You are on page 1of 11

NeuroImage 52 (2010) 1059–1069

Contents lists available at ScienceDirect

NeuroImage
j o u r n a l h o m e p a g e : w w w. e l s e v i e r. c o m / l o c a t e / y n i m g

Complex network measures of brain connectivity: Uses and interpretations
Mikail Rubinov a,b,c, Olaf Sporns d,⁎
a

Black Dog Institute and School of Psychiatry, University of New South Wales, Sydney, Australia
Mental Health Research Division, Queensland Institute of Medical Research, Brisbane, Australia
c
CSIRO Information and Communication Technologies Centre, Sydney, Australia
d
Department of Psychological and Brain Sciences, Indiana University, Bloomington, IN 47405, USA
b

a r t i c l e

i n f o

Article history:
Received 1 August 2009
Revised 1 October 2009
Accepted 2 October 2009
Available online 9 October 2009

a b s t r a c t
Brain connectivity datasets comprise networks of brain regions connected by anatomical tracts or by
functional associations. Complex network analysis—a new multidisciplinary approach to the study of
complex systems—aims to characterize these brain networks with a small number of neurobiologically
meaningful and easily computable measures. In this article, we discuss construction of brain networks from
connectivity data and describe the most commonly used network measures of structural and functional
connectivity. We describe measures that variously detect functional integration and segregation, quantify
centrality of individual brain regions or pathways, characterize patterns of local anatomical circuitry, and test
resilience of networks to insult. We discuss the issues surrounding comparison of structural and functional
network connectivity, as well as comparison of networks across subjects. Finally, we describe a Matlab
toolbox (http://www.brain-connectivity-toolbox.net) accompanying this article and containing a collection
of complex network measures and large-scale neuroanatomical connectivity datasets.
© 2009 Elsevier Inc. All rights reserved.

Introduction
Modern brain mapping techniques—such as diffusion MRI,
functional MRI, EEG, and MEG—produce increasingly large datasets
of anatomical or functional connection patterns. Concurrent technological advancements are generating similarly large connection
datasets in biological, technological, social, and other scientific fields.
Attempts to characterize these datasets have, over the last decade, led
to the emergence of a new, multidisciplinary approach to the study of
complex systems (Strogatz, 2001; Newman, 2003; Boccaletti et al.,
2006). This approach, known as complex network analysis, describes
important properties of complex systems by quantifying topologies of
their respective network representations. Complex network analysis
has its origins in the mathematical study of networks, known as graph
theory. However, unlike classical graph theory, the analysis primarily
deals with real-life networks that are large and complex—neither
uniformly random nor ordered.
Brain connectivity datasets comprise networks of brain regions
connected by anatomical tracts or by functional associations. Brain
networks are invariably complex, share a number of common
features with networks from other biological and physical systems,
and may hence be characterized using complex network methods.
Network characterization of structural and functional connectivity
data is increasing (Bassett and Bullmore, 2006, 2009; Stam and

⁎ Corresponding author.
E-mail address: osporns@indiana.edu (O. Sporns).
1053-8119/$ – see front matter © 2009 Elsevier Inc. All rights reserved.
doi:10.1016/j.neuroimage.2009.10.003

Reijneveld, 2007; Bullmore and Sporns, 2009) and rests on several
important motivations. First, complex network analysis promises to
reliably (Deuker et al., 2009) quantify brain networks with a small
number of neurobiologically meaningful and easily computable
measures (Sporns and Zwi, 2004; Achard et al., 2006; Bassett et al.,
2006; He et al., 2007; Hagmann et al., 2008). Second, by explicitly
defining anatomical and functional connections on the same map of
brain regions, network analysis may be a useful setting for
exploring structural–functional connectivity relationships (Zhou et
al., 2006; Honey et al., 2007, 2009). Third, comparisons of
structural or functional network topologies between subject
populations appear to reveal presumed connectivity abnormalities
in neurological and psychiatric disorders (Stam et al., 2007, 2009;
Bassett et al., 2008; Leistedt et al., 2009; Ponten et al., 2009; Wang
et al., 2009b).
In this article, we provide a non-technical introduction to complex
network analysis of brain connectivity and outline important
conceptual issues associated with its use. We begin by discussing
the construction of structural and functional brain connectivity
networks. We then describe the most commonly used measures of
local and global connectivity, as well as their neurobiological
interpretations. We focus on recently developed network measures
(Boccaletti et al., 2006; Costa et al., 2007b) and provide a freely
available Matlab toolbox, containing these measures, as well as their
weighted and directed variants (Table A1). Finally, we discuss some of
the issues associated with comparing structural and functional
connectivity in the same subject and comparing connectivity patterns
between subjects.

