The non-random arrangement of genes in the genome is intimately connected to the pattern of gene expression across the genome 1. While the connection between gene location and expression has been known for some time in prokaryotes 2, a similar genome-wide connection between gene order and gene expression has relatively recently been identified in eukaryotes 3. Clusters of physically adjacent genes that are co-expressed are now known to be common in eukaryotic genomes and have been reported in yeast 45, plants 6, worms 78910, fruit flies 111213, mice 141516, and humans 1617181920.

A number of mechanisms have been proposed to explain the existence of these co-expression clusters including the presence of duplicate genes that are in close physical proximity, shared regulatory regions, chromatin-level regulation, and common pathway or tissue regulated expression of physically clustered genes 1821. Similarly, a number of hypotheses have been proposed concerning the interplay between transcriptome evolution and genome organization that can explain the existence of co-expression clusters including positive selection for genomic rearrangements leading to close physical proximity of co-expressed genes 22 and purifying selection against genomic rearrangments that break-up co-expression clusters 518. The possibility that co-expression results in correlated rates of sequence evolution among cluster genes has also been proposed 23 and a recent analysis has found evidence of co-evolution of tissue specific expression of adjacent genes 24. Although not fully resolved, these analyses have led to a clearer picture of both the pattern of co-expression clusters within species and explanations concerning why gene expression is often coordinated among physically adjacent genes.

If there is a physical clustering of coordinated gene expression within species, then it is likely that gene expression can also evolve in a coordinated manner 24. A block of consecutive genes where expression evolves in a coordinated manner will leave an evolutionary signature that can be detected by non-zero expression correlation among neighboring genes when analyzing multiple species. Therefore, just as correlated expression profiles are used to identify co-expression among genes within species 412 the same approaches can be used to analyze co-evolution of expression in gene clusters when comparing gene expression among species. Instead of analyzing the correlations in gene expression among developmental stages, environments, tissue localizations, etc. 1, we consider the correlations among mean gene expression levels estimated for each species. This approach will identify clusters of genes where the evolution of expression is not a gene independent process.

Our goal is to identify clusters of genes that show consistent patterns of coordinated expression evolution among species of Drosophila. We assayed genome-wide gene expression levels using Affymetrix GeneChip Arrays in three-day-old male adults under standardized environmental conditions 25. The following seven species in the D. melanogaster subgroup were analyzed: D. melanogaster, D. simulans, D. sechellia, D. mauritiana, D. santomea, D. teissieri, and D. yakuba. The relationships among these species are well established (Fig. 1) and represents a taxonomic sampling that spans ~6 million years 26. Clusters identified for these species are therefore of value for understanding how the transcriptomes of species evolve across time scales on the order of one-hundred thousand to several million years.

A number of methods have been applied to the identification of co-expression clusters within species using microarray expression data 41223272829. The most common of these is calculation of a statistic based on the estimated correlation matrix for blocks of consecutive genes, generally the mean of N*(N-1)/2 correlations when considering N consecutive genes 12. For the current study, we use a different statistic to summarize the correlations among N genes: the fraction of variation explained by the leading eigenvalue of the correlation matrix. This statistic describes the maximum fraction of variation that can be explained by a linear function of the original variables after scaling the variance of each variable to one. This statistic therefore provides an intuitive description of the degree to which a set of genes act as a single unit because the closer this ratio is to "1" the greater the degree that expression of the genes are completely correlated, regardless of whether the correlations among any gene pair are positive or negative. While this statistic has not been explicitly applied to the analysis of co-expression, the leading eigenvalue(s) of a correlation or covariance matrix are commonly used to summarize the structure of correlated variation 3031.

Our analysis shows that a considerable proportion of transcriptome evolution among species in the D. melanogaster subgroup occurs via co-evolution of expression in clustered genes. Comparison of the locations of clusters that reflect coordinated evolution of gene expression across taxa to clusters of coordinated expression within the species D. simulans demonstrated a lack of correspondence in locations. This implies that different mechanisms may be responsible for producing co-expression clusters within species and those producing co-expression clusters that evolve in a coordinated manner. We additionally analyze a number of genome organization, functional, and evolutionary aspects to identify over-(under-) representation with clusters displaying coordinated expression within D. simulans or coordinated expression evolution among the seven species. Of these, the most interesting are genes that show a signature of adaptive evolution in their coding sequences. A previous analysis of tissue co-expression within mice and humans did not find a significant positive correlation between rates of non-synonymous substitutions (K_A) and co-expressed genes 23, although the ~75 million years since humans and mice diverged likely limited the power of this analysis 32. Here, we find that co-expression clusters that vary within D. simulans are not enriched for adaptive evolving loci. However, genes with an adaptive evolutionary signature are over-represented in clusters where expression is co-evolving among species. This result points to a connection between coordinated gene expression evolution and adaptive evolution in coding regions of genes, although the exact nature of this connection is still unknown.

