Open Access

Research Article

• Download Article XML
• Download Article PDF
• Download Citation
• E-mail this Article
• Order Reprints
• Print this Article

Introduction

Several recent reports indicate that the transcriptional complexity of mammalian genomes has been significantly underestimated. Large-scale sequencing of full-length transcripts, expressed sequence tags (ESTs), and shorter tags [1] and transcriptional maps constructed by the use of tiling arrays [2–5] demonstrate that human and mouse genomes contain an abundance of complex loci with overlapping transcription on the two DNA strands. Although individual complex loci have been described in detail [6–8], a global description of the general properties of gene arrangements within such complex loci is lacking.

Two types of cis-coupling of genes have been reported to be widespread in mammalian genomes. (1) More than a thousand pairs of divergently transcribed, nonoverlapping genes spaced by less than 1,000 bp have been found in the human genome, comprising 9% of known genes [9]. The genes in such a pair typically share a bidirectional promoter. (2) Numerous pairs of oppositely transcribed genes whose exons overlap in the genome (cis–antisense pairs) have been identified in human and mouse genomes [10,11]. Human cDNA and EST data indicate that 22% of transcripts are involved in cis–antisense pairs [12]. Data from tiling array experiments and sequencing of short tags representing 5′- and 3′-ends of transcripts suggest that cis–antisense pairs might be even more widespread, perhaps involving more than 60% of all loci [4,13]. For both bidirectionally promoted pairs and cis–antisense pairs, there is evidence that paired genes tend to be coexpressed [9,11,13–15]. In some bacteria, it is well established that natural antisense transcripts from cis–antisense pairs can regulate expression of the gene encoded on the opposite strand (for review, see [16]). Numerous case studies suggest that cis-encoded natural antisense transcripts are important regulators in eukaryotes as well, potentially affecting a range of processes including transcription, imprinting, DNA methylation, and RNA splicing, editing, and degradation (for review, see [17,18]).

Cross-species genome comparisons can reveal conserved genomic features that are likely to be functionally important, and species-specific features that might underlie phenotypic differences. For the great majority (81%) of human bidirectionally promoted pairs where the genes have mouse orthologs, the bidirectional arrangement is conserved, suggesting that it is functionally important [9]. Similarly, for human cis–antisense pairs where the genes have orthologs in pufferfish, proximity and orientation of the paired genes is conserved in pufferfish significantly more often than for pairs of neighboring genes on the same strand [19]. However, evidence for cross-species conservation of actual overlapping arrangements of genes has been more limited. Searches for human–mouse orthologs that form cis–antisense pairs in both organisms have previously reported at most 347 gene pairs [19–21], a very small number compared to the thousands of species-specific pairs found. In addition, the actual exon overlaps within cis–antisense pairs have been reported to lack elevated conservation in general [20], contrary to the hypothesis that blocks of conservation in untranslated regions (UTRs) and extreme conservation in translated regions of transcripts indicate antisense regulation [22]. A limitation of the aforementioned comparative studies of cis–antisense pairs might have been their exclusive focus on protein-coding genes: recent unbiased surveys of mouse transcripts have indicated that cis–antisense pairs most frequently consist of one coding and one noncoding transcript [11,13].

We define complex loci as genomic regions in which multiple genes share transcribed regions in antisense orientation and/or bidirectional core promoters. In this study we construct comprehensive and highly reliable genome-wide datasets of cis–antisense and bidirectionally promoted gene pairs from human and mouse transcript sequence data and present an analysis of the higher-level organization of these pairs in complex loci. We further explore human–mouse conservation of complex loci at both sequence and structure levels, taking into account both coding and noncoding transcripts. We describe a widespread occurrence of “chains” of overlapping transcriptional units (TUs), a several times greater number of human–mouse conserved cis–antisense pairs than previously reported, and additional species-specific complex arrangements. We perform sampling to reach an estimate of the total fraction of genes in cis–antisense arrangement, and of the fraction of such arrangements that are conserved between human and mouse. We study the sequence composition of bidirectional promoters and its relation to the positioning of transcriptional start sites (TSSs). Finally, we take a closer look at a number of homeotic genes, to assess the extent of transcription from the opposite strand at these loci and the conservation of the transcripts.