such as distribution shape. Parcellation schemes that lump heterogeneously connected brain regions into single nodes may be less meaningful. particularly in functional or effective networks (Figs. 2003). 2007) and may be especially useful in filtering the influence of weak and potentially non-significant links.. small-scale brain networks could consist of neurons linked by synapses or of cortical columns linked by intercolumnar connections. Network measures are often represented in multiple ways. characterize patterns of local anatomical circuitry. Nodes should ideally represent brain regions with coherent patterns of extrinsic anatomical or functional connections. directed anatomical networks constructed from tract tracing studies (e. in general. network measures also have binary and weighted. a measure of directed information flow. The effective network represents patterns of causal interactions.. On the other hand. Fig. The nature of nodes The nature of nodes and links in individual brain networks is determined by combinations of brain mapping methods. The use of MEG and EEG sensors may be problematic in this regard. 2009). or coherence of anatomical tracts. The order of nodes in connectivity matrices has no effect on computation of network measures but is important for network visualization (Fig. 2 shows illustrative anatomical. Measures of brain networks An individual network measure may characterize one or several aspects of global and local brain connectivity. anatomical and effective connections may conceptually be represented with directed links. while links represent anatomical. while Table A1 contains mathematical definitions of all measures. O. be quantitatively compared. as collated from histological tract tracing studies. On the other hand. Depending on the measure. structural and functional networks may only be meaningfully compared. Many recent studies discard link weights. 2B. Weighted and directed variants of measures are typically generalizations of binary undirected variants and . quantify importance of individual brain regions. For instance.g. 2009). with appropriate redefinitions... 2A). as the nature of nodes and links largely determines the neurobiological interpretation of network topology (Butts. or effective connections (Friston. Anatomical connections typically correspond to white matter tracts between pairs of brain regions. functional. directed and undirected variants. functional connectivity may reflect linear or nonlinear interactions. Threshold values are often arbitrarily determined. Many combinations occur in various experimental settings (Horwitz. while matrix entries denote links. Networks constructed using distinct parcellation schemes may significantly differ in their properties (Wang et al. while weighted links also contain information about connection strengths (Figs. 2B. functional. In addition. The functional network represents patterns of cross-correlations between BOLD signals estimated from these dynamics. C—bottom panels). Fig.. current neuroimaging methods are unable to directly detect anatomical or causal directionality. This distribution is most commonly characterized by its mean. functional or effective) and measure-specific (e. a parcellation scheme should completely cover the surface of the cortex.. 3 illustrates some basic concepts underlying these measures.. Functional and effective connectivity networks were constructed from time series of brain dynamics simulated on this anatomical network. 2A) indicate the existence of a large proportion of reciprocal connections in the cortex. our discussion is equally applicable to smaller scale connectivity. These links tend to obscure the topology of strong and significant connections and as a result are often discarded. and test resilience of networks to insult. anatomical parcellation schemes. Links may also be differentiated by the presence or absence of directionality. weighted characterization usually focuses on somewhat different and complementary aspects of network organization (e. as well as interactions at different time scales (Zhou et al. 2B.g. Measurement values of all individual elements comprise a distribution. Effective connections represent direct or indirect causal influences of one region on another and may be estimated from observed perturbations (Friston et al.g. 2A). measures of individual network elements (such as nodes or links) typically quantify connectivity profiles associated with these elements and hence reflect the way in which these elements are embedded in the network. 2009a) and cannot. Nodes in large-scale brain networks usually represent brain regions. all self-connections or negative connections (such as functional anticorrelations) must currently be removed from the networks prior to analysis. as binary networks are in most cases simpler to characterize and have a more easily defined null model for statistical comparison (see below). The anatomical network represents large-scale connection pathways between cortical regions in the macaque. Thus. Thus. The choice of a given combination must be carefully motivated.1060 M. adapted from the study by Honey et al. Weak and non-significant links may represent spurious connections. Binary links denote the presence or absence of connections (Fig. as computed with transfer entropy. and individual nodes should not spatially overlap. or of the entire brain. Specifically. density. Directed patterns of effective connectivity may be inferred from changes in functional activity that follow localized perturbations. Unfortunately. if these networks share the same parcellation scheme (Honey et al. Functional connections correspond to magnitudes of temporal correlations in activity and may occur between pairs of anatomically unconnected regions. or a proportional weight threshold (Figs. 2009). (2007). 2007). and measures of connectivity. Future network methods may be able to quantify the role of negative weights in global network organization. although other features. we describe measures that variously detect aspects of functional integration and segregation. and effective connectivity networks. Construction of brain networks A network is a mathematical representation of a real-world complex system and is defined by a collection of nodes (vertices) and links (edges) between pairs of nodes (Fig. given that sensors may detect spatially overlapping signals and are generally not aligned with boundaries of coherent regions (Ioannides. Rubinov. In addition to these different representations. Sporns / NeuroImage 52 (2010) 1059–1069 Note that while we concentrate on the analysis of large-scale connectivity. time scale) features of connectivity.. and networks should ideally be characterized across a broad range of thresholds. Saramaki et al. 2003). which provides a more global description of the network. Weights in anatomical networks may represent the size. C). The nature of links In addition to the type of connectivity (anatomical. Rows and columns in these matrices denote nodes. In this section. links are also differentiated on the basis of their weight and directionality. while weights in functional and effective networks may represent respective magnitudes of correlational or causal interactions. 1). All networks are represented by their connectivity (adjacency) matrices. by applying an absolute. C—top panels). 1994). Fig. depending on the dataset. Independently. which provides some validity for the use of undirected anatomical networks. may be more important if the distribution is nonhomogeneous.

the significance of network statistics should be established by comparison with statistics calculated on null-hypothesis networks. while Figs.. known as the transitivity. Consequently. Rubinov. while the global efficiency is primarily influenced by short paths. Other algorithms detect overlapping modular network structure. Sporns / NeuroImage 52 (2010) 1059–1069 therefore reduce to the latter when computed on binary undirected networks. 2003). and hence acknowledge that single nodes may simultaneously belong in multiple modules (e.. One notable algorithm (Newman. 2005). or the total “wiring cost” of the network. Individual values of the degree therefore reflect importance of nodes in the network. functional.. sometimes termed the strength. This composition. 2009). on average. network measures based on functional paths are less straightforward to interpret. The degree to which the network may be subdivided into such clearly delineated and nonoverlapping groups is quantified by a single statistic. and the degree distribution. the modularity (Newman. while the presence of clusters in functional networks suggests an organization of statistical dependencies indicative of segregated neural processing.. Note. 2005). Locally. The mean clustering coefficient is normalized individually for each node (Table A1) and may therefore be disproportionately influenced by nodes with a low degree. The average inverse shortest path length is a related measure known as the global efficiency (Latora and Marchiori..g. Paths are sequences of distinct nodes and links and in anatomical networks represent potential routes of information flow between pairs of brain regions. such as the number of nodes and links. Watts and Strogatz. On the other hand. To illustrate the different representations and variants of a network measure. Unlike most other network measures. as previously determined by physiological recordings. Measures of integration characterize this concept by estimating the ease with which brain regions communicate and are commonly based on the concept of a path. Both of these algorithms have been generalized to detect modular structure in weighted (Newman.. with a high number of triangles implying segregation (Fig. Palla et al. Newman. the characteristic path length is primarily influenced by long paths (infinitely long paths are an illustrative extreme). The most commonly used null-hypothesis network has a random topology but shares the size. known as clusters or modules. The mean clustering coefficient for the network hence reflects. Measures of functional integration Functional integration in the brain is the ability to rapidly combine specialized information from distributed brain regions. but also find the exact size and composition of these individual groups. Paths in functional networks represent sequences of statistical associations and may not correspond to information flow on anatomical connections. as paths between disconnected nodes are defined to have infinite length. 2004a) and directed (Leicht and Newman. density and binary degree distribution of the original network (Maslov and Sneppen. Consequently. The average shortest path length between all pairs of nodes in the network is known as the characteristic path length of the network (e. 2006) is known to be quite accurate and is sufficiently fast for smaller networks.g. 2008) networks. Measures of segregation have straightforward interpretations in anatomical and functional networks. within the network. 2001). especially if the weight distribution is nonhomogeneous.. More sophisticated measures of segregation not only describe the presence of densely interconnected groups of regions. 2005) and directed (Fagiolo. 3).. 2008) performs much faster for larger networks and is also able to detect a hierarchy of modules (the presence of smaller modules inside larger modules). The presence of clusters in anatomical networks suggests the potential for functional segregation in these networks. is defined as the sum of all neighboring link weights. Optimization algorithms generally sacrifice some degree of accuracy for computational speed. functional connectivity data. 2B and C shows functional and effective networks reordered by this modular structure. 2007) networks. and correspondingly zero efficiency. Figure 2A shows the modular structure of the anatomical connectivity network of a large portion of the macaque cortex. which may significantly contribute to integration in larger and sparser networks (Estrada and Hatano. 1998). is revealed by subdividing the network into groups of nodes. known as the network's modular structure (community structure). The directed variant of the degree distinguishes the number of inward links from the number of outward links. 2007). as discussed below. the prevalence of clustered connectivity around individual nodes. Such measures may be easier to interpret when information about anatomical connections is available for the same subjects (e. O. The degrees of all nodes in the network comprise the degree distribution. 2004b). The mean network degree is most commonly used as a measure of density. is normalized collectively and consequently does not suffer from this problem (e. while the weighted variant of the degree. Lengths of paths consequently estimate the potential for functional integration between brain regions. with shorter paths implying stronger potential for integration. the global efficiency may be meaningfully computed on disconnected networks.. we consider a basic and important measure known as the degree. 2008). Measures of segregation primarily quantify the presence of such groups.g. Anatomical. Simple measures of segregation are based on the number of triangles in the network. 2B —top panel). Honey et al. Measures of functional segregation Functional segregation in the brain is the ability for specialized processing to occur within densely interconnected groups of brain regions.M. . already contain such information for all pairs of brain regions (Fig. Some authors have argued that this may make the global efficiency a superior measure of integration (Achard and Bullmore. Null-hypothesis networks have simple random or ordered topologies but preserve basic characteristics of the original network. These anatomical modules correspond to groups of specialized functional areas. and a minimally possible number of between-group 1061 links (Girvan and Newman. by its nature. 2002). which is an important marker of network development and resilience. rather than computed exactly (Danon et al. and effective modules in this network show extensive overlap. with a maximally possible number of withingroup links. Both the clustering coefficient and the transitivity have been generalized for weighted (Onnela et al. such as the visual and somatomotor regions. the fraction of triangles around an individual node is known as the clustering coefficient and is equivalent to the fraction of the node's neighbors that are also neighbors of each other (Watts and Strogatz. that this network may have a different weighted degree distribution. however. the optimal modular structure for a given network is typically estimated with optimization algorithms. 1998) and is the most commonly used measure of functional integration. The degree of an individual node is equal to the number of links connected to that node. 2002). Another more recently developed algorithm (Blondel et al. which in practice is also equal to the number of neighbors of the node. Note that measures such as the characteristic path length and the global efficiency do not incorporate multiple and longer paths. It is important to note that values of many network measures are greatly influenced by basic network characteristics. Unlike the characteristic path length. More generally. A classical variant of the clustering coefficient.g.

