GraphCrunch: A tool for large network analyses
 Tijana Milenković^{1},
 Jason Lai^{1} and
 Nataša Pržulj^{1}Email author
DOI: 10.1186/14712105970
© Milenković et al; licensee BioMed Central Ltd. 2008
Received: 03 June 2007
Accepted: 30 January 2008
Published: 30 January 2008
Abstract
Background
The recent explosion in biological and other realworld network data has created the need for improved tools for large network analyses. In addition to well established global network properties, several new mathematical techniques for analyzing local structural properties of large networks have been developed. Small overrepresented subgraphs, called network motifs, have been introduced to identify simple building blocks of complex networks. Small induced subgraphs, called graphlets, have been used to develop "network signatures" that summarize network topologies. Based on these network signatures, two new highly sensitive measures of network local structural similarities were designed: the relative graphlet frequency distance (RGFdistance) and the graphlet degree distribution agreement (GDDagreement).
Finding adequate nullmodels for biological networks is important in many research domains. Network properties are used to assess the fit of network models to the data. Various network models have been proposed. To date, there does not exist a software tool that measures the above mentioned local network properties. Moreover, none of the existing tools compare realworld networks against a series of network models with respect to these local as well as a multitude of global network properties.
Results
Thus, we introduce GraphCrunch, a software tool that finds wellfitting network models by comparing large realworld networks against random graph models according to various network structural similarity measures. It has unique capabilities of finding computationally expensive RGFdistance and GDDagreement measures. In addition, it computes several standard global network measures and thus supports the largest variety of network measures thus far. Also, it is the first software tool that compares realworld networks against a series of network models and that has builtin parallel computing capabilities allowing for a user specified list of machines on which to perform compute intensive searches for local network properties. Furthermore, GraphCrunch is easily extendible to include additional network measures and models.
Conclusion
GraphCrunch is a software tool that implements the latest research on biological network models and properties: it compares realworld networks against a series of random graph models with respect to a multitude of local and global network properties. We present GraphCrunch as a comprehensive, parallelizable, and easily extendible software tool for analyzing and modeling large biological networks. The software is opensource and freely available at http://www.ics.uci.edu/~bionets/graphcrunch/. It runs under Linux, MacOS, and Windows Cygwin. In addition, it has an easy to use online web user interface that is available from the above web page.
Background
Motivation
The recent technological advances in experimental biology have yielded large amounts of biological network data. Many other realworld phenomena have also been described in terms of large networks (also called graphs), such as various types of social and technological networks. Thus, understanding these complex phenomena has become an important scientific problem that has lead to intensive research in network modeling and analyses.
Analogous to genetic sequence comparison, comparing large cellular networks will revolutionize biological understanding. However, comparing large networks is computationally infeasible due to NPcompleteness of the underlying subgraph isomorphism problem [1]. Note that even if the subgraph isomorphism was feasible, it would not find a practical application in biological network comparisons, since biological networks are extremely unlikely to be isomorphic. Thus, large network comparisons rely on heuristics, commonly called network parameters or properties. These properties are roughly categorized into global and local. The most widely used global properties are the degree distribution, the clustering coefficient, the clustering spectrum, the average diameter and the spectrum of shortest path lengths [2] (defined in Section "Global network properties"). Local properties include network motifs, small overrepresented subgraphs [3–5], and two measures based on graphlets, small induced subgraphs of large networks: the relative graphlet frequency distance (RGFdistance), which compares the frequencies of the appearance of graphlets in two networks [6], and the graphlet degree distribution agreement (GDDagreement), which is a graphletbased generalization of the degree distribution [7] (details are given in Section "Local network properties").
Various network models have been proposed for realworld networks. Starting with ErdösRényi random graphs [8], network models have progressed through a series of versions designed to match certain properties of realworld networks. Examples include random graphs that match the degree distribution of the data [9], network growth models that produce networks with scalefree degree distributions [10] or small network diameters [11], geometric random graphs [12], or networks that reproduce some biological and topological properties of real biological networks (e.g., stickiness model [13]).
Currently, there does not exist a software tool that measures all of the above mentioned global and local network properties. Computing global network properties is computationally and conceptually easy and various software tools are available for this purpose. However, none of them have builtin capabilities to compare realworld networks against a series of network models, based on these properties. The computational challenge is in finding local properties. Currently available software packages focus on searching for network motifs [14–16]. Thus far, there does not exist a publicly available, opensource software tool that computes local properties other than network motifs. For this reason, GraphCrunch primarily focuses on graphlets and has the unique capabilities of computing RGDdistance and GDDagreement measures of local network structure. We introduce GraphCrunch as a comprehensive, parallelizable, and easily extendible opensource software tool for analyzing and modeling of large realworld networks.
Methods
Network properties currently supported by GraphCrunch.
Global Properties: 

Degree distribution 
Clustering coefficient 
Clustering spectrum 
Average diameter 
Spectrum of shortest path lengths 
Local Properties: 
Relative graphlet frequency distance (RGFdistance) 
Graphlet degree distribution agreement (GDDagreement) 
Network models currently supported by GraphCrunch.
