Coexpression of linked gene pairs persists long after their separation.

In many eukaryotes, physically linked gene pairs tend to be coexpressed. However, it is still controversial to what extent this neighbor coexpression is maintained by selection and to what extent it is nonselective, purely mechanistic "leaky expression." Here, we analyze expression patterns of gene pairs that have lost their linkage in the evolution of Saccharomyces cerevisiae since its last common ancestor with Kluyveromyces waltii or that were never linked in the S. cerevisiae lineage but became neighbors in a related yeast. We demonstrate that coexpression of many linked genes is retained long after their separation and is thus likely to be functionally important. In addition, unlinked gene pairs that recently became neighbors in other yeast species tend to be coexpressed in S. cerevisiae. This suggests that natural selection often favors chromosomal rearrangements in which coexpressed genes become neighbors. Contrary to previous suggestions, selectively favorable coexpression appears not to be restricted to bidirectional promoters.

Introduction

A gene's expression pattern is influenced by its genomic location, both in prokaryotes and eukaryotes. In prokaryotes, neighboring genes often form operons, resulting in tight coexpression of neighboring genes. In eukaryotes, physically linked gene pairs also show higher coexpression than randomly chosen gene pairs (Cohen et al. 2000; Kruglyak and Tang 2000; Lercher et al. 2002, 2003; Williams and Hurst 2002; Hurst et al. 2004; Singer et al. 2005; Lercher and Hurst 2006; Semon and Duret 2006; Batada et al. 2007; Kensche et al. 2008). For example, in yeast, adjacent gene pairs show correlated expression regardless of their relative orientation (Cohen et al. 2000; Kruglyak and Tang 2000), and this coexpression relationship spans up to 30 neighboring genes (Lercher and Hurst 2006). In the worm Caenorhabditis elegans, many coexpressed genes are organized into operons (Lercher et al. 2003). In the mouse genome, both immune system genes and tissue-specific genes are found to be expressed in clusters (Williams and Hurst 2002). In the human genome, housekeeping genes also show strong clustering (Lercher et al. 2002). Based on their apparent evolutionary conservation, it has been proposed that such coexpression clusters are selectively favorable in mammals (Singer et al. 2005). However, a later report found that highly coexpressed gene pairs are more likely to be broken up by rearrangements, concluding that neighbor coexpression is in fact generally disadvantageous in mammals (Liao and Zhang 2008).

The coexpression of neighboring genes in prokaryotic operons is conceptually simple. In eukaryotes, a range of mechanisms has been proposed to be responsible for the coexpression of closely spaced genes. Coexpressed neighboring genes in divergently transcribed orientation suggest that bidirectionally active promoters play a role in regulating coexpression (Cohen et al. 2000; Kruglyak and Tang 2000; Kensche et al. 2008), although such “bipromoters” may also serve to reduce stochastic gene expression noise (Wang et al. 2011). Chromatin structure also likely has an impact on the coexpression of closely located genes (Hurst et al. 2004; Batada et al. 2007; Chen et al. 2010). Finally, gene pairs that share the same transcription factors or that may be prone to a failure of transcription termination (“transcriptional read-through”) were also reported to be responsible for coexpression of neighboring genes (Semon and Duret 2006; Batada et al. 2007; Michalak 2008). Neighboring genes with similar functions have lead to the proposal that the coexpression of linked genes may be related to gene function (Cohen et al. 2000; Michalak 2008).

Thus, neighboring eukaryotic genes tend to be coexpressed. But is this coexpression really selectively favorable, or is it a nonselective, purely mechanistic by-product of genomic neighborhood, as suggested at least for mammals (Liao and Zhang 2008)? If neighbor coexpression is indeed functional, then coexpression should be maintained even if the neighborhood is broken up by genomic rearrangements. Here, we compare the effects of current and ancestral gene order on current gene expression patterns in the yeast Saccharomyces cerevisiae. In particular, we show that gene pairs that were genomic neighbors in the evolutionary past, but are separated now, show higher coexpression than randomly chosen gene pairs. As nonselective neighbor coexpression should seize after breaking up the neighborhood, our results indicate a significant role of natural selection in the coexpression of linked yeast genes.