1994). networks are commonly represented by their connectivity matrices. 2006). The structural. with rows and columns representing nodes and matrix entries representing links. functional. and symmetrizing. Rubinov. In comparison. Structural networks are commonly extracted from histological (tract tracing) or neuroimaging (diffusion MRI) data. through thresholding. Small-world brain connectivity Anatomical brain connectivity is thought to simultaneously reconcile the opposing demands of functional integration and segregation (Tononi et al. O. binarizing. Functional networks are commonly extracted from neuroimaging (fMRI) or neurophysiological (EEG. 1.. most studies examining functional brain networks also report various degrees of small-world organization. It is commonly thought that such an organization reflects an . To simplify analysis. and effective connectivity networks in Fig. While a binary path length is equal to the number of links in the path. 2007) and share a high global efficiency. a weighted path length is equal to the total sum of individual link lengths. MEG) data. networks are often reduced to a sparse binary undirected form.1062 M. 3). as large weights typically represent strong associations and close proximity. Such a design is commonly termed small-world and indeed appears to be a ubiquitous organization of anatomical connectivity (Bassett and Bullmore. Sporns / NeuroImage 52 (2010) 1059–1069 Fig. Paths are easily generalized for directed and weighted networks (Fig. Furthermore. and consequently a lower global efficiency.. A well-designed anatomical network could therefore combine the presence of functionally specialized (segregated) modules with a robust number of intermodular (integrating) links. Link lengths are inversely related to link weights. Connection lengths are typically dimensionless and do not represent spatial or metric distance. For computational convenience. 2 differ in their values of the global efficiency. Structural and effective networks are similarly organized (Honey et al. Construction of brain networks from large scale anatomical and functional connectivity datasets. functional networks have weaker connections between modules.

while cool colors represent negative correlations. bottom) The same network. In comparison. and effective connectivity networks. functional. but it may also falsely report a smallworld topology in highly segregated. top) Effective connectivity network. (C. of the same number as in the structural network. Sporns / NeuroImage 52 (2010) 1059–1069 1063 Fig. and therefore weaker small-world attributes. and consequently have small-world topologies. For instance. Network motifs Global measures of integration and segregation belie a rich repertoire of underlying local patterns of connectivity. (C. The significance of a motif in the network is determined by its frequency of occurrence. 2. The frequency of occurrence of different motifs in the whole network correspondingly represents the characteristic motif profile of the network. but poorly integrated networks. 2008). Functional activations on a local anatomical circuit may at any given time utilize only a portion of that circuit. Warm colors represent positive correlations. small-world networks should be simultaneously highly segregated and integrated. Illustrative anatomical. 2 are simultaneously highly segregated and integrated. These patterns of local connectivity are known as (anatomical) motifs (Fig. This measure may be useful for snapshot characterization of an ensemble of networks. thresholded by removing negative and self-self correlations. as well as groups of somatosensory and somatomotor regions. Rubinov. and bidirectional loops. 2004). A more complete understanding of the relationship between brain dynamics and functional connectivity will help to clarify this issue. representing cross-correlation of the regional BOLD signal. constructed by computing inter-regional transfer entropy. the functional network is also highly segregated but has a lower global efficiency. as estimated from simulated neural mass model dynamics. 2006). usually normalized as the motif z-score by comparison with ensembles of random null-hypothesis networks (Milo et al. thresholded to preserve the strongest connections. (2007) for a further exploration of these networks. More generally. anatomical triangles may consist of feedforward loops. Area names and abbreviations are provided in the supporting information. optimal balance of functional integration and segregation (Sporns and Honey. bottom) The same network. See Honey et al. Such local patterns are particularly diverse in directed networks. feedback loops.. 3). (B. 2004). The frequency of occurrence of different motifs around an individual node is known as the motif fingerprint of that node and is likely to reflect the functional role of the corresponding brain region (Sporns and Kotter. form largely contiguous (color-coded) anatomical regions. (B. (A) Large-scale anatomical connection network of the macaque cortex. Recently. 2002).M. yet have approximately the same characteristic path length as random networks (Watts and Strogatz. defined as possible subsets of connection patterns embedded within anatomical motifs (Sporns and Kotter. O. with distinct frequencies of individual loops likely having specific functional implications. Anatomical and effective networks in Fig. top) Functional connectivity network. This network includes the ventral and dorsal streams of visual cortex. Consequently. Network modules representing these groups. 1998). the somewhat abstract nature of functional paths (see above) makes interpretation of the small-world property less straightforward in functional networks. from the simulated fast time-scale dynamics. a measure of small-worldness was proposed to capture this effect in a single statistic (Humphries and Gurney. Functional motif significance may be optimally characterized by weighted . While this is a plausible hypothesis. this measure should not in general be regarded as a substitute for individual assessments of integration and segregation. Small-world networks are formally defined as networks that are significantly more clustered than random networks. This observation motivated the introduction of functional motifs.

