Dysregulated mechanisms underlying Duchenne muscular dystrophy from co-expression network preservation analysis

Duchenne Muscular Dystrophy (DMD) is an X-linked recessive disorder with its primary insult on the skeletal muscle. Severe muscle wasting, chronic inflammation and fibrosis characterize dystrophic muscle. Here we identify dysregulated pathways in DMD utilizing a co-expression network approach as described in Weighted Gene Co-expression Network Analysis (WGCNA). Specifically, we utilize WGCNA’s “preservation” statistics to identify gene modules that exhibit a weak conservation of network topology within healthy and dystrophic networks. Preservation statistics rank modules based on their topological metrics such as node density, connectivity and separability between networks. Raw data for DMD was downloaded from Gene Expression Omnibus (GSE6011) and suitably preprocessed. Co-expression networks for each condition (healthy and dystrophic) were generated using the WGCNA library in R. Preservation of healthy network edges was evaluated with respect to dystrophic muscle and vice versa using WGCNA. Highly exclusive gene pairs for each of the low preserved modules within both networks were also determined using a specificity measure. A total of 11 and 10 co-expressed modules were identified in the networks generated from 13 healthy and 23 dystrophic samples respectively. 5 out of the 11, and 4 out of the 10 modules were identified as exhibiting none-to-weak preservation. Functional enrichment analysis identified that these weakly preserved modules were highly relevant to the condition under study. For instance, weakly preserved dystrophic module D2 exhibited the highest fraction of genes exclusive to DMD. The highly specific gene pairs identified within these modules were enriched for genes activated in response to wounding and affect the extracellular matrix including several markers such as SPP1, MMP9 and ITGB2. The proposed approach allowed us to identify clusters of genes that are non-randomly associated with the disease. Furthermore, highly specific gene pairs pointed to interactions between known markers of disease and identification of putative markers likely associated with disease. The analysis also helped identify putative novel interactions associated with the progression of DMD.


Background
Duchenne muscular dystrophy (DMD), is a lethal form of dystrophinopathy characterized by marked deficiency or absence of subsarcolemmal cytoskeletal protein-dystrophin. Absence of this protein is caused due to frame shift mutations of the dystrophin gene [1]. Dystrophin, part of the dystroglycan complex plays a crucial role in maintaining the integrity of the muscle fiber. Absence of dystrophin causes uneven mechanical force transmissions leading to sarcolemmal ruptures and subsequent atrophy. Clinical manifestations of DMD occur by second year of birth and progressively degrade with time. The first decade of life is marked by developmental delays, and steady decreases in the strength of the limbs and torso with subsequent loss of ambulation. Respiratory and cardiac complications arise by the second decade of life leading to death [2]. Here we utilize a co-expression networks approach to gain insights into molecular interactions dysregulated in dystrophic skeletal muscle with respect to healthy muscle.
Co-expression networks are being increasingly used for deciphering disease mechanisms and providing systems level views of dysregulated pathways [3,4]. The basic premise of co-expression analysis is that strongly correlated genes are likely to be functionally associated. Weighted Gene Co-expression Network Analysis (WGCNA) is an open source tool that performs co-expression analysis using a network theoretic approach. WGCNA integrates expression differences across samples into a higher order network structure, elucidating relationships among genes based on their co-expression profiles [5,6].
Here, we propose to utilize a set of statistics implemented in WGCNA, called preservation statistics, to elucidate global differences in mechanisms underlying the early phase of DMD [7]. Traditionally, these statistics have been utilized to identify modules of genes that are topologically preserved between two networks. In contrast to this approach, we propose to utilize these statistics to identify modules that do not exhibit a preservation of topology between networks. This is based on the premise that such modules would represent a cohort of gene interactions that are vastly different between conditions and point to dysfunctional pathways and interactions.
In our current study we utilize a previously published dataset on DMD containing a cohort of healthy and affected individuals (mostly children) -representing the early phase of DMD development [8]. Briefly, we evaluated differential mechanisms between dystrophic and healthy skeletal muscle using the following approach; first, co-expression networks were generated independently for healthy and dystrophic samples; second, clustering each of the co-expression networks resulted in several groups of biologically relevant genes (modules) for each condition; and finally, preservation of modular topology from one condition was detected with respect to the second condition, allowing us to identify differences in gene connectivity patterns between conditions.
The results of our differential analysis reveal convergent molecular mechanisms consistent with published studies in addition to providing us novel hypothesis on gene interactions associated with the early phase of DMD.