Models: 

ErdösRényi random graphs 
Random graphs with the same degree distribution as the data 
Scalefree BarabásiAlbert model graphs 
Ndimensional geometric random graphs 
Stickiness model graphs 
Network models
GraphCrunch currently supports five different types of random graph models: (1) ErdösRényi random graphs (henceforth denoted by "ER") [8]; (2) random graphs with the same degree distribution as the data (henceforth denoted by "ERDD") [9]; (3) BarabásiAlbert type scalefree networks (henceforth denoted by "SFBA") [10]; (4) ndimensional geometric random graphs for all positive integers n (henceforth denoted by "GEOn D") [12]; and (5) stickiness model networks (henceforth denoted by "STICKY") [13].
All generated model networks have the number of nodes and edges within 1% of those in the realworld networks. The model network generators are implemented as follows. ErdösRényi random graphs are generated by using the LEDA [17] implementation of G_{n,m}, a random graph G with n nodes and m edges. Random graphs with the same degree distribution as the data are generated by using the "stubs method" (see section IV.B.1 of [2] for details): the number of "stubs" (to be filled by edges) is assigned to each node in the model network according to the degree distribution of the realworld network; edges are created between pairs of nodes picked at random; after an edge is created, the number of "stubs" left available at the corresponding "endnodes" of the edge is decreased by one. Scalefree networks are generated by using the BarabásiAlbert preferential attachment model [10]. Geometric random graphs are defined as follows: nodes correspond to uniformly randomly distributed points in a metric space and edges are created between pairs of nodes if the corresponding points are close enough in the metric space according to some distance norm. A variant of geometric random graphs supported by GraphCrunch uses ndimensional Euclidean boxes and the Euclidean distance norm. The default is set to be the 3dimensional Euclidean space ("GEO3D"), but user specified dimensions are also supported. Finally, "stickiness network model" is based on stickiness indices, numbers that summarize node connectivities and thus also the complexities of binding domains of proteins in proteinprotein interaction (PPI) networks (see [13] for details).
Local network properties
Due to NPcompleteness of the underlying subgraph isomorphism problem, large network comparisons rely on network properties. These heuristics need to encompass large number of constraints, in order to reduce degrees of freedom in which networks being compared can vary. GraphCrunch has the unique capability to compute two highly constraining measures of local structural similarities between two networks: RGFdistance [6] and GDDagreement [7]. These measures are based on graphlets, small connected nonisomorphic induced subgraphs of large networks [6]. Note that graphlets are different from network motifs since they must be induced subgraphs (motifs are partial subgraphs) and since they do not need to be overrepresented in the data when compared with "randomized" networks. An induced subgraph of a graph G on a subset S of nodes of G is obtained by taking S and all edges of G having both endpoints in S; partial subgraphs are obtained by taking S and some of the edges of G having both endpoints in S. Since the number of graphlets on n nodes increases superexponentially with n, GraphCrunch currently bases its RGFdistance and GDDagreement computations on 3–5node graphlets (presented in Figure S1 in Additional file 1).
RGFdistance compares the frequencies of the appearance of all 3–5node graphlets in two networks (see [6] for details). Note that the networks being compared by GraphCrunch always have the same number of edges, and thus the frequencies of occurrence of the only 1node graphlet, a node, and the only 2node graphlet, an edge, are also taken into account by this measure. Thus, RGFdistance encompasses 31 similarity constraints by examining the fit of 31 graphlet frequencies.
GDDagreement generalizes the notion of the degree distribution to the spectrum of graphlet degree distributions (GDDs) in the following way [7]. The degree distribution measures the number of nodes of degree k, i.e., the number of nodes "touching" k edges, for each value of k. Note that an edge is the only graphlet with two nodes (graphlet denoted by G_{0} in Figure S2 in Additional file 1). GDDs generalize the degree distribution to other graphlets: they measure for each graphlet G_{ i }, i ∈ 0, 1,..., 29, (illustrated in Figure S2 in Additional file 1) the number of nodes "touching" k graphlets G_{ i }at a particular node. A node at which a graphlet is "touched" is topologically relevant, since it allows us to distinguish between nodes "touching", for example, a copy of graphlet G_{1} in Figure S2 in Additional file 1 at an endnode, or at the middle node. This is summarized by automorphism orbits (or just orbits, for brevity), as illustrated in Figure S2 in Additional file 1: for graphlets G_{0}, G_{1},..., G_{29}, there are 73 different orbits, numerated from 0 to 72 (see [7] for details). For each orbit j, we measure the j^{ th }GDD, i.e., the distribution of the number of nodes "touching" the corresponding graphlet at orbit j. Thus, the degree distribution is the 0^{ th }GDD. The j^{ th }GDDagreement compares the j^{ th }GDDs of two networks (see [7] for details). The total GDDagreement between two networks is the arithmetic or the geometric average of the j^{ th }GDDagreements over all j (henceforth arithmetic and geometric averages are denoted by "amean" and "gmean", respectively). GDDagreement is scaled to always be between 0 and 1, where 1 means that two networks are identical with respect to this property. By calculating the fit of each of the 73 GDDs of the networks being compared, GDDagreement encompasses 73 similarity constraints. Furthermore, each of these 73 constraints enforces a similarity of two distributions, additionally restricting the ways in which the networks being compared can differ. (Note that the degree distribution is only one of these 73 constraints.) Therefore, GDDagreement is a very strong measure of structural similarity between two networks. Both of the RGFdistance and GDDagreement measures were used to discover a new, wellfitting, geometric random graph model of PPI networks [6, 7]. In Section "A case study", we illustrate the biological importance of choosing a wellfitting network null model of PPI networks.
