CHAOS 21, 016109 (2011)

Analyses of antigen dependency networks unveil immune system reorganization between birth and adulthood
Asaf Madi,1,2 Dror Y. Kenett,1 Sharron Bransburg-Zabary,1,2 Yifat Merbl,3 Francisco J. Quintana,3 Stefano Boccaletti,1,4 Alfred I. Tauber,5 Irun R. Cohen,3,a) and Eshel Ben-Jacob1,6,a)
School of Physics and Astronomy, Tel Aviv University, 69978 Tel Aviv, Israel Faculty of Medicine, Tel Aviv University, 69978 Tel Aviv, Israel 3 Department of Immunology, Weizmann Institute of Science, Rehovot 76100, Israel 4 CNR, Institute of Complex Systems, Florence, Italy and the Italian Embassy in Israel, Israel 5 School of Medicine, Boston University, Boston, Massachusetts 02215, USA 6 The Center for Theoretical and Biological Physics, University of California San Diego, La Jolla, California 92093, USA
2 1

(Received 28 October 2010; accepted 29 December 2010; published online 29 March 2011) Much effort has been devoted to assess the importance of nodes in complex biological networks (such as gene transcriptional regulatory networks, protein interaction networks, and neural networks). Examples of commonly used measures of node importance include node degree, node centrality, and node vulnerability score (the effect of the node deletion on the network efciency). Here, we present a new approach to compute and investigate the mutual dependencies between network nodes from the matrices of node-node correlations. To this end, we rst dene the dependency of node i on node j (or the inuence of node j on node i), D(i, j) as the average over all nodes k of the difference between the i k correlation and the partial correlations between these nodes with respect to node j. Note that the dependencies, D(i, j) dene a directed weighted matrix, since, in general, D(i, j) differs from D( j, i). For this reason, many of the commonly used measures of node importance, such as node centrality, cannot be used. Hence, to assess the node importance of the dependency networks, we dene the system level inuence (SLI) of antigen j, SLI( j) as the sum of the inuence of j on all other antigens i. Next, we dene the system level inuence or the inuence score of antigen j, SLI( j) as the sum of D(i, j) over all nodes i. We introduce the new approach and demonstrate that it can unveil important biological information in the context of the immune system. More specically, we investigated antigen dependency networks computed from antigen microarray data of autoantibody reactivity of IgM and IgG isotypes present in the sera of ten mothers and their newborns. We found that the analysis was able to unveil that there is only a subset of antigens that have high inuence scores (SLI) common both to the mothers and newborns. Networks comparison in terms of modularity (using the Newmans algorithm) and of topology (measured by the divergence rate) revealed that, at birth, the IgG networks exhibit a more profound global reorganization while the IgM networks exhibit a more profound local reorganization. During immune system development, the modularity of the IgG network increases and becomes comparable to that of the IgM networks at adulthood. We also found the existence of several conserved IgG and IgM network motifs between the maternal and newborns networks, which might retain network information as our immune system develops. If correct, these ndings provide a convincing demonstration of the effectiveness of the new approach to unveil most signicant biological information. Whereas we have introduced the new approach within the context of the immune system, it is expected to be effective in the studies of other complex biological social, C nancial, and manmade networks. V 2011 American Institute of Physics. [doi:10.1063/1.3543800] The immune system is a dynamic network whose complexity is comparable to that of the central nervous system. Being an adaptiveresponsive complex system, it stores latent information about body conditions. This information can in principle be deciphered, provided proper analyses of the immune network state are used. Here, we introduce a new approach to investigate the immune state based on the construction of network of
a)

Author to whom correspondence should be addressed. Electronic mail: irun.cohen@weizmann.ac.il and eshelbj@gmail.com.

mutual dependencies between antigens. These antigen dependency networks are computed from antigen microarray data of the reactivity of hundreds of autoantibodies. We used this approach to investigate the sera of ten mothernewborn pairs. Inspection of the local topological organization revealed a higher topological similarity between the IgG networks at birth (newborns) and adulthood (mothers). Analyses of the global network organization revealed that: 1. At birth, the IgM modularity of the newborns is higher than that of their IgG network. 2. During the development of the immune system, the
C V 2011 American Institute of Physics

1054-1500/2011/21(1)/016109/11/$30.00

21, 016109-1

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-2

Madi et al.

Chaos 21, 016109 (2011)

modularity of the IgG network increases and becomes comparable to that of the IgM networks at adulthood. Dening a new measure, which we termed the antigen inuence score, we found that the lists of the 20 most inuential antigens include three antigens common to the newborns and mothers both for the IgG and IgM networks. The analysis further unveils the existence of several conserved network motifs (between the newborns and maternal), which might retain important immune information as the immune system develops. If correct, these results provide a convincing demonstration that the new approach can unveil important biological information within the context of the immune system. We expect that it will be as effective in exposing hidden information in a wide range of other complex biological networks (e.g., gene networks, protein interaction networks, and neural networks), as well as social, nancial and manmade networks.

