Introduction

Discriminating between functional and junk sequences in the non-coding fraction of a genome has become one of the major challenges of functional and evolutionary genomics (Bird et al. 2006). Indeed, functional non-coding DNA is involved in the regulation of gene expression and thus in the evolution of novelties and adaptation between species (Castillo-Davis 2005). Functional non-coding sequences fall into two main categories: protein binding sites such as transcription factor binding sites (TFBSs), enhancers, and silencers, which are involved in the control of gene expression, and sequences that control chromatin organization such as insulators and matrix attachment regions (Cooper and Sidow 2003).

At least two different in silico approaches have been developed and may be combined to extract functional non-coding elements from the bare genome sequence data. Word count approaches search for over or under-represented motifs, but the main shortcoming of these methods is the high rate of false positives they generate (Hampson et al. 2002). The comparative approach, or “phylogenetic footprinting” as defined by Tagle et al. (1988), relies on homologous sequence data from at least two species combined with evolutionary theory, which states that substitutions accumulate much faster at non-functional DNA bases than at functionally constrained base positions. The comparative approach, using pairwise and multiple genome comparisons, has been successfully applied to identify conserved elements in mammals (Dermitzakis et al. 2004; Xie et al. 2005), vertebrates (Bejerano et al. 2004; Siepel et al. 2005), Drosophila (Bergman and Kreitman 2001; Halligan et al. 2004; Siepel et al. 2005), Caenorhabditis (Shabalina and Kondrashov 1999; Siepel et al. 2005), and the Saccharomyces sensu stricto group (Chin et al. 2005; Cliften et al. 2001; Kellis et al. 2003). In these groups of organisms, species divergence is such that pairwise aligned segments still retain “false positives,” that is, sequence identity because of shared ancestry, not because of selective constraint on a DNA sequence. As a consequence, the proportion of constrained non-coding sites has to be calibrated by the rate of neutral substitution, to correct for conserved but neutrally evolving sequences in Saccharomyces sensu stricto (Chin et al. 2005) or human–chimpanzee comparisons (Keightley et al. 2005).

This dense genome data coverage for evolutionary close groups of eukaryotic model organisms is not available for species belonging to the other four eukaryotic supergroups of the eukaryotic tree of life (Keeling et al. 2005), once the phylum of Unikonts (Fungi and Metazoans) has been removed. These organisms account for most of the eukaryotic diversity and whole genome data are still scattered along highly divergent branches, so that when two genomes from the same phylum are sequenced (as in Chlorophyta, Ciliates or Apicomplexan), the evolutionary distance reaches saturation on neutrally evolving sites. As a consequence, the methodological problem of discriminating between the phylogenetic footprints generated by selective constraints, and the footprints generated by shared ancestry, shifts the problem to that of discriminating the footprints generated by selective constraints and the footprints generated by the alignment algorithm itself.

The genus Ostreococcus belongs to the prasinophytes, an ecological important group dominating marine photosynthetic picoeukaryotes (Vaulot et al. 2008). They are the smallest eukaryotic free-living photosynthetic organisms identified to date, with a size of 1 μm, and are found worldwide in the marine environment (Rodriguez et al. 2005) and in the Sargasso Sea shotgun metagenome sequence data (Piganeau et al. 2008). Ostreococcus tauri and Ostreococcus lucimarinus cells are morphologically similar, even at electron microscopy level, and are characterized by a single chloroplast, a single mitochondrion and a cytoplasm bounded by a membrane lacking any detectable cell wall or scales. These species show specific adaptations to different environments as depth and/or light intensity (Rodriguez et al. 2005). The genome sequences of O. tauri (Derelle et al. 2006) and O. lucimarinus (Palenik et al. 2007) have recently been completed and revealed very short intergenic regions, raising the issue of the structure of its regulatory elements. The analysis of their protein coding genes revealed high levels of divergence, as measured by synonymous, non-synonymous and intronic rates of molecular evolution (Jancek et al. 2008; Piganeau and Moreau 2007), raising in turn a methodological problem to detect conserved intergenic regions.

