The gene expression datasets originate from four separate, previously-published skeletal muscle studies (one on mouse and three on humans) performed on the Affymetrix GeneChip platform (Table).

Dataset M (mouse, n = 36) was derived from six different skeletal muscle groups: diaphragm, extensor digitorum longus, gastrocnemius, quadricep, soleus, and tibialis anterior of eight-week-old male mice. The mice represented three genetic strains: wildtype (C57BL/10SnJ), mdx (C57BL/10ScSn-Dmd^mdx/J), and mdx^5cv (B6Ros.Cg-Dmd^mdx-5cv), with their transcriptome profiles assayed in duplicate on the Affymetrix U74Av2 platform [¹⁷,¹⁸].

Dataset H1 (human, n = 24) consists of transcriptome profiles of surgically-biopsied quadriceps from 12 control and 12 DMD subjects assayed on the Affymetrix U95Av2 platform [⁸,¹⁹].

Dataset H2 (human, n = 32) contains transcriptome profiles of four different human muscle groups: deltoid, gastrocnemius, quadriceps, and tibialis anterior obtained at autopsy from each of 8 individuals ? three pediatric and five geriatric ? assayed on the Affymetrix U133A platform [²⁰].

Dataset H3 (human, n = 24) contains transcriptome profiles of 12 distinct normal organ-tissues pooled from 10?25 individuals ? bone marrow, brain, heart, liver, lung, kidney, skeletal muscle, pancreas, prostate, spinal cord, spleen, and thymus ? assayed on the Affymetrix U95Av2 microarrays in duplicate [²¹,²²].

Curated and predicted National Center for Biotechnology Information (NCBI) Entrez-identified mouse-human orthologue gene pairs were obtained from NCBI HomoloGene, (data freeze October 07, 2005) [²³]. To ensure a one-to-one and well-defined mapping, cross-species matches were restricted to those with the highest reciprocal percentage sequence homology. For instance, the human gene ACAT2 (EntrezID 39) has two distinct mouse orthologues: Acat2 (87.12% sequence homology, EntrezID 110460) and Acat3 (85.10% sequence homology, EntrezID 224530). Acat2 was considered the human ACAT2's unique mouse-orthologue since it had the higher percentage homology with ACAT2.

The Affymetrix GeneChip platforms contain the following number of Entrez-identified genes (data freeze May 17, 2005): U74Av2, 8,894 mouse genes; U95Av2, 8,978 human genes; U133A, 12,963 human genes. Between the mouse U74Av2 and human U95Av2 platforms, there exist 5,306 mouse-human orthologue pairs. Similarly, mouse U74Av2 and human U133A platforms have 6,547 orthologue pairs.

The primary gene set for cross-species analyses are the 5,306 orthologue gene pairs between the mouse Affymetrix U74Av2 (dataset M) and human Affymetrix U95Av2 (dataset H1 and H3) platforms. Of these 5,306 genes, 5,288 (99.7%) are shared in common with the 6,547 U74Av2/U133A orthologue pairs, and were thus used in cross-species analyses between datasets M and H2 (Affymetrix U133A). Each sample gene expression profile is represented as algebraic N-vector signifying the reported expression intensities for N (>0) distinct genes measured in that sample.

Principal component analysis (PCA) was used to obtain the subsets of genes that were dominant contributors to the global sample variation in datasets M and H1, respectively [see Additional file 1] [see Additional file 2] [see Additional file 3] [^24-27]. We define the g_j's with |a_j| > 0.03 in any one of principal components 1?3 (PC1-3) following PCA of datasets M (271 genes) and H1 (234 genes) to be the dominant contributors to global sample variation for datasets M and H1 respectively. Eighty-six genes are common to both sets.

Linear (Pearson) correlation was used to quantify the linear similarity between mouse and human samples [²⁴]. Each mouse or human sample is a gene expression profile vector of length N, with the j^th mouse vector component being orthologous to the j^th human vector component. When calculating the mouse-human profile correlations at the whole transcriptome-scale, N = 5,306, the total number of orthologue gene pairs between the mouse U74Av2-human U95Av2 platforms; and N = 234 (respectively, N = 271 ? see preceding paragraph) when calculating the mouse-human profile correlations based on the subset of genes that are dominant contributors to global sample variance in dataset H1 (respectively, dataset M).

The non-parametric Wilcoxon (Mann-Whitney) rank-sum test [²⁸] was used to assess the probability that the distribution of linear correlations of human skeletal muscle to mouse soleus is similar to the distribution of linear correlations of human skeletal muscle to mouse non-soleus muscles. The underlying expression data distributions were non-Gaussian, thus a parametric test was not used.

Whole transcriptome scale (5,306-gene profile) similarities between skeletal muscle datasets M (mouse) and H1 (human) were investigated. Linear correlation was used as a measure of similarity. Two additional human transcriptome datasets were considered as in silico positive controls: (H2) transcriptome profiles from four skeletal muscle groups of eight human subjects from autopsies measured on the Affymetrix U133A platform [²⁰]; and (H3) transcriptome profiles from 12 distinct human tissues (two muscle, ten non-muscle) [²¹] measured on the same microarray platform as H1.