I. INTRODUCTION

Antibody networks have been studied in the past based on the connectivity between idiotypes and anti-idiotypes antibodies that bind one another. Here, we call attention to a different network of antibodies, antibodies connected by their reactivities to sets of antigensthe antigenreactivity network. The recent development of antigen microarray chip technology for detecting global patterns of antibody reactivities makes it possible to study the immune system quantitatively using network theory methods. In this immune network, the nodes represent the antigens spotted on the chip, and the links between the nodes represent the relationships between the antibody reactivities calculated for each group of subjects. In other words, an immune network for a given group of subjects corresponds to the network of similarities between antigen reactivities within that group of subjects. In an earlier paper,1 we demonstrated the use of the widespread Pearson correlation coefcient as the similarity measure to differentiate between maternal and newborns datasets. More specically, we evaluated the antigenantigen correlations by analysing antigen microarray data of the reactivity of IgM and IgG autoantibodies present in the sera of pairs of mothers and their newborns. We reported that the IgG repertoires of each mother and her newborn were very closely related and distinct for each mothernewborn pair.1 The IgM repertoires, in contrast, differed markedly between mothers and newborns; each mother manifested a different pattern of IgM reactivities that was distinct from her newborns cord IgM repertoire. However, the IgM reactivities of each of the newborns manifested very similar antigen-binding proles indicating that in utero each developing fetus produced autoantibodies to a similar set of self-molecules. A subsequent analysis of the data revealed that the reactivity proles to certain self-molecules were highly correlated as sets of functional antigen-reactivity cliques.1 More recently, we have extended the aforementioned study by applying graph and network theory analysis methods24; specically, we evaluated the minimum spanning tree (MST) for the networks of antigen correlations com-

puted from the correlation matrices.16 We thus made it possible to identify communities and their relations according to the topology of the network. While investigations of the antigenantigen correlations from microarray data have proven to be an efcient bioinformatics approach to decipher important features related to functional relations between antigens, it does not provide information about the causal relations between antigens. Detailed assessments of the properties that can be inferred from studying correlations have been well investigated in the past.79 Recently, Kenett et al.10 introduced a new method to study relationships of inuence, or dependency, by using partial correlations to construct a new type of networks. In their study, they applied this approach to the analysis of stock relationships and were able to uncover important information regarding the underlying dependency relationships between stocks traded in the New York Market. Here, we adopted and generalized the aforementioned to develop a system level analysis of antigen dependency networks as a step toward inference of causal relations between antigens. The analysis is based on partial correlations, which are becoming ever more widely used to investigate complex systems. Examples range from studies of biological systems, such as gene networks,11,12 to nancial systems in inspections of the market index effect on stock correlations.13 In simple words, the partial (or residual) correlation is a measure of the effect (or contribution) of a given antigen, say j, on the correlations between another pair of antigens, say i and k. To be more specic, the partial correlations of the (i, k) pair, given j is the correlations between them after proper subtraction of the correlations between i and j and between k and j. Dened this way, the difference between the correlations and the partial correlations provides a measure of the inuence of antigen j on the (i, k) correlation. Therefore, we dene the inuence of antigen j on antigen i, or the dependency of antigen i on antigen j - D(i, j), to be the sum of the inuence of antigen j on the correlations of antigen i with all other antigens. Next, we dene the system level inuence (SLI) of antigen j, SLI( j), as the sum of the inuence D(i, j) of j on all other antigens i. To demonstrate that the antigen dependency network analysis can unveil important biological information, we used the new approach to reanalyze autoantibodies of the IgM and IgG isotypes present in the sera of mothers and their newborns. To this end, we constructed the antigen dependency networks for the groups of mothers and newborns. We rst constructed the IgG, and IgM and combined IgG and IgM correlation networks and networks of antigen dependencies for the two groups. Next, the networks of the two subject groups were compared by employing two measures that were developed in the context of network theory and are widely usedthe divergence rate14 and modularity score.15 The rst method was used to test for signicant differences between the topological organizations of the two networks by quantication of their differences. The second method was used to assess the differences in the modular organization between the two networksthe ability to partition the network into modules such that the number of edges between modules is signicantly less than expected by chance. We found a higher modularity for the maternal networks,

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-3

Antigen dependency networks

Chaos 21, 016109 (2011)

specically for the IgG network. In terms of topological similarities, we detected that the IgG networks of the two groups have higher similarities than those of the IgM networks. Our results might indicate that the immune network development from birth to adulthood is accompanied with an increase of the networks modularity and with a slight increase in the similarities between the isotypes networks. Furthermore, we proceeded to identify, analyze, and compare the networks of the two subject groups in terms of the most inuential antigens. We found that the most inuential antigens in the mothers and newborns are composed of both isotypes IgG and IgM, which points to the additional role of the transferred IgG in inuence over the network. We also found a few subnetworks of highly connected antigen that are conserved between the two groups in both IgG and IgM isotypes. The above ndings illustrated that the analysis of antigen dependencies networks enable to unveil possible inuential antibodies in maturation of the immune system.
II. METHODS A. The antigenantigen correlations

di; kjj  Ci; k PCi; kjj:

(3)

This avoids the trivial case were antigen j appears to strongly effect the correlation C(i, k), mainly because C(i, j), C(i, k), and C( j, k) have small values. We note that this quantity can be viewed either as the correlation dependency of C(i, k) on antigen j (the term used here) or as the correlation inuence of antigen j on the correlation C(i, k).10
D. Antigen dependencies

Next, we dene the total inuence of antigen j on antigen i or the dependency D(i, j) of antigen i on antigen j to be: Di; j
N 1 1 X di; kjj: N 1 k6j

(4)

First, we computed the antigenantigen correlations from the antibody reactivities data obtained by the antigen microarray technology as was done in the past.1 The correlations between the antigen-reactivity proles (the reactivities of the antigen in all subjects) were calculated by Pearson formula:16
 Xi n li Xj n lj n Ci; j : (1) ri rj where Xi(n) and Xj(n) are the reactivity of antigens i and j of subject n and ri and rj are the standard deviation of the reactivity proles of antigens i and j. Note that the antigenantigen correlations (or for simplicity the antigen correlations) for all pairs of antigen dene a symmetric correlation matrix whose (i, j) element is the correlation between antigens i and j.
B. Partial correlations