Global network properties
Global network properties currently supported by GraphCrunch are the degree distribution, the average network diameter, the spectrum of shortest path lengths, the average clustering coefficient, and the clustering spectrum. The degree of a node is the number of edges incident to the node. The degree distribution, P(k), describes the probability that a node has degree k. The smallest number of links that have to be traversed to get from a node x to a node y in a network is called the distance between nodes x and y and a path through the network that achieves this distance is called the shortest path between nodes x and y. The average of shortest path lengths over all pairs of nodes in a network is called the average network diameter. The spectrum of shortest path lengths is the distribution of shortest path lengths between all pairs of nodes in a network. The clustering coefficient of a node z in a network, C_{ z }, is defined as the probability that two nodes x and y which are connected to the node z are themselves connected. The average of C_{ z }over all nodes z of a network is the clustering coefficient, C, of the network; it measures the tendency of the network to form highly interconnected regions called clusters. The distribution of the average clustering coefficients of all nodes of degree k in a network is the clustering spectrum, C(k).
Implementation
GraphCrunch can be downloaded from the GraphCrunch home page (see Section "Availability and requirements"). It runs under Linux, MacOS, and Windows Cygwin. GraphCrunch is implemented in C++, Perl, and Bourne Shell scripts. The programs for generating model networks and calculating network properties are implemented using C++ and the LEDA library for combinatorial and geometric computing [17]. Shell scripts collect the output of these programs to create the final results of the analyses, as well as the graphical and tabular representation of the statistics summarizing the results. The programs and the shell scripts are organized in the directories presented in Table S1 in Additional file 1. This organization allows for easy extendibility of the software to include additional network models and properties (see Section 1 in Additional file 1 for details). Furthermore, GraphCrunch has builtin parallel computing capabilities: it distributes its processes over a user specified cluster of machines. Installation details are described in Section 1 in Additional file 1.
GraphCrunch interfaces
There are three ways of running GraphCrunch: via the commandline interface, the rundialog interface, and the online web interface. Upon installation, a user can choose either the commandline or the rundialog interface. The commandline interface allows for specifying all of the following in a single command: the realworld network (input graph) to be processed, the random graph models against which the data is to be compared, the number of networks to be generated per random graph model, the network properties and comparisons between the data and the model networks, and the name of the output data file (see Section 2.1 in Additional file 1 for details). The rundialog interface is available for the Linux and MacOS versions of GraphCrunch. It provides the same functionality as the commandline interface in a perhaps more userfriendly manner: it guides a user through the sequence of screens, step by step, as presented in Section 2.2 in Additional file 1 (also see Figure S3 in Additional file 1). The input and output formats of these two interfaces are similar (described in Sections "Input format" and "Output format and results" below) and they both allow for largescale scientific computing network analysis and modeling projects. The GraphCrunch online web user interface (available from the GraphCrunch webpage) is provided for nonexpert users and those with less intensive processing needs; we recommend that firsttime users start with online GraphCrunch. The online GraphCrunch is described in more details in Section 2.3 in Additional file 1 (also see Figures S4, S5, and S6 in Additional file 1).
Input format
GraphCrunch supports two input graph formats: the LEDA graph format (.gw) and the "edge list" format (.txt). The specifics of the LEDA graph format are given at the GraphCrunch web page. The edge list format is simply the graph adjacency list, i.e., the list of node pairs (edges of the network) separated by tabs or spaces, with one node pair per line. The current implementation of GraphCrunch deals with undirected, simple (i.e., no loops or multiple edges), unweighted graphs. Thus, for either of the above two formats, GraphCrunch automatically removes all selfloops, multiple edges, and edge directions. Dealing with directed weighted edges is a subject of future work (see Section "Future directions").
Output format and results
GraphCrunch creates three types of output: the tabular output file, the set of intermediate files, and the visualized output.
The tabular output file is a spreadsheet of tabseparated values (.tsv) that contains summarized output statistics. An example of a GraphCrunch output file is presented in Table S2 in Additional file 1 (also see Section 2.4 in Additional file 1).
The set of intermediate files includes generated model networks corresponding to the input network, in LEDA graph (.gw) format, and the files containing the network properties (e.g., clustering spectra, graphlet counts, graphlet degree distributions etc.). The intermediate files allow for additional analyses of the results without performing any additional computeintensive processing. Also, the tabular output file contains only the statistics of network parameter similarities between the data and the model networks, but not the results from which the statistics were computed – these results are contained in the intermediate files. For example, the GDDagreement presented in the output file is a single number between 0 and 1 while the intermediate files contain the actual GDDs for each of the 73 orbits. Thus, intermediate files are needed if one needs to analyze a particular GDD. Similar holds for other network properties.