Results

cis–Antisense Pairs and Bidirectional Promoters Are Abundant in Mammalian Genomes

We inferred TUs from genomic mappings of EST and full-length cDNA sequences from FANTOM3 and the public databases. Particularly rigorous criteria were applied to thoroughly eliminate the inclusion of artificially reversed sequences (see Materials and Methods). This is straightforward for spliced sequences, since their orientation can be verified by sequence motifs at splice junctions. We assessed the performance of the part of the procedure that handles mappings of unspliced sequences. Mappings based on EST sequence only (EST mappings) were treated more stringently than mappings with cDNA support (cDNA mappings), because of the higher quality and higher methodological reliability of transcript strand annotation of the latter set. We estimate that the procedure, when applied to human data, correctly determined the orientation of 99.8% of unspliced cDNA mappings and 99.8% of unspliced EST mappings. Tests on mouse data gave very similar estimates (Figure S1). Only 0.09% of unspliced human and mouse cDNA mappings were rejected, but 52% of unspliced human EST mappings and 34% of unspliced mouse EST mappings were rejected because of insufficient information on original strand orientations. The lower rejection rate for mouse EST mappings largely reflects higher availability and consistency of read direction (5′/3′) annotation for mouse EST sequences, where a higher proportion of ESTs were produced using cap trapping technology [23].

Starting from 161,805 human and 140,769 mouse cDNA sequences and a total of ~8 million ESTs, we obtained 42,887 human and 36,606 mouse TUs (Table 1). Using the estimated reversal rates and the empirical genomic distribution of mappings, we performed simulations that indicated that the frequency of false TUs due to mapping reversal was about one in 1,000 for human and about one in 800 for mouse (see “Accuracy Assessment of Orientation Procedure” in Materials and Methods). This would yield about 40 false human and 50 false mouse TUs, which is an acceptable rate that would not impact the conclusions of any of the further analyses we performed.

Table 1.

Numbers of cDNA and EST Sequences Used and Resulting TUs

doi:10.1371/journal.pgen.0020047.t001

cDNA and EST sequences support involvement of at least 25% of all TUs in cis–antisense pairs and 9% of all TUs in bidirectionally promoted pairs.

Nearly half of all TUs were involved in one or more of the types of bidirectional transcription defined in Figure 1: cis–antisense pairs, non-exon-overlapping antisense pairs, and bidirectionally promoted pairs (Table 2). The most common arrangement was the cis–antisense pair, which involved 25% of TUs in both human and mouse. Putative bidirectionally promoted pairs involved 9%–10% of TUs and were also roughly equally frequent in the two organisms. On the other hand, the raw frequency of non-exon-overlapping antisense pairs differed between human and mouse. Since TUs in non-exon-overlapping antisense pairs need not share exon sequence similarity, this dataset may contain a number of TUs that are artificially nested because of genome assembly and transcript sequence mapping errors. For this reason, the subsequent analysis focused on cis–antisense pairs and bidirectionally promoted pairs.

Figure 1. TU Pairs Searched For

We defined a cis–antisense pair as two oppositely transcribed TUs that share at least 20 bp of exon sequence, a non-exon-overlapping antisense pair as two oppositely transcribed TUs that overlap by at least 20 bp, but not within exons, and a bidirectionally promoted pair as two divergently transcribed TUs that overlap by less than 20 bp and are less than 1,000 bp apart.

doi:10.1371/journal.pgen.0020047.g001

Table 2.

Numbers of TU Pairs Detected and TUs Involved

doi:10.1371/journal.pgen.0020047.t002

To further investigate the reliability of the cis–antisense pair dataset, we categorized the pairs based on the types of mappings supporting exon overlaps (Figure S2). Most pairs (human: 60%; mouse: 58%) were supported by spliced mappings on both strands. The great majority of cis–antisense pairs (human: 78%; mouse: 88%) were supported by sequence types generally considered to be of high quality (either cDNA or spliced EST mappings), indicating that our set of cis–antisense pairs is highly reliable.

RT-PCR validation of a sample of cis–antisense pairs suggests that at least 80% are expressed from both strands.