Materials and Methods

Data Sources

The S. cerevisiae gene order as well as the ancestral gene order were taken from the yeast gene order database (Byrne and Wolfe 2005). We only retained genes with known positions in both data sets for further analysis.

Saccharomyces cerevisiae genome sequence data were downloaded via ftp from the Saccharomyces Genome Database (ftp://genome-ftp.stanford.edu/pub/yeast/). Ancestral gene order from eight reconstructed chromosomes was obtained from Gordon et al. (2009). Information on gain and loss of neighborhood along the yeast phylogeny was taken from Kensche et al. (2008).

The divergent gene pairs that share a promoter (bipromoter pairs) were taken from genome-wide tiling array experiments (Xu et al. 2009).

Expression Data

We employed the same expression data used by Batada, Urrutia, and Hurst to assess the influence of chromatin remodeling on the coexpression of neighboring yeast genes (Batada et al. 2007). Briefly, coexpression was averaged across 23 large-scale time course messenger RNA (mRNA) expression data sets, each covering at least 10 different time points. For each data set, Pearson's product moment correlation coefficient was calculated between the expression vectors of any two genes; for a given gene pair, coexpression was then defined as the mean value across data sets (for details, see supplementary information 1 of Batada et al. 2007).

Some of these data sets were obtained using cDNA microarrays, which may have spotted chromosomal neighbors onto neighboring microarray spots. It is hence possible that coexpression of genes neighboring in the current S. cerevisiae genome was overestimated due to experimental artifacts (Lercher and Hurst 2006). To address this issue, we repeated part of our analyses using coexpression derived from a set of 1,370 Affymetrix microarray experiments from the National Center for Biotechnology Information Gene Expression Omnibus (GEO) database (GEO accession IDs are listed in supplementary table S1, Supplementary Material online). We renormalized the log2-transformed expression values across all microarrays using the “aroma.light” package in BioConductor (Gentleman et al. 2004). Pairwise mRNA coexpression between two genes was again calculated as Pearson's product moment correlation coefficient across experiments. Results based solely on Affymetrix microarrays were qualitatively very similar to those presented in the main text (see supplementary results S1, Supplementary Material online).

Tandemly duplicated genes can lead to overestimation of the coexpression of neighboring genes. To avoid biases caused by tandem duplications, we removed all such pairs from our analyses. Tandem duplicates were identified as neighbors in the S. cerevisiae genome with Blast e value <0.01 (Batada et al. 2007).

Evolutionary Conservation of Bipromoter Gene Pairs

To test if bipromoter gene pairs are more conserved than other divergent gene pairs, we employed the reconstructed ancestral gene order published by the Wolfe lab (http://wolfe.gen.tcd.ie/ygob/). Only genes annotated in both the S. cerevisiae genome and the ancestor genome were used. Divergent gene pairs in the S. cerevisiae genome were marked as “ancestral” if they were direct chromosomal neighbors in the reconstructed ancestor and as “new” otherwise.

A simple logistic regression model was utilized to determine if bipromoter gene pairs are still more conserved than non-bipromoter pairs after controlling for coexpression level and intergenic distance, which is the strongest known predictor for linkage breakup (Poyatos and Hurst 2007). We used the following model:with

z: 1 = ancestral pair, 0 = new pair;

orient: 0 = bipromoter divergent pair, 1 = non-bipromoter divergent pair;

coexpr: coexpression level;

igd: intergenic distance.

Intergenic distance was measured as the distance in base pairs between the transcription start sites of genes. Calculations were performed in the R environment for statistical computing. The model shows that after controlling for coexpression and intergenic distance, bipromoter status still has a highly significant effect on gene pair conservation (P = 2.9 × 10−7); the effect of orientation was indeed much stronger than the influence of the other two variables (data not shown).

Dollo Parsimony Method to Calculate Ancestral States

Phylogenetic relationships of 19 yeasts and the neighboring gene pairs of orthologs in these fungi were downloaded from the supplementary file of Kensche et al. (2008). We used the Dollo parsimony method implemented in PAUP* (Wilgenbusch and Swofford 2003) to calculate the ancestral neighborhood state of each gene pair. This algorithm provided us with gain and loss information of gene neighborhood relationships at each internal node.