As dened, D(i, j) is a measure of the average inuence of antigen j on the correlations C(i, k), over all antigens k not equal to j. The antigen dependencies dene a dependency matrix D whose (i, j) element is the dependency of antigen i on antigen j. It is important to note that while the correlation matrix C is a symmetric matrix, the dependency matrix D is nonsymmetricalD(i, j) = D( j, i)since the inuence of antigen j on antigen i is not equal to the inuence of antigen i on antigen j. For this reason, some of the methods used in the analyses of the correlation matrix [e.g., the principle component analysis (PCA)] have to be replaced or are less efcient. However, other methods, similar to ones presented here, can also account for the nonsymmetric nature of the dependency matrix.
E. The antigen SLI

Next we sorted the antigens according to the system level inuence of each antigen on the correlations between all other antigen pairs. The system level inuence of antigen j, SLI( j), is simply dened as the sum of the inuence of j on all other antigens i not equal to j, that is: SLIj
N 1 X i6j

Next we use the resulting antigen correlations to compute the partial correlations. The rst order partial correlation coefcient is a statistical measure indicating how a third variable affects the correlation between two other variables.13 Since here the variables are the antigen reactivities, we will proceed with the presentation for this case. The partial correlation between antigens i and k with respect to a third antigen jPC(i, k j j)17 is dened as: Ci; k Ci; jCk; j PCi; kjj p : 1 C2 i; j1 C2 k; j (2)

Di; j:

(5)

For completeness of the presentation we note that in a recent publication in which we studied the symmetric antigenantigen correlations, we employed the widely used eigenvalue centrality6,18 as a measure of the antigen importance. Here, we employed the system level inuence, which can better account for the nonsymmetric nature of the antigen dependencies, as the measure of the antigen importance.
F. Network representation of the antigen dependencies

where C(i, j), C(i, k), and C( j, k) are the antigen correlations dened above.
C. The correlation influence and correlation dependency

The relative effect of the correlations C(i, j) and C( j, k) of antigen j on the correlation C(i, k),10 is given by:

Similar to the correlation matrix, the dependency matrix can also be presented as a weighted matrix whose nodes are the antigens and the edges are the antigen dependencies. However, the nonsymmetric nature of the dependency matrix is reected by the fact that the antigen dependency network is a directed graph. The latter means that the edges direction

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-4

Madi et al.

Chaos 21, 016109 (2011)

determines their value, i.e., value of the edge directed from node i to node j can differ from the value of the edge directed from node j to node i.
G. Informative subgraphs of the dependency network

H. Construction of the PMFG informative subgraph

The complete dependency network for N antigens contains N(N 1) edges. Since most of the edges have small values (weak dependencies), the relevant information about the network (e.g., topology, organization, and the most inuential antigens) can be obscured. Several methods have been developed to overcome this obstacle by constructing from the complete network a subgraph that captures the most relevant information embedded in the original network. A widely used method to construct informative subgraph of a complete network is the MST.1925 Another informative subgraph that retains more information (in comparison to the MST) is the planar maximally ltered graph (PMFG),26 which is used here. Both methods are based on hierarchical clustering, and the resulting subgraphs include all the N nodes in the network whose edges represent the most relevant antigen dependencies. The MST subgraph contains (N 1) edges with no loops while the PMFG subgraph contains 3(N 2) edges. In a recent publication,5 we have constructed the MST subgraph to study the symmetric antigenantigen correlations. Here, for a better account of the nonsymmetric nature of the antigen dependencies, we employed the PMFG subgraph.

To construct the PMFG), we rst order the N(N 1) values of the dependencies matrix D in decreasing rank. We then start from the pairs of nodes, say i and j, with the highest dependency and draw a directed link j ! i between them. The process continues according to the rank order, while at each iteration a directed link is added if and only if the resulting graph (network) is still planar, i.e., it can be drawn on the surface of a sphere without link crossing.26 In the resulted directed subgraph, referred to as {G}, the original values of the dependencies are not retained (i.e., all the directed links have a weight 1). We also note that the subgraph {G} contains (for N ) 1), 3(N 2) edgesthe maximum number of directed edges for planar graph.26
I. Hybrid presentation of the informative subgraphs

We developed a hybrid presentation in which the information about the antigens (nodes) system level inuence namely, the inuence score, is superimposed on the informative subgraphs by coloring each node j according to its SLI( j) value. It is important to emphasize that since the SLIs were calculated on the original dependency network, the SLIs contain additional information that might have been lost in the reduction of the complete network to a subgraph. Hence, the hybrid representation can unveil additional features beyond the features that can be obtained by each analysis on its own. In Fig. 1, we show examples of the informative subgraphs for

FIG. 1. (Color) The hybrid presentation of the informative subgraphs of the correlationvs dependency network. (a) Maternal dependency network, and (b) maternal correlations. (c) and (d), the same as (a) and (b) for the newborns. To simplify the presentation, we show the subgraphs for a selected subset of only 20 antigens that are the top separators between the two groups of subjects using t-test ranking between the two datasets. Arrows indicate the directionality of the inuence. The antigen (nodes) inuence scores (SLI) is color coded from dark blue for the least inuential antigens to dark red for the most inuential ones.