and play a key role in network resilience to insult. closeness centrality is defined as the inverse of the average shortest path length from one node to all other nodes in the network. Measures of centrality may have different interpretations in anatomical and functional networks. measures of motif occurrence. facilitate functional integration. However. Measures of centrality Important brain regions (hubs) often interact with many other regions. Weighted and directed variants of centrality measures are in most cases based on weighted and directed variants of degree and path length (Table A1). 2008). while measures of segregation are often based on triangle counts (blue) but also include more sophisticated decomposition into modules (ovals). or by the capacity of anatomical . and the motif intensity z-score (Onnela et al. and consequently act as important controls of information flow (Freeman. measures of integration are based on shortest path lengths (green). as discussed above. There is a source of possible confusion in motif terminology. These measures are typically based on basic properties of network connectivity (in bold type).. due to the paucity of local undirected connectivity patterns. of which two are shown—one motif containing dashed links. and consequently enable functional links between anatomically unconnected regions. Rubinov. Nodes with a high within-module degree but with a low participation coefficient (known as provincial hubs) are hence likely to play an important part in the facilitation of modular segregation. 1978). The notion of betweenness centrality is naturally extended to links and could therefore also be used to detect important anatomical or functional connections. 2005). The calculation of betweenness centrality has been made significantly more efficient with the recent development of faster algorithms (Brandes. Sporns / NeuroImage 52 (2010) 1059–1069 Fig. are likewise sometimes used to detect central brain regions. nodes with a high participation coefficient (known as connector hubs) are likely to facilitate global intermodular integration. anatomically central nodes often facilitate integration. 2001. 2005). Such links in turn make central nodes less prominent and so reduce the sensitivity of centrality measures in functional networks. within-module version of degree centrality (Table A1). The degree has a straightforward neurobiological interpretation: nodes with a high degree are interacting. The degree may be a sensitive measure of centrality in anatomical networks with nonhomogeneous degree distributions. Many measures of centrality are based on the idea that central nodes participate in many short paths within a network. Patterns of local connectivity are quantified by network motifs (yellow). There are many measures of centrality. 3. defined as the fraction of all shortest paths in the network that pass through a given node. the extent of functional deterioration is heavily determined by the affected anatomical region in a stroke. In addition. Bridging nodes that connect disparate parts of the network often have a high betweenness centrality (Fig. An illustration of key complex network measures (in italics) described in this article. An example three-node and four-link anatomical motif contains six possible functional motifs. Motif intensity takes into account weights of all motif-comprising links and may therefore be more sensitive in detecting consistently strong functional configurations.1064 M. Kintali. Measures of network topology. path-based measures of centrality in functional networks are subject to the same interpretational caveats as path-based measures of integration (see above). The degree. Measures of network resilience Anatomical brain connectivity influences the capacity of neuropathological lesions to affect functional brain activity. In modular anatomical networks. discussed in other sections. and one motif containing crossed links. degree-based measures of within-module and between-module connectivity may be useful for heuristically classifying nodes into distinct functional groups (Guimera and Amaral. On the other hand. is one of the most common measures of centrality. Measures of node centrality variously assess importance of individual nodes on the above criteria. structurally or functionally. motifs are generally not used in the analysis of undirected networks. The complementary participation coefficient assesses the diversity of intermodular interconnections of individual nodes. O. Hub nodes (black) often lie on a high number of shortest paths and consequently often have high betweenness centrality. Motifs (“structural” and “functional”) were initially considered only in the context of anatomical brain networks (Sporns and Kotter. 2004). We also note that motif and resilience measures. we describe the more commonly used measures. On the other hand. 3). For instance. A related and often more sensitive measure is betweenness centrality. Thus. and in this section. The within-module degree z-score is a localized. known as the motif intensity. with many other nodes in the network. For instance. motif measures may also be meaningfully applied to some effective connectivity networks. Measures of centrality may be based on node degree (red) or on the length and number of shortest paths between nodes. For instance.

as anatomical connectivity in such networks becomes increasingly sparse. Complex network analysis is able to characterize such network resilience properties directly and indirectly. Another useful measure of resilience is the assortativity coefficient (Newman. more highly resolved networks. we provide weighted and directed variants for all our measures—many of these variants are likewise not yet available in other software. 2007a). 2003. For instance. 2006) remain meaningful despite differences in density. 2008). A number of features distinguish our toolbox from most other software packages. Alstott et al. the global efficiency would be preferable to the characteristic path length as a measure of integration. Network comparison Conclusion Complex network analysis may be useful for exploring connectivity relationships in individual subjects or between subject groups. Related measures of assortativity computed on individual nodes include the average neighbor degree (Pastor-Satorras et al. we developed a freely available and open source Matlab toolbox (http://www. Most real-life networks do not have perfect power–law degree distributions. Leung and Chau. The increased emphasis on structure–function and between-subject comparisons in studies of brain networks will require the development of accurate statistical comparison tools. while functional connectivity remains comparatively dense.. Direct measures of network resilience generally test the network before and after a presumed insult. In individual subjects. or in the emergent simulated functional connectivity or dynamical activity (e. comparisons between anatomical and functional modular structure (e.R. To accompany this article. 1999). Functional networks are likely to be denser than anatomical networks.S. The significance of such comparisons will become more obvious with increased knowledge about causal relationships between brain regions. Alternatively. insults may be computationally simulated by random or targeted removal of nodes and links. On the other hand. O. The open source nature of our toolbox allows researchers to customize individual functions for their needs. Zhou et al. We hope that the brain mapping community will be able to benefit from and contribute to these tools.g.. while others have powerful network visualization capabilities. and O. Some of these packages are especially suitable for the analysis of large networks containing thousands of nodes. comparisons may detect abnormalities of network connectivity in various brain disorders (Bassett and Bullmore. as they will typically contain numerous connections between anatomically unconnected regions (Damoiseaux and Greicius. . 2006). One such feature is the degree distribution (Barabasi and Albert. For instance. The effects of such lesions on the network may then be quantified by characterizing changes in the resulting anatomical connectivity. Honey et al. The accompanying brain connectivity toolbox allows researchers to start exploring network properties of complex structural and functional datasets. Preliminary steps have already been taken in this direction (Costa et al. Finally. were supported by the J. 2002). M.g.. M. Sporns / NeuroImage 52 (2010) 1059–1069 connectivity to withstand degenerative change in Alzheimer's disease. as mediated by direct anatomical connections. was supported by CSIRO ICT Centre scholarship. The assortativity coefficient is a correlation coefficient between the degrees of all nodes on two opposite ends of a link. the toolbox contains datasets for large scale neuroanatomical networks of the mammalian cortex of several species. 2008) and Python (Hagberg et al. and Chris Honey for their contributions to the brain connectivity toolbox.. 2009).R..net). The nontrivial relationship between structural and functional brain connectivity and the consequent interpretational difficulties of some functional measures (see above) make between-subject comparisons of functional networks difficult. but highly vulnerable to disruptions of high-degree central nodes. 1065 to become more pronounced in larger. but are not yet widely available. complex networks with power-law degree distributions may be resilient to gradual random deterioration. the toolbox provides functions for network manipulation (such as thresholding) and includes algorithms for generating null-hypothesis networks of predetermined (random.g.. 2009). 2004. Across subject populations. ordered and other) topologies. Indirect measures of resilience quantify anatomical features that reflect network vulnerability to insult. such as Matlab (Gleich. Patric Hagmann.M.S. 2009). Differences in density between anatomical and functional networks make global comparisons between these networks less straightforward. When testing resilience in such a way. In addition to these features. 2005). Jonathan Power and Vassilis Tsiaras for suggesting valuable improvements to our toolbox functions.R.. In addition. similar to those in cellular and molecular biology (Sharan and Ideker. Acknowledgments We thank Rolf Kötter. Other factors that may affect comparisons of network topology include degree and weight distributions. Networks with a positive assortativity coefficient are therefore likely to have a comparatively resilient core of mutually interconnected high-degree hubs. is grateful to Michael Breakspear for his supervision and support during this project. many networks have degree distributions that locally behave like a power–law—the extent to which these distributions fit a power–law may hence be a useful marker of resilience (Achard et al. 2006). The development of a detailed anatomical map of the human brain is an important step in this direction (Sporns et al. patients with a progressive neurodegenerative disease may be imaged over a longitudinal period. These packages include command-line toolboxes in popular languages. and Alain Barrat for definitional clarifications. Individual central nodes that score lowly on these measures may therefore compromise global network function if disrupted. For instance. 2007). Notably. NWB-Team... McDonnell Foundation Brain NRG JSMF22002082. Brain connectivity analysis software Multiple network analysis software packages are freely available on the Web. We described a collection of measures that quantify local and global properties of complex brain networks.. On the other hand. These differences in density are likely Complex network analysis has emerged as an important tool for characterization of anatomical and functional brain connectivity. comparisons of structural and functional networks may provide insights into structural– functional connectivity relationships (e.brain-connectivity-toolbox. Weighted and directed variants of assortativity measures are based on the respective weighted and directed variants of the degree (Barrat et al. M.. 2007). and to incorporate functions into larger analysis protocols. Rubinov.. Aviad Rubinstein. Our toolbox includes many recently developed network measures.. Here we discuss some general issues associated with network comparison. 2006). 2001) and the recently introduced local assortativity coefficient (Piraveenan et al. which are discussed in this article. it is prudent to use measures that are suitable for the analysis of disconnected networks. as well as standalone graphical user interface software (Batagelj and Mrvar. 2008). networks with a negative assortativity coefficient are likely to have widely distributed and consequently vulnerable high-degree hubs.