To experimentally assess the validity of our cis–antisense pair dataset, we performed orientation-specific RT-PCR as previously described [12,24]. We investigated the expression in adult mouse brain of complementary transcripts corresponding to 20 randomly selected cis–antisense pairs supported by at least one cDNA or EST from adult mouse brain on each strand. As negative controls, we selected five highly expressed genes for which we could find no evidence of antisense transcription in sequence databases. We were able to detect the coexpression of sense and antisense transcripts in brain for 16 of the 20 cis–antisense pairs (Figure 2). For one of the remaining pairs, the result was ambiguous because of the presence of many additional bands of unexpected sizes. One of the negative controls (Rps27) also reproducibly showed evidence of antisense transcription. In retrospect, this control was ill-chosen: there are several copies of Rps27 pseudogenes in the mouse genome, so the complementary transcripts need not be transcribed from the same loci as Rps27 itself. In conclusion, the RT-PCR results suggest that at least 80% of the cis–antisense pairs in our dataset are expressed from both strands, and that there might exist a significant number of antisense transcripts that are yet to be discovered.

Figure 2. Validation of the Expression of Randomly Selected cis–Antisense Pairs by RT-PCR

To confirm the expression of complementary transcripts, we performed orientation-specific RT-PCR as described previously [12,24]. Primers were designed to amplify regions of exon overlap. For each candidate or control, four RT-PCR reactions (corresponding to the four lanes in each gel image) were carried out using adult mouse brain RNA as template. Orientation specificity was achieved by restricting which primers were present during reverse transcription single-strand synthesis: no primer (first lane), only sense primer (second lane), only antisense primer (third lane), and both sense and antisense primers (fourth lane). In all reactions, both primers were present during the subsequent PCR reactions. For candidates, sense and antisense primers were designed with respect to the genomic plus strand. For controls, primers were designed with respect to the control transcript. Out of five highly expressed control genes with no evidence of antisense transcription in sequence databases, we detected antisense transcription for one (Rps27). We reproducibly observed evidence of anti-Rps27 transcripts using two different primer pairs (unpublished data). We tested 20 cis–antisense pairs from our computationally constructed dataset and detected expression of both strands for 16 (underlined). For one additional cis–antisense pair (number 11), the result was ambiguous because of the presence of many bands of unexpected size. The 20 cis–antisense pairs were selected at random from the mouse dataset, with the requirements that exon overlaps be at least 200 bp (to allow amplicons of at least 100 bp) and that there be at least one cDNA or EST from adult brain supporting the exon overlap on each strand.

doi:10.1371/journal.pgen.0020047.g002

Properties of cis–antisense overlaps are highly similar between human and mouse.

Many cis–antisense pairs in our set (34% of human pairs and 34% of mouse pairs) had multiple distinct exon-to-exon overlaps. Both genomes had on average 1.6 distinct exon-to-exon overlaps per cis–antisense pair. The size distributions for exon overlaps ranged from 1 bp to 5,200 bp with medians of 159 bp (human) and 172 bp (mouse), and distinct peaks around 100 bp. The average repeat content within exon overlaps was 9.5% in human and 5.9% in mouse. For comparison, we measured the repeat content within the entire exonic sequence of each TU. The average repeat content of entire TUs was nearly three times as high for TUs in general (25% and 17% of exon sequence for human and mouse, respectively), and nearly two times as high for TUs involved in cis–antisense pairs (human: 17%; mouse: 12%). We therefore concluded that cis–antisense pairs tend to involve TUs with low repeat content, and that exon overlaps tend to be located in repeat-poor regions of those TUs. However, for a subset of cis–antisense pairs (human: 213; mouse: 53), more than 90% of the exon overlap was repeat sequence. (Here it should be noted that there might be an underrepresentation of repeat-rich transcripts in the dataset because of the inherent difficulty of unambiguously mapping them onto the genome.)

We classified the cis–antisense pairs—based on transcriptional direction of participant TUs—as divergently transcribed (head-to-head overlapping), convergently transcribed (tail-to-tail overlapping), or fully overlapping (one TU completely spanned by the other). In agreement with previous observations on the FANTOM3 dataset [13], we found these three classes to be roughly equally common in mouse (Table S1). In human, divergent and convergent cis–antisense pairs were also roughly equally common, but fully overlapping pairs were more frequent, constituting 42% of all pairs. A significant number of these fully overlapping pairs might represent actual divergent/convergent cases that were not detected as such because of the lower availability of full-length cDNA sequence for human.

Over 40% of all TUs might be involved in cis–antisense pairs.