From the comparison of these two genomes, we investigated (i) how we could estimate the degree of sequence conservation in intergenic sequences, (ii) whether some of these footprints are conserved through the green lineage by comparing the footprints with functional footprints previously identified in Arabidopsis and (iii) whether gene orientation of flanking genes influenced footprint size.

Methods

Datasets

Whole genome sequences and gene annotations for O. tauri and O. lucimarinus were downloaded from http://bioinformatics.psb.ugent.be/genomes/ and http://genome.jgi-psf.org/euk_home.html. When several gene annotations were available, we chose the Eugene annotation method that predicted shorter intergenic regions (to reduce positives due to unannotated coding sequences). Saccharomyces bayanus and Saccharomyces cerevisiae sequences were downloaded from the Saccharomyces genome database (ftp://genome-ftp.stanford.edu/pub/yeast/data_download/sequence/fungal_genomes/) and Candida glabrata sequences from the Genolevure database (http://cbi.labri.fr/Genolevures/download/GL2_index.php).

CDS could be mapped on chromosomes by BLAST (Altschul et al. 1990): positions of each gene on chromosomes were extracted using codes implemented in C language.

Orthologous intergenic regions (OIR) were defined as intergenic sequences between two genes having orthologs (defined as reciprocal best blast hits) in the same order and orientation in the two genomes compared. These OIRs were then extracted from the whole genome sequence data to be aligned with different alignment tools.

Alignment Software and Processing

We propose a permutation based validation scheme to estimate the significance of detected conserved sequences. Briefly, for each pair of orthologous intergenic sequences, we apply an alignment method on the real dataset and on 100 randomized datasets. As such, it is possible to compare the observed degree of conservation to an empirically determined distribution. This is then used to correct the observed identity between two sequences, to get an estimate of the proportion of constrained sites in each intergenic sequence.

This approach has been first applied to the S. cerevisiae and S. bayanus genomes to test and validate our method, because both our approach and the calibration by the neutral substitution rate approach can be used in this genus, as a consequence of the wealth of genomes available. We then applied our method to the two unicellular green algae O. tauri and O. lucimarinus.

ACANA (Huang et al. 2006) is a pairwise sequence alignment algorithm that uses a Smith–Waterman-like dynamic programming algorithm. ACANA has been shown to be highly accurate for divergent sequences, as compared to BLASTZ (Schwartz et al. 2003), CHAOS (Brudno et al. 2003a) and DIALIGN (Pohler et al. 2005) for local alignments (Huang et al. 2006). After benchmarking additional alignment algorithms that have been especially designed to align non-coding DNA sequences, Lagan 1.1 (Brudno et al. 2003b) and Mavid 2.0 build 4 (Bray and Pachter 2004), we found that ACANA 1.10 used in local mode (shortly ACANAL) with parameters -C 5 -T 1 best discriminated real from random sequences, as most OIRs with at least one significant footprint were obtained with this algorithm (results not shown). For both Ostreococcus and Saccharomyces OIR alignment software comparison led us to use ACANAL for all further analysis to study the nature and the level of constraint in OIR of Ostreococcus and Saccharomyces. Local footprints reported by ACANA have a minimal length of 9 bp.

Screening Footprints for Functional cis-Regulatory Elements

In order to investigate the putative function of our footprints, we assigned a significance level to each local alignment, hereafter “footprint.” The problem of assessing the significance of an alignment is complex (Altschul et al. 1994). To estimate the significance of footprints obtained for ACANAL, we ran the alignment method to the real and to 100 re-shuffled datasets, keeping the mononucleotide frequencies equal (i.e., for each re-shuffled dataset, all positions of both sequences were randomly permutated). We then calculated a score, defined as the number of nucleotide matches multiplied by the percent identity of the match, for each footprint, and estimated a P-value by counting how many times a bigger score was observed in the re-shuffled dataset. Note that stretches of ‘N’ (sequence gaps in final genome assembly) were maintained as fixed blocks during the shuffling process. This part was computer intensive because of running each alignment program over 400,000 times (number of intergenic regions times 100 times) and took approximately 3 weeks per whole genome comparison on a 2.8 GHz, Intel Computer with 1.5 GB of RAM.