Results and discussion
Network construction and modularity detection WGCNA was utilized to construct unsigned weighted co-expression networks from 13 healthy and 23 DMD muscle samples across 4000 most varying genes (see Methods). Briefly, unsigned network adjacency matrices were obtained by raising the Pearson correlation matrices to a power β =5 for each condition [5]. The adjacencies were transformed to similarity matrices for subsequent clustering. When represented as networks, each entry of the similarity matrix ij corresponds to weight on the edge between genes ij. The strength of similarity between two genes depends not only on the correlation but also on their shared network neighborhood [5]. Clustering based on such a similarity allowed for identification of gene groups that were biologically relevant.
Hierarchical clustering of the two weighted networks resulted in eleven modules for the network from healthy samples (N1-N11, see Methods) and ten modules from the network derived from dystrophic samples (D1-D10, see Methods). Additional file 1 shows the clustering dendrograms and corresponding modules identified in both networks. Genes that did not cluster were excluded from further analysis for the purposes of this study.

Functional characterization of modules identified in healthy and dystrophic networks
Modules identified using WGCNA have been repeatedly shown to be biological relevant to the condition under study [4,9]. We utilized functional enrichment analysis as a method to assess the functional coherence of modules identified within each of the networks.

Characterizing modules of the healthy network
Enrichment of modules from the healthy network revealed several functions routinely associated with healthy skeletal muscle such as striated muscle contraction, energy generation and extracellular matrix organization (Table 1).
Skeletal muscle contraction occurs via the coordinated movement of several proteins particularly the actinmyosin complex within the sarcomere, incident upon a changing Ca 2+ flux. Several genes encoding the sarcomeric proteins such as MYH2, MYH6, MYH7, MYBPC2 TPM1, TPM3, TNNC1/2, TNNI1, TNNT1, MYOM2, MYOZ1, MYOZ2, and MYOZ3 were identified in modules N1 and N8 [10]. The extracellular matrix (ECM) surrounding the skeletal muscle plays an important role in force transmission and affects the mechanical properties of the skeletal muscle. Several genes associated with the ECM and focal adhesion such as COL4A1, COL4A2, COL5A2, COL6A1, COL6A2, ITGA6, ITGB1, CAV1, CTNNB1, ACTB and LAMN4 were identified in modules N1 and N8 [11]. Muscle contraction and relaxation depend primarily upon energy derived from hydrolysis of adenosine triphosphate (ATP) within the mitochondria. Glycogen/glucose and lipid metabolism serve as major sources of ATP within muscle. Several genes associated with such metabolism were identified within modules N6 and N9 with genes such as NDUFB3, NDUFB5, FABP4, AACS, ADIPOQ, SDHA, SDHB and SCD.

Characterizing modules of the dystrophic network
Though the same 4000 genes were used to construct the co-expression network in each case, modules cluster differently based on their co-expression. Subsequent enrichment of modules from the dystrophic network revealed functions particularly associated with dystrophic muscle such as response to wounding ( Table 2).