To estimate the true proportion of TUs that are involved in cis–antisense pairs, we recomputed the TU and cis–antisense pair datasets using random subsets of all available transcript sequences. Figure 3A and 3B show the fraction of TUs we observed to be involved in cis–antisense pairs as a function of the number of transcript sequences used. For both human and mouse, a saturation curve

fitted almost perfectly to the sampled data: here a is the fraction of TUs involved in cis–antisense pairs at saturation, and c (the equivalent of the Hill coefficient) is a measure of sequence redundancy in the set and depends on the choice of sampled set (e.g., all transcripts or only one sequence per cDNA clone). The saturation curves predicted that the fraction of TUs involved in cis–antisense pairs approaches 0.45 for human and 0.43 for mouse as the number of transcript sequences increases. Using two other sampling approaches, we obtained closely similar estimates (Figure S3). Thus, based on the current data, over 40% of human and mouse TUs might eventually be found to be involved in cis–antisense pairs if transcript sequencing continues.

Figure 3. Estimating the Extent and Conservation of Antisense Transcription

(A and B) Estimation of proportion of TUs involved in cis–antisense pairs. Open circles indicate the fraction of all human TUs on the plus strand (A) and all mouse TUs on the plus strand (B) that were found to be involved in cis–antisense pairs when the minus-strand TUs were recomputed starting from random transcript sequence samples of different sizes. Filled circles represent the full datasets based on all available transcript sequences. The saturation curves (see Equation 1) indicated by the lines fit almost perfectly to the sampled data. Fitted human and mouse saturation curves approach 0.45 and 0.43, respectively, as the number of transcript sequences increases, indicating that more than 40% of all TUs might be involved in cis–antisense pairs. Similar estimates were obtained by other sampling approaches (Figure S3).

(C) Estimation of the proportion of human cis–antisense pairs that are conserved in mouse. Open circles indicate the proportion of human cis–antisense pairs found to be conserved in mouse when the full human dataset was compared to mouse datasets recomputed from random mouse transcript sequence samples of different sizes. The same type of saturation curve as in (A) was fitted to the data. Here, a model with c = 1 (i.e., hyperbolic saturation) was preferable as it provided an equally good fit while being simpler. The fitted curve approaches 0.25 as the number of mappings grows, indicating that about 25% of human cis–antisense pairs are conserved in mouse.

doi:10.1371/journal.pgen.0020047.g003

Nearly 1,000 cis–antisense pairs are conserved between human and mouse.

As noted above, we found a striking agreement between human and mouse in prevalence and general properties of cis–antisense pairs. We proceeded to assess the agreement between the human and mouse datasets at the individual pair level. First, we counted the number of human and mouse cis–antisense pairs that had exon overlaps in corresponding positions in a BLASTZ net alignment of the two genomes (alignments were obtained from the University of California Santa Cruz (UCSC) Genome Browser Database [25]; see Materials and Methods). There were 962 such pairs in human, and 943 corresponding pairs in mouse, constituting 16% and 18% of all human and mouse cis–antisense pairs, respectively (Table S2). The human and mouse numbers differ slightly because a small proportion of mouse pairs corresponded to several human pairs and vice versa. We consider this a strict assessment of conservation, because exon overlaps were required to be in corresponding places (implying conserved structure). However, we did not set any explicit sequence conservation threshold, since sequence might not be of primary importance for antisense regulation and previous work has indicated that antisense overlaps do not tend to have elevated sequence conservation [20]. The majority of cis–antisense pairs (69% of human pairs and 82% of mouse pairs) had more than 90% of their exon overlap sequence within BLASTZ net alignments, indicating that the implicit requirement for sequence similarity imposed by the use of precomputed alignments did not severely limit our ability to detect conserved cis–antisense pairs. However, it is likely that a large number of truly conserved pairs were not detected as such because of transcript sequences that have not been discovered yet. We attempted to estimate the true extent of conservation of cis–antisense pairs by a sampling approach equivalent to the one we employed above to estimate the fraction of TUs involved in cis–antisense pairs. To estimate how observed conservation grows with increasing transcript sequence data, we compared the entire human dataset against mouse datasets computed from different-sized random samples of all available mouse transcript sequences (Figure 3C). The same type of saturation curve as used above fit well to the data. Here, a curve with c = 1 (i.e., a hyperbolic saturation model) was preferable as it provided an equally good fit while being simpler. The curve predicts that up to about 25% of human cis–antisense pairs are conserved in mouse per the definition of conservation employed here. To estimate whether mouse cis–antisense pairs are likely to be conserved at a similar rate in human, we repeated the analysis in an analogous manner, sampling human transcripts instead of mouse transcripts. A hyperbolic saturation model again fit well to the data and predicted that about 26% of mouse cis–antisense pairs are conserved in human at the saturation level (unpublished data).