We first investigated whether some of the significant footprints (P < 0.05) could be due to incomplete or mis-annotations. We therefore screened the footprints identified by ACANAL for open reading frames (blastx against GenBank, E-value < 1× 10−3) and found that putative ORFs are negligible in our data (Saccharomyces ~1.02% and Ostreococcus ~1.66%). Next, we screened the footprints for the presence of RNA genes (using the 116 RNA annotations of O. tauri and 132 for Saccharomyces) and found no OIR containing RNA genes in Ostreococcus and 36 in Saccharomyces out of 6 and 52 RNA loci present in the complete OIR dataset, respectively. These RNAs were removed for further analysis.

We compiled a list of 589 yeast regulatory elements described in the literature (referred to as reference motifs): 160 from Kellis et al. (2003), 50 from SCPD (Zhu and Zhang 1999), and 379 from Elemento and Tavazoie (2005) (Elemento and Tavazoie 2005). Similarly, we collected all plant motif instances from AGRIS (Davuluri et al. 2003) and PLACE (Higo et al. 1999), yielding 986 (partially redundant) reference motifs. We retrieved all footprints and compared these sequences with the reference motifs using DNA-pattern from rsa-tools (http://rsat.ulb.ac.be/rsat/) allowing zero substitutions. The fold enrichment for reference motifs was calculated by taking the ratio of the observed over the expected frequency of motifs located in footprints, where the latter was computed by counting motif instances on mono-nucleotide reshuffled sequences (i.e., maintaining base composition).

Estimating the Proportion of Nucleotide Sites Under Constraint

For each alignment with i segments containing Id i identical nucleotides per segment, we defined a conservation score, S, that gives the fraction of conserved nucleotides in the alignment.

$$ S = \sum\limits_{i} {{\frac{{{\text{Id}}_{i} }}{{{\text{length\_of\_alignment}}}}}} $$

The proportion of conserved nucleotides in an alignment j, S j , can be expressed as the sum of the proportion of nucleotides conserved as a result of constraint, F j , that share 100% identity, and the proportion of nucleotide conserved by chance, that equals 1 − F j , by the average identity observed in the re-shuffled random sequences, S random,j :

$$ S_{j} = F_{j} \times 1 + (1 - F_{j} ) \times \overline{S}_{{{\text{random}},j}} $$

Thus for each intergenic alignment, the proportion of nucleotides under constraint, F j can be estimated as:

$$ F_{j} = {\frac{{S_{j} - \bar{S}_{{{\text{random}},j}} }}{{1 - \bar{S}_{{{\text{random}},j}} }}} $$

Footprints for each OIR in Ostreococcus and Saccharomyces are available as Supplementary material.

Results

Estimation of the Level of Constraint on Intergenic Regions

We first benchmarked our method on Saccharomyces genome data, in order to compare our estimate of the proportion of phylogenetic footprints with other estimates based on multiple genomes alignments containing species showing less divergence. We focused on the two most distant Saccharomyces sensu stricto, S. cerevisiae and S. bayanus, showing saturated substitution patterns (Ks ~ 1.2), before performing a whole genome comparison of the two marine unicellular algae O. tauri and O. lucimarinus.

We extracted 2,758 OIRs from Ostreococcus and 2,203 from Saccharomyces. An OIR is defined as a pair of intergenic regions flanked by the same orthologous genes (showing similar relative transcriptional orientations) in both species. Note that this definition is more stringent compared to the frequently applied definition of orthologous promoters, which does not consider the conservation of both flanking genes. Sequences features of OIRs together with features of the compared genomes are described in Table 1. Since the S. bayanus genome sequence assembly is distributed over 1,098 contigs, it is not meaningful to compare the overall percent of OIRs (number of OIRs divided by total number of orthologous genes) between Ostreococcus and Saccharomyces. Strikingly, Ostreococcus intergenic regions are, on average, 25% shorter than Saccharomyces intergenic regions, suggesting greater compaction of regulatory elements in Ostreococcus, and/or smaller regulatory elements in OIRs.