Rubinov. P 2t w i Weighted transitivity⁎. we assume that weights are normalized. aij : jaN Shortest path length: a basis for measuring integration Shortest path length (distance).j∈N wij. (Weighted) meanof triangles around i. n−1 P P . Lw = 1n iaN Characteristic path length of the network (e. such that 0 ≤ wij ≤ 1 for all i and j. lw is the sum of all weights in the network. = ∑j∈Naij. aij = 0 otherwise (aii = 0 for all i).1066 M. CY = Transitivity of the network (e.g. w dw ij = ∑auv∈gi↔j f(wuv).haN. iaN n n−1 Directed characteristic path length. Note that dij = ∞ for all disconnected pairs i. Watts and Strogatz. (i. computed as lw = ∑i. auv agi X j where gi↔j is the shortest path (geodesic) between i and j. kout i (Directed) in-degree of i. and n is the number of nodes. 1X 1X = E = n iaN loc.haN 1=3 : Number of directed  triangles  around i. an inverse) auv .haN  aij + aji ðaih + ahi Þ ajh + ahj . and l is number of links. X ki = Links (i. C= C = n iaN i n iaN ki ðki − 1Þ Directed clustering coefficient (Fagiolo. Y Eloc = 1 2n P iaN P  −1  Y  −1 Y + dhj ðN i Þ ðaij + aji Þðaih + ahi Þ djh ðNi Þ  out  out  P : in in j. Measure Binary and undirected definitions Weighted and directed definitions Basic concepts and measures Basic concepts and N is the set of all nodes in the network. Global efficiency Global efficiency of the network (Latora and Marchiori. Degree: number of links connected to a node Degree of a node i. (i. 1X 1X E= E = n iaN i n iaN P jaN. Henceforth. between nodes i and j.j≠i aij aih djh ðNi Þ P ki ðki − 1Þ out ki  in + ki out ki tY iaN i in + ki − . 2007). We compute the number of links as l = ∑i. in directed networks aij does not necessarily equal aji. j≠i w wij wih ½djh ðN i Þ −1 ki ð ki − 1 Þ iaN P 1 − v2 Weighted local efficiency⁎. j. dij = X Shortest weighted path length between i and j. See Saramaki et al.i n iaN h − 1Þ aij aji P Directed transitivity⁎. j) is a link between nodes i and j. EY = 1n iaN . aij is the connection status between i and j: aij = 1 when link (i. where gi→j is the directed shortest path from i to j. where f is a map (e. (2007) for n other variants. Number of triangles: a basis for measuring segregation Measures of integration Characteristic path length Number of triangles around a node i.j≠i dij n iaN i n iaN . Newman.. notation L is the set of all links in the network. j ≠ i  −1 P jaN.   P j. L= Shortest directed path length from i to j. (Directed) out-degree of i. 2003). LY = 1X 1 XP L = jaN. Local efficiency Eloc h i−1 j. w dij .. O. P Weighted characteristic path length.  P tiY = 1 2 j. j) are ordered from i to j.j≠i ki + ki ki + ki − 1 − 2 a a jaN ij ji . j≠i . P tiY P iaN ðkout + kin Þðkout + kin − 1Þ − 2 i i i i where Ci is the clustering coefficient of node i (Ci = 0 for ki b 2). Directed links (i. w Eloc = P 1 2 : . Sporns / NeuroImage 52 (2010) 1059–1069 Appendix A.haN. as aij and as aji). TY = P ðki jaN a a jaN ij ji i: 1 = 3 : Directed local efficiency⁎.  P geometric tiw = 1X ti = a a a : 2 j. 2001). 2005). w from weight to length and gi↔j is the shortest weighted path between i and j. Ew = n jaN.haN ij ih jh 1 2 wij wih wjh j. and djh (Ni) is the length of the shortest path between j and h..g. n − 1 where Ei is the efficiency of node i. where Li is the average distance between node i and all other nodes.g.. j≠i dw ij n  − 1 iaN P Directed global efficiency. that contains only neighbors of i. 1998). P 2t w 1 i iaN ki ðki − 1Þ. P n − 1Y d P jaN. Transitivity Weighted clustering coefficient (Onnela et al. kw i = ∑j∈Nwij. Local efficiency of the network (Latora and Marchiori. 1998).j∈N aij (to avoid ambiguity with directed links we count each undirected link twice. −1 dY ij jaN. Consequently. T w = P iaN 1 n k iaN i P iaN 2ti T= P : k iaN i ðki − 1Þ iaN Note that transitivity is not defined for individual nodes. j ∈ N).i is the local efficiency of node i. j ≠ i ij 1 . 2001). Weighted degree of i. Cw = 1X 1X 2ti .j≠i −1 dij n−1 P 1 Weighted global efficiency. kin i = ∑j∈Naji. Measures of segregation Clustering coefficient Clustering coefficient of the network (Watts and Strogatz. where Eloc. j) exists (when i and j are neighbors). j) are associated with connection weights wij. d→ ij = ∑aij∈gi→j aij. Table A1 Mathematical definitions of complex network measures (see supplementary information for a self-contained version of this table).haN.