Several conserved genes are in cis-antisense or bidirectionally promoted arrangement with nonconserved TUs.

Our saturation estimates indicated that most cis–antisense pairs are not conserved between human and mouse. Accordingly, detailed inspection of homeotic and other transcription factor loci provided several examples of nonconserved cis–antisense and bidirectionally promoted transcripts, some with experimentally supported regulatory roles (see below). We therefore wanted to examine the genome-wide occurrence of nonconserved transcripts in cis–antisense or bidirectionally promoted arrangement with known genes. To find such TUs in the cis–antisense pair dataset, we focused on the subset of cis–antisense pairs where one member (the known gene) had detectable conservation outside the region of antisense overlap, and the other member (the nonconserved TU) showed no conservation outside the region of overlap. Among all 3,442 divergent and convergent cis–antisense pairs in mouse (Table S1), there were only 50 pairs that fulfilled this criterion. We applied the same analysis to bidirectionally promoted pairs. Of the 1,638 bidirectionally promoted pairs in mouse, 40 fulfilled our criterion for conservation of one member only. (We did not perform this analysis on fully overlapping cis–antisense pairs because of difficulties in attributing conservation to individual genes that are completely overlapped by another gene.) Thus, we identified a total of 90 nonconserved TUs in bidirectionally promoted or cis–antisense arrangement with known genes.

These TUs may represent either lineage-specific transcripts, or instances where the location of transcription is conserved between human and mouse, but the transcribed sequence is not. We use the term positional equivalents to refer to the latter: human and mouse TUs that are at genomically equivalent locations relative to well-annotated genes at orthologous loci, but that do not share sequence similarity (Figure 4A). The evidence for positional equivalents was limited, resulting in 16 manually curated positional equivalents involved in cis–antisense pairs, and a further 17 sharing bidirectional promoters with known genes (Table S3). A representative mouse TU with a human positional equivalent is shown in Figure 4B. The 33 identified positional equivalents showed no or weak evidence of protein-coding potential (Table S3). Additionally, their transcribed regions had often been modified substantially after species divergence, via species-specific insertions of repeat elements: 17/33 (52%) mouse TUs with human positional equivalents contained rodent-specific B1–B4 SINEs, and 13/33 (39%) human TUs contained primate-specific Alu SINEs and MER1 elements. In six cases there were both primate-specific repeats in human transcripts and rodent-specific repeats in the corresponding mouse positional equivalents.

Figure 4. Positional Equivalents

(A) Schematic depiction of positional equivalents. By positional equivalents (red arrows), we mean mouse and human TUs that are at genomically equivalent locations relative to well annotated genes at orthologous loci (blue arrows), but that do not share sequence similarity.

(B) Positional equivalents divergently transcribed with the putative tumor suppressor RNH1 [48]. Two transcript isoforms of RNH1 are shown for both mouse (top) and human (bottom). A mouse TU supported by cDNA AK020472 shares a putative bidirectional promoter with Rnh1. The human equivalent (cDNA AK095144) is head-to-head cis–antisense to RNH1. Regions with gray background are within a BLASTZ net alignment of the two genomes. For Rnh1 and RNH1, protein-coding sequence is indicated in dark blue and UTRs in light blue. The positional equivalents lack sequence conservation, assessed by BLASTZ net coverage and BL2SEQ alignment of transcripts, demonstrate gene structure differences, and contain lineage-specific repeats (indicated in black).

doi:10.1371/journal.pgen.0020047.g004

Broad Transcriptional Start Regions and a Mirror Sequence Composition Define Midpoints of Bidirectional Promoters

Bidirectional promoters are associated with broad transcriptional start regions.