Identifying functional differences between healthy and dystrophic muscle-a systems approach
The functional annotation clustering results above suggested a mutual exclusivity of certain functions between dystrophic and healthy muscle implying a difference in the topology of connections for genes within these networks.
In order to systematically assess and quantify differential gene co-expression, we performed a "preservation" analysis. This allowed us to identify modules that were fairly unique in terms of their gene co-expression within a given network compared to another. We utilized a method implemented in WGCNA called "modulePreservation" [7]. In contrast to the idea of the original paper which aimed at identifying modules preserved between conditions, we aimed to identify modules "weakly preserved" across conditions (see Methods). We hypothesized that modules that were either weakly preserved or non-preserved in either condition might point to dysregulated pathways in disease that were either acquired or lost with respect to a healthy skeletal muscle.
Assessing differential co-expression in healthy muscle with respect to dystrophy In order to evaluate how the topology of the healthy network differed from the dystrophic network, we computed the preservation (density, connectivity and separability statistics) of modules from the healthy network (reference network) as compared to the dystrophic network (test network). Lower preservation statistics suggested a loss of co-expression structure between these gene pairs in the dystrophic network.
Based on the median preservation score, we identified a total of 5 interesting modules. Two modules N1 and N8 from the healthy network were non-preserved in the dystrophic network while three other modules N2, N3 and particularly N7 exhibited weak preservation ( Figure 1A). A table of the observed preservation statistics for all modules of the healthy network is provided in Additional file 2. Z summary , a permutation statistic (see Methods) defined for assessing significance of the observed preservation also revealed a low preservation of these modules (Table 3). Broadly, loss of healthy muscle function and weakened contractility in dystrophic muscle triggers the activation of atrophic pathways leading to severe muscle wasting, changes to the extracellular matrix, fibrosis and necrosis over time [16]. Accordingly, unpreserved modules N1 and N8 were Figure 1 Differential co-expression in healthy muscle with respect to dystrophy. A: Scatter plot identifying the median rank of module preservation between test (dystrophic) and reference (healthy) networks. B: Co-expression between genes identified in module N1-The co-expression patterns for a subset of genes identified within module N1 are shown here. All interactions have a specificity of >0.95. Darker the line, stronger is the strength of co-expression between the gene pairs. associated with genes necessary for striated muscle contraction, while the weakly preserved modules (N2, N3 and N7) were associated with ECM and cytoskeletal framework of the skeletal muscle.
We utilized a co-expression specificity measure [17] (see Methods) to elucidate co-expressed genes pairs (edges) from these 5 modules. We observed that modules exhibiting none-to-low preservation in the healthy network consistently had a higher fraction of gene pairs exclusive to the healthy network than their preserved counterparts (Table 4). For instance,~35% of the gene pairs considered (599/1738) within N1 were specific to the healthy network ( Figure 1B). Several of the genes involved are known markers influencing skeletal muscle contraction such as ANKRD2, TNNC2, SMAD3, HSPB1, CRYAB, SDC4, and MYOD1.
It was interesting to observe however that a majority of the genes identified as being part of these interactions were ion-binding as per GO's molecular function ontology (zinc and copper, p < 10 −4 ). A visual inspection of subset of the exclusive genes pairs identified reveals strong co-expression between several zinc binding genes such as metallothioneins (MT1E/F/H/X), ZNF593, and genes affecting muscle contraction (ANKRD1, MYOD1, SMAD3, HSPB1). Metallothioneins have been postulated to be associated with a host of functions ranging from chaperones for synthesis of metalloproteins, to reservoirs of essential metals (Zn and Cu) in healthy tissue [18]. Specifically, metallothioneins (MTs) exhibit specific redox properties and have been speculated to selectively control release and uptake of Zinc [18]. However, it is interesting to note that MTs were co-expressed with genes affecting muscle contraction only within the healthy network, suggesting a link between zinc homeostasis, and muscle contraction in healthy muscle. The exclusivity of connections to the healthy network further emphasizes the possibility of an aberration in zinc homeostasis and its effect on contraction in DMD.
Assessing differential co-expression in dystrophic muscle with respect to healthy tissue A similar analysis with dystrophic network as the reference network, allowed us to identify gene pairs that were not conserved in healthy tissue. As proposed earlier, we speculated that identifying non/weakly preserved modules in the dystrophic network could point to gene associations that are gained in dystrophy. We identified two modules-D1 and D8 that exhibited no preservation in the healthy network while two other modules D3 and to a greater extent D2 were weakly preserved (Figure 2A). A table of the observed preservation statistics for all modules of the dystrophic network is provided in Additional file 3. The Z summary scores (see Methods) likewise revealed a low preservation of these modules via permutation testing (Table 5).
Juvenile dystrophic muscle, in general, exhibits atrophy and is pre-necrotic, with pathways associated with wounding and inflammation being subsequently activated. The functional enrichment identified within these four modules (Table 2) corroborated our approach, highlighting functions that are more pronounced in dystrophic muscle compared to healthy tissue.
These modules also exhibited higher specificity of connections to the dystrophic network than their preserved counterparts (Table 6). For instance, the highest specificity was observed for module D2 with nearly 45% of its gene pairs as being specific to dystrophy (specificity >  0.95). Interestingly, the dystrophic-specific gene interactions identified in module D2 corresponded with interactions categorized as a part of the inflammatory and tissue repair/remodeling repertoire of genes, as witnessed in models of skeletal muscle injury, particularly dystrophy ( Figure 2B). For instance, expression of SPP1, a multifunctional cytokine (also called early T-cell activation-1 (Eta-1), osteopontin), is linked with macrophage infiltration, resulting in a chronic inflammatory response observed in dystrophic muscle [13,19]. VSIG4, a regulator of T-cell activation expressed mostly in macrophages is strongly coexpressed within D2 [19]. Though the exact mechanisms by which skeletal muscle attracts and allows entry of neutrophils and macrophages in dystrophic muscle are not well understood, there is evidence suggesting that ITGB2 is required to control the functional activities of neutrophils and macrophages within muscle [20]. Fibrosis observed in DMD, is largely activated in response to chronic inflammatory processes initiated within dystrophic muscle [13] and broadly refers to the accumulation of excess connective tissue (ECM) [11]. SPP1 which is also expressed in fibrotic lesions is considered a marker for disease severity in DMD [21]. SPP1 is required for differentiation of myofibroblasts [22], an important class of fibroblastic cells required for wound healing, present abundantly within dystrophic muscle. Fibronectin (FN1) serves as a marker for fibroblast activation in muscle [23].
SPP1, in addition to modulating fibrotic responses, promotes cell-cell and cell-matrix adhesions through its interaction with integrins, and CD44 [24]. Mature focal adhesion complexes containing genes such as ACTN1, fail to form in the absence of SPP1 (SPP1 −/− ) within myofibroblast cultures [22] suggesting similar pathways for adhesion in dystrophy. Interestingly, within this module we also identify MMP9-a matrix metalloproteinase whose increased expression, particularly in the pathology of DMD, is associated with breakdown of cytoskeleton-ECM components leading to sarcolemmal damage and fiber necrosis [25,26] Additionally, MMP9 is also suggested to act as an inflammatory stimulus for mediating neutrophil and macrophage infiltration within the dystrophic skeletal muscle [27,28].
SPP1 is subject to extensive posttranslational modification via glycosylation, phosphorylation and sulphation. Specific posttranslational modifications have been associated with altered properties and function of SPP1 [29]. Interestingly, ACP5, a phosphatase required for mineralization of cartilage and bone matrix resorption [28] was recently demonstrated to be responsible for phosphorylation of SPP1 in endometrial tissue. Though no evidence for role of ACP5 or post-translation modification of SPP1 in dystrophic muscle exists, the co-expression of ACP5 with SPP1 suggests a possible role in dystrophy warranting further investigation.
Additionally, SPP1 shows high specificity interactions with certain ECM markers including CTSK, LUM, VCAN and VIM ( Figure 2B). Though there is no direct evidence for the interaction of these markers with SPP1, the extant understanding of the ECM markers combined Figure 2 Differential co-expression in dystrophic muscle with respect to healthy muscle. A: Scatter plot identifying the median rank of module preservation between test (healthy) and reference (dystrophic) networks. B: The Pearson correlation between a subset of genes identified in module D2 from the dystrophic network-For the same set of genes from module D2, we also identify correlation patterns in the healthy network (left). The size of the node is proportional to the sum of all correlation strengths at the node in the network shown. Red lines indicate positive correlation and blue indicate negative correlation.
with the high specificity of co-expression observed in our network module suggest possible associations with SPP1 in dystrophic muscle.
Overall, our results indicate that the modules exhibiting low preservation statistics contain several gene pairs that are likely to be associated with the disease progression. Though it is conceivable that not all the genes identified within these less-preserved modules play a role in disease, several high specificity gene pairs identified were noted and hypothesized to play a significant role in pathogenesis of DMD.