M. i = 1− Out-degree participation coefficient. iaN where h′ is any nh node motif.. zh = where 〈Jrand. deviation for the number of occurrences of h in an ensemble of random networks. lh link subnetwork.g. whereuthe sum is over all occurrences of h in the network. i.i . l′h ≥ lh link subnetwork. kout i 2 P kin ðmÞ 1 − maM i kin . An equivalent alternative formulation of the modularity  P kk i j (Newman.h〉 and σ σ J rand. σ kðmi Þ −1  Y −1 Directed closeness centrality. jaN h ≠j. 2002).h are the respective mean and standard deviation for the intensity of h in an are the respective mean and standard ensemble of random networks.h i σ Irand. Within-module degree z-score of node i (Guimera and Amaral.mj : !2 # X Qw = . Network motifs Anatomical and functional motifs Motif z-score Jh is the number of occurrences of motif h in all subsets of the network (subnetworks).i ðh VÞ = X Jh V. Intensity z-score of motif h (Onnela et al. Participation coefficient Participation coefficient of node i (Guimera and Amaral.mj : w − w w ij i. yin i =  P P  maM where M is the set of modules (see modularity). Q = X " euu − uaM Weighted and directed definitions Weighted modularity  (Newman.jaN aij − i l i δmi . w w Weighted within-module degree z-score. j ≠ i. zout out i in Within-module in-degree z-score. j≠i ij − 1 Y. ki (mi) is the within-module degree of i (the number of links between i and all other nodes in mi).h . Pintensity P I   iaN Ih V.h ≠ i. i σ k ðmi Þ out k ðmi Þ − kout ðmi Þ = i . 1978). 2004). kin m σ ð iÞ where mi is the module containing node i. 2006) is given by Q = 1l δmi . i. zin i = ki k ð mi Þ σ ðmi Þ − kin ðmi Þ .jaN aij − l where mi is the module containing node i.i (h′) is the nh node motif fingerprint for node i.mj = 1 if mi = mj. z-Score of motif h (Milo et al. 2005).h Ih = P  u Πði. Fnh ðhVÞ = X iaN Fnh . where ρhj is the number of shortest paths between h and j. 1978). yw i = 1− maM  kw ðmÞ 2 i . Note that motifs are directed by definition. w w P ki kj 1 δmi .i (h′) is the nh node motif intensity fingerprint for node i. d jaN. h will occur as an anatomical motif in an nh node. kw i  ðmÞ 2 kout i . zi = ki ðmi Þ − kðmi Þ . and kðmi Þ and σk(mi) are the respective mean and standard deviation of the within-module mi degree distribution. lh link. zIh = . and 0 otherwise. yout i In-degree participation coefficient. if at least one combination of lh links in the subnetwork matches links in h (Sporns and Kotter. Ih − hIrand. and ki (m) is the number of links between i and all nodes in module m. and Ih′. Measures of centrality Closeness centrality −1 Li Betweenness centrality  n−1 = P : d jaN. Freeman. F nh. Lw i Closeness centrality of node i (e. = Pn = Pn − 1 w. (continued on next page) . and ρhj (i) is the number of shortest paths between h and j that pass through i. bi = Within-module degree z-score  Weighted closeness centrality.. Rubinov. 2008).j≠i ij P 1 ðn − 1Þðn − 2Þ hj ðiÞ ρ h.i is the intensity of FnI h ðh VÞ = motif h′ around node i.i is the number of occurrences of motif h′ around node i. and Jh′. Fnh. euv vaM where the network is fully subdivided into a set of nonoverlapping modules M. directed connected pattern. zw = ki ðmi Þ w− k ðmi Þ. h is an nh node. iaN F nh . 2004b).h Jrand. Freeman. 2004).mj .g. 2005).h i (Weighted) intensity of h (Onnela et al. and δmi. yi = 1 − X k ðmÞ 2 i maM ki . Weighted participation coefficient. 2004). Within-module out-degree z-score. where 〈Irand.jaN l l Directed modularity Newman. O.i h V = I where h′ is any nh node motif. 2005).h〉 and σ I rand. Sporns / NeuroImage 52 (2010) 1059–1069 1067 Table A1 (continued) Measure Binary and undirected definitions Modularity Modularity of the network (Newman.. Motif fingerprint nh node motif fingerprint of the network (Sporns and Kotter. h will occur (possibly more than once) as a functional motif in an nh node.. provided that path lengths are computed on respective weighted or directed paths. if links in the subnetwork match links in h (Milo et al. h and L is the set of links in the uth occurrence of h. ρ hj . d jaN. jÞaL u wij h 1 lh . 2002). Li Betweenness centrality of node i (e. and euv is the proportion of all links that connect nodes in module u with nodes in module v.  (Leichtoutand in P k k Q Y = 1l i. j≠i ij Betweenness centrality is computed equivalently on weighted and directed networks. 2005).. Jh − h Jrand. nh node motif fingerprint of the network.

