Currently, the MADMuscle database collects all transcriptome data sets related to muscle studies from the public repository Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) 13 of the National Center for Biotechnology Information (NCBI) at the National Institutes of Health (NIH). These data sets were identified using the following keywords: "muscle", "myo", "heart" or "cardio". This represents a total of 535 analyzed data sets (see Additional File 1) corresponding to 447 unique GEO series, 116 different microarray platforms and 7 different species (Homo sapiens Hs, Mus musculus Mm, Rattus norvegicus Rn, Canis familiaris Cf, Gallus gallus Gg, Drosophila melanogaster Dm, and Caenorhabditis elegans Ce).

Raw data from GEO correspond most of the time to already pre-processed and/or normalized data. This step thus aims at normalizing the pre-normalized microarray data to remove artefacts and to ensure that each entry of the database follows the same procedure (see details in part 2 of the Additional File 3 and also Additional File 4). For each data set the K-nearest neighbors method 59 was used to evaluate the missing values in the microarray data sets. Then, non-linear effects such as background or saturation were corrected by LOWESS 60, as previously described 6162, using a channel by channel procedure 63, each array being individually normalized to the median profile of all arrays.

Hierarchical classification 55 was used to investigate relationships between samples and relationships between genes both on raw and re-normalized data (Figure 1A, B and Additional File 1). An average linkage clustering, using Pearson's correlation as similarity metric, was performed with Cluster 3.0 64 and applied to data that were log-transformed and median-centred on genes 65.

Only expression matrixes with at least ten columns/samples (semi-empirically determined threshold) were kept for subsequent analysis (see part 3 of the Additional File 3 for detailed information). Clusters of co-expressed genes were identified using an iterative k-means 66 procedure (Figure 1C and Additional File 1) based on the Forgy's algorithm 67 and implemented in the statistical software package R 68. Genes conserved in 95% of the 1000 independent k-means for a same cluster were retained (for a more detailed description of the method, see also part 4 of the Additional File 3). A total of 4,432 clusters of co-expressed genes were generated using the stable k-means procedure. An average two-way linkage clustering (i.e. genes and samples) was also applied to each of these clusters (Figure 1D and Additional File 1). The resulting heatmaps (hierarchical clustering and k-means) and trees were generated with the Linux command line clustergram image creation utility Slcview http://slcview.sourceforge.net.

We defined an outlier as an aberrant gene expression value in one sample. We detected these aberrant values using a conventional method 69. Considering a particular gene and its vector of expression values across samples, we defined Q1 and Q3 to be the first and third quartile of this vector. We defined IQR to be the interquartile range (Q3 - Q1), and the interval [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR] was defined as the "normal" range of expression values. We thus considered outliers as gene expression values that lie outside this range. For a vector of normally distributed data, this threshold corresponds to approximately 5% of the data. For a particular sample, these aberrant expression values are often observed for numerous genes. K-means clustering is sensitive to such repeated aberrant values, leading to a high number of irrelevant gene signatures. To avoid such artifacts, we then identified for each gene signature, samples having gene expression values that are drastically different from the rest of the samples. In this context, samples with a proportion of outliers that exceeds 20% are rejected and the cluster is marked as "outlier".

The quality of a cluster (Figure 1D and Additional File 1) is determined by a test statistic based on the Pearson's product-moment coefficient. Considering a pair of gene expression profiles, we computed the p-value from the Student's t-statistic used to test the null hypothesis of positive correlation. The quality q of the cluster c was defined as the geometric mean of the p-values of all the gene pairs belonging to c. For large clusters (more than 100 genes), q was estimated using a resampling technique: the computation was performed on a subset of 10,000 gene pairs randomly extracted from the cluster. Note that for clusters marked as "outlier", their quality was estimated from the remainder of the data set, excluding the samples with a proportion of outliers >20%. Five quality classes were defined as a function of the cluster p-value. The quality of a study (Figure 1A and Additional File 1) was inferred as the mean quality of its clusters (see part 5 of the Additional File 3 for more details on the quality classes associated to the clusters and to the studies).