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-5

Antigen dependency networks

Chaos 21, 016109 (2011)

the maternal and newborns datasets and a comparison to correlation based PMFG networks. In Appendix A, we show a comparison between the informative subgraphs in which the antigen inuence score (SLI) was calculated for the entire network and the case in which they were calculated within the informative subgraphs. Note that in the latter case, since the directed links are not weighted, the measure of the system level inuence of each antigen j is simply the total number of directed edges from j to other antigens. We emphasize that all the analyses described below were performed on the informative subgraphs {G}.
J. Network comparison based on divergence rate

Note that in the informative subgraphs studied here, each directed link from node i to node j corresponds to a topological distance 1 from i to j, and we take the neighborhoods nodes {i}NN to be the nodes that have a topological distance 1 with node i. The topological distance between two nodes that are not directly connected by an edge is the number of directed edges of the shortest path connecting the two nodes.
K. Network comparison based on network modularity

We have used two measures, the divergence rate presented here and the network modularity presented next, for quantitative comparison between the dependency networks of the mothers and newborns. We emphasize that both the divergence rate and network modularity comparisons were performed on the corresponding informative subgraphs and not on the complete networks. The comparisons were performed on the informative subgraphs corresponding to the dependency networks of both the IgG and the IgM isotypes of the maternal and newborns datasets. The divergence rate measure developed by Lee and Kim,27 is based on the idea of quantication of the information difference between two process (variables) based on the notion of conditional entropy. In information theory, the specic conditional entropy h(X j Y y) is the entropy of a process (variable), under the condition that another process Y is assigned the value y. The conditional entropy H(X j Y) is then the average of h(X j Y y) over all possible y that Y can take. It can be shown that H(X j Y) H(X j Y) H(X), where H(X j Y) is the combined entropy of processes X and Y and H(X) is the entropy of process X. The conditional entropy28 has been used to dene the metric distance or information distance ID(X, Y) between two processes X and Y as ID(X, Y) : H(X j Y) H(Y j X). Motivated by this notion, Lee and Kim27 dene the notion of the metric distance (MD) MD(GX, GY) between two graphs {GX} and {GY} to be: MDGX; GY  CDivGX j GY CDivGY j GX (6)

Modular organization is characteristic of many biological and social networks alike and it is a hallmark of systems that perform multiple-parallel tasks.18,2932 Modular organization means that the network is composed of subgroups of nodes that are strongly connected (e.g., strong correlations in the case of correlation networks or strong dependencies in the case studied here), with sparser or weaker connections between the modules. The ability to detect such groups could be of signicant practical importance. For instance, groups within the correlation of gene expression network might correspond to sets of genes with related functions.31 Several methods have been developed to identify and quantify network modularity. Here we employ Newmans partitioning algorithm15 as a quantitative measure of the network modularity score. Then we compare the dependency networks of the mothers and newborns based on their modularity score. The partitioning algorithm is based on computing a modularity matrix [M] from the adjacency matrix [A] associated with a graph {G}. The idea is to subtract from the adjacency matrix a shufed adjacency matrix. The latter corresponds to a shufed graph (network) in which the edges of the original network are distributed randomly between the nodes. More specically, since the element A(i, j) of the adjacency matrix is the number of edges connecting nodes i and j (note that it can be 0, 1, or 2 in the case studied here), the element M(i, j) of the modularity network is dened to be: Mi; j  Ai; j kiKj 2NLinks (8)

where CDiv(GX j GY) and CDiv(GY j GX) can be viewed as conditional divergences and are calculated as follows: First we dene DGX(i) to be the sum of the topological distances from a node i to all its neighborhoods nodes {i}NN. Then we dene the conditional distances CDiv(GX j GY)(i) to be the sum of the topological distances in graph {GY} from node i to the group of nodes {i}NN dened in graph {GX}. Note that these nodes, which are in the neighborhood of i in graph {GX}, need not be in the neighborhood of i in the graph {GY}. We also note that DGY(i) and CDiv(GY j GX)(i) are dened in a similar way. With these denition at hand, CDiv(GX j GY) is dened to be: CDivGXjGY   N CDivGXjGYi 1X  log10    N i1 DGXi (7)

where k(i) and k( j) are the degrees of nodes i and j, and NLinks are the total number of edges in the original graph. Once the modularity matrix is computed, the next step is to nd the leading (most positive) eigenvalue and its corresponding eigenvector, and the graph is portioned to two groups, one contains the positive elements and the other contains the negative ones. Then, generally speaking, the process continues while at each iteration a generalized modularity matrix (after proper subtraction from of the separated groups from the previous modularity network). The process halts when the generalized modularity matrix has no positive eigenvalues. To calculate the modularity score, we used the generalized modularity matrix and averaged the sum of all its elements. We note that the use of Newmans algorithm provides a size invariant modularity measure and thus enables us to study the role of network size on modularity as an independent, interesting organization variable.
III. RESULTS

We studied the antigen correlation matrices for the antigen-reactivity data of mothers and their newborns. We start with the combined correlation matrices of both IgG and IgM

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-6

Madi et al.