1. j1. Deuker.F.. 2009. 2005. A. G.. Average neighbor degree Average degree of neighbors of node i (Pastor-Satorras et al. Phys. E.. S. U. V.. 2002). small-world human brain functional network with highly connected association cortical hubs. R.. Small-world brain networks.... Lefebvre. Sociol. 163–177. 3747–3752.1068 M. A. BMC Syst. S. P : kw nn.. Centrality in social networks: conceptual clarification.. Opin.T. L = Lrand Directed network small-worldness. = Cumulative in-degree distribution. Stat.. 2001. A. 1999). Suckling. Greicius. Travieso. 512–523. Sci. Costa. j1). (Eds. Meyer-Lindenberg.S.. Bullmore. Barthelemy. Rodrigues. M. 2006. j2) ∉ L. Kaiser. j2) ∈ L. Salvador. Bassett..i = jaN aij kj ki Average weighted neighbor degree (modified from Barrat et al. Bullmore. jÞaL 2 ij Þ w ðki i2 Directed assortativity coefficient (Newman. h l−1 rw = l−1 P P w w ði. B. 215–239. Small-world networks often have S ≫ 1.i = w jaN wij kj : w ki Average directed neighbor degree⁎. Springer. Brandes. 2001).. j1). 2009. 2008. Measure `of network small-worldness. Rev. Efficiency and cost of economical brain functional networks. Hagmann. e17.d. 525–533 (Epub 2009 Jun 30). 77–103. Human brain networks in health and disease. 026107.A... 2007). Damoiseaux. Fagiolo. at doi:10. Nonlinear Soft Matter Phys.S. 1. e1000408. E. .. Biol.  out P in a + aji Þ ki + ki jaN ð ij Y  knn..neuroimage. E Stat. Diaz-Guilera. Rockstroh. k Vzk Cumulative degree distribution of the network (Barabasi and Albert. in this case.S.. h P 2002).-L. Graph Drawing Software. 25.S.L.  i2 P out in 1 kout + kin l−1 k kj − l − 1 j ði. D. 2009. D. Meyer-Lindenberg.jÞaL 12 ki + kj −1 P out 2 ki in + ki  . Bullmore. Blondel. out Cumulative out-degree distribution..jÞaL ki kj − l ði. 1999. Complex networks: Structure and dynamics.jÞaL 2  out 2 ki +   i h in 2 kj − l−1  P 1 ði. j2).. 76.. Breakspear.. Complex brain networks: graph theoretical analysis of structural and functional systems. V. Acad. Boas. Emergence of scaling in random networks. P10008. Boccaletti. 9239–9248. L.R. M. Bassett.. the weighted degree distribution is not preserved. Lambiotte. where C and Crand are the clustering coefficients. F. R. Reproducibility of graph metrics of human brain functional networks. D. A.F. Rubinov. jÞaL 2 ij i 1 − l −1 P w + kj 1w ði. 2004). Butts. Nathan. j1) ∈ L and (i1. lowfrequency. A resilient. P. J.1016/j. Latora. LY = C Y rand In both cases.D. pp. P09008. Neurosci. E. 2006.. Rev. |i1+j2| + |i2+j1| b |i1+j1| + |i2+j2| (Sporns and Kotter..2009. The links are then rewired such that (i1.. Predicting the connectivity of primate cortical networks from topological and spatial node properties. 3. L. 2002). Moreno. Weighted assortativity coefficient (modified from Leung and Chau. R. D. S. S. Adaptive reconfiguration of fractal small-world human brain functional networks. 2009.J. P ðkw ÞP = k Vzk  p k V . Science 286. E. 16. Science 325. M. Duke. A. A. Neuroscientist 12. Bullmore.10. Biol.. Mech. Costa.. 77. L. R. 2007a. N.D. w Þ w C = C rand w ... Sw = S= C = Crand .. NeuroImage 47. S. weights may be switched together with links. S. O.V.. E. Sporns / NeuroImage 52 (2010) 1059–1069 Table A1 (continued) Measure Measures of resilience Degree distribution Binary and undirected definitions Weighted and directed definitions P ðkÞ = X  P  w Cumulative weighted degree distribution. J. Bassett..jÞaL 12 k2i + k2j − l − 1 ði.. Lw = C rand Y Y C = C rand . 509–512. 2009. Proc. i2. (Maslov and Sneppen.. Verchinski... Freeman.. M. D.. B.. Soc. Pajek — analysis and visualization of large networks In: Jünger.. References Achard. Alstott. 28. W. Danon. C. Neurol. 2007. Rep. Bullmore. jÞaL 2 i rY = l−1 Other concepts Degree distribution preserving network randomization. Stat. Arenas. Duch. E.. Comparing community structure identification. M. Mattay. 63–72. Mech. Sci. 213. PLoS Comput. Sporns. The architecture of complex weighted networks. 2004). j2). J. Supplementary data Supplementary data associated with this article can be found. Bullmore. G. Mrvar. Communicability in complex networks. P. Hatano. Batagelj. Nonlinear Soft Matter Phys. P knn. E Stat. 424. A faster algorithm for betweenness centrality. J. V. Phys. such that links (i1. Phys.. Weinberger. U. (i2.i = Assortativity coefficient Assortativity coefficient of the network (Newman. L. Berlin. 175–308.. pðkVÞ. Rev. Mutzel.. S. small-world networks often have S ≫ 1. Barabasi. L. Achard.. J. jÞaL i ði.. P  ðkin  Þ =P k Vzkout p k V . Adv. PLoS Comput.. jÞaL  1 w ði.jÞaL 2 ki + kj r=   h  i2 : P P l − 1 ði. Guillaume. J. 340–347. and L and Lrand are the characteristic path lengths of the respective tested network and a random network.T. h   i2 −1 P 1 l ði.. J. Hierarchical organization of human cortical networks in health and schizophrenia. Clustering in complex directed networks. V. D.U. Vespignani. P h 1 ði. Proc... in the online version.... Network small-worldness (Humphries and Gurney. Barrat. Christensen. Revisiting the foundations of network analysis. O.. J. Sporns. j2 ∈ N at random. 22. Weighted network small-worldness. Estrada.S. Albert. Generalizations that have not been previously reported (to the best of our knowledge) are marked with an asterisk (⁎). In weighted networks. Bassett. J. 1978... E.. B. Neurosci. All binary and undirected measures are accompanied by their weighted and directed generalizations. 036111. 167–242.T. O.. Brain Struct Funct. 26. E. 2003. L..jÞaL 2 w + kj i2 ... but may be subsequently approximated on the topologically randomized graph with a heuristic weight reshuffling scheme. 1460–1468. Greater than the sum of its parts: A review of studies combining structural connectivity and resting-state functional connectivity. Hilgetag. Netw. (i2. Natl.. kout + kin i j Degree-distribution preserving randomization is implemented by iteratively choosing four distinct nodes i1. Modeling the impact of lesions in the human brain. T.. jÞaL 2 ij wij ki kj − l w 2 ðki Þ −1 w 2 + ðkj Þ P  h w w ðk ði. Phys. Pastor-Satorras..S. Nat.S. M. Natl. A. 2008. E. E.003.R. Bassett. 101. j1) ∉ L.. (i2.. “Latticization” (a lattice-like topology) results if an additional constraint is imposed. 2009. 5.. D. Achard.. (i2.. C. 2006. Biol. 10. 103. P. Bullmore. Whitcher.. Chavez. 2007b. Bullmore. 2008..C. Cammoun. Characterization of complex networks: a survey of measurements. Math. M. Appendix B. 2008. 2004. 56. The algorithm is equivalent for weighted and directed networks. A.D. k Vzk where p(k′) is the probability of a node having degree k′..A. J. Y. P. SY = i2 . 186–198. 414–416. 2005. 2006. Neurosci. 19518–19523. while links (i1.. Smith. P k in p k V . 2007.). 2008). Acad. Curr. The Brain Connectivity Toolbox contains Matlab functions to compute most measures in this table. A.. U. Fast unfolding of communities in large networks. Hwang...