Conclusions
An analysis of modules exhibiting a low preservation between dystrophic and healthy conditions showed that these modules showed a higher specificity among gene pairs pertinent to the condition under study. We illustrated the application of using preservation statistics to detecting modules functionally associated with dysregulated pathways in disease, as exemplified by the inflammatory module D2. This approach enabled identifying putative biomarkers, such as ACP5 identified within module D2, likely to be associated with the disease.
In summary, our method provided a simple approach to identifying differences between conditions, which can be utilized for exploratory analysis of dysregulated pathways in disease using a published set of statistics.

Data processing
The data set was preprocessed using Bioconductor/R packages affy and WGCNA. Data was MAS 5.0 normalized using functions in affy and any array with an average inter sample correlation <2 SDs (σ) below the mean was removed [9]. This resulted in the removal of a single array -GSM139506.CEL (2.54 σ below mean) from the study. All probes with missing Entrez gene identifiers were excluded from this study, resulting in a data set comprising of the expression values for 11101 probes. Multiple probes were collapsed into a gene based on variance resulting in a final reduced expression data set comprising of 7996 genes and 36 samples. A subset of 4000 most varying genes was used to construct the coexpression networks, in an effort to minimize computational complexity and eliminate low varying genes that may contribute minimally to the co-expression matrix. The number 4000 was chosen as it represents roughly half the total number of genes (7996 genes) identified after pre-processing. This method of gene list selection is agnostic to their pathophysiological role in the muscle.

