reproduce. In the future, this concept may be replaced by the more detailed continuous approach with the full spectrum of gene fitness. A similar school of thinking was shown in our previous paper [22], which presented multiexponential fitting that described the full spectrum of contributions from different classes of proteins. This method also indicated the domination of the two basic subpopulations.
Conclusions
The current model leads to a number of predictions that we can hope to test in the notsodistant future. The most interesting findings are the following:

A small sample of synchronized proteins decreases and differentiates; the degree of a single protein node expands.

The evolution of a node degree is slower than the evolution of the proteome.

The evolution of the total proteome stabilizes.

Entering the class of proteins that are essential for basic biological processes is approximately 50fold slower than leaving it.

Large dynamic changes, involving new protein emergence and inactivation in class X, do not disturb the steady state of the entire system.

There is a parabolic relationship between the total number of interactions and the total number of interacting proteins.

The connectivity of the oldest part of the interaction network is dense; the node degree distribution follows the sum of the two shifted powerlaw functions.
We hope that the above paper presents a helpful advance in this interesting area.
Methods
Mathematical formulation of a kinetic model of the evolution of a protein interaction network
Proteome formation
where f_{ 0 } is the rate of origination of entirely new proteins of class X, k_{ i } is the rate of protein inactivation, k_{ XY } and k_{ YX } are the rates of protein migration between classes X and Y, k_{ 2 } is protein duplication rate and t is the time.
Interactome formation
where ξ_{ 0 } is the degree of an entirely new protein, μ is the rate of an emerging new interaction within the proteome, Δε is the increase in the rate per link resulting from the preference effect, r is the rate of interaction loss and μ is the protein age. The meaning of the other symbols is the same as that previously stated. N_{∞} and κ are described by equations A.3 and A.5.
Node degree distribution
Total number of links
Declarations
Acknowledgements
This work has been supported by grant 772/NCOST/2010.
Authors’ Affiliations
References
 Barabasi AL, Reka A: Emergence of scaling in random networks. Science. 1999, 286: 509512. 10.1126/science.286.5439.509.View ArticlePubMedGoogle Scholar
 Jun SH, Jun WJ: Evolution of network from node division and generation. Chinese Phys. 2007, 16: 15811585. 10.1088/10091963/16/6/015.View ArticleGoogle Scholar
 Qin H, Lu HHS, Wu WB, Li WH: Evolution of the yeast protein interaction network. Proc Natl Acad Sci USA. 2003, 100: 1282012824. 10.1073/pnas.2235584100.PubMed CentralView ArticlePubMedGoogle Scholar
 Dam TJP, Snel B: Protein complex evolution does not involve extensive network rewiring. PLoS Comput Biol. 2008, 4 (7): e100013210.1371/journal.pcbi.1000132.PubMed CentralView ArticlePubMedGoogle Scholar
 Teichmann SA, Babu MM: Gene regulatory network growth by duplication. Nat Genet. 2004, 36: 492496. 10.1038/ng1340.View ArticlePubMedGoogle Scholar
 Zhao J, Ding GH, Tao L, Yu H, Yu ZH, Luo JH, Cao ZW, Li YX: Modular coevolution of metabolic networks. BMC Bioinforma. 2007, 8: 31110.1186/147121058311.View ArticleGoogle Scholar
 Kunin V, PereiraLeal JB, Ouzounis CA: Functional evolution of the yeast protein interaction network. Mol Biol Evol. 2004, 21: 11711176. 10.1093/molbev/msh085.View ArticlePubMedGoogle Scholar
 Tamames J, Moya A, Valencia A: Modular organization in the reductive evolution of proteinprotein interaction networks. Genome Biol. 2007, 8: R9410.1186/gb200785r94.PubMed CentralView ArticlePubMedGoogle Scholar
 Ortutay C, Vihinen M: Efficiency of the immune protein interaction network increases during evolution. Immunome Res. 2008, 4: 410.1186/1745758044.PubMed CentralView ArticlePubMedGoogle Scholar