C. Hum. K.C..E.D.A. W. Onnela. Fast algorithm for detecting community structure in networks.. E Stat. B. M. Cortex 17.. Rev. Rev. 2005. Daffertshofer... V.. M. PLoS Biol. 466–470. Pastor-Satorras.J. O. 161–170. Matlabbgl (version 4.W.P. T.1906v0802. and University of Michigan. 213–224. Kotter. Graph theoretical analysis of magnetoencephalographic functional connectivity in Alzheimer's disease. Proc. Ponten. B..C.. Proc... I. Kurths. Horwitz. A. Y.. J. Honey. G.. Newman. C. 2005...W. C.J. Curr. U. Predicting human resting-state functional connectivity from structural connectivity. Dynamic functional connectivity. 145–162. J. Chen. Zamora. I.. Verbunt. 1. Thiran.. A 378. 2001. Altered sleep brain functional connectivity in acutely depressed patients.. Nolte.M.. Hum. Nonlinear Biomed.. Phys. Epub 2009 Jun 8 . Brain Mapp.47(4):1590-1607. S. S. Sporns.. Jones. Maslov.. S.. 2003. L. PLoS Biol.. G. G. 2005. Rev.. L. Sharan..K.. 2002. Tononi. Matlab toolbox for functional connectivity.. D.... Hierarchical organization unveiled by functional connectivity in complex brain networks.D. 30.. Chklovskii. 45. M. H. Strogatz.F. Lanquart. I. Sporns.. Proc.J.. Analysis of weighted networks.. Modeling cellular machinery through biological network comparison.. Stam.H. Amaral. Sporns. L.J.. In: Varoquaux. 2009. K. Millman. Science 298. J. K.. Evans.P.. M. Neuroimage. C. 198701.. Stam.. Exploring network structure. van Dijk.... Wang.. 2005. A. U. Natl. Science 296. Lett.J.. D. L.J. P02001. He. 2007. Network workbench tool.. A.. R. Local assortativeness in scale-free networks.. Natl... Watts. Uncovering the overlapping community structure of complex networks in nature and society... Reijneveld. Efficient behavior of small-world networks. Stam. Piraveenan. Reijneveld.. Kertesz. 104.N. Cortex 17. 2004. J. Proc. Hagberg. Derenyi.. Newman.J. Harrison. Berendse. S. Vespignani. Nature 410. A. 167–256. P.. U. 2007.. J. H. Zhang.F.. Wang. Rev. E Stat. 84. The small world of the cerebral cortex. V.J. 2009. Natl... Milo. 70.. Acad. J. M. Leistedt. S. Vaught. 591–602. Vazquez.. Community structure in social and biological networks. Kotter.A.J.. Leicht. 2006. Neuroinformatics 2. 2006. Siegle.. D. S. Newman. Mech. 100. Sporns. Hum. dynamics. Sporns. Breakspear. A. e42. C.. S. J. Chen. Dynamical and correlation properties of the internet. Gigandet. Small-world anatomical networks in the human brain revealed by cortical thickness from MRI. Z.. R. C. (Eds. G. 106. Girvan. Chau. Stam. 2008. M. J. Gurney. X. The human connectome: a structural description of the human brain. P..Y. Rev.C. Penny.. Specificity and stability in topology of protein networks. P. Natl. A.. 2. O. 1511–1523.. M. U. Graph theoretical analysis of complex networks in the brain. Cammoun. M.M..J. Generalizations of the clustering coefficient to weighted complex networks. Cereb. L. 7821–7826.E. 17. U. Newman. Kotter.C. 065103. 2009.. 1994. H. 30.. pp... G. H.. 2005. R. K... A. 0809. T. 2004a. Northeastern University. Phys. Nonlinear Soft Matter Phys. SIAM Rev. J. Bartolomei.. Acad. 2207–2219. Thompson. J.. Kaski. van Cappellen van Walsum. de Munck. 10240–10245. C. 217. Neurol. Newman.P.. C. 103. 197–204.. Lett. T. H. Honey. Altered small-world brain functional networks in children with attention-deficit/hyperactivity disorder. Network structure of cerebral cortex shapes functional connectivity on multiple time scales. Friston. C. Vicsek. Q. O. 1273–1302. Exploring complex networks. Humphries. Exp..H. J. Pasadena. J. 19219–19220.. Saramaki. Nature 435. Small worlds inside big brains...J. W. Phys. The elusive concept of brain connectivity. Betweenness centrality: algorithms and lower bounds. 268–276. Natl. B. C. 638–649. Tang. Proceedings of the 7th Python in Science Conference (SciPy2008). NeuroImage 19. 2002. Newman. 027105.J. Wang.. L... N.. Kashtan. Modularity and community structure in networks. 258701. M. Marchiori. O.E. Phys. A. Brain Mapp. S. Zemanova. Zhu. S. Leung.. Coumans. Cammoun. 8577–8582. Farkas. Zomaya.. K. Sporns / NeuroImage 52 (2010) 1059–1069 Friston. Opin.A. 2006. Gleich. Sporns. 814–818. e369. Acad.. R. Palla.P. Rev.. Scheltens..J. 75. Prokopenko. W. A.J. M.. D. A. 1. E. M... 2009a. 2007.. Biol. 1998.J.C. Mapping the structural core of human cerebral cortex.. Y. M. Zhu. 103. O.J. M. G. Y. 2007. L. F. Parcellation-dependent small-world brain functional networks: a resting-state fMRI study. C. The structure and function of complex networks. Y. Z. NeuroImage 19.. Saramaki... S. N. 5033–5037.. J. Nonlinear Soft Matter Phys.. C. Kaski.J. Y.. Newman.. Meuli..... 2009 Oct 1. PLoS Comput. G. A measure for brain complexity: relating functional segregation and integration in the nervous system. 97. Small-world networks and functional connectivity in Alzheimer's disease. Brain 132. Zwi. Kintali.. Zhou. Q. 2008. S. R. 2407–2419..A.C.. Indiana university. Rubinov. Q.. G. Wang. K. 056131. 440–442. 2008.. Cartography of complex networks: modules and universal roles. 91. and function using networkx. Schult. Network motifs: simple building blocks of complex networks. Collective dynamics of ‘small-world’ networks. Nonlinear Soft Matter Phys. Acad. Phys. Cao. Biotechnol. M.. arXiv. U. 1069 NWB-Team. Acad. P.. Nat. Ioannides. 2006.. R.. Functional and effective connectivity in neuroimaging: a synthesis. Sporns. Douw. Breakspear. L. Motifs in brain networks.. 2008. 2001..J. E 69. Phys. R. 89. Onnela. e0002051. Yang.M. Network ‘small-world-ness’: a quantitative method for determining canonical network equivalence. Shen-Orr. Nature 393. Zang. 2002. Weighted assortative and disassortative networks model. O. Sci.. Latora. Assortative mixing in networks. 11–15. e159. C. Alon. Zhou. Sneppen. 2004. 2003. 2008. D. Edelman. 2008. 910–913. Hagmann.). P. 824–827. Y.. A.. 2001.. A. Phys. 1994.. 2004b. P. Sporns.. CA USA.E. 2009.. J. R. 2087011–2087014. Natl..F.. 2006. 2007. S. Lett. Indications for network regularization during absence seizures: weighted and unweighted graph theoretical analyses. Honey. Stam. M. M.E. Hum... Meuli. Rev. 427–433. Brain Mapp. Lett.E.. 2008. 2. Guimera. 2009. Manshanden. C.. 238103. 2002. Wedeen. Community structure in directed networks. Neurobiol. Sci. Rev. 2007. Acad. J. 6. O. Sci. A. 2003.. Rev. B. Lett. U.. 118703. 30.J.E. Kivela. 2009b. Linkowski. 56–78... 87. R. Honey. S. Proc.. S... Swart. Physica.. Phys. 28002. O. Lett. T. Gigandet. Dynamic causal modelling. He... 87. Dumont. 2007. J. J.. Ideker. Jones.C. Cereb.J. Proc. de Haan. 066133.J. E Stat. 71. Hagmann. Tononi.. Zang. 3. Kertesz. 2035–2040. Phys. Stat. Sci. Europhys. Montez. Sci. Itzkovitz. A... He.0). PLoS ONE 3..J. 99. Intensity and coherence of motifs in weighted complex networks. 24. Scheltens.A. Zhong.P. Sci. Brain Mapp... Strogatz.. O. Phys. Gong.. X. Hilgetag.. 92–99.