The visualized output is a set of files (in .ps format) that contain userfriendly graphical interpretations of the results presented in the tabular output file. One graphical file (plot) is created per network property. A plot illustrates the fit of the network models to one or more realworld networks with respect to the given property. Thus, in a single plot, it is possible to simultaneously illustrate the fit of network models to many realworld networks with respect to one property. Examples of plots are presented in Figure S7 in Additional file 1.
Further details about output results are available in Section 2.4 in Additional file 1.
Results
A case study
Motivation
Since modeling biochemical networks is a vibrant research area, we analyze proteinprotein interaction (PPI) networks to illustrate functionality of GraphCrunch. In a PPI network, nodes correspond to proteins and undirected edges represent physical interactions between the proteins. The choice of an appropriate null model can have important implications for any graphbased analysis of PPI networks. For example, the use of an adequate null model is vital for structural motif discovery, which requires comparing realworld PPI networks with randomized ones [18, 19]. Using an inappropriate network null model may identify as overrepresented (underrepresented) subgraphs that otherwise would not have been identified. Another example is that a good null model can be used to guide biological experiments in a time and costoptimal way and thus minimize the costs of interactome detection [20].
Here, we illustrate a way in which GraphCrunch could be used to help answer some of the PPI network modeling questions. The results presented in this section are not intended for suggesting the bestfitting null model for PPI networks, but rather as an illustration of a possible application of GraphCrunch. Some of the questions that one might ask are the following. What is the bestfitting null model for PPI networks? Do all PPI networks belong to the same graph family, or are they categorized into graph subfamilies? How does the incompleteness of the PPI data affect the choice of an appropriate null model? What is the effect of the noise in PPI networks? Can a unique null model (or a model family) be used for motif search in all PPI networks?
Analysis
A wellfitting null model should generate graphs that closely resemble the structure of real networks. This closeness in structure is reflected across a wide range of statistical measures. Thus, testing the fit of a model entails comparing modelderived random graphs to real networks according to these measures. The more constraining the measures are, the fewer degrees of freedom exist in which the compared networks can vary. Thus, by using highly constraining measures a betterfitting null model can be found. GraphCrunch supports various global and local measures of network structure. Global network properties, such as the degree distribution, may not be detailed enough to capture the complex topological characteristics of PPI networks. Thus, here we focus on RGFdistance and GDDagreement that impose a large number of constraints on the networks being compared (as explained in Section "Local network properties").
We analyze PPI networks of the eukaryotic organisms yeast Saccharomyces cerevisiae, fruitfly Drosophila melanogaster, nematode worm Caenorhabditis elegans, and human. To illustrate the process of finding the bestfitting null model for a given realworld network, we first generate three random networks belonging to each of the five random network models currently supported by GraphCrunch. Next, we compare the realworld network with model networks according to RGFdistance and GDDagreement measures. Finally, we report the statistics of the fit of each of the network models to the realworld network.
Completeness of interactomes
First, we analyze higherconfidence ("core") parts of PPI networks of the four organisms mentioned above. We denote by "yeastcore" the yeast highconfidence PPI network described by von Mering et al. [21], consisting of 988 proteins and 2,455 interactions among them. We denote by "flycore" the fruitfly D. melanogaster highconfidence PPI network published by Giot et al. [22], containing 4,602 proteins and 4,637 interactions. We denote by "wormcore" the worm C. elegans core PPI network described by Li et al. [23], containing 1,356 proteins and 1,983 interactions. Finally, we denote by "humancore" the human core PPI network described by Stelzl et al. [24], having 363 proteins and 756 interactions.
It is possible that PPI networks of different organisms belong to different graph families. It is also possible that they belong to the same graph family, but that the current incompleteness of the interactome data for these organisms produces apparent PPI network structural differences [25, 26]. Since the two PPI networks of yeast and fruitfly are more extensively studied and thus are more complete, and since they belong to the same graph family, it is possible that the latter interpretation is correct. If we accept this hypothesis, then the improved fit of the two degreepreserving network models (ERDD and STICKY) to the two less complete PPI networks of worm and human may be attributed to their incompleteness rather than to their true biological structure. Regardless of the reasons for the difference in wellfitting null models of these four PPI networks, a question of whether the same random graph model should be used to identify structural motifs in PPI networks remains.
Noise in interactomes
A case study: conclusions
There are many open questions when analyzing and modeling PPI networks. As demonstrated, GraphCrunch is a tool for large network analyses and modeling, capable of finding wellfitting random network models for realworld networks according to various local (illustrated above) and global network properties. It outputs the analysis and modeling results in the form that allows for easy comparisons of different realworld networks. As such, it can be a useful tool in the process of gathering the knowledge necessary to answer many biological questions.