Analysis of cap analysis of gene expression (CAGE) data has confirmed two major types of TSS regions associated with different types of core promoters (P. Carninci, A. Sandelin, B. Lenhard, D. A. Hume, Y. Hayashizaki, et al., unpublished data). TATA-box promoters typically initiate transcription from a single position in the genome, while TATA-less promoters can initiate transcription within an interval of 100 bp or more that often coincides with a CpG island [26]. We assembled a dataset of putative bidirectional promoters in the mouse genome well supported by CAGE tag data, and analyzed their genome-wide sequence properties. Bidirectional promoters were identified by scanning for pairs of divergently oriented CAGE tag clusters (TCs) (see Materials and Methods). Our final set consisted of 766 bidirectional promoters, each defined by a divergent TC pair at a separation up to 500 bp. Compared to a control set of 8,056 unidirectional promoters, the bidirectional promoter TCs showed a markedly larger dispersion of CAGE-determined TSS locations (Figure S4). Consistent with this finding, bidirectional promoters were associated with CpG islands more often than were unidirectional promoters (94% of bidirectional promoter TCs were CpG-island-associated, compared to 60% of unidirectional promoter TCs; p < 2.2 × 10⁻¹⁶, Chi-squared test). In addition, CpG islands associated with bidirectional promoters were significantly larger than CpG islands associated with unidirectional promoters (median CpG island sizes of 760 and 557 bp, respectively; p < 2.2 × 10⁻¹⁶, Wilcoxon rank sum test). To experimentally confirm the observed size of transcriptional initiation regions in bidirectional promoters, we used quantitative real-time PCR (qRT-PCR) to measure expression levels in mouse brain RNA samples of different regions near the 5′-ends of transcripts from the genes Ddx49 and Cope, which share a bidirectional promoter (Figure 5). For Ddx49, we could confirm a very low level of expression of the longest transcripts, and much higher expression levels of transcripts initiated further downstream. For Cope, we could confirm great variability within the canonical TSS region, and the existence of an alternative upstream TSS region. Figure 5 demonstrates that the real-time PCR results support the observed distribution of CAGE tags, confirming the breadth of transcription initiation regions and relative TSS usage within them.

Figure 5. TSS Variability at the Ddx49/Cope Bidirectional Promoter in Mouse

(A) The charts show the distribution of CAGE tag 5′-ends over the first five exons of each of the two genes Ddx49 and Cope, and over their intergenic region. CAGE tag mappings indicate that transcription of Cope can start within two wide regions in the first exon of the gene. The initial part of this first exon (hatched) has support from several ESTs, but no cDNA sequences. The three large TCs at the Ddx49/Cope locus span 79, 114, and 150 bp, indicating great variability of transcriptional initiation within each cluster. To confirm the existence of such variability by qRT-PCR, primers (connected boxes) were designed to measure expression of selected regions of the Ddx49 (primer pairs A1–A4) and Cope (primer pairs B1–B5) transcripts.

(B) Detailed view of CAGE tag frequencies and primer locations over the three transcription initiation regions indicated by CAGE tags. Gray lines show cumulative CAGE tag frequencies.

(C) Expression levels of different regions of the Ddx49 and Cope transcripts in adult brain RNA as measured by qRT-PCR. Primer pairs A1 and A2 confirmed low level of expression of the longest Ddx49 transcripts indicated by CAGE (copy numbers in 12.5 ng of total RNA were 3.2 [standard deviation = 1.1] and 5.1 [standard deviation = 3.0] for A1 and A2, respectively). Primer pair B1 confirmed transcription of Cope from upstream of the canonical initiation region. Primer pairs B2–B4 supported variability of transcriptional initiation within the canonical region.

doi:10.1371/journal.pgen.0020047.g005

Bidirectional promoters display a mirror sequence composition.

The two divergently oriented transcription start regions identifying a bidirectional promoter were generally closely spaced, but for only 12% of bidirectional promoters did the TCs overlap by one or more bases (Figure S4). To investigate an association between the separation of divergent TCs and the sequence composition of bidirectional promoters, we aligned the entire set of bidirectional promoter sequences at the midpoint between the TCs and visualized the result as a compositional sequence logo (Figure 6). On the genomic plus strand, there was an apparent excess of cytosines to the left of the midpoint, and a corresponding excess of guanines to the right of the midpoint. There was also small excess of adenines to the left of the midpoint and a corresponding excess of thymines to the right of the midpoint. This mirror-image sequence composition is a landmark of bidirectional promoters, making them markedly different from unidirectional CpG-island-overlapping promoters or random genomic regions (Figure S5).

