For each group of orthologous genes global multiple alignments among the human, mouse, rat, and dog loci were performed using MLAGAN 25. We took into consideration all genes for which there were predicted othologs within Ensembl 31 in the mouse genome, human genome, and any third mammalian species, which led us to analyze 9,749 groups of orthologous genes (36% of the annotated mouse genes). Most genes (about 88%) were found to be conserved in all four species considered, with only about 12% found in three out of four species (about 6% in each triplet; Figure 1). For each locus we took into account the whole genomic repeat-masked sequence containing the transcriptional unit as well as the complete flanking sequences up to the preceding and following gene. This lead us to analyze 37% of the murine genome sequence overall. The alignments were parsed using VISTA (visualizing global DNA sequence alignments of arbitrary length) 32 searching for segments of minimum 100 base pairs (bp) length and 70% identity. We further selected these regions by only taking into account those regions that were found at least in mouse, human, and a third mammalian species and which overlapped by at least 50 bp, which resulted in a set of 364,358 rCNEs (Table 1). These were then filtered stringently to distinguish 'genic' from 'nongenic' (see Materials and methods, below). This analysis classified 22.7% of the resulting rCNEs as 'genic', while 281,644 nongenic elements account for about 46 megabases, or 1.77%, of the murine genome.

We further annotated mammalian rCNEs based on their position in the mouse genome with respect to the gene locus in order to define whether they were located before the annotated transcription start site (TSS; 'pre-gene'), within the intronic portion of the gene, or posterior to the transcriptional unit ('post-gene'). Approximately 54% of rCNEs were found to fall within intergenic regions, of which 37% were post-gene and 63% pre-gene (Table 1).

We searched for conservation of rCNEs in teleost genomes using CHAOS 24, selecting regions that presented at least 60% identity over a minimum length of 40 bp as compared with the mouse sequence of the rCNEs. This method allowed us to identify regions that are reversed or moved in the fish locus with respect to the corresponding mammalian locus. For each locus in every species analyzed we took into account the whole genomic repeat-masked sequence containing the transcriptional unit as well as the complete flanking sequences up to the preceding and following gene. We defined as SCEs those regions of the mouse genome that were conserved at least in the fugu orthologous locus and filtered out any sequence shorter than 20 bp as a result of the overlap analysis with zebrafish and tetraodon (see Materials and methods, below, for details). Our analysis identified 21,427 nonredundant nongenic SCEs, which were found in about 30% of the genes analyzed (2,911; Table 2). The distribution of their length and percentage identity is shown in Figure 2e,f. The median length and percentage identity (45 bp and 67%, respectively) reflect closely the cut offs provided to CHAOS in the alignment (40 bp and 60% identity), although there is a significant number of outliers whose length is equal to or greater than 200 bp (223 elements whose maximum length is 669 bp) and whose median percentage identity is 74%. No elements were identified that were completely identical to their mouse counterpart (the maximum percentage identity found was 97%).

We decided to investigate further the extent to which the elements identified, which are still retained within the locus analyzed, have shuffled in terms of relative position and orientation relative to the transcriptional unit, and would thus be missed by a simple regional global alignment (such as MLAGAN). The results of this revealed that only 28% of elements identified have retained the same orientation and the same position with respect to the transcriptional unit taken into account (that is to say, have remained pre-gene, intronic, or post-gene. Labeled as 'collinear'; Figure 2a), whereas others have shifted in terms of orientation ('reversed'; Figure 2b), position ('moved'; Figure 2c), or both ('moved-reversed'; Figure 2d). Thus, almost two-thirds of the SCEs identified would have been missed by a global, albeit regional, alignment approach.

A possible explanation for the large number of noncollinear elements is that they could appear shuffled owing to assembly artifacts. In order to assess whether the large number of elements identified as noncollinear were merely due to assembly artifacts, we analyzed the number of SCEs containing a single hit in fugu and not classified as collinear that also had a match in tetraodon. If the shuffling were merely due to assembly artifacts, then we would expect approximately half of the noncollinear hits in fugu also to be noncollinear in tetraodon. The results, however, were significantly different, because more than 80% of the elements were not collinear in both species (P < 2.2 × e^-16 obtained by performing a χ² comparison between the proportion obtained and the expected 0.5/0.5 proportion). These findings emphasize that shuffling is a mechanism of particular relevance when searching for short, well conserved elements across long evolutionary distances and that its true extent can only be detected by using a sensitive global-local alignment approach, as opposed to a fast genome-wide approach 25.