GraphCrunch performance
To evaluate the performance of GraphCrunch and its behavior with increasing network complexity, we perform the following analyses:

First, we observe GraphCrunch's running times for input networks with different number of nodes and edges, but with constant edge densities. We determine the edge density E/V (where V is the number of nodes and E is the number of edges in a network) to be 2.94, by taking the average edge density of five yeast PPI networks. We use the yeast PPI networks since the yeast interactome has been extensively studied and thus is more complete compared to interactomes of other organisms. We find the average edge density of the following yeast PPI networks: the network containing only highconfidence interactions described by von Mering et al. [21], the network containing the topranked 11,000 interactions described by von Mering et al. [21], the "core" subset of the yeast PPI network from DIP [27, 28], and the entire yeast PPI networks from DIP [27] and MIPS [29] databases downloaded in April 2007. We generate input networks consisting of 100, 500, 1,000, 2,000, 3,000, 4,000, 5,000, 7,500, and 10,000 nodes, and thus with 294, 1,471, 2,941, 5,882, 8,820, 11,764, 14,706, 22,050, and 29,411 edges, respectively, enforcing the same edge density in each of these networks.

Next, we measure and compare the running times of GraphCrunch for networks with the same number of nodes, but with different edge densities. We observe edge densities smaller, equal to, and larger than the average edge density in yeast PPI networks described above. We generate input networks with 1,000 nodes using edge densities of about 2, 3, 4, 7, and 11, resulting with 2,273, 2,941, 4,167, 7,143, and 11,100 edges amongst 1,000 nodes, respectively.

Finally, we measure the influence of different properties of input networks on performance of GraphCrunch, by testing as input data networks belonging to different graph models. We use three different graph model generators to create the input networks: ER, GEO3D, and SFBA (introduced in Section "Network models"). These input networks differ in the following: GEO3D model networks have high clustering coefficients and Poisson degree distributions; ER model networks have low clustering coefficients and Poisson degree distributions; and SFBA model networks have low clustering coefficients and powerlaw degree distributions. Therefore, we are able to evaluate the effect of the degree distribution and the clustering coefficient measures on the performance of GraphCrunch, by directly comparing running times for input networks drawn from different network models.
We run GraphCrunch once for each of the input networks. We used the following GraphCrunch settings for each run. GraphCrunch compares the input network against all five network models that it currently supports. It generates three random networks per network model and thus, it analyzes 15 model networks per input network. When comparing input and model networks, GraphCrunch computes all network properties that it currently supports.
Since GraphCrunch exhaustively finds all graphlets in the input and model networks, the exponential increase in the running time is expected (Figure 3). Note however, that the largest currently available PPI networks have about 6,000 nodes, which coupled with GraphCrunch's parallel computing capabilities gives reasonable processing times (also see Section "Future directions" for planned future speedups). The degree distribution has an influence on the performance of GraphCrunch. Since input scalefree SFBA networks contain highly connected nodes (hubs), during the exhaustive graphlet enumeration process GraphCrunch severely overcounts graphlets in the vicinity of hubs (of course, the overcounting is corrected after the completion of the exhaustive graphlet enumeration and GraphCrunch returns correct graphlet counts) which degrades its time performance (see [30] for details). Thus, if a large dense network with scalefree degree distribution is being compared to a large number of random graphs mainly belonging to scalefree network models, a user may want to consider distributing the processing over a cluster of machines. Clustering coefficients do not have an influence on the performance of GraphCrunch.
In addition to the analyses described above, we evaluate GraphCrunch's performance against the performances of other network analysis software tools. See Section "Comparison with the existing tools" for details.
GraphCrunch applications
GraphCrunch is primarily intended for analyzing and modeling of biological networks, although it can be used for other realworld networks as well. GraphCrunch has been applied to and validated through several research projects, some of which have already been published in prestigious refereed journals such as Science, Bioinformatics, Journal of the Royal Society Interface, and PLoS Computational Biology [6, 7, 13, 31, 32]. For example, GraphCrunch was used for analyzing global network properties of a mammalian TGFβ signaling network [31]. Also, it was used for computing RGFdistances between several realworld PPI networks and model networks [6], as well as for finding GDDagreements between PPI and model networks [7]. These studies demonstrated that PPI networks of several eukaryotic organisms are better modelled by geometric random graphs then by scalefree networks [6, 7]; this was the first time that geometric random graphs were used to model PPI networks. GraphCrunch was also used to demonstrate the superiority of the fit of the newly introduced stickinessindex based network model of PPI networks over other network models [13]. Furthermore, GraphCrunch was used to compare graphlet frequency distributions of the preferential attachmentbased and the duplicationbased scalefree model networks that were created by starting with different seed graphs and to demonstrate that different seed graphs influence the topologies of the resulting networks [32]. In addition to analyzing and modeling PPI networks described above, GraphCrunch has recently been successfully applied to other types of biological networks such as biomolecular interaction and cellular interaction networks (work in progress).
Discussion
Comparison with the existing tools
The summary of the functionalities of GraphCrunch and similar purpose software tools. "Yes" and "No" denote that a software package supports or does not support a given function, respectively.