 Evlampiev K, Isambert H: Modeling protein network evolution under genome duplication and domain shuffling. BMC Syst Biol. 2007, 1: 4910.1186/17520509149.PubMed CentralView ArticlePubMedGoogle Scholar
 Veron AS, Kaufmann K, BornbergBauer E: Evidence of interaction network evolution by wholegenome duplications: A case study in MADSbox proteins. Mol Biol Evol. 2007, 24: 670678.View ArticlePubMedGoogle Scholar
 Fraser HB, Hirsh AE, Steinmetz LM, Scharfe C, Feldman MW: Evolutionary rate in the protein interaction network. Science. 2002, 296: 750752. 10.1126/science.1068696.View ArticlePubMedGoogle Scholar
 Makino T, Gojobori T: Evolution of proteinprotein interaction network. Genome Dyn. 2007, 3: 1329.View ArticlePubMedGoogle Scholar
 Almaas E: Biological impacts and context of network theory. J Exp Biol. 2007, 210: 15481558. 10.1242/jeb.003731.View ArticlePubMedGoogle Scholar
 Wuchty S: Evolution and topology in the yeast protein interaction network. Genome Res. 2004, 14: 13101314. 10.1101/gr.2300204.PubMed CentralView ArticlePubMedGoogle Scholar
 Noirel J, Simonson T: Neutral evolution of proteinprotein interactions: a computational study using simple models. BMC Struct Biol. 2007, 7: 7910.1186/14726807779.PubMed CentralView ArticlePubMedGoogle Scholar
 Ispolatov I, Krapivsky PL, Yuryev A: Duplicationdivergence model of protein interaction network. Phys Rev E Stat Nonlin Soft Matter Phys. 2005, 71: 061911PubMed CentralView ArticlePubMedGoogle Scholar
 Kim WK, Marcotte EM: AgeDependent evolution of the yeast protein interaction network suggests a limited role of gene duplication and divergence. PLoS Comput. Bio. 2008, 4 (11): e100023210.1371/journal.pcbi.1000232.View ArticleGoogle Scholar
 Denny P, Preiser P, Williamson D, Wilson I: Evidence for a Single Origin of the 35 kb Plastid DNA in Apicomplexans. Protist. 1998, 149 (1): 5159. 10.1016/S14344610(98)700094.View ArticlePubMedGoogle Scholar
 Wilson RJ, Denny PW, Preiser PR, Rangachari K, Roberts K, Roy A, Whyte A, Strath M, Moore DJ, Moore PW: Complete gene map of the plastidlike DNA of the malaria parasite Plasmodium falciparum. J Mol Biol. 1996, 261 (2): 155172. 10.1006/jmbi.1996.0449.View ArticlePubMedGoogle Scholar
 Batada NN, Hurst LD, Tyers M: Evolutionary and physiological importance of hub proteins. PLoS Comput Biol. 2006, 2 (7): e8810.1371/journal.pcbi.0020088.PubMed CentralView ArticlePubMedGoogle Scholar
 Pawlowski PH, Kaczanowski S, Zielenkiewicz P: Protein interaction network. Double exponential model. J Proteomics Bioinform. 2008, 1: 061067. 10.4172/jpb.1000010.View ArticleGoogle Scholar
 PastorSatorras R, Smith E, Solec RV: Evolving protein interaction networks through gene duplication. J Theor Biol. 2003, 222: 199210. 10.1016/S00225193(03)000286.View ArticlePubMedGoogle Scholar
 Deutschbauer AM, Jaramillo DF, Proctor M, Kumm J, Hillenmeyer ME, Davis RW, Nislow C, Giaever G: Mechanisms of haploinsufficiency revealed by genomewide profiling in yeast. Genetics. 2005, 169 (4): 19151925. 10.1534/genetics.104.036871.PubMed CentralView ArticlePubMedGoogle Scholar
 Hillenmeyer ME, Fung E, Wildenhain J, Pierce SE, Hoon S, Lee W, Proctor M, St Onge RP, Tyers M, Koller D, Altman RB, Davis RW, Nislow C, Giaever G: The chemical genomic portrait of yeast: uncovering a phenotype for all genes. Science. 2008, 320: 362365. 10.1126/science.1150021.PubMed CentralView ArticlePubMedGoogle Scholar
 Shestopaloff YK: General law of growth and replication, growth equation and its applications. Biophys Rev Let. 2012, 07 (71): 10.1142/S1793048012500051.
 Eisenberg E, Levanon EY: Preferential attachment in the protein network evolution. Phys Rev Lett. 2003, 91: 138701View ArticlePubMedGoogle Scholar
 Wagner A: Asymmetric functional divergence of duplicate genes in yeast. Mol Biol Evol. 2002, 19 (10): 17601768. 10.1093/oxfordjournals.molbev.a003998.View ArticlePubMedGoogle Scholar
 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: 399403.View ArticlePubMedGoogle Scholar
 Hamilton WD: Geometry for the selfish herd. J Theor Biol. 1971, 31: 295311. 10.1016/00225193(71)901895.View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.