Linear correlations between 5,306-gene transcriptome profiles of each sample from dataset M to each sample from dataset H1 were computed. The human correlation group averages and standard deviations relative to each of the six mouse muscle groups ? three genetic strains are shown in Figure 1A?B. All 24 human quadriceps samples had a significantly higher correlation to mouse soleus than to non-soleus mouse muscles (p < 0.02, Wilcoxon), independent of mouse genetic background and human myopathic condition.

To determine whether the higher correlation of human quadriceps to mouse soleus could be reproduced with other human muscles, similar correlation analyses were performed with human datasets H2 and H3 relative to dataset M. Cross-species linear correlation analyses of the 5,288-gene pairs selected between datasets M and H2 showed a significantly higher correlation in 28 of the 32 human samples to mouse soleus (p < 0.02, Wilcoxon) than to non-soleus mouse muscles (Figure 1C?D).

To determine whether the higher correlation of human skeletal muscles to mouse soleus is specific to muscle tissue, similar correlation analyses were performed with the 5,306-gene transcriptome profiles of human tissue not primarily composed of muscle (bone marrow, brain, liver, lung, kidney, pancreas, prostate, spinal cord, spleen and thymus) in dataset H3 relative to dataset M. The non-muscle human profiles did not return a significantly higher correlation to any mouse muscle, whereas the pooled human heart and skeletal muscle profiles in dataset H3 showed a significantly higher correlation to the mouse soleus than to non-soleus mouse muscles (p < 0.02, Wilcoxon) (Figures 1A, 1C). Indeed, non-muscle human tissue-to-mouse muscle correlations were significantly lower than human muscle-to-mouse muscle correlations.

We identified candidate subsets of genes (by their ontologic category or contribution to sample variation in silico) that might underwrite the transcriptomic resemblance of human muscle to mouse soleus. Correlation analyses of mouse-human sample profiles were performed as above, relative to four subsets of the 5,306-gene whole transcriptome separately: ST1, the 733-gene sub-transcriptome (13.8% of the whole transcriptome) recorded by the Gene Ontology Consortium [²⁹] to be integral to cell plasma membrane ? note that dystrophin and other sarcolemma-related structural proteins involved in myopathies belong to this ontologic category; ST2, the 271-gene sub-transcriptome (5.1% of the whole transcriptome) of dominant contributors to global sample variance in dataset M; ST3, the 234-gene sub-transcriptome (4.4% of the whole transcriptome) of dominant contributors to global sample variance in dataset H1; and ST4, the 4,188-gene sub-transcriptome complement to ST1-3 as a negative in silico control ? these are genes neither integral to the cell plasma membrane nor dominant contributors to global sample variance in the muscle datasets M and H1. Twenty-one genes in ST2, and 16 in ST3, belong to the cell plasma membrane category/ST1.

Using ST1-3 to characterize sample profiles, we again observed significantly higher correlations of human muscles in datasets H1-3 to mouse soleus than to non-soleus mouse muscles (p < 0.02, Wilcoxon) [see Additional file 4] [see Additional file 5] [see Additional file 6]. Similarly, correlations of non-muscle human tissue in dataset H3 to mouse dataset M are lower and do not show a significantly higher correlation to any mouse muscle group ? with one exception in the ST3 sample profile characterization where among non-muscle human tissue, brain and spinal cord had significantly higher correlation with mouse soleus than with non-soleus muscle profiles. In contrast, when ST4 was used to characterize sample profiles, all human muscle and non-muscle samples from dataset H1-3 have lower correlations to mouse muscle profiles in dataset M compared to their ST1-3 sample profile characterization, and human muscle profiles show no significant correlation to any mouse muscle group (Figure 2).

We investigated whether the myopathic state (specifically DMD) of a human muscle (specifically quadriceps in dataset H1) was reflected in its correlation against wildtype mouse skeletal muscle groups (dataset M).

Mouse-human correlation analyses were performed between datasets M and H1, relative to the above sample profile characterizations: whole transcriptome and sub-transcriptomes ST1-4. In all except the ST4 characterization, control human quadriceps to wildtype mouse muscle correlations were significantly higher than DMD human quadriceps to wildtype mouse muscle correlations (p < 0.02, Wilcoxon) (Figure 3).

The authors express their appreciation for support from the following sources: NIH P01 NS40828-01A1 (ATK, ISK, LMK), NIH R01 NS047527 (ATK), R21 NS052498 (ATK), NIH U01 HL066582-01 (ISK), NINDS K08 NS048180 (PBK), the Bernard F. and Alva B. Gimbel Foundation (LMK), and the Joshua Frase Foundation (LMK). LMK is an Investigator of the Howard Hughes Medical Institute.