The sequences of known protein-coding genes for five fully sequenced metazoan genomes (H. sapiens, M. musculus, D. rerio, D. melanogaster, C. elegans) were retrieved from several sources (RefSeq v.10, UCSC mm7 assembly, WormBase WS140, Flybase r4.2, Riken Fantom 3.0). From each dataset, we filtered splice variants and removed non-coding transcripts, pseudogenes and purely computational gene predictions, and mapped each cDNA on the corresponding genome to extract the Overlapping Gene Clusters (OGCs). OGCs were detected when there was partial or total overlap between the genomic coordinates of two or more genes. Gene boundaries were defined as the start and the end of the longest transcript (the complete list and features of OGCs are provided in Additional files 1 and 2). Our selection criteria allowed the detection of OGCs laying both on the same (parallel) and on opposite (antiparallel) DNA strand (Figure 1). Although we started from restrictive datasets, our estimates of overlapping protein-coding genes (Table 1) were consistent with previous analyses in human, mouse and Drosophila 1327293031. According to our results, overlap involves 4–8% of protein-coding genes, with the exception of Drosophila, where the percentage of OGCs is higher (26.2%, Table 1).

We compared the observed data on overlapping genes to a null model that simulates the distribution of expected events in case of neutrality. For each species, we re-assigned random positions to the individual genes within each chromosome and counted the resulting number of overlaps.

In all species the overall number of observed OGCs was significantly lower than randomly expected (Table 1), suggesting selection against the retention of overlap as a general mechanism of gene arrangement. There are at least two reasons possibly explaining the counter selection of gene overlap in Metazoa. First, each mutation occurring within the overlapping regions would affect two or more sequences at the same time, and would likely reduce the ability of the involved genes to become optimally adapted 32. Second, overlap can result in transcriptional 3334 or translational 35 interference between overlapping reading frames. Both these reasons help to explain why OGCs formed by several genes, as well as those involving coding sequences, are particularly selected against (see below).

Although overlap of protein-coding genes is generally counterselected, some classes of overlap are preferentially retained. Comparison to random expectation showed that observed OGCs display a non-random distribution in terms of their abundance, reciprocal orientation, and overlap pattern (Table 1 and Figure 2).

While the number of random OGCs varied according to the different gene density of the analyzed species (Table 1 and Additional file 2), this tendency was not maintained in the observed data. Observed OGCs in human and mouse were around 4–5 times less than expected, while they were ~2 times less in fly and ~12 times less in worm. In agreement with our observation, a remarkable abundance of antisense transcripts in fly and a paucity in worm have been recently reported 1214. The different rates of overlapping genes in fly and worm could be due to species-specific features. The higher proportion of overlapping genes in fly might be partly explained by the high gene density and the extended UTR length (Additional File 1). The low number of OGCs in worm may be instead a consequence of the presence of operons, which involve at least 15% of C. elegans genes 36. Each operon contains from two to eight genes which are cotranscribed from the same strand as a polycistronic RNA and trans-spliced 36. It is conceivable that such feature might place a constraint on the plasticity of the worm genome, disfavoring the retention of specific overlap types, such as antiparallel and partial arrangements. Similar genomic constraint has been recently proposed to explain the paucity of duplicated genes in operons 37.