Correlation between Age of Separation and Coexpression

To test if coexpression is higher for ancestrally linked gene pairs that were separated more recently, we also used the information on ancestral gene order and separation ages derived from Kensche et al. (2008) as described above. We did not find a significant correlation between separation age and coexpression level (Pearson's product moment correlation coefficient: R = 0.087, P = 0.14; Spearman’s rank correlation coefficient: ρ = −0.044, P = 0.46).

Neighborhood Conservation Predicts Coexpression only for Divergent Gene Pairs

We used the recently reconstructed gene order of the pre–whole-genome duplication yeast ancestor (Gordon et al. 2009), which is believed to be about 100–150 My old (Sugino and Innan 2005). We first compared the coexpression of gene pairs that are conserved between the ancestor and S. cerevisiae with the coexpression of gene pairs newly formed in S. cerevisiae. Here, coexpression of two genes is defined as the correlation of gene expression values across a large data set of time series experiments (Kafri et al. 2005; Batada et al. 2007). To avoid potential biases caused by tandemly duplicated genes, such gene pairs were removed prior to all analyses.

Three possible scenarios exist: 1) If the conserved gene pairs are less likely to be coexpressed compared with newly formed gene pairs, then highly coexpressed neighboring gene pairs may be generally disadvantageous, as was observed recently in mammals (Liao and Zhang 2008). 2) If conserved gene pairs share similar coexpression profiles with newly formed gene pairs, then neighbor coexpression is likely to be largely selectively neutral. 3) If the conserved gene pairs generally show higher coexpression levels compared with newly formed gene pairs, then this indicates that neighbor coexpression is generally advantageous, as previously suggested (Singer et al. 2005).

Table 1 shows the results of this comparison. For divergently oriented S. cerevisiae gene pairs (←→), those that were already in this orientation in the ancestral genome show higher coexpression compared with newly formed divergent gene pairs. No such difference between conserved and new pairs was found for convergent or cooriented gene pairs. This indicates that in yeast, only divergent gene pairs are under selection for high coexpression. Surprisingly, there is no difference between the coexpression of newly formed divergent gene pairs and convergent gene pairs (P = 0.59 comparing new divergent gene pairs with conserved convergent gene pairs and P = 0.59 comparing new divergent gene pairs with newly formed convergent gene pairs, Brunner–Munzel tests). Thus, divergent gene pairs do not always show higher coexpression compared with other types of adjacent gene pairs in yeast. These results are not a consequence of variation in intergenic distance, which is known to be the strongest predictor of gene neighborhood conservation in yeast (Poyatos and Hurst 2007): the effect of neighborhood conservation status remains qualitatively unchanged when using both conservation status and intergenic distances as predictors in a logistic regression model (table 1).

Table 1

Only Divergent Gene Pairs Show Higher Coexpression in Ancient Compared with New Neighbors (P Values from Brunner–Munzel Tests)

	Neighbors in Ancestor and Saccharomyces cerevisiae			Neighbors in S. cerevisiae only
	Mean	N	Distance (bp)	Mean	N	Distance (bp)	Pa	Pb
Divergent (←→)	0.14	738	372	0.11	502	682	0.00064	0.010
Convergent (→←)	0.11	708	3,558	0.11	561	3,709	0.95	0.61
Cooriented (→→)	0.084	1,140	2,085	0.084	1,090	2,226	0.85	0.78

P value from Brunner–Munzel test.

P value from a logistic regression model to test if coexpression still has an impact on gene pair conservation after controlling for intergenic distance (details analogous to the model described for bipromoter pairs in Materials and Methods).

The observed difference between ancestral and young divergent gene pairs is likely related to the activity of bidirectionally active promoters (bipromoters) (Kruglyak and Tang 2000). Gene pairs newly formed by rearrangements will rarely be controlled by bipromoters. We identified gene pairs regulated by bipromoters based on published data (Xu et al. 2009). We hypothesized that strong coexpression of divergent gene pairs is found predominantly for bipromoter pairs. Consistent with this prediction, we found that coexpression of bipromoter divergent pairs is significantly stronger than for non-bipromoter divergent pairs (0.260 ± 0.009 vs. 0.180 ± 0.0104; P = 6.0 × 10−9, Brunner–Munzel test). Furthermore, we found that 469 out of 660 bipromoter gene pairs (71.1%) were already present in the ancestral genome, whereas the same is true for only 48.9% of the non-bipromoter divergent gene pairs. This result is again not an artifact of intergenic distance (Poyatos and Hurst 2007) or of the coexpression level of these genes (P = 2.9 × 10−7 in a logistic regression model, see Materials and Methods for details). More importantly, there is no difference between the coexpression level of conserved bipromoter gene pairs and new bipromoter gene pairs (0.261 ± 0.011 vs. 0.256 ± 0.178; P = 0.94, Brunner–Munzel test).