Chaos 21, 016109 (2011)

FIG. 2. (Color) The hybrid presentation of the informative subgraphs of the dependency network. The selected layout, FruchtermanReingold threedimensional (3D), visualizes the differences between the maternal (a, c) and the newborns (b, d) networks. In (a, b), the nodes were colored according to the isotypes, IgG: green and IgM: red. In (c, d), the colors indicate the strength of the SLI of each antigen on the correlations between all other antigen pairs, from most affecting antigen (dark red) to least affecting antigen (dark blue), The arrows indicate the directionality of the inuence. Note the wide dispersal of highly affecting antigens in the networks.

isotypes. The dependency networks for the mothers and newborns were calculated separately and then compared. Several standard algorithms for automatic graph drawing were implemented. Here we used two main network layouts: spring embedders based on minimization of the total energy of the system (KamadaKawai33 and FruchtermanReingold34). In Fig. 2, we show results of the hybrid presentation of the informative subgraphs of the dependency networks. We start by showing the differences between the isotypic organization (IgG=IgM) [Figs. 2(a) and 2(b)]. Comparing Figs. 2(a)

and 2(b) indicates slightly a higher isotype integration in the maternal combined IgG and IgM network, as is reected by the more homogeneous distribution of the IgG and IgM isotypes in this network. Quantitative comparison shows 684 edges in the mothersversus only 511 edges in the newborns between nodes from different isotypes. Another interesting point comes from looking at the general direction of the inuence of these specic connections. It shows more inuence of IgG over IgM in the maternal dataset (385 vs 299), in contrast, in the newborns dataset, we see more inuence of IgM over IgG (318 vs 193). Next, we recolored the networks according to the strength of the SLI of each antigen on the correlations between all other antigen pairs, from most inuential antigens (dark red) to least inuential ones (dark blue). We note the wide dispersal of highly inuence scored antigens in the networks. Finally, a quantitative comparison between the networks (see Sec. II) revealed low similarities (MD 1.39) between the overall topology of the networks. Thus, the informative subgraphs of the dependency networks for mothers and newborns yield a relatively distinct topological organization of their reactivities to the different antigens. We continue to investigate the relationships between the analyzed antigens and more specically about the most inuencing antigens. As a proxy of antigen inuence, we use the SLI of each antigen on the correlations between all other antigen pairs (see Sec. II). The results of the top 20 most inuential antigens for the networks of mothers and newborns are summarized in Fig. 3. As we can see, both isotypes are preset in the top inuential antigen in the mothers and also in the newborns. In addition, from these lists of 20 most inuential antigens, there is only one exact common antigen, GroEL-13. We note also the existence of many peptides of HSP60, GroEL, and HSP70 in both lists.
A. Separation to isotypes

We continue to test our network analysis approach on the separated isotypes. We note that the produced networks

FIG. 3. (Color online) Top 20 most inuential antigens in the dependency networks of the (a) mothers and (b) newborns. The bars indicate the SLI of each antigen on the correlations between all other antigen pairs in the network. For visualization proposes, inuence score (SLI) values were rescaled between 0 and 1.

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-7

Antigen dependency networks

Chaos 21, 016109 (2011)

FIG. 4. (Color) The hybrid presentation of the informative subgraphs of the dependency isotypic network presented with the FruchtermanReingold 3D layout. (a) Mothers IgG, (b) mothers IgM, (c) newborns IgG, and (d) newborns IgM. The colors indicate the strength of the SLI of each antigen on the correlations between all other antigen pairs, from most affecting antigens (dark red) to least affecting antigens (dark blue). The arrows indicate the directionality of the inuence.

FIG. 5. (Color) The hybrid presentation of the informative subgraphs of the dependency network for the separate isotypes. The selected layout, Kamada Kawai, was partitioned by Newmans algorithm, and each cluster was assigned a different color. (a) Mothers IgG, (b) mothers IgM, (c) newborns IgG, and (d) newborns IgM networks. Arrows indicate the directionality of the inuence. Note that the same colors between two networks do not indicate similar members.

(Fig. 4) are highly signicant in comparison to randomly generated networks with equal distribution.
B. Modularity of the antigen network

Modularity is considered to be one of the main organizing principles of biological networks.35,36 A biological network module consists of a set of elements (e.g., proteins/reactions) that form a coherent structural subsystem and have a distinct function. Several studies have explored the role of modularity and network organization in various protein interaction and regulatory cellular networks.3739 It was suggested that there is a positive selection favoring modularity because it enhances development by enabling evolutionary changes to take place in conned modules while preserving global functions.40 Focusing specically on modularity in immune networks, the immune antigen dependencies networks of mothers and newborns were reconstructed. We then used Newmans algorithm (Sec. II) to partition the network and quantied the modularity of the each network (Fig. 5). Observing the network modularity Q for the different isotypes reveals a noteworthy lead to the maternal IgM (Q 0.752) and newborns IgM (Q 0.751) networks, followed by the rest: maternal IgG (Q 0.737) and newborns IgG (Q 0.666). It seems that the maturation of our immune system involves a higher segregation of the network into modules. A quantitative MD comparison of the different networks, namely the divergence rate measure (see Sec. II) reveals that the IgG networks are more similar than the IgM networks (MDIgG 1.14 vs MDIgM 1.37). Furthermore, when observing similarities between the isotypes within each group, the newborns networks are slightly more similar (MDNewborns 1.26) than those of the mothers (MDMothers 1.29). As expected the IgG networks were the