Table 1 General features of the Saccharomyces and Ostreococcus genomes and sequence features of orthologous gene set and OIRs used in analysis
Full size table

For each OIR, we estimated the proportion of sites under constraint, F, as the excess of identity in the observed alignment as compared to the expected alignment for the shuffled intergenic region (see Methods section). Our estimate of the average F in yeast is 30.4%, consistent with previous estimates of 30% based on conservation calibration with the local neutral substitution rate (Chin et al. 2005). However, F is not normally distributed, suggesting that the average level of constraint is not a good descriptor of the genome wide level of constraint on intergenic regions (Fig. 1). It appears that 12% (Saccharomyces) to 23% (Ostreococcus) of intergenic regions contain no detectable footprint at all (F = 0), and that there is a large variation in the level of constraint for the remaining OIR. Excluding the intergenic regions with F = 0, the average level of constraint in intergenic regions raises from 30.4 to 34.7% in yeast and from 13 to 17% in Ostreococcus.

Fig. 1
figure 1

F distribution in Ostreococcus (Ota-Olu) and in Saccharomyces (Sce-Sba)

Full size image

Moreover, F is strikingly different with regard to gene orientation of neighboring genes (Table 2). Head-to-head OIR (divergent gene pairs) are on average longer in both species, followed by head-to-tail OIR and tail-to-tail OIR (converged gene pairs), hereafter HH, HT, and TT. However, if following (Chin et al. 2005), we define conserved sequences as regulatory elements, the regulatory element structure seems to be different between Saccharomyces and Ostreococcus. In yeast, the longer regulatory elements are in HT OIR (200 bp) whereas in Ostreococcus, the longer regulatory elements are in HH OIR (53 bp) (Table 2). In addition, the low F value for Ostreococcus TT compared to HH OIR (0.08 and 0.18, respectively) suggests that 3′ regulatory elements are rare. This is in contrast with Saccharomyces, where the average F value for HH and TT is identical (0.17).

Table 2 Average F differs between gene orientation
Full size table

From these results, we can test null models of regulatory element structure for the three different types of OIR, as defined in Hermsen et al. (2008) (Fig. 2). First, let us define a simple model of regulatory element structure, the additive regulatory element model. In this model, a HH region contains two 5′-regulatory element sequences of total length l HH , a TT region contains two 3′-regulatory element sequences of total length l TT, and a HT region contains one of each type of regulatory element sequences, of total length l HT = l TT + l HH. We have three distributions to estimate the distribution of two variables, l TT and l HH. We can thus use the remaining distribution to test whether the expected distribution equals the observed distribution. We estimate l TT, l HT, and l HH from the total number of constrained nucleotides in each OIR, excluding OIR with no conserved site (F = 0). Note that this is different from the Hermsen et al. (2008) study that is based on the total length of the intergenic regions.

Fig. 2
figure 2

Regulatory element structures. Black arrows give the 5′->3′ orientation of neighboring genes, shaded box: 3′ regulatory elements, of total length l TT, gray box: 5′ regulatory elements, of total length l HH

Full size image

Second, we defined the bidirectional regulatory element structure, where a HH region contains one 5′-regulatory element sequences of length l HH, a TT region contains one 3′-regulatory element sequences of length of length l TT, and a HT region contains one of each type of regulatory element sequences.

The additive regulatory element structure model cannot be rejected from the Ostreococcus data (Table 3), whereas this model does not fit the Saccharomyces data (P < 10−5), because of the greater than expected length of regulatory elements in HT regions. Strikingly, even the bidirectional regulatory element structure model, predicting an average l HT of 154 bp cannot account for the observed length of regulatory elements in HT regions (Table 3).

Table 3 Estimates of the average lengths of the conserved regulatory elements under the additive and the bidirectional models (see text and Fig. 2)
Full size table

Benchmarking Method on Saccharomyces Data: Comparison with Known Motifs and Gene Orientation Effect

To evaluate the power of our phylogenetic footprinting approach to identify cis-regulatory elements, we compared all significant footprints (as defined in “Methods” section) against a reference set of regulatory motifs (589 for fungi and 936 for green plants; see “Methods” section). For 97% of the Saccharomyces footprints (4559/4710) and 99% of the Ostreococcus footprints (2329/2350), there was a perfect match with a reference motif. Although the degenerate nature of several of the reference regulatory elements, in general, hinders the identification of biologically functional motif instances (Vavouri and Elgar 2005), it is interesting to note that for both species a large number of reference motifs occur much more frequently in our footprints than expected by chance. The enrichment for reference motifs was calculated by taking the ratio of the observed over the expected frequency of motifs located in footprints, where the latter was computed by counting motif instances on reshuffled sequences. Considering the 50 motifs from the Promoter Database of S. cerevisiae (SCPD), 74% are twofold or more enriched in the significant footprints. Examples in yeast are the ESR1 binding site (GATGAG Kellis_g036), the SCPD_GCN4 motif regulating biosynthetic genes in response to amino acid starvation, the PAC binding site (CTCATCGCA Elemento_21; involved in rRNA transcription) and the Met4 binding site (AACTGTGGC Kellis_g057; involved in amino acid metabolism) all showing a greater than threefold enrichment.

To verify if some of the reference motifs could resemble 3′ regulatory signals captured in our footprints, we analyzed the frequencies of all reference motifs over the HH, HT, and TT OIRs. Whereas most motifs are most frequent in the HT class and nearly absent in the TT class—indicative of 5′ promoter regulatory motifs, we did find some motifs that occur with high frequencies in TT OIRs. One example in yeast is the TATATA upstream efficiency element (nataTATATAyATATATAnn, 4% HH, 41% HT, and 56% TT; n = 27), an mRNA 3′-end processing element appearing upstream of the poly(A) cleavage site (Graber et al. 1999). Another motif WTATWTACADG described by Kellis and co-workers is also depleted in HH OIRs (4% HH, 69% HT, and 27% TT; n = 48) and resembles a down-stream element identified in a set of co-expressed genes whose product localizes to the cytosolic translational machinery, the mitochondrial DNA translational machinery or the mitochondrial outer membrane (Kellis et al. 2003). Also in Ostreococcus, we found some motifs, including the plant poly(A) signal AATAAT (4% HH, 16% HT, and 80% TT; n = 25), enriched in TT and depleted in HH OIRs (Fig. 3). These results demonstrate that our footprints also capture 3′ regulatory elements most probably playing a regulatory role in mRNA splicing, localization, or stability.

Fig. 3
figure 3

Gene orientation frequencies for known plant motifs enriched in conserved Ostreococcus footprints. The series refer to the frequency of motifs in conserved footprints counted per OIR class (HH head-to-head, HT head-to-tail, and TT tail-to-tail). Numbers in parenthesis indicate the number of significant footprints matching a known motif used to calculate the frequencies. The data for all motifs are provided as Supplementary material

Full size image

Searching for Function of Footprints in Ostreococcus: Investigating Motif Conservation for Four cis-Regulatory Elements