In all genomes except zebrafish, OGCs formed by two genes occurred at a frequency significantly higher than expected (Figure 2A). In addition, OGCs in human, mouse, and fly were mostly formed by antiparallel convergent pairs which overlapped only partially, while in zebrafish and more markedly in worm nested overlaps were preferred (Figures 2B and 2C). However, the results in zebrafish should be taken carefully, since they are probably affected by the poor coverage of the corresponding gene set. Likewise, the annotation of 5' and 3' untranslated regions appears particularly incomplete in worm (Additional file 1), which may contribute to an underestimation of some overlap classes (i. e. partial overlap, CDS/UTR and UTR/UTR overlaps, Figure 2). In all species overlaps between genes lying on the same strand and those sharing coding regions are strongly selected against (Figures 2C and 2D). Overlap between UTRs is preferentially retained in all organisms, while the overlap between coding regions and introns is common in zebrafish, drosophila and worm (Figure 2D). The non-random features of observed OGCs suggest that different overlap types are under different selective pressures. The retention of specific overlapping classes might be allowed when it provides selective advantages: in the case of genes on opposite strands the advantage could be represented by antisense regulation. Human, mouse and fly are significantly enriched in overlapping pairs potentially able to form antisense, which include all antiparallel overlaps sharing exons (H. sapiens 55%, p < 0.001; M. musculus 58%, p < 0.001; D. melanogaster 53%, p < 0.001, chi-squared test). This result suggests that, at least in these species, positive selection might act to preserve antisense regulation. It cannot be excluded, however, that part of the positive effect could be a consequence of the negative selection towards parallel and CDS/CDS overlaps.

We next evaluated the conservation of OGCs across metazoan evolution by verifying both the presence of orthologous genes and the overlap conservation. For each pair of analyzed species, we assigned pairwise orthology for all sequence entries, extracted the orthologs involved in OGCs, and verified whether the overlapping arrangement was conserved (Figure 3A). Most overlapping genes in one species had their corresponding orthologs in the others (Figure 3B), but very few overlaps were maintained (Figure 3C). In total, ~40% of human OGCs were also present in mouse -a higher percentage than previous estimates (6.6–17%) 6133338, but lower than the rate of orthologous genes between the two species (75.6%).

Among OGCs conserved between human and mouse, the antiparallel arrangement was represented the most (~88%), highlighting again the tendency to maintain possible sense-antisense regulation. Interestingly, convergent and nested antiparallel arrangements were significantly enriched in the conserved set (chi-square = 22.47, p = 2.14e-6 and chi-square = 23.55, p = 1.2e-6, respectively), when compared to divergent overlaps (Table 2). This result supports previous observations that 3'-3' (convergent) overlapping pairs are significantly more conserved than 5'5' (divergent) ones, and indicates a prevalent role for 3'UTRs in antisense regulation 1439.

Parallel OGCs did not show any significant enrichment in the conserved set (Table 2). Since same-strand overlaps are strongly selected against (Table 1), we investigated whether the ones that are conserved are more likely to be functional. Indeed, we found that several parallel OGCs conserved between human and mouse might be functionally related on the basis of the available literature data (Additional data file 3).

Although the vast majority of overlap is not conserved over long evolutionary distances, we found evidence of few ancient overlaps. Overall, three OGCs were conserved between Ecdysozoa (nematodes and arthropods) and Deuterostomia (vertebrates). Interestingly, the only OGC that is conserved from C. elegans to human was lost in arthropods, while two different OGCs are conserved from D. melanogaster to human. All of these OGCs are formed of two genes with a nested antiparallel arrangement. One of the two clusters conserved in D. melanogaster (Cluster 77, Additional File 2) involves the synapsin (Syn) and an inhibitor of metalloproteinase (Timp) genes. According to the model proposed for the evolution of the Syn-Timp cluster 40, the locus containing the ancestral nested genes has undergone gene duplications and losses in vertebrates, followed by function partitioning among the resulting paralogs. A comparable succession of events is compatible also with the evolution of the only OGC conserved between vertebrates and worm (Cluster 371, Additional File 2). In this case, the ancestral OGC locus seems to have undergone duplication after the split between Protostomia and Deuterostomia, followed by function partitioning among the resulting paralogs (Additional file 4).

The poor evolutionary conservation of gene overlap in Metazoa suggests that its occurrence is species-specific. Such species-specificity was not due to a recent origin of the overlapping genes, as previously suggested 21332. We found that most overlapping genes in one species had orthologs in the other species, although they did not overlap (Figures 3B and 3C). In addition, 30.2% of human overlapping genes and 25.8% of mouse overlapping genes remained physically adjacent in the compared genome, although the superimposition was lost (see below).