most similar, as IgG crosses the placenta during pregnancy and is transferred from the mother to her fetus. However, the marked differences between the IgM networks suggest different network organization as was shown previously in terms of antibody proles.1 We continue to investigate the relationships between the analyzed antigens and more specically about the most inuencing antigens. The results of the top 20 most inuential antigens for the maternal and newborns isotype networks are summarized in Table I. As we can see from these lists, we have only a few common antigens;: between the networks of IgG, we have HSP70-6, oligo C, and beta 2 macroglobulin (marked in yellow); between the networks of IgM we have HSP60-10, HSP60-13, and GroEL-21 (marked in green). A comparison between the isotypes within the groups of mothers and newborns shows a common antigen only in the maternal list (marked in bold): HSP60-22 and beta melanocytestimulating hormone. This reemphasizes the larger similarities between the organizations of the maternal isotype networks. To quantify the differences in the ordered lists of inuential antigens for each of the networks, we used a heuristic method that measures the Euclidian distance between the indexing of the sorted lists of antigens. These results support our previous ndings as it shows a higher conservation between the IgG lists (Euclidian distanceIgG 1936) versus IgM lists (Euclidian distanceIgM 1995) and slightly a higher conservation within the maternal networks (Euclidian distancemothers 2014)versus the newborns one (Euclidian distancenewborns 2018).
C. Conserved elements subnetworks

To nd conserved networks elements between the datasets of the mothers and newborns, we applied a heuristic

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-8

Madi et al. TABLE I. Top 20 most inuential antigens in the isotype networks. Mothers IgG HSP60-22 HSP60-25 Myeloperoxide Actin Factor II GroEL-31 HSP70-20 GroEL-18 HSP70-4 GroEL-3 HSP70-28 HSP70-23 HSP70-6 Beta MSH Beta 2 microglobulin HSP70-43 HSP70-26 HSP60-32 Oligo C BETA 2 macroglobulin Mothers IgM HSP60-26 HSP70-37 PBS HSP70-8 GroEL-23 GroEL-13 Poly Asp IL 4 HSP60-35 C9 nonST HSP60-10 GroEL-21 HSP70-36 HSP60-28 HSP60-22 HSP60-13 GroEL-19 Ins A Beta MSH MT 256 Newborns IgG MT 3 HSP70-12 Oligo C HSP70-36 GroEL-25 HSP70-6 GroEL-22 HSP60-19 E. coli 27 HSP60-277 Lactoferin Beta 2 macroglobulin MMP3 PTH HSP70-11 MT 180 Pepstatin A HSP60-10 GroEL-10 GroEL-2

Chaos 21, 016109 (2011)

Newborns IgM SOD HSP70-3 pG LPD HSP60-13 HSP70-39 HSP60-10 HSP70-43 GroEL-21 Peroxidase Spectrin HSP60-27 HSP70-42 GroEL-14 HSP60-23 HSP70-41 HSP70-18 HSP60-18 GroEL-16 Pepstatin HSP60-9

method that nds such subnetworks between the informative subgraphs of the dependency network for the separated isotypes (see Appendix C). A comparison between the maternal and newborns networks showed a few such conserved elements in the IgG (56 antigens) and IgM (50 antigens) isotypes with the largest subnetwork between the IgG networks consists of 6 antigens and the largest subnetwork between the IgM networks consists of 7 antigens (Fig. 6).
IV. DISCUSSION

We present here a new approach to investigate antigen dependency networks computed from matrices of antigenantigen correlations. The latter are calculated from antigen microarray data of autoantibody reactivity of IgM and IgG isotypes present in the sera of ten mothers and their newborns. We used the antigen dependencies to construct a new quantitative measure of the system level inuence of antigen j or inuence score of j, SLI( j) as the sum of the inuence of j on all other antigens i. While we have introduced the new approach and its ability to unveil important biological information within the context of the immune system, it is expected to be applicable to a wide range of other complex biological networks (e.g., gene networks, protein interaction networks, and neural networks), as

FIG. 6. (Color) The hybrid presentation of the largest conserved element between the (a) IgG informative subgraphs of mothers and newborns. (b) The same for IgM isotype. Arrows indicate the directionality of the inuence.

well as social, nancial, and manmade networks. We expect that, in all of these examples, our method will be able to unveil important hidden information (see also Ref. 41). Partial correlations were employed to construct the antigen dependencies from the antigenantigen correlations. More specically, we rst dene the dependency of node i on node j (or the inuence of node j on node i), D(i, j) as the average over all nodes k of the difference between the correlations C(i, k) and the partial correlations PC(i, k j j). Using these denitions, the inuence score SLI( j) is the sum of D(i, j) over all nodes i. Investigating the IgG and IgM combined networks of the mothers and newborns, we found that, in the two networks, the top ranked antigens (antigens with high inuence score) include both isotypes. We also found that the most inuential antigens vary from birth to adulthood as is reected by the fact that only a few antigens are included both among the top inuential antigens of the newborns networks and among the top ones in the maternal network. These ndings are both for the combined networks (only one common antigen) and the separated IgG and the IgM networks (only three common antigens in each). Furthermore, the combined networks of the mothers and newborns were found to have a higher isotype integration in the maternal network. This is in agreement with previous ndings in the context of correlation networks.5 Since the antigen dependency networks are directed, we could also inspect the directionality of the isotypes inuence. Doing so, we found a higher IgG ! IgM inuence in the maternal networks while the newborns networks exhibit a higher IgM ! IgG inuence. This result is somewhat unexpected considering the fact that IgG antibodies are transferred from mother to her fetus via the placenta during pregnancy. We also found, in both the IgG and the IgM networks, the existence of subnetworks exist at birth (the newborn networks) and are persist in the maternal networks. The conserved networks have a total of 56 antigens in the IgG networks and 50 antigens in the IgM networks. These conserved subnetworks