For Ostreococcus, the verification of known regulatory elements is hampered by our limited knowledge about transcriptional control in these green algae. Therefore, we selected four well studied cis-regulatory elements (E2F, TELO, I-box, and CRE/DRE motif) from land plants (i.e., Arabidopsis thaliana) and investigated whether these motifs are conserved in green algae and, if so, whether they are present in our Ostreococcus footprints. For the E2F box, involved in the regulation of DNA replication genes during S-phase, we initially selected 38 Arabidopsis genes containing a consensus WTTSSCSS motif in their promoter sequence (Vandepoele et al. 2005) and annotated as involved in DNA replication. For the corresponding orthologous Ostreococcus genes present in our data set, we found that half (7/14) of the OIRs contained a (W)TTSSCSS motif, and in each of these seven cases, the motif was located in a footprint. Examples of genes with a conserved E2F site are two DNA polymerase subunits (alpha and beta), a MCM subunit and an ORC subunit (with P < 0.01, 0.01, 0.51, and 0.86, respectively hypergeometric distribution; Pilpel et al. 2001), well-described E2F target genes in plants and animals. Complementary, mapping the location of the E2F footprints in relationship to the gene orientation reveals that 97% of all motifs occur in HH or HT OIRs, confirming its role as a promoter cis-regulatory element (Fig. 3). The TELO box (AAACCCTA) is frequently found in the promoter of cytosolic ribosomal proteins in Arabidopsis (Tremousaygue et al. 1999). We selected 26 Ostreococcus genes orthologous to known TELO target genes from Arabidopsis that were annotated as involved in ribosome biogenesis and assembly. However, only one Ostreococcus lucimarinus gene in this dataset contains the (A)AAACCCT(A) motif, which was not located in the OIR of the O. tauri gene. The I-box (CTTATC) is a promoter element frequently found in light-responsive genes. Starting from a set of 297 Arabidopsis I-box genes (Vandepoele et al. 2006), we found that 33 orthologous Ostreococcus genes contain the (C)TTATC motif, of which only three motif instances are located in footprints. Although this result might suggest that our identified footprints only partially detect these cis-regulatory elements, an alternative scenario is that these boxes—as they are defined in Arabidopsis—are not functional in green algae. To investigate this possibility, we first identified the set of Ostreococcus genes (orthologous to the Arabidopsis target genes) that contain the motif (discarding if it is located in a footprint or not) and then assessed whether the functional annotation linked to the Arabidopsis reference genes is conserved in Ostreococcus. Whereas for the E2F box a strong (conserved) GO enrichment toward DNA replication was found (hypergeometric distribution P < 0.01), the GO enrichment for photosynthesis (P < 0.01) observed in Arabidopsis I-box target genes is not observed in Ostreococcus. Consistent with this, the presence of an I-box in the promoter of Ostreococcus rbcS, a known I-box target gene in higher plants (Altschul et al. 1990), could not be confirmed.

The (G/a)(T/c)CGAC CRE/DRE transcription factor-binding site has been shown to be involved in the Low Temperature response in Hordum (barley) and Arabidopsis (Xue 2003; Sharma et al. 2005 for review). The GTCGAC motif is overrepresented in our significant footprints and frequently found in HH OIR in Ostreococcus (67% HH vs. 30% HT and 3% TT; Fig. 3).

The low temperature response involves transcriptional activation and repression of several pathways, and activated pathways include those involved in the accumulation of osmoprotectants like sugars, amines, and compatibles solutes (Sharma et al. 2005). We investigated whether we could identify regions of similarity between the 66 upregulated A. thaliana genes under cold response containing this motif in their 1 kb upstream sequence (Vogel et al. 2005) and the Ostreococcus genes also having this motif in a significant footprint. To assess whether the number of genes with regions of similarity was higher than expected by chance, we compared it with the number of genes with regions of similarity having a significant footprint and no GTCGAC motif. We found a significant excess of homologues with the Ostreococcus genes also having a GTCGAC motif as compared with the Ostreococcus genes not having this motif (Fisher exact test, P = 0.01).

We also investigated that Ostreococcus contains a gene homologue to the transcription factors associated to this motif in A. thaliana, the CRT/DRE binding factors. These transcription factors belong to the large AP2 multigene family of DNA-binding proteins (Riechmann and Meyerowitz 1998). There are only two homologues to these genes in both Ostreococcus genomes, annotated as transcription factors as they contain the AP2 like DNA-binding domain (JGI protein ID 23938 and 9237 in O. lucimarinus; 23659 and 30470 in O. tauri).