Figure 6. Landmark Sequence Composition of Bidirectional Promoters

We defined the midpoint of a bidirectional promoter as the midpoint between the most 5′ TSS in each of the two divergently oriented TCs defining the bidirectional promoter. Sequences corresponding to the region spanned by the TCs were extracted from the genomic plus strand. All bidirectional promoter sequences were aligned at their midpoint and the logo created with WebLogo [49]. The logo displays the four nucleotides ranked by their frequency at each position, so that more common nucleotides appear above less common ones. The charts above the logo show the distribution of CAGE tag 5′-ends mapping to the plus strand (upper chart) and minus strand (lower chart) around bidirectional promoter midpoints. The CAGE tag distribution was computed as the sum of tag counts at each position over all bidirectional promoters. The peak of nearly 5,000 tags on the plus strand is due to the Rps2 gene, which appears to be most highly expressed from a single TSS.

doi:10.1371/journal.pgen.0020047.g006

CpG islands often contain multiple binding sites for the transcription factor Sp1 [26]. Considering that the Sp1 binding consensus motif is GGGGCGGGGT [27], the bias in guanine and thymine frequencies we observed across the midpoints of bidirectional promoters would be consistent with a corresponding bias in directionality of Sp1 binding. We scanned the region to the right of bidirectional promoter midpoints for putative Sp1 sites and found 43% more sites on plus strands than on minus strands. The relationship was reversed to the left of the midpoint (60% more binding sites on minus strands than on plus strands). On both sides of the midpoint, both plus and minus strands were significantly enriched for putative Sp1 binding sites compared to random sequences with the same lengths and background nucleotide frequencies (emitted from a first-order Markov chain to preserve dinucleotide composition) (Figure S6). This supports the idea of Sp1 as the probable key general transcriptional factor that binds to CpG island promoters [26].

Coexpressed cis–Antisense Pairs Display Conserved Overlaps Containing Noncanonical TSSs

To pinpoint cis–antisense pairs with regulatory interactions between pair members, we concentrated on 242 pairs from mouse with available microarray expression data for 61 tissues (GNF1M data; [28]). We selected only probesets that mapped to regions of exon overlap between cis–antisense partners, in order to avoid detecting mixed signal from antisense-overlapping and nonoverlapping transcript isoforms. The majority (84%) of pairs with available probesets were convergently (tail-to-tail) overlapping. We found a significant positive correlation across the entire panel of tissues for 58/242 (24%) cis–antisense pairs at the 0.05 level, and a significant negative correlation for only 14/242 (6%) pairs. After correcting for multiple testing, 17/242 (7%) pairs remained significantly positively correlated, and no pairs remained significantly negatively correlated. We assessed how likely it would be to obtain this result if TUs were paired at random. In only three out of 10,000 sets of 242 random TU pairs did we obtain 17 or more significantly correlated pairs, and none of the sets contained more than 14 significant positive correlations (Figure 7). By the same methodology, members of bidirectionally promoted pairs were also found to have positively correlated expression profiles more often than would be expected by chance (unpublished data). The 17 cis–antisense pairs identified as positively correlated all belonged to the convergent class (Table S4). For 15 of these pairs, the overlap included UTRs at the 3′-ends of both TUs. Of these exon overlaps at apparent noncoding regions, 11 contained stretches of high conservation. In three cases, this conservation was clearly limited to the overlap region (Table S4), indicating possible functional importance of exon overlaps [22]. Inspection of CAGE tag mappings to the 17 positively correlated cis–antisense pairs revealed that for 13 pairs there was evidence of TSSs at one or both of the 3′-ends involved in the overlap (Table S4). The largest 3′-end TC (supported by 18 tags) among coexpressed cis–antisense pairs was observed in the 3′-UTR of Ppp1ca, which encodes a catalytic subunit of a protein phosphatase required for cell division. The 3′-UTR of Ppp1ca is conserved in human, and in both genomes overlaps by about 220 bp the 3′-UTR of Rad9, which encodes a cell-cycle checkpoint protein required for DNA damage repair [29].

Figure 7. Members of cis–Antisense Pairs Have Positively Correlated Expression Profiles More Often than Expected by Chance

Out of 242 murine cis–antisense pairs with expression data for 61 tissues, 17 showed significant positive correlation across the entire set of tissues after correction for multiple testing, and no pairs showed significant negative correlation (red squares). The same test was applied to 10,000 sets of 242 random TU pairs (box plots, with circles indicating outliers), demonstrating that members of cis–antisense pairs have positively correlated expression profiles more often than expected by chance.