These results have two important implications. On one hand, they suggest that coexpression per se cannot explain the conservation of bipromoter structures. On the other hand, the results indicate that there is a selective advantage for the retention of bipromoter structures. Besides coexpression, another conserved function of bipromoters could be to reduce transcriptional noise (Wang et al. 2011).

Gene Pairs that Used to be Neighbors Are Still Coexpressed

We next analyzed the 2,765 ancestrally neighboring gene pairs that are located on different chromosomes in the current S. cerevisiae genome. On average, these separated pairs are significantly more coexpressed compared with 10,000 randomly chosen gene pairs (P = 5.3 × 10−6, Wilcoxon rank sum test). Except for shared cis-regulatory sites, none of the proposed mechanistic reasons for neighbor coexpression appear capable of explaining the persistence of coexpression after separation. Thus, coexpression is likely selectively favorable for these ancestrally linked gene pairs and was thus kept or restored after their separation.

We do not find significant differences between the coexpression of gene pairs that were ancestrally linked in different relative orientations (P = 0.084 between divergent and convergent pairs, P = 0.13 between divergent and cooriented gene pairs, and P = 0.70 between cooriented and convergent pairs; Brunner–Munzel tests). Of the known factors that specifically affect the coexpression of genomic neighbors, only chromatin remodeling acts independently of gene orientation. That we find no differences between orientations thus appears consistent with the idea that ancestral coexpression of these pairs was caused by local chromatin remodeling and is now maintained by shared cis-regulatory sequences; these cis-regulatory sequences may affect transcription factor binding as well as local chromatin remodeling. This maintenance further suggests that low-level coexpression caused by chromatin remodeling may in many cases be selectively favorable.

The results presented above suggest that at least part of the high coexpression level of neighboring yeast gene pairs is due to natural selection on coexpression. Thus, genes still need to be coexpressed when pairs are separated through a genomic rearrangement. To further verify this hypothesis, we used recent data based on gene pair conservation across 19 fungi (Kensche et al. 2008). We reconstructed the gene order in the common ancestor of these species using Dollo parsimony as implemented in PAUP* (Wilgenbusch and Swofford 2003).

We only analyzed genes that were direct neighbors in the ancestral genome but that are now located on different chromosomes because genes located nearby on a yeast chromosome still show similar expression profiles even when separated by tens of genes (Lercher and Hurst 2006).

As already observed in our first data set, separated gene pairs show slightly higher coexpression compared with random gene pairs (fig. 1; P = 0.00020, Brunner–Munzel test). Again, there is no difference between the coexpression of divergent, convergent, and cooriented ancestral gene pairs after their separation (P = 0.18 between divergent and convergent pairs, P = 0.32 between divergent and cooriented gene pairs, and P = 0.76 between cooriented and convergent pairs; Brunner–Munzel tests).

F

Gene pairs located on different Saccharomyces cerevisiae chromosomes but neighboring in the ancestral genome (red) or in another yeast lineage (green) show slightly higher coexpression than random gene pairs (black), although coexpression is lower than in gene pairs that are neighbors in the current S. cerevisiae genome (blue). Average coexpression is significantly higher than random expectations for all three types of neighbors (P ≤ 0.0002 in each comparison).

If gene neighborhood is under positive selection for genes that need to be coexpressed, then we would further expect that orthologs of non-neighboring coexpressed S. cerevisiae genes are more likely to become genomic neighbors in other yeast lineages; consequently, genes neighboring in at least one other yeast species but located on different chromosomes in both S. cerevisiae and in the common ancestor should show higher coexpression than random gene pairs. This is indeed the case (fig. 1; P = 3.3 × 10−5, Brunner–Munzel test).