Discussion

We investigated the level of constraint in non-coding regions from pairs of divergent streamlined genomes using a phylogenetic footprinting approach. This methodology is specifically adapted to three genomic features that seem to characterize recent unicellular eukaryotic genome projects. First, these genomes are gene dense and thus contain short intergenic regions that can be analyzed over their full length. Second, pairwise genome comparison shows a well-preserved short scale synteny, enabling a stringent definition of OIRs by shared adjacent orthologous genes. Third, genomes available for comparison share a high level of divergence, as measured by complete saturation at neutral evolving sites, making constraint estimates by calibration with neutral substitution rates practically unsuitable. The method applied in this study is expected to produce a very low fraction of false positives. First, because it has been developed for highly divergent species, where all neutral sites have been overwritten and there is no sequence similarity due to shared ancestry alone. Second, the observed alignments were benchmarked against alignments obtained from randomized sequence dataset to quantify the amount of conservation expected by chance and to determine significance values for the different footprints.

On the other hand, our method has three main limitations. The first one is common to all phylogenetic footprinting methods, that is that they are only able to detect the highly constrained functional sequences, which is only a fraction of the actual functional non-coding sequences of an organism (Li et al. 2007; Samanta et al. 2006; Wittkopp 2006). The second one is that our method will not detect constrained sequences of low complexity, as mono-nucleotide repeats. This is because the randomized sequences (based on the real dataset) will be too much like the real dataset to allow discrimination. The third one is that we will not detect any conserved sequence shorter than 9 bp, the minimum size of segments reported by the local alignment algorithm ACANA applied in this study.

Using the S. cerevisiae—S. bayanus genome comparison as a benchmark, we show that our method effectively detects regulatory elements previously identified by other methods relying on multiple species comparison. Our estimation of the proportion of constrained sites in yeast non-coding regions, 30%, is consistent with previous estimates relying on multispecies comparison and calibration by the neutral substitution rate (Chin et al. 2005). We then applied our method to the two available genomes of Ostreococcus, the smallest free-living eukaryotic photoautotrophic cells, to investigate regulatory element structure. There are approximately half as many constrained sites in intergenic regions in Ostreococcus as compared to Saccharomyces, with an average proportion of constrained sites, F, of 13%. This is consistent with the higher divergence of the two Ostreococcus genomes (Table 1), and reflects the degradation of some evolutionary information with time. On the other hand, this degradation enables to come closer to the sequence backbone of gene expression control, that remain conserved between even more distantly related species, as exemplified by the conserved non-coding sequences of master control genes in development, conserved between fly and chicken (Blanco et al. 2005).

We also found that the level of constraint depends on the type of intergenic region considered, given by the 5′ or 3′ orientation of the neighboring genes. This has not been reported for multicellular genomes as such, because OIRs are not defined for adjacent orthologous gene pairs, but are rather defined by taking a region upstream and downstream one gene. However, 3′ UTR sequences have been shown to be more constrained than 5′ UTR in vertebrates and to a lesser extent in Drosophila, possibly reflecting widespread post-transcriptional regulation by microRNAs (Siepel et al. 2005). Since microRNA interference is not believed to occur in to the two organisms we analyzed (Cerruti and Casas-Mollano 2006 for review), the different trend we observe is not surprising.