For each microarray platform, gene annotation was performed with the MADGene tool 70 (see also part 6 of the Additional File 3 for more details on this database). For each identified cluster of co-expressed genes, functional annotation (Figure 1D and Additional File 1) was performed using Gene Ontology (GO) 71 and GoMiner 72. Significance of over- or under-representation of GO terms was calculated using Fisher's exact test.

The user can upload his gene list as a tab-delimited text file. Thanks to the MADGene resource 70 (see also Additional File 2 and part 6 of the Additional File 3), the user need not specify the identifier types that are used in his own list. When a list is submitted, the user can select the studied species and then the tool displays the gene annotation that has been performed for confirmation.

MADMuscle statistically assesses the overlap between the input gene list and the lists relative to all the clusters of the database (see Additional File 2). The concordance between the two lists is calculated as the number of genes in common divided by the number of genes in the input list. Although a single gene is allowed to occur multiple times in the same list, it is counted/considered only once in the comparisons. When the lists derive from different species, a gene from the input list is considered if its homolog is found in the cluster gene list, as reported in the NCBI HomoloGene database. MADMuscle performs pair-wise comparisons (Fisher's test) and produces a summary table that reports the number of common genes between the paired lists along with the p-value. For each comparison, the names of the common genes can be retrieved.

The objective was to identify common transcriptional programs of muscle development across a wide range of microarray data sets. Therefore, we systematically compared each cluster of the database to all others, using the statistical test presented in the section "Cluster comparison". Instead of one input gene list, we successively introduced all the gene lists contained in the cluster database. This pairwise comparison of clusters is illustrated in Figure 2. Each of the 1,247 "good" labeled clusters of the database was systematically compared to all others using the Fischer exact test to detect significant similarity (strong overlap) or dissimilarity (poor or no overlap). The resulting p-values were computed to generate the diagonally symmetric similarity matrix for the 1,247 clusters (Figure 2A).

Compiling large numbers of microarray studies in a single database simplifies the analysis as it integrates many conventional assays into a single genome-wide expression profile. Thus the results from different studies can be compared directly 23. However, microarray studies often generate gene signatures consisting of hundreds of genes, making it difficult to distinguish which gene expression features are critical. For instance, in some data sets, a gene may show little variation and may have no co- expressed genes, whereas in other data sets the same gene may be part of a robust cluster of co-expressed genes 22. In this type of situation, it is useful to compare the results of different studies - by means of meta-analysis - to determine which results are most robust and most consistent across a range of studies.

The result of our analysis is a global map showing the clusters that are shared under a wide variety of physiological conditions (Figure 2A). The classification of this map identified 6 major, distinct and wide meta-clusters (M1 to M6). Each of them contained at least 15 similar clusters from various studies, microarray platforms or species. Many other small meta-clusters, gathering less than 10 clusters, could also be identified but will not be discussed in the present study. Whereas M1 gathered 16 clusters, 79 clusters were found in M2, 25 in M3, 46 in M4, 59 in M5 and 19 in M6.

It is likely that conserved co-expression patterns between studies have a functional relationship 22. A sufficiently large and diverse set of profiles, obtained from various conditions, results in a relatively comprehensive identification of co-expressed transcript groups. This allows additional hypotheses to be drawn regarding the functions of genes based on the regulatory characteristics of their transcripts. Such an approach better reflects specific biological processes 555657. To assess the reliability of these meta-clusters, we evaluated their functional annotation by examining the overlap of GO terms for each cluster of co-expressed genes (Figure 2B). By taking into account both enrichment and p-values, we clearly showed that each of the identified meta-clusters is associated with specific GO terms. The functional annotation of the meta-clusters is discussed in part 9 of the Additional File 3.