Two examples of SCEs that were identified in our study are shown in Figure 3. Example A shows the locus of Sema6d, a semaphorin gene that is located in the plasma membrane and is involved in cardiac morphogenesis. This locus represents a conserved element that is found after the transcriptional unit at the 3' end of the gene in all mammals analyzed, whereas it is located upstream in fish genomes and reversed in orientation in the fugu and tetraodon genomes. Example B shows the locus of the tyrosine phosphatase receptor type G protein, a candidate tumor suppressor gene, which has a conserved element in the first intron of all mammalian loci analyzed, which is found in reversed orientation in all fish genomes, downstream of the gene in the fugu and tetraodon genomes, and in the second intron in the zebrafish genome.

We analyzed the type of genes that are associated with SCEs by assessing the distribution of Gene Ontology (GO) terms 33 using GOstat 34 (see Materials and methods, below). Although the results indicate significant over-representation of gene classes typical of genes harboring noncoding conservation ('trans-dev' enrichment) as reported previously (Additional data file 1), the number of genes within our analysis containing nongenic SCEs (2,911) is approximately an order of magnitude greater than that of the number of genes containing CNEs (330). The overlap between the two datasets is 291 genes, and so almost all (>88%) genes containing SCEs also contain CNEs. A GO analysis comparing genes containing CNEs and those containing SCEs (Figure 4) revealed that there are several GO categories that are significantly under-represented in the CNE dataset as compared with ours. These categories were not seen in the previous analysis (Additional data file 1) because they are not over-represented in our dataset as compared with the entire genome.

The most striking difference is found in the analysis by cellular components; there is an approximate 54-fold enrichment in genes belonging to the extracellular regions that contain SCEs as compared with genes in the same class that contain CNEs. In fact SCEs are present in more than 50% of the genes we were able to classify as belonging to the extracellular matrix and in 35% of those belonging to the extracellular space, whereas CNEs are only found in six and two such genes, respectively. These gene sets differ significantly in both extracellular regions and membrane GO cellular component categories (P < 0.001; Additional data file 1). Enrichments in the order of 10-fold to 13-fold are seen when comparing genes involved in physiological and cellular processes, respectively. For both of these categories our analysis was able to identify SCEs in more than 30% of the genes belonging to this class. The differences, although substantial (about sevenfold) are not as extreme when comparing 'trans-dev' genes (genes categorized as belonging to the 'regulation of biological process' and 'development' using GO) because the CNE dataset has a stronger bias for those genes (P < 0.001; Additional data file 1). Finally, although we identified SCEs in 40% of genes assigned to the 'behavior' class, none of the genes in this class has CNEs. The data thus suggest that there are both quantitative and qualitative differences between the two datasets.

Because a large proportion of our dataset undergoes shuffling, we decided to investigate whether shuffling is a property that is dependent on proximity to the transcriptional unit. To address this question we divided our dataset of nongenic SCEs between collinear (as discussed above) and noncollinear (all other categories discussed above taken together) elements, and analyzed the distribution of their distances from the TSS (pre-gene set), the intron start (intron start), the intron end (intron-end set) and the 3' end of the transcript (post-gene). This analysis demonstrated that collinear elements were distributed significantly closer to the start and the end of the transcriptional unit compared with noncollinear elements, whereas no differences were observed in terms of proximity to the intron start and intron end (Additional data file 2).

In order to investigate this phenomenon at higher resolution, we subdivided all loci analyzed in our dataset into 1,000 bp windows within the areas, and verified whether the proportion of collinear versus noncollinear elements deviated significantly from the expected proportions in any of these windows (see Materials and methods, below, for details). The results of the analysis are shown in Figure 5. The only window that exhibited a high χ² result with significantly less shuffled elements than collinear ones (P = e^-08), was the 1,000 bp window immediately upstream of the TSS. No similar results were found in any other 1,000 bp windows across the gene loci analyzed. Similar results were obtained when deploying other window sizes (data not shown). To ascertain whether the result observed was due to annotation problems, we inspected the GO classification of the genes that presented nongenic collinear elements in the 1,000 bp window discussed above and observed significant enrichment (P < 0.001) for 'trans-dev' genes, whereas the same test conducted on genic collinear elements in the same window revealed no significant GO enrichment (Additional data file 3).