Previous studies have already pinpointed some qualitative inter-kingdom differences between genome structure, as the positive correlation between first intron length and expression in Arabidoposis and rice as opposed to animals (Ren et al. 2006). We observed that the regulatory element structure estimated from the Ostreococcus genomes is markedly different from the yeast regulatory element structure. Indeed, the regulatory element structure in Ostreococcus is consistent with an additive regulatory element structure model, where the total length of constrained sites in HT intergenic regions corresponds to the sum of constrained sites in a 3′ regulatory element and in a 5′ regulatory element. This also may suggest some kind of optimization of intergenic sequence space in Ostreococcus, because the length of each type of intergenic region is proportional to the level of constraint. On the other hand, the regulatory element structure in yeast is biased toward a higher level of constraint in HT regions. There are at least three possible explanations for this trend. First, cases of bidirectional regulatory element structure in HH intergenic regions have been reported experimentally in yeast (Ishida et al. 2006), and a bidirectional regulatory element structure suggests a greater compaction of regulatory elements in HH non-coding regions. However, if we assume that all HH and TT intergenic regions contain bidirectional regulatory elements, we show that there is still an average excess of regulatory element sequences in HT region (46 bp, Table 3), so that this cannot be the sole factor responsible for this trend, in addition to the fact that it is unlikely that all HH regions contain bidirectional regulatory elements. Recently, a study on the distribution of TT, HT, and HH total intergenic sequence lengths in S. cerevisiae, ignoring sequence conservation, suggested that about 30% of HH regions contain a bi-directional transcriptional regulatory regions (Hermsen et al. 2008). Second, yeast HT intergenic regions could evolve at a slower pace than HH and TT intergenic regions. This could be a consequence of a recent partial loss of an interleaving orthologous gene leading to a present HT OIR. Indeed, it has been shown that before the whole genome duplication (WGD), gene order orientation in Saccharomyces was more biased toward HH and TT occurrences, and that single gene deletions in pairs of HH and TT genes are responsible for the present distribution of adjacent gene orientation (Byrnes et al. 2006). To test this scenario, we investigated whether HT OIR prior to WGD are shorter and have a lower F than more recent HT OIR, using the available genome data of A. gossypii, a pre-WGD species. We re-estimated F in the 245 adjacent gene pairs of the S. cerevisiae—S. bayanus that were already adjacent and in same orientation between Sc and Ag. We found a higher average F in these pre-WGD OIR than in the more recent OIR (preWGD:postWGD OIR:HH: F = 18.6:16.8, TT: F = 47.8:44.7, and HT: F = 15.2:17.2). However, the higher level of constraint in HT regions remains in the preWGD gene pairs and there is thus no detectable effect of single gene deletion on the pattern we observe in yeast.

Third, transcriptional interference, the perturbation of one transcription unit by another, could exercise different selective pressures on different gene orientations. There is experimental evidence for differential strengths of transcriptional interference as a consequence of gene orientation in mammalian cells, with more interference in tandems of HT genes (Eszterhas et al. 2002), but to our knowledge, this has not been investigated in yeast.

Estimating the proportion of sites under constraint in non-coding regions is often independent from a functional analysis non-coding regulatory elements, and this may lead to apparent conflicting interpretations (Bush and Lahn 2005; Keightley et al. 2006). Comparing our footprints against a reference set of yeast and plant motifs indicates that a high proportion of conserved footprints identified in this study captures known cis-regulatory elements. A detailed analysis of the E2F pathway targeting DNA replication genes in animals and higher plants reveals that several Ostreococcus genes are orthologous to bona fide replication genes containing an E2F binding site in their promoter. This finding indicates that this pathway is also evolutionary conserved in green algae. Interestingly, all E2F binding sites present in Ostreococcus DNA replication that were conserved between species were detected using our pairwise footprinting approach. We have also established a three level homology based correspondence between (i) a cis-regulatory motif overrepresented in our footprints (the cold response CRE/DRE element), (ii) the transcription factors binding this motif, and (iii) the genes up regulated by this motif in Arabidopsis. Experimental analysis is now required to demonstrate that the cis-regulatory motif proposed is a binding site upregulating gene transcription under low temperature, to conclude about the conservation of the low temperature regulation pathway between Ostreococcus and Arabidopsis.

Altogether, these findings indicate that the E2F pathway is conserved in green algae and that there may be conservation of the Low Temperature pathway involving the CRE/DRE transcription binding site. In contrast, the cis-regulatory control of cytosolic ribosomal proteins and light-regulated genes in green algae might be driven by other promoter elements than the TELO or the I-box, respectively. These results indicate that, like observed between different yeast species (Tanay et al. 2005), the underlying cis-regulatory network has evolved substantially within the green plant lineage, even for highly conserved processes like photosynthesis or ribosome biogenesis. However, these results show that the footprints we describe provide a starting point to characterize other cis-regulatory elements in green algae and may thus enable to unravel very ancient regulatory pathways. Deciphering the genomic information in free-living unicellular eukaryotes will enable us to picture the toolbox of functional non-coding sequences in ancestor eukaryotic cells. On the other hand, it will enable to pinpoint cis-regulatory divergence and major regulatory novelties implied in the evolution of multicellularity in plants and animals.