Software tool  Global properties  Local properties  Visualization  Number of network models  Output of model networks  Input graph format 

GraphCrunch  Yes  Yes  Yes (Results)  5  Yes  LEDA graph format (.gw); Edge list (.txt) 
mfinder  No  Yes  Yes, with mDraw  3  Yes (not by default)  Edge list (.txt) 
MAVisto  No  Yes  Yes  1  No  Pajek format (.net) and .gml 
FANMOD  No  Yes  Yes  3  No  Edge list (.txt) 
tYNA  Yes (limited)  Yes (limited)  Yes  0  No  Edge list (.txt) and .sif 
pajek  Yes  Yes (limited)  Yes  2  Yes  Pajek format (.net) 
Compared to the tools mentioned above, GraphCrunch provides the largest variety of network properties. The main purpose of mfinder, MAVisto, and FANMOD is motif search; they do not compute global network properties. On the other hand, pajek focuses on global network properties and has very limited local network analysis capabilities; its search for subgraphs is limited to 3–4node rings. tYNA's global and local network analyses are limited: it calculates the statistics of global network properties and focuses on three network motif types only. Unlike any of these software packages, GraphCrunch uses all of the 2–5node graphlets for computing its two highly constraining graphletbased local network properties, GDDagreement [7] and RGFdistance [6] (described in Section "Local network properties"), along with five standard global properties (described in Section "Global network properties"). We chose to use all 2–5node graphlets for the following reasons. There are only nine 2–4node graphlets, so if only these would be used, the complexity of the local network topologies of networks being compared would not be captured. On the other hand, using graphlets with 6 or more nodes would substantially increase the computational complexity. Since we provide GraphCrunch as an opensource software package, users could choose to modify this or any other part of the code to fit their particular needs.
Furthermore, GraphCrunch uses all of these properties for comparing realworld networks against a series of network models. Five network models are currently supported by GraphCrunch. Although Mfinder, FANMOD and pajek offer more than one network model (MAVisto does not), none of these tools support such a variety of network models as GraphCrunch does (see Table 3). Note that tYNA does not generate random models at all and it searches for subgraphs in realworld networks only. Furthermore, GraphCrunch determines the fit of the various network models to the realworld networks with respect to an array of various global and local network properties; none of the other currently available network analysis software tools have this functionality.
Out of the above mentioned software packages that generate random graphs corresponding to the realworld networks, only mfinder can keep the generated random graphs after the output is created, but not by default. Pajek is capable of generating random networks, but not for the purpose of comparing them with the data. In contrast, GraphCrunch creates, compares with the data, and outputs all of the random graphs that it generates. While some of the software packages that compute global network properties provide only simple statistics (e.g., averages) without outputting the actual data used for computing them, GraphCrunch provides such raw data; this enables the users to perform additional analyses without having to do any further computing on the networks being analyzed.
Mfinder and FANMOD both include an option of using random subgraph sampling heuristics for speeding up the computing time, a feature that GraphCrunch currently does not support (see Section "Future directions"). However, GraphCrunch's exhaustive graphlet counting is very competitive. We tested mfinder, FANMOD, and GraphCrunch on the E. coli transcriptional regulation network [4] (we ignored the directionality of this network) on a 3.0 GHz Xenon processor with 1 GB of RAM. Using the exhaustive search mode for undirected subgraphs of size 5, mfinder spent 8 minutes 50 seconds counting, FANMOD spent 7 seconds, and GraphCrunch spent a second.
Furthermore, GraphCrunch has distributed computing capabilities for running the graphletbased analyses in parallel over a user specified cluster of machines, a feature that is not supported by Mfinder, FANMOD, MAVisto, tYNA, or pajek. This functionality will become crucial as the biological network data sets grow.
Future directions
Due to an increasing volume of biological network data that is becoming available, new models for representing these networks and new techniques for analyzing and comparing them will unquestionably be needed. Thus, we will keep extending GraphCrunch with new functions. One of the first extensions will be enabling GraphCrunch to deal with directed and weighted networks. Also, we will extend it to support additional network models and properties.
At the moment, GraphCrunch compares a realworld network with a series of network models with respect to network properties. We will upgrade GraphCrunch with a capability for comparing two or more realworld networks.
We will also implement two samplingbased heuristic approaches for estimating graphlet frequency distributions in PPI networks: Targeted Node Processing (TNP) and Neighborhood Local Search (NLS) [30]. TNP achieves accurate graphlet frequency distribution estimates, 300–690 times faster than the exhaustive searches. NLS performs 95–377 times faster than the exhaustive searches. We will also extend these heuristics to produce fast approximate GDDs and implement these extensions in GraphCrunch. Due to their high efficiency, these heuristics will be crucial as biological network data sets become larger and more complete.
Conclusion
GraphCrunch is a freely available, opensource software tool that implements the latest research in analyzing and modeling of biological networks. GraphCrunch analyzes large biological and other realworld networks and finds bestfitting network models by comparing realworld networks against random graph models with respect to global and local network properties. GraphCrunch supports a variety of network models and properties. Moreover, it is easily extendible to include additional network models and properties. GraphCrunch has builtin parallel computing capabilities, a feature that will become crucial as biological network data sets grow.
As biological data become larger and more complete, the need for improving the network analysis, modeling, and comparison techniques will continue to rise. We will keep including into GraphCrunch the latest research results in network analysis and modeling. Furthermore, our longerterm goal is to make GraphCrunch capable of performing biological network alignments.