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-9

Antigen dependency networks

Chaos 21, 016109 (2011)

motifs might serve to maintain information about the healthy network organization upon the immune network reorganization from childhood to adulthood state described next. We investigated the networks topological similarity using the divergence rate measure. These investigations revealed a higher similarity between the IgG network of the mothers and the IgG network of the newborns (in comparison to the similarity between the two IgM networks). These results are consistent with the fact that considerable amounts of the IgG cross the placenta during pregnancy and transfer from the mother to her fetus. Investigating the networks modularity, we found that the networks of IgM exhibit a higher modularity in comparison to the networks of the IgG and that these differences are more profound in the newborns. It may indicate that a healthy immune state has modular organization, which affords an efcient performance of many tasks as a part of its normal physiological role. We note that the modular organization of the healthy antigen dependency network is consistent with earlier ndings of Madi et al.1 about the existence of antigen cliques in the correlation network of healthy subjects. Put together, we found that the immune system at birth is associated with a higher modular reorganization of the IgG network and more pronounced topological reorganization of the IgM network. We note that the network divergence rate topology

is associated with local organization (the nearest neighbors of each antigen), while the modularity is associated with the global network organization. These ndings may indicate a profound and intricate response associated with local and global reorganizations of the immune networks system between the adult and infant immune states: the IgG networks exhibit a more profound global reorganization while the IgM networks exhibit a more profound local reorganization. If correct, these ndings provide a dramatic demonstration of the effectiveness of the new approach to unveil the most signicant biological information.
ACKNOWLEDGMENTS

We are thankful to Alexandra Sirota-Madi for her technical help in crucial times. This research has been supported in part by the Maugy-Glass Chair in Physics of Complex Systems and by the National Science Foundation-sponsored Center for Theoretical Biological Physics (CTBP) Grant Nos. PHY-0216576 and 0225630, and by the University of California at San Diego.
APPENDIX A: SLI

As a proxy of antigen inuence, we use the SLI of each antigen on the correlations between all other antigen pairs (see Sec. II). We start by calculating the SLI scores of each

FIG. 7. (Color) Top 100 most inuential antigens in the dependency networks of the (a) mothers, (b) newborns, (c) mothers IgG, (d) newborns IgG, (e) mothers IgM, and (f) newborns IgM. The antigens were ranked according to their SLI scores and divided by the maximum score to get a relative SLI score. Note the change in SLI scores around 20 antigens mainly in the integrated diabetic and healthy IgM networks.

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-10

Madi et al.

Chaos 21, 016109 (2011)

antigen in the maternal and newborns networks and in the separate isotype networks (Fig. 7). For the purpose of comparison, we divided the SLI scores by the largest SLI value. We plotted the rst 100 sorted relative antigen SLI values and observed for a notable change in the steady decreasing plots. As we can see the strongest change is in the integrated IgG and IgM maternal and newborns networks is around 20 antigens the same with less extent is seen in the newborns IgG network. In addition, we also compare between the informative subgraphs in which the antigen system level inuence was calculated for the entire network and the case in which they were calculated within the informative subgraphs. Note that in the latter case, since the directed links are not weighted, the measure of the system level inuence of each antigen j is simply the total number of directed edges from j to other antigens. For this type of comparison, we simply measured the correlation between the two score vectors. Results show a signicant change between the two calculations (maternalIgG 0.25, maternalIgM 0.33, newbornsIgG 0.19, and newbornsIgM 0.33), pointing to the importance of adding this information from the complete dependency network (D).
APPENDIX B: COMPARISON TO CORRELATION BASED PMFG

was then partitioned by Newmans algorithm,32 and each cluster was assigned a different color. We then superimposed this cluster coloring on the informative subgraphs dependency network of the same dataset [Fig. 8(b)]. As we can see, it is not only the external topological construction of the networks but also a great difference between the internal relationships.
APPENDIX C: CONSERVED SUBNETWORKS

To nd conserved networks elements between the datasets of the mothers and newborns, we applied a heuristic method that nds such subnetworks between the two networks. The method nds these conserved subnetworks as follows: First we dene DGX(i) to be the topological distances from a node i to all its neighborhoods nodes {i}NN. Then we dene the conditional distances CDiv(GX j GY)(i)to be the

To farther demonstrate the differences between a correlation based network and a partial correlation based dependency network, we constructed the informative subgraph, a correlation PMFG based network, for 290 antigens of the mothers [Fig. 8(a)]. The selected layout (KamadaKawai)

FIG. 8. (Color) The hybrid presentation of the informative subgraphs of the correlation vs dependency network. (a) Maternal IgG correlation, (b) maternal IgG dependency network, (c) maternal IgM correlations, and (d) maternal IgM dependency network. (a, c) The selected layout, Kamada Kawai, for the correlation based PMFG was partitioned by Newmans algorithm, and each cluster was assigned a different color. (b, d) Superimposing the cluster coloring on the dependency networks. Arrows indicate the directionality of the inuence. Note that the same colors between two networks do not indicate necessarily similar nodes but members of the same cluster.

FIG. 9. (Color) Conserved subnetworks. (a) The IgG conserved subnetworks between the maternal and newborns networks. (b) The same for IgM. (c) The maternal IgG dependency network with all conserved elements marked in red. (d) The same for the IgM network.