doi:10.1371/journal.pgen.0020047.g007

Chains of Overlapping TUs Occur in Gene-Dense Areas with Antisense Transcription

Next we investigated whether local gene density was related to the incidence of antisense transcription. To avoid bias due to the fact that cis–antisense pairs will always have an average density higher than that of individual random genes, we examined 100-kbp regions directly flanking each pair, rather than the region covered by the pair itself. Regions flanking cis–antisense pairs had roughly 30% higher TU density and 30% more exon sequence than regions flanking TUs not involved in cis–antisense pairs (Table S5). Many TUs in our dataset formed cis–antisense pairs and/or bidirectionally promoted pairs with several other TUs. To quantify this phenomenon, we searched for chains of bidirectional transcription, where we defined a chain as a group of three or more TUs associated by cis–antisense and/or bidirectionally promoted arrangement. Since TUs represent clusters of transcript sequences that in cases of incomplete coverage might not correspond to entire genes, and in rare instances contain sequence from adjacent genes [1], we applied strict rules on TU structure in order not to overestimate the occurrence and extent of chains (see Materials and Methods). In human we detected 1,480 chains, containing 5,263 TUs (12 % of all TUs). In mouse, there were 1,153 chains, containing 3,987 TUs (11% of all TUs) (Table 3). The largest computationally predicted chain involved 11 TUs: the human gene encoding the giant muscle protein titin, nine antisense TUs overlapping titin exons, and one TU that might represent an alternative 3′-end of titin transcripts. The titin chain aside, the largest human chains involved eight TUs, and the largest mouse chains involved seven TUs. An example of a five-TU chain from mouse is given in Figure 8. This region contains a gene encoding a well studied transcriptional regulator (Hsf1) and three metabolic genes, allowing the possibility of cis-regulation of genomically adjacent genes of diverse functions and resulting effects on their downstream targets. The genomic distribution of chains, cis–antisense pairs, and bidirectionally promoted pairs is illustrated in Figure S7. Human Chromosome 19, which has the highest gene density of all human chromosomes [30], also had the highest densities of cis–antisense pairs, bidirectionally promoted pairs, and chains.

Table 3.

Numbers of Chains Detected and Their Sizes

doi:10.1371/journal.pgen.0020047.t003

Figure 8. A Five-TU Chain on Mouse Chromosome 15

TUs on the genomic plus and minus strands are shown in dark gray and light gray, respectively (boxes represent exons). CpG islands are shown as black boxes. From left to right, the chain contains a member of the aminoacyl tRNA transferase class II family (D330001F17Rik), which has two cis–antisense transcripts: fully overlapping (cDNA AK034666) and convergent (Bop1). The latter encodes a ribosome biogenesis protein and shares a CpG-island bidirectional promoter with the heat-shock-induced transcription factor 1 gene (Hsf1). Hsf1, in turn, is convergently cis–antisense to the diacylglycerol O-acyltransferase 1 gene (Dgat1).

doi:10.1371/journal.pgen.0020047.g008

Nonconserved and Noncoding TUs Are Transcribed Antiparallel to Many Homeotic Genes

Antisense transcripts to the HOXA11 gene have been found to be conserved between human and mouse [31], and we have recently reported chains at the HOXA cluster in both organisms [13]. In mouse, the Hoxa3 and Hoxa7 genes formed a chain together with two uncharacterized TUs. In human, there were three chains of three TUs each, including HOXA3, HOXA4, HOXA9, HOXA10, HOXA11, and four uncharacterized TUs. We have further shown that several noncoding transcripts in the human HOXA cluster are coexpressed with adjacent coding HOXA genes in various human tissues, and are likely to be involved in the opening and closing of chromatin and sequential transcriptional activation of HOX cluster members (L. Sessa, A. Breiling, G. Lavorgna, L. Silvestri, V. Orlando, et al., unpublished data). Based on these findings, we selected homeotic genes as a group to focus on. HOX genes are arranged into four clusters in both human and mouse genomes [32]. Within each of the four clusters, the HOX genes are transcribed in the same direction. To assess the extent of transcription from the opposite strand at HOX loci and around dispersed homeotic genes (structurally and functionally related to the HOX genes), we specifically searched for EST sequences that mapped to the opposite strand at such loci and either overlapped homeotic genes or were intergenically located. We detected 232 human and 46 mouse ESTs on the opposite strand at HOX loci (Table 4), and a total of 445 opposite-strand ESTs distributed over 53 out of 95 dispersed human and mouse homeotic loci analyzed (Figure 9). The detected ESTs did not display any significant open reading frames or similarities to known proteins. Thus, transcription from the noncod