Availability and requirements

Project name: GraphCrunch.

Project home page: http://www.ics.uci.edu/~bionets/graphcrunch/.

Operating systems: Linux, MacOS, and Windows Cygwin.

Programming language: C++, Perl, and Bourne Shell scripts.

Other requirements: Windows Cygwin version requires the LEDA 5.0.1 Cygwin license and the gcc 3.4 compiler. We recommend that Perl 5.6+ as well as dialog 0.3+ or Xdialog are also installed for each of the three operating systems.

License: None.

Any restrictions to use by nonacademics: None.
Declarations
Acknowledgements
We thank the Institute for Genomics and Bioinformatics (IGB) at UC Irvine for providing computing resources. This project was supported by the NSF CAREER IIS0644424 grant.
Authors’ Affiliations
References
 Cook S: The complexity of theoremproving procedures. In Proc 3rd Ann ACM Symp on Theory of Computing: 1971. New York, Association for Computing Machinery; 1971:151–158.Google Scholar
 Newman MEJ: The structure and function of complex networks. SIAM Review 2003, 45(2):167–256.View ArticleGoogle Scholar
 Milo R, ShenOrr SS, Itzkovitz S, Kashtan N, Chklovskii D, Alon U: Network motifs: simple building blocks of complex networks. Science 2002, 298: 824–827.View ArticlePubMedGoogle Scholar
 ShenOrr SS, Milo R, Mangan S, Alon U: Network motifs in the transcriptional regulation network of Escherichia coli. Nature Genetics 2002, 31: 64–68.View ArticlePubMedGoogle Scholar
 Milo R, Itzkovitz S, Kashtan N, Levitt R, ShenOrr S, Ayzenshtat I, Sheffer M, Alon U: Superfamilies of Evolved and Designed Networks. Science 2004, 303: 1538–1542.View ArticlePubMedGoogle Scholar
 Pržulj N, Corneil DG, Jurisica I: Modeling Interactome: ScaleFree or Geometric? Bioinformatics 2004, 20(18):3508–3515.View ArticlePubMedGoogle Scholar
 Pržulj N: Biological Network Comparison Using Graphlet Degree Distribution. Bioinformatics 2006, 23: e177e183.View ArticleGoogle Scholar
 Erdös P, Rényi A: On random graphs. Publicationes Mathematicae 1959, 6: 290–297.Google Scholar
 Molloy M, Reed B: A critical point of random graphs with a given degree sequence. Random Structures and Algorithms 1995, 6: 161–180.View ArticleGoogle Scholar
 Barabási AL, Albert R: Emergence of scaling in random networks. Science 1999, 286(5439):509–512.View ArticlePubMedGoogle Scholar
 Watts DJ, Strogatz SH: Collective dynamics of 'smallworld' networks. Nature 1998, 393: 440–442.View ArticlePubMedGoogle Scholar
 Penrose M: Geometric Random Graphs. Oxford University Press; 2003.View ArticleGoogle Scholar
 Pržulj N, Higham D: Modelling proteinprotein interaction networks via a stickiness index. Journal of the Royal Society Interface 2006, 3(10):711–716.PubMed CentralView ArticleGoogle Scholar
 Kashtan N, Itzkovitz S, Milo R, Alon U: Efficient sampling algorithm for estimating subgraph concentrations and detecting network motifs. Bioinformatics 2004, 20: 1746–1758.View ArticlePubMedGoogle Scholar
 Schreiber F, Schwobbermeyer H: MAVisto: a tool for the exploration of network motifs. Bioinformatics 2005, 21: 3572–3574.View ArticlePubMedGoogle Scholar
 Wernicke S, Rasche F: FANMOD: a tool for fast network motif detection. Bioinformatics 2006, 22: 1152–1153.View ArticlePubMedGoogle Scholar
 Mehlhorn K, Naher S: Leda: A platform for combinatorial and geometric computing. Cambridge University Press; 1999.Google Scholar
 ArtzyRandrup Y, Fleishman SJ, BenTal N, Stone L: Comment on "Network Motifs: Simple Building Blocks of Complex Networks" and "Superfamilies of Evolved and Designed Networks". Science 2004, 305: 1107c.View ArticleGoogle Scholar
 Milo R, Itzkovitz S, Kashtan N, Levitt R, Alon U: Response to Comment on "Network Motifs: Simple Building Blocks of Complex Networks" and "Superfamilies of Evolved and Designed Networks". Science 2004, 305: 1107d.View ArticleGoogle Scholar
 Lappe M, Holm L: Unraveling protein interaction networks with nearoptimal efficiency. Nature Biotechnology 2004, 22: 98–103.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(6887):399–403.View ArticlePubMedGoogle Scholar
 Giot L, Bader J, Brouwer C, Chaudhuri A, Kuang B, Li Y, Hao Y, Ooi C, Godwin B, Vitols E, Vijayadamodar G, Pochart P, Machineni H, Welsh M, Kong Y, Zerhusen B, Malcolm R, Varrone Z, Collis A, Minto M, Burgess S, McDaniel L, Stimpson E, Spriggs F, Williams J, Neurath K, Ioime N, Agee M, Voss E, Furtak K, Renzulli R, Aanensen N, Carrolla S, Bickelhaupt E, Lazovatsky Y, DaSilva A, Zhong J, Stanyon C, Finley RJ, White K, Braverman M, Jarvie T, Gold S, Leach M, Knight J, Shimkets R, McKenna M, Chant J, Rothberg J: A protein interaction map of Drosophila melanogaster. Science 2003, 302(5651):1727–1736.View ArticlePubMedGoogle Scholar
 Li S, Armstrong C, Bertin N, Ge H, Milstein S, Boxem M, Vidalain PO, Han JD, Chesneau A, Hao T, Goldberg N, Li DS, Martinez M, Rual JF, Lamesch P, Xu L, Tewari M, Wong S, Zhang L, Berriz G, Jacotot L, Vaglio P, Reboul J, HirozaneKishikawa T, Li Q, Gabel H, Elewa A, Baumgartner B, Rose D, Yu H, Bosak S, Sequerra R, Fraser A, Mango S, Saxton W, Strome S, van den Heuvel S, Piano F, Vandenhaute J, Sardet C, Gerstein M, DoucetteStamm L, Gunsalus K, Harper J, Cusick M, Roth F, Hill D, Vidal M: A Map of the Interactome Network of the Metazoan C. elegans. Science 2004, 303: 540–543.PubMed CentralView ArticlePubMedGoogle Scholar
 Stelzl U, Worm U, Lalowski M, Haenig C, Brembeck F, Goehler H, Stroedicke M, Zenkner M, Schoenherr A, Koeppen S, Timm J, Mintzlaff S, Abraham C, Bock N, Kietzmann S, Goedde A, Toksoz E, Droege A, Krobitsch S, Korn B, Birchmeier W, Lehrach H, Wanker E: A Human ProteinProtein Interaction Network: A Resource for Annotating the Proteome. Cell 2005, 122: 957–968.View ArticlePubMedGoogle Scholar
 Stumpf MPH, Wiuf C, May RM: Subnets of scalefree networks are not scalefree: Sampling properties of networks. Proceedings of the National Academy of Sciences 2005, 102: 4221–4224.View ArticleGoogle Scholar
 Han JDH, Dupuy D, Bertin N, Cusick ME, M V: Effect of sampling on topology predictions of proteinprotein interaction networks. Nature Biotechnology 2005, 23: 839–844.View ArticlePubMedGoogle Scholar
 Xenarios I, Salwinski L, Duan J, Higney P, Kim S, Eisenberg D: DIP, the Database of Interacting Proteins: a research tool for studying cellular networks of protein interactions. Nucleic Acids Research 2002, 30: 303–305.PubMed CentralView ArticlePubMedGoogle Scholar
 Deane C, Salwinski L, Xenarios I, Eisenberg D: Protein interactions: two methods for assessment of the reliability of high throughput observations. Molecular and Cellular Proteomics 2002, 1: 349–356.View ArticlePubMedGoogle Scholar
 Mewes HW, Frishman D, Guldener U, Mannhaupt G, Mayer K, Mokrejs M, Morgenstern B, Munsterkotter M, Rudd S, Weil B: MIPS: a database for genomes and protein sequences. Nucleic Acids Research 2002, 30: 31–34.PubMed CentralView ArticlePubMedGoogle Scholar
 Pržulj N, Corneil DG, Jurisica I: Efficient Estimation of Graphlet Frequency Distributions in ProteinProtein Interaction Networks. Bioinformatics 2006, 22(8):974–980.View ArticlePubMedGoogle Scholar
 BarriosRodiles M, Brown KR, Ozdamar B, Liu Z, Donovan RS, Shinjo F, Liu Y, Bose R, Dembowy J, Taylor IW, Luga V, Przulj N, Robinson M, Suzuki H, Hayashizaki Y, Jurisica I, Wrana JL: HighThroughput Mapping of a Dynamic Signaling Network in Mammalian Cells. Science 2005, 307: 1621–1625.View ArticlePubMedGoogle Scholar
 Hormozdiari F, Berenbrink P, Pržulj N, Sahinalp C: Not All Scale Free Networks are Born Equal: the Role of the Seed Graph in PPI Netwok Emulation. PLoS Computational Biology, in press.Google Scholar
 Yu H, Zhu X, Greenbaum D, Karro J, Gerstein M: TopNet: a tool for comparing biological subnetworks, correlating protein properties with topological statistics. Nucleic Acids Research 2004, 32: 328–337.PubMed CentralView ArticlePubMedGoogle Scholar
 Yip K, Yu H, Kim P, Schultz M, Gerstein M: The tYNA platform for comparative interactomics: a web tool for managing, comparing and mining multiple networks. Bioinformatics 2006, 22: 2968–2970.View ArticlePubMedGoogle Scholar
 Batagelj V, Mrvar A: Pajek – Program for Analysis and Visualization of Large Networks. Timeshift The World in TwentyFive Years: Ars Electronica 2004, 242–251.Google 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.