This study enabled us to identify a transcriptional landscape of skeletal muscle. It correctly infers a number of known fundamental biological processes within a skeletal muscle context. We found that a substantial number of correlated expression patterns occur in multiple independent data sets (Meta-clusters). Since tight correlation may imply common regulation 75 the identification of the precise role of transcription factors (e.g. ESRRA) in the coordination of gene expression patterns will be followed by studies on muscle pathology as was recently done for other tissues 76.

To illustrate the capacity of MADMuscle to perform such comparisons, we chose two clusters (Figure 1C and 1D) from the re-analyzed "GSE1004_GPL91" data set 8, which discriminate normal muscle from muscle affected by Duchenne Muscular Dystrophy (DMD). DMD is a severe genetic myopathy caused by the lack of the sarcolemmal protein dystrophin, and is clinically characterized by progressive and irreversible degeneration of muscle tissue. The first cluster (cluster 1, DMD+) contains 1,107 genes over-expressed in DMD muscle and involved in extracellular cell-matrix adhesion and the inflammatory-immune response (Figure 1D). The second cluster (cluster 5, DMD-) corresponds to 896 genes under-expressed in DMD muscle and involved in translation or coding for mitochondrion proteins (Figure 1D). Our results confirm the authors' initial observations 8 that many of the differentially expressed genes reflect changes in infiltration by inflammatory cells and connective tissue. Taken together, these two clusters recapitulate the major pathological feature of DMD which is abnormal connective tissue proliferation following myofiber degeneration.

These histological observations, although dramatically exacerbated in the DMD muscle, should also be present in other pathological or physiological conditions. For instance, skeletal muscle fibrosis, along with connective tissue proliferation, is a major pathological hallmark of chronic myopathies. In these pathologies, myofibers are replaced by progressive deposition of collagen and other extracellular matrix proteins produced by muscle fibroblasts. It is thus reasonable to argue that the coordinated gene expression patterns that reflect these histological changes will be conserved across studies. Comparing genomic expression profiles across species can reveal evolutionary conserved mechanisms, as illustrated in McCarrol et al. 36. Based on the MADMuscle meta-analysis tool, the 50 best hits for each of the two input gene lists (cluster 1 and cluster 5) were identified and analyzed (see Figure 3). These 100 clusters differ largely in their gene composition: 50 clusters display significant resemblance with the DMD+ cluster 1, while the 50 others contain significantly more genes observed in the DMD- cluster 5 (Figure 3A). Among the 100 top clusters showing significant overlaps with either cluster 1 or cluster 5, we identified 5 distinct species (Homo sapiens Hs, Mus musculus Mm, Rattus Norvegicus Rn, Canis familiaris Cf and Gallus gallus Gg), 18 distinct microarray platforms and various physiological conditions (including other muscle pathologies). This result underlines the capacity of the tool to highlight co-expressed genes across studies. We chose some typical examples among the 100 output results to illustrate this feature (Figure 3B). We found for instance that the two DMD gene signatures, namely DMD+ (cluster1) and DMD- (cluster 5), strongly resemble their counterparts (respectively II and VI: GSE466_GPL81) in the 16-wk-old mouse mdx muscle - the animal model of DMD - in spite of real discrepancies pointed out in the study 77. Additional information can be found in part 10 of the Additional File 3.

Finally, MADMuscle allows the identification of genes that change repeatedly in different studies, even when the studies are on different species or microarray platforms. The deregulation of these genes is caused by similar histological changes in the studied tissues. Reinforcing this idea is the fact that the GO terms (Figure 3C), supported by each of the 100 resulting clusters, fit very well with the direct functional annotation deduced from the two signatures DMD+ and DMD-. The GO terms associated with the 50 clusters resembling the DMD+ signature converged towards the extracellular matrix and the immune response; those associated with the 50 clusters resembling the DMD- signature identified muscle and mitochondrion markers. It is remarkable that functional annotation of the two signatures could thus have been deduced a posteriori from the meta-analysis. This may simplify the interpretation of lists of genes with altered expression, a critical and time-consuming part of microarray research.

We used our own transcriptional analysis of muscle (fascia lata tensor or paravertebral) affected by DMD. This study is presented in part 12 of the Additional File 3. We obtained two lists of differential genes (Figure 4A): 483 genes were over-expressed (gene list "+") in the DMD muscle, 473 genes showed a clear down-regulation (gene list "-"). The two gene lists (+: up and -: down) from our study (Figure 4A) were computed using MADMuscle. Because of the conserved co-expression across studies 22, these two gene lists enabled us to retain clusters of co-expressed genes from 5 distinct independent GEO series related to DMD studies (Figure 4B) (GSE465 78, GSE1004 8, GSE1007 79, GSE3307 80, and GSE6011 81),

The resulting clusters from each study were merged to create a meta-matrix (Figure 4B) in which genes differentially expressed between DMD and normal muscles were identified. The p-values can be calculated in each individual study and then combined, yielding an overall estimate of gene significance 23. Another approach is to apply p-value combination only after the construction of a meta-profile, defined as the trimmed median expression profile of all the equally annotated features 82. The combination of results from different studies partially solves the problem of a small sample number (inherent to microarray experiments) and thus helps to detect the truly differentially expressed genes. We therefore applied a Student's t-Test (p-value <0.01) on each gene expression meta-profile to identify significant variations between the control meta-group and the DMD meta-group. Reliably deregulated genes (meta-clusters of either up-regulated or down-regulated in DMD muscle) as well as invariant genes were identified.

The results from the meta-matrix were confronted to those from our study (Figure 4C). As expected from the small number of samples explored in our data set, part of the results could not be validated (false positives). Actually, some DEGs were not present (NP) in MADMuscle clusters, indicating that they did not significantly vary in any data set (+/NP: 87 genes; -/NP: 113 genes). Other DEGs were not significant in most MADMuscle clusters (+/NS: 155 genes; -/NS: 179 genes). Some DEGs even varied significantly in the opposite sense to that observed in our study (+/-: 39 genes; -/+: 72 genes). These discrepancies in the results could reflect a muscle-type effect since in our study we used paravertebral and fascia lata tensor muscles whereas in the public data sets, mainly quadriceps muscle biopsies were investigated. On the contrary, 202 DEGs were found to be significantly up-regulated both in MADMuscle clusters and in our study (+/+ group) while 109 DEGs showed a clear down-regulation (-/- group).

Finally, to assess the differences between these two complete gene lists (+/+ and -/-), we analyzed their expression patterns, given by the 1,247 "good" labeled clusters of the database, with a Multidimensional scaling (MDS) 83. This approach is particularly pertinent for the visualization of the similarities and differences observed in the data (Figure 4D and part 13 of the Additional File 3). The results, represented as a 2 D plot, show a clear separation between DMD+ (red) and DMD- (green) genes, achieved along one of the two components of the plot. One can note that DMD+ genes spread more widely than DMD- genes. This is probably due to the fact that in the DMD muscle, most of the up-regulated genes correspond to different invading cell types. Interestingly, genes from these two groups are also preferentially associated with specific GO terms (e.g. "mitochondrion" for DMD- genes; "extracellular region" and "cell adhesion" for DMD+ genes).

These observations are in agreement with the conclusions from previous studies on DMD muscle. In addition, among the perturbed biological functions, new interesting functions were identified by this meta-analysis and will be the scope of further studies. For instance, among the DMD+ genes, we clearly found an over-representation of GO terms associated with muscle (GO:0007517) and neuron (GO:0022008; GO:0048699) development. This could be explained by the following situation: while the satellite cell pool is quickly exhausted by repeated cycles of degeneration and regeneration 84, other resident muscle cell populations 85 may also contribute to muscle fiber regeneration, along with reinnervation 8687 in the dystrophic muscle. These findings are now under investigation to clarify the precise role of these biomarkers in the context of the pathology.