There are examples of functional processes whose poor conservation during evolution is part of their functional role, alternative splicing being the most striking one 41. Although approximately two-thirds of human genes are alternatively spliced 42, only 10–20% of them conserve the spliced exons in the orthologous genes in mouse 43. Hence we can propose a species-specific usage of gene overlap similarly to what seems to happen for alternative splicing 41.

In order to infer possible mechanisms for overlap formation, we compared the gene structure (gene length and exon number) of conserved and non-conserved overlapping genes in human and mouse. In particular, we analyzed the gene structure of human and mouse overlapping genes whose orthologs lie adjacent (i. e. without any gene between them) but do not overlap in the other genome. We found that 226 human overlapping gene pairs (corresponding to 30.2% of the total) and 171 mouse overlapping gene pairs (25.8% of the total) had orthologs that do not overlap but remain adjacent in the genome of the other species (Table 3). The 226 human overlapping gene pairs were significantly longer (z' = 2.53, p = 5.7e-3, Mann-Whitney U-test 44, Table 3) and had more exons (z' = 2.72, p = 3.3e-3) than the mouse orthologs, when compared to the set of conserved overlapping genes (Table 3). Similarly, the 171 human orthologs of mouse overlapping gene pairs were shorter (z' = 2.95, p = 1.6e-3) and were formed with fewer exons (z' = 2.28, p = 1.1e-2) than the conserved overlapping pairs. In addition, non-conserved overlapping gene pairs tended to significantly overlap in their UTRs for both human (chi-square = 23.4, p = 1.3e-6) and mouse (chi-square = 24.2, p = 8.9e-7), when compared to the conserved set (Table 3).

The structural analysis of orthologs of human and mouse overlapping genes that remain adjacent but lack the superimposition shows that the overlap formation is frequently associated with an increase in gene size and exon number. We therefore suggest that the overlap between adjacent genes may originate by species-specific acquisition of additional, non-coding exons. In agreement with our results, most of the loci analyzed by the ENCODE consortium were found to possess distal 5' non-coding exons which map into neighboring genes and tend to be tissue- or cell-line-specific 45.

In order to evaluate whether the presence of overlap is a mechanism for regulation of gene expression, we used the human OGC dataset to cross-examine a human breast cancer transcriptome obtained by massive pyrosequencing 46. To be able to detect the expression of transcripts normally expressed at low levels, we used a normalized cDNA library (see Methods). For this reason, our analysis is mostly qualitative and aims to detect the reciprocal expression of genes involved in overlap. Although global gene expression can result quantitatively altered by the tumorous condition, a significant modification in the pattern of reciprocal expression between overlapping genes is unlikely. We defined three patterns of reciprocal expression: co-expression, when both genes were represented in the library; discordant expression, for OG pairs in which expression is observed for only one gene in the pair; and no expression, for OGs whose expression was not detected. Figure 4 shows the frequencies of these three expression patterns in the breast cancer library, by grouping the OG pairs according to the type of overlap.

The observed rate of co-expression in the whole dataset was 27.6%, while the percentage of discordant expressed OGs was 42.5%. Taking into account the overall coverage of known genes in our cDNA library, the co-expression rate is four times higher than expected by the random probability of having any two genes expressed at the same time in the library (7.3%). Therefore, OGs showed a significant tendency to be co-expressed (upper cumulative distribution function, p = 6.7e-102). It should be noted that we obtained significant co-expression even though we removed all sequences mapping to more than one gene in the same cluster (see Methods). Such filtering step likely led to an underestimation of the level of co-expression of overlapping genes, but it did not influence the final result. By contrast, the percentage of discordantly expressed genes is not significantly different from random expectation (upper cumulative distribution function, p = 0.043). Previous studies reported higher co-expression rates, ranging from 35.1% to 44.9% 1047, with the differences likely due to experimental design (i. e. differences in the starting dataset) and in the number of analyzed tissues.

Considering the different overlapping arrangements, we also observed that co-expression was significantly higher for both convergent (chi-square = 4.69, p= 3.03e-2) and divergent OGs (chi-square = 4.28, p= 3.85e-2), when compared to the frequency of the complete overlaps. On the opposite, we observed no statistically significant differences among overlapping arrangements when considering discordantly expressed OGs. Taken together, these results further support the hypothesis that gene overlap might be used to co-ordinate expression of adjacent genes.