Clusters of genes evince coordinated evolution of expression across the seven species (Figure 2). On all chromosomes at all scales, there were far more windows with significant coordinated expression evolution than expected at random (Table 1). Many of these significant windows were identified across multiple window sizes and likely reflect a single larger block of genes where expression evolution is coordinated 12 (we present combined significant windows at different cutoffs in Additional files 1, 2, 3, 4, 5, 6, 7, 8, 9, 10). These windows were also robust to the removal of individual species and therefore were not being driven by evolution in a specific lineage (results presented in Additional files 1, 2, 3, 4, 5, 6, 7, 8, 9, 10). In addition, there was no detectable difference between the absolute level of transcript abundance as measured by the arrays for neighboring genes where expression displays co-evolution compared to other groups of neighboring genes (p-values > 0.05 for all window sizes). Because we used the D. melanogaster as the reference genome for ordering genes along chromosomes there is the possibility that genome re-arrangements would lead to some of these significant clusters to include non-physically adjacent genes in some of the species. Our results are, however, robust to this issue as we found that breakpoints between D. simulans-D. melanogaster or D. yakuba-D. melanogaster interrupted clusters where expression is co-evolving no more frequently than expected by chance, which is consistent with a previous analysis of co-expression within species 33.

The use of a sliding window approach to identify co-evolution of expression in gene clusters means that tests at a given scale will be correlated with neighboring windows and these tests will also be correlated across window sizes. There is not a clear optimal approach for dealing with the multiple testing problem and the properties of strategies such as estimation of False Discovery Rates (FDRs) 3435 are not clear in such cases. To assess whether there was a clear genome-wide tendency for coordinated expression evolution, we therefore used a permutation approach using total number of significant tests at a given window size (2, 5, 10, and 20) as a test statistic to assess the null hypothesis that there are no more gene clusters where expression is co-evolving than we would expect at random 12. With only seven species (samples) these tests are not expected to be particularly powerful. However, the test was still rejected at a window size of 2 (p-value < 0.02) and at a window size of 10 (p-value < 0.04) (although not for window sizes of 5 and 20) indicating that at least on genomic scales spanning 2 to 10 genes, there is genome level co-evolution of gene expression in neighboring genes.

Similar results were obtained for the analysis of within species co-expression for D. simulans (Table 2, Figure 3). Many windows on all chromosomes at all scales were significant. Interestingly, the test of a genome-wide pattern produced significant results for window sizes of 2 (p-value < 0.04) and 10 (p-value < 0.01). While this could be interpreted as an artifact of microaray design 3637, there is no regular spacing to the distribution of significant windows 28. Interestingly, there was little overlap between the significant windows identified as evolving across species and being co-expressed within D. simulans (Table 3). The number of overlapping windows obtained when comparing repeated analysis of mean expression levels for all species and the number of overlapping windows for repeated analysis of the D. simulans data (i.e. non-overlap due to permutation effects) are presented for comparison. Given that many of the co-evolving expression clusters and the co-expression clusters identified within D. simulans may reflect false positives, a small fraction of overlap between these cluster types might be expected. However, even at a conservative cutoff (p-value < 0.001) the absolute number of overlapping clusters is still very low (Table 3) indicating that there is little correspondence. It therefore appears that completely different sets of genes are involved in the pattern of co-expression within species compared to those where expression evolves in a coordinated manner across species.

While the mechanisms underlying the existence of clusters of co-evolution of expression among species and co-expression clusters within species cannot be resolved from these data, the existence of paralogous genes in close proximity can be ruled out as the major factor for the observed pattern. Paralogous genes in close proximity may be expected to produce the evolving clusters or within species co-expression clusters as a result of shared regulatory elements and/or maintained common functions. However, paralogs may also produce the pattern by cross-hybridizing to common probes on the microarray. If the second of these possibilities can explain a considerable proportion of clusters, this could mean the observed pattern was an artifact of the microarray assay. However, we find that very few paralogous gene sets are within either evolving clusters or co-expression clusters (between 1.4%–12.0% of evolving clusters identified using window sizes 2–20 and a p-value cutoff of 0.001 contain paralogous genes and 2.1%–8.5% of co-expression clusters within D. simulans contain paralogous genes; see Additional file 9). The majority of evolving clusters and co-expression clusters cannot therefore be explained by paralogous genes.