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp

016109-11

Antigen dependency networks

18

Chaos 21, 016109 (2011) E. Fuchs, A. Ayali, E. Ben-Jacob, and S. Boccaletti, Phys. Biol. 6, 36018 (2009). C. Coronnello, M. Tumminello, F. Lillo, and R. N. Mantegna, Acta. Phys. Pol. B 36, 2653 (2005). 20 H. de Jong, J. Comput. Biol. 9, 67 (2002). 21 R. N. Mantegna, Eur. Phys. J. B 11, 193 (1999). 22 R. N. Mantegna, Comput. Phys. Commun. 121122, 153 (1999). 23 G. J. Ortega, R. G. Sola, and J. Pastor, Neurosci. Lett. 447, 129 (2008). 24 ` M. Tumminello, C. Coronnello, F. Lillo, S. Micciche, and R. N. Mantegna, Int. J. Bifurcation Chaos Appl. Sci. Eng. 17, 2319 (2007). 25 D. B. West, An Introduction to Graph Theory (Prentice-Hall, Englewood Cliffs, NJ, 2001). 26 M. Tumminello, T. Aste, T. Di Matteo, and R. N. Mantegna, Proc. Natl. Acad. Sci. U. S. A. 102, 10421 (2005). 27 S.-G. Lee and S. Kim, Phys. Rev. E 73, 041924 (2006). 28 Non-linear Structures in Physical SystemsPattern Formation, Chaos and Waves, edited by L. Lam and H. C. Morris (Springer-Verlag, New York, 1990). 29 S. Boccaletti, V. Latora, Y. Moreno, and M. Chavez, Phys. Rep. 424, 175 (2006). 30 L. Danon, A. Daz-Guilera, J. Duch, A. Arenas, J. Stat. Mech.: Theory Exp. 2005, P09008. 31 A. Madi, Y. Friedman, D. Roth, T. Regev, S. Bransburg-Zabary, and E. Ben Jacob, PLoS One 3, e2708 (2008). 32 M. E. J. Newman, Eur. Phys. J. B 38, 321 (2004). 33 T. Kamada and S. Kawai, Inf. Process. Lett. 31, 7 (1989). 34 T. M. J. Fruchterman and E. M. Reingold, Softw.: Pract. Exp. 21, 1129 (1991). 35 L. H. Hartwell, J. J. Hopeld, S. Leibler, and A. W. Murray, Nature (London) 402, C47 (1999). 36 D. M. Wolf and A. P. Arkin, Curr. Opin. Microbiol. 6, 125 (2003). 37 E. Segal, N. Friedman, D. Koller, and A. Regev, Nat. Genet. 36, 1090 (2004). 38 E. Segal, M. Shapira, A. Regev, D. Peer, D. Botstein, D. Koller, and N. Friedman, Nat. Genet. 34, 166 (2003). 39 B. Snel and M. A. Huynen, Genome Res. 14, 391 (2004). 40 C. O. Wilke and C. Adami, Mutat. Res. 522, 3 (2003). 41 Y. Louzoun, S. Solomon, H. Atlan, and I. Cohen, Physica A 297, 242 (2001).
19

topological distances in graph {GY} from node i to the group of nodes {i}NN dened in graph {GX}. Finally, we select those subsets of nodes that have the same distance in both graphs {GX} and {GY}. Note that these nodes, which are in the neighborhood of i in graph {GX}, need not be in the neighborhood of i in the graph {GY}.
A. Madi, I. Hecht, S. Bransburg-Zabary, Y. Merbl, A. Pick, M. ZuckerToledano, F. J. Quintana, A. I. Tauber, I. R. Cohen, and E. Ben-Jacob, Proc. Natl. Acad. Sci. U. S. A. 106, 14484 (2009). 2 R. Albert and A.-L. Barabasi, Rev. Mod. Phys. 74, 47(2002). 3 R. L. Graham and P. Hell, IEEE Ann. Hist. Comput. 7, 43 (1985). 4 R. N. Mantegna and H. E. Stanley, An Introduction to Econophysics: Correlations Complexity in Finance (Cambridge University Press, Cambridge UK, 2000). 5 A. Madi, D. Y. Kenett, S. Bransburg-Zabary, Y. Merb, F. J. Quintana, A. I. Tauber, I. R. Cohen, and E. Ben-Jacob, PLoS One 6, e17445 (2011). 6 M. E. J. Newman, SIAM Rev. 45, 167 (2003). 7 J. Pearl, Causality: Models, Reasoning, and Inference (Cambridge University Press, London, United Kingdom, 2000). 8 W. Shipley, Cause Correlation in Biology: A Users Guide to Path Analysis, Structural Equations, and Causal Inference (Cambridge University Press, Cambridge, United Kingdom, 2000). 9 P. Spirtes, C. Glymour, and R. Scheines, Causality, Prediction, and Search (Springer Verlag, New York, 2000). 10 D. Y. Kenett, M. Tumminello, A. Madi, G. Gur-Gershgoren, R. N. Mantegna, and E. Ben-Jacob, PLoS ONE 5(12), e15032 (2010). 11 A. de la Fuente, N. Bing, I. Hoeschele, and P. Mendes, Bioinformatics 20, 3565 (2004). 12 A. Reverter and E. K. F. Chan, Bioinformatics 24, 2491 (2008). 13 Y. Shapira, D. Y. Kenett, and E. Ben-Jacob, Eur. Phys. J. B 72, 657 (2009). 14 U. Lee and S. Kim, Phys. Rev. E 73, 041920 (2006). 15 M. E. Newman, Proc. Natl. Acad. Sci. U. S. A. 103, 8577 (2006). 16 J. L. Rodgers and W. A. Nicewander, Am. Stat. 42, 59 (1988). 17 K. Baba, Aust. N. Z. J. Stat. 46, 657 (2004).
1

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://cha.aip.org/cha/copyright.jsp