In order to verify the ability of SCEs to predict functional enhancer elements, we conducted an overlap analysis (see Materials and methods, below) of SCEs with 98 mouse enhancer elements deposited in Genbank. We compared the overlap of SCEs with that of two other datasets that present conservation in fish genomes, namely CNEs and UCEs. The results presented in Figure 6 show that although CNEs and UCEs are able to detect only one and two known enhancers from our dataset, respectively, SCEs detect 18 of them successfully.

In order to validate the cis-regulatory activity of SCEs we chose a subset of SCEs to be tested for in vivo enhancer activity by amplifying them from the fugu genome and co-injecting them in zebrafish embryos with a minimal promoter-reporter construct yielding transient transgenic zebrafish embryos. Twenty-seven SCEs were tested, of which four overlapped known mouse enhancers for which activity had not previously been reported in fish, and the remaining 23 (from 12 genes, of which four were not trans-dev genes, for a total of eight fragments not associated with trans-dev genes) did not overlap any known feature. Detailed information on each SCE tested, including diagrams of their localization in mammalian and fish genomes as well as multiple alignments, is shown in Additional data file 4. As a control set 12 noncoding, nonrepeated, and nonconserved fragments were also chosen for co-injection assays, of which nine were from the same genes from which SCEs had been picked and three were from random genes (see Materials and methods, below, for details). Owing to the mosaic expression patterns that are obtained with this technique, results were recorded in two ways: by counting the number of cells stained for X-Gal and recording, where possible, the tissue in which the LacZ-positive cells were found; and by plotting LacZ-positive cells on expression maps that represent a composite overview of the LacZ-positive cells of all the embryos tested. Results of the cell counts are shown in Table 3 (For greater details, see Additional data file 3) and the expression maps are shown in Figure 7. The cell counts were used to define statistically which fragments exhibited tissue-restricted enhancer activity or generalized enhancer activity (see Materials and methods, below).

As a positive control a published regulatory element from the shh locus, ar-C 27, was coinjected with the HSP:lacZ fragment. From a total of 27 SCEs, 22 (about 81%) were able to enhance significantly the activity of the HSP:lacZ construct in comparison with the embryos injected with HSP:lacZ only (see Materials and methods, below, for details). Of these, three out of the four tested known mouse enhancers that were found to be conserved in fish were confirmed to act as enhancers in fish. A similar percentage of positive results (82.6%) was obtained excluding these enhancers in the count. The enhancer effect in 20 out of the 22 positive SCEs was not generalized but observed in a tissue-restricted manner.

The expression patterns obtained in our experiments were compared with expression data retrieved from the Zebrafish Information Network 3536. Multiple SCEs found within a single gene locus gave similar tissue-restricted enhancer activity. For example, all four SCEs tested from the ets-1 locus gave expression that was highly specific to the blood precursors (SCE 1646 in Figure 7c). This result is in accordance with reported data, which showed ets-1 expression in the arterial system and venous system. Moreover, both elements tested from the zfpm2 (also described as fog2 37) gene gave central nervous system (CNS) specific enhancer activity, which is in accordance with a recent report showing that the expression of both fog2 paralogs is restricted to the brain 37. Similarly, elements tested from the mab-21-like genes gave CNS and eye specific enhancer activity (SCE 4939; Figure 7f). This pattern of expression corresponds with the patterns reported in the brain, neurons, and eye 3839. The SCEs that were found in the pax6a and hmx3 genes were shown to give CNS specific enhancement, which is in accordance with the reported expression of these genes in the CNS 35. Finally, SCE 3121 from the gene jag1b gave specific expression in the CNS and in the eye (Figure 7d), which is in partial agreement with reported expression of this gene (expressed in the rostral end of the pronephric duct, nephron primordia, and the region extending from the otic vesicle to the eye 40).