Co-expression network generation and modularity detection
The topological overlap measure (TOM) in WGCNA between genes i and j is defined as follows Where A is the weighted adjacency matrix given by A ij = |cor(x i, x j) | β and β ≥ 1 is the soft thresholding power.  TOM takes continuous values between 0 and 1, where 0 for a gene pair indicates no similarity between the genes while 1 indicates a direct link. The soft thresholding power β for each dataset in our study was ascertained as prescribed in the original publication [5].
Co-expression networks from the adjacency matrices of healthy and dystrophic samples were generated using the "TOMsimilarity" function available via WGCNA. Hierarchical clustering on the topological dissimilarity (1-TOM) was performed using the function "flashClust". The tree cut height was dynamically determined using the function "cutreeDynamic" in WGCNA, for identifying modules in each of our networks. Additional files 4 and 5 provide a list of all genes identified in each of the networks (healthy and dystrophic respectively) and their corresponding module assignments.
Preservation of modules "Module preservation" or preservation statistics implemented in WGCNA allows us to detect the conservation of gene pairs between two networks (test and reference) [7]. Briefly, three types of network based module preservation statistics have been identified by this method, namely 1) Density based preservation statistics: determine if nodes remain highly connected in the test network. Four independent measures account for this statistic. 2) Connectivity based preservation statistics determine the extent to which connectivity patterns between nodes in the reference network are similar to the test network. Three independent measures of the network account for this statistic. 3) Separability based preservation statistics determine if network modules remain distinct from one another in the test network.
Network based statistics employed by WGCNA do not require identification of modules within the test network to ascertain the conservation of reference network modules within the test network. This is in contrast to several existing methods that ascertain module preservation as discussed in the original publication. The authors of the original publication have shown that using this method it is possible to identify sets of preserved co-expression across species.
As these statistics measure distinct aspects of module preservation, two composite measures have been defined 1) Median rank: A composite measure that is based on observed preservation values and is less dependent on module sizes. It is defined as the mean of median ranks computed for connectivity and density measures of each module (0.5 (medianRank connectivity + medianRank density ). 2) Z summary : A permutation based composite Z statistic that is used to assess the significance of observed statistics and is defined as the mean of Z scores computed for density and connectivity measures (0.5(Z denstiy + Z connectivity )). An associated empirical p-value is also calculated by the algorithm.
We utilize median rank to identify module preservation and Z summary to assess significance of module preservation via permutation testing. Based on the number of modules within each of our networks, a median rank of 8 was chosen as a cutoff to detect weak preservation. Permutation was performed 200 times given the computational complexity involved for our network sizes. Based on the thresholds prescribed in [7], modules with a Z summary score >10 indicate preservation, 2 to 10 indicate weak to moderate preservation and <2 indicate no preservation in the permutations.

Network specific gene pairs
Condition specific interactions for a given pair of genes i and j was defined as [17]: Where, TOM ij (cond). is the normalized TO for the gene pairs i-j in the given condition (healthy or disease).
We considered gene pairs to be condition specific, if the specificity was >0.95 and were in the top 1% of the gene pairs ranked on TOM similarity in any given module. Number of edges in an undirected network is computed as n(n-1)/2, where n is the number of nodes. Considering the top 1% allowed us to focus only on the strongest co-expression patterns within the module, rather than noise.

Enrichment analysis and visualization
The results presented correspond to the top term identified in the highest-ranking cluster (as of this analysis) using the annotation clustering feature available in DA-VID [31], with Gene Ontology's Biological process functional annotations. Additional files 6 and 7 provide the top 3 functional annotation clusters identified for each of the modules within the healthy and dystrophic networks respectively. Cytoscape [32] and Bioconductor [33] were utilized for generating the figures in this paper.

Availability of supporting data
The data sets supporting the results of this article are published data sets available through the Gene expression omnibus repository, [GSE6011 http://www.ncbi.nlm.nih. gov/geo/query/acc.cgi?acc=GSE6011].