Discussion

Using ancestral gene order information gained from the yeast gene order browser (Byrne and Wolfe 2005), we confirmed that among neighboring gene pairs, divergently oriented pairs are the ones that were most likely to be conserved during genome evolution (Kensche et al. 2008). More specifically, this conservation is mostly due to bipromoter gene pairs. The conservation implicates stabilizing selection on the relative positioning of this subset of the divergently arranged gene pairs.

After separation of neighboring gene pairs through genomic rearrangements, we no longer found any difference between divergent and convergent or cooriented gene pairs; all three types of ancestrally neighboring gene pairs show higher than expected coexpression in S. cerevisiae after their separation through genomic rearrangements. It is possible that the two genes in these coexpressed separated pairs had part of their cis-regulatory apparatus in common even before their separation, so that coexpression could be partially maintained after the rearrangement; conversely, it may be that coexpression was initially lost in the rearrangement and was reinstated through cis-regulatory changes afterward.

The coexpression of ancestrally neighboring gene pairs that are now located on different chromosomes is sharply reduced compared with the pairs that are neighbors in the current yeast genome (fig. 1). This observation is expected, as factors such as chromatin remodeling are known to strongly influence the coexpression of linked genes in yeast (Batada et al. 2007). Thus, although part of the neighbor coexpression is likely maintained by natural selection, it is likely that a substantial component of neighbor coexpression is nonselective “leaky” expression of one or both neighbors.

Could it be that all coexpression is in fact nonselective (purely mechanistic), and separated pairs only show coexpression because part of the mechanistic apparatus shared between the two genes is maintained through the separation? In particular, many of the linkage losses may be a consequence of the whole-genome duplication experienced in the S. cerevisiae lineage after the common ancestor of the yeasts analyzed here. Assume that the common ancestor contained a neighboring gene pair A,B together with a cis-regulatory region c that affects both genes (c-A-B). The whole-genome duplication will duplicate the complete set, resulting in c1-A1-B1 and c2-A2-B2. If subsequently A1 and B2 (or A2 and B1) are lost, the now separated genes A and B retain their identical cis-regulatory region c.

However, it is unlikely that such a scenario explains our observations, for at least two reasons. First, if selection would play no role in the maintenance of coexpression, then coexpression should fade with increasing age of the separation. This is not the case (Spearman's ρ = −0.044, P = 0.46; for details, see Materials and Methods). Second, and most importantly, a nonselective, purely mechanistic model cannot explain why unlinked gene pairs that only recently became neighbors in other yeast species are coexpressed in S. cerevisiae. That these pairs show coexpression very similar to ancestrally linked pairs (fig. 1) seems only compatible with the hypothesis that natural selection promotes chromosomal rearrangements that bring together coexpressed genes.

When discussing the properties of neighboring gene pairs, these are usually classified by their relative orientation into three categories—divergent gene pairs (head to head), convergent gene pairs (tail to tail), and cooriented gene pairs. Those three types of gene pairs appear to have different properties—divergent gene pairs are the most conserved and show stronger coexpression than the other two orientations (Kensche et al. 2008). Here we show that as far as coexpression is concerned, there are essentially only two types of neighboring gene pairs in the genome—bipromoter gene pairs and non-bipromoter gene pairs. Bipromoter gene pairs show strong signals of conservation and coexpression, whereas non-bipromoter gene pairs do not. After separation through genomic rearrangements, ancestral divergent gene pairs no longer exhibit higher coexpression compared with other ancestral gene pairs, supporting the view that chromatin remodeling dominates the coexpression of most neighboring gene pairs (Batada et al. 2007).

In conclusion, we have shown that not only gene neighborhood in the current S. cerevisiae genome but also gene order in the ancestral genome and gene order in related yeasts are predictive of coexpression. These results support a role for natural selection in the establishment and maintenance of neighbor coexpression in yeast and argues against a purely mechanistic view that considers neighbor coexpression as a neutral (or even slightly deleterious) phenomenon.

Supplementary Material

Supplementary results, table S1, and figure S1 are available at Genome Biology and Evolution online (http://www.gbe.oxfordjournals.org/).