Novel enhancer functions were also detected for SCEs neighboring lmx1b1, which showed CNS specific activity, and SCEs neighboring four genes not belonging to the trans-dev category, such as mapkap1 (Figure 7e), tmeff2 and 3110004L20Rik (producing proteins integral to the membrane), and elmo1 (associated with the cytoskeleton), which exhibited strong generalized and/or tissue specific activity. No endogenous expression data are available for these genes for comparison. In contrast to the results with SCE elements, only two out of 12 (about 17%) of the genomic control fragment set derived from the same loci of the SCEs exhibited significant enhancement of LacZ activity (Table 3).

Taken together, these data demonstrate that SCEs act as bona fide enhancers that can drive tissue-restricted as well as generalized expression during embryo development.

In this study we demonstrate, using a unique combination of tools aimed at obtaining regional, global-local sensitive alignments applied at the genome level, that the number of conserved non-coding sequences shared between mammalian and fish genomes is at least an order of magnitude higher than was previously proposed and is spread across thousands of genes. In fact, approximately 30% of the genes analyzed presented at least one SCE. Our GO analysis results indicate a 'trans-dev' bias similar to those described in previous studies addressing genes exhibiting noncoding conservation 1415. On the other hand, the significant increase in the sheer number of elements identified and in the number of genes exhibiting SCEs enabled us to detect conserved nongenic elements in a third of the genes studied, indicating that conservation of cis-regulatory modules is a widespread phenomenon in vertebrates, and is not limited to a few hundred genes, as suggested by previous studies. The GO analysis also revealed that certain classes of genes, such as those located in the extracellular space and extracellular matrix, exhibit conserved non-coding sequences, which were not identified with previous approaches and indicate that noncoding elements conserved across vertebrates are present in a larger and more diverse set of genes than was previously thought. Although we also observed a larger number of genes involved in cellular and physiological processes, many of them are also assigned to 'trans-dev' categories, and so their involvement in development and regulation of transcription cannot be excluded. Indeed, it is important to note that eight out of the 23 randomly selected fragments were not associated with trans-dev genes by GO classification, and that six of these fragments exhibited significant enhancer activity in our co-injection assays (Table 3). This confirms that conservation is not an exclusive characteristic of regulatory regions associated with trans-dev genes.

That shuffling plays an important role in the identification of conserved non-coding sequences is illustrated by the fact that 72% of our dataset was observed to be either inverted or moved, or both, in the fish locus with respect to the mouse locus. Assembly artifacts are unlikely to be an important factor in the elements identified as shuffled because they would also affect gene structures and therefore correct gene prediction and ortholog detection, which is at the basis of our dataset. We were reassured about this by our tetraodon-fugu comparison, which indicated that most elements found to be shuffled in one species were also shuffled in the other. A notable exception to the general shuffling bias in the elements found was a 1,000 bp window immediately upstream of the TSS. Taking into account that the proximal promoter region is considered to be approximately -250 bp to +100 bp from the TSS 41, and assuming that TSS annotations in the mouse genes analyzed are precise, this finding suggests that there is a class of enhancer elements that are more constrained in both position and orientation, perhaps working in tight connection to the promoter complex. The fact that the genes containing nongenic collinear elements in this window show the 'trans-dev' bias associated with our overall SCE dataset, as well as with previous analyses of noncoding conservation, reassures us that this result is not a mere product of bad annotation of the first exon in these genes. It is particularly reassuring that performing the same analysis on SCEs found in the same window but classified as 'genic' (and thus more likely to be real evidence of annotation problems) did not exhibit this bias.

Lack of conservation can also be due to the fact that the evolution of regulatory motifs involves constant de novo creation and destruction of them over time because of their short sequences and plastic nature 42 (for review 43). The dissection of cis-regulatory elements from different species, however, indicates clearly that there are cases in which although the same transcription factors are involved in the regulation of a gene, all sequences that are not responsible directly for the binding of transcription factors are not preserved and so overall sequence conservation is very poor 2. Thus, the quest to identify regulatory conservation must be complemented by a more thorough understanding of the inherent grammar of regulatory sequences, which would lead to improved alignment models specifically tailored to regulatory sequences 23.

During the past few years several strategies hav