The RefSeq cDNA sets 48 for five organisms (H. sapiens, M. musculus, D. rerio, D. melanogaster, and C. elegans) were downloaded from the UCSC ftp site (RefSeq v.10, March 2005) 49. We also retrieved mouse cDNAs from the RIKEN database (Fantom 3.0) and the UCSC collection of mouse cDNAs (Mm7 assembly), while for fly and worm we used Flybase (FlyBase r4.2) and Wormbase (WormBase WS140), respectively 50515253.

The genomic position of each sequence was mapped on the corresponding genome by using BLAT 54 (human Build 35; mouse Build 34; zebra fish Zv4; fly Release 4; worm WS120). The pairs of genes whose genomic coordinates partially or totally overlap were extracted and grouped in OGCs. Filters were adopted to avoid (a) splice variants of the same gene, and (b) artifacts due to the position mapping. We considered each pair of cDNAs sharing three or more exons as splice variants of the same gene if more than 20% of the exon number overlapped. In the case of cDNAs with two or less exons, we considered them as splice variants if at least one residue overlapped at the exon level. For each group of predicted splice variants, only the longest gene was taken as gene representative. Artifacts such as the inclusion of the mRNA poly-A in the gene mapping were avoided by excluding all the 3' exons composed of more than 70% of one single nucleotide.

For all five species analyzed, the gene positions of the unique gene sets were randomly reassigned within the corresponding chromosomes with no constraints in the type of overlaps, the reciprocal arrangement, and the number of genes per cluster. The analysis was repeated for 10 rounds and the resulting number of overlapping genes, overlapping gene pairs, and overlapping gene clusters were counted at each round. The average number was considered for comparison with the observed dataset. Features of the OGCs, such as the reciprocal arrangement, the component distribution and the type of overlapping region were also analyzed.

The fraction of overlaps that results in sense/antisense complementarity at the mRNA level were calculated by extracting all overlap that occur on opposite strands and involve exons of both genes. The statistical significance of the difference between the observed and the random set was assessed by applying a chi-squared test (degree of freedom = 1) to the resulting 2 × 2 contingency matrix 44.

To test the specificity of the data produced, we performed a manual analysis of the D. rerio dataset (108 OGCs). No obvious false positive due to the methodology could be found. The sensitivity of our method was assessed by benchmarking the derived set against an extensive collection of overlapping genes previously reported. We included 8 independent large-scale screenings of human antisense transcripts/nested genes 913272930555657 and about 100 experimental studies on specific overlapping gene pairs (Additional files 5 and 6). OGCs reported in the literature with no match in our dataset were checked manually. The main reasons for the lack of coverage were due to the selection criteria (i. e. we deliberately excluded pseudogenes or non-coding RNAs which were instead included in some large-scale screenings). Only 5 cases were found to be false negatives, giving an estimate specificity of 99%.

The orthology relationships between the overlapping genes in the five analyzed species were assessed by using a two-step procedure (Figure 3A). First, for all pairs of species we carried out all-against-all tBLASTx 58 between the corresponding cDNA sets. The best reciprocal hits between two species were assigned as orthologous genes. Secondly, we derived orthologous overlapping genes by extracting all overlapping genes conserved between each pair of species.

We compared the gene structure of the conserved OGCs between human and mouse with human and mouse overlapping genes whose orthologs do not overlap but are adjacent in the genome of the other species. The first set (conserved overlapping genes between human and mouse; the first column in Table 3) was composed of 282 pairs of overlapping genes, while the second (overlapping in human but adjacent in mouse chromosomes; the second column in Table 3), and the third (overlapping in mouse but adjacent in human chromosomes; the third column in Table 3) were composed of 226 and 171 gene pairs, respectively. For each gene, we measured the gene length, defined as the genomic coordinates on the corresponding chromosome, and the exon numbers, as derived from the BLAT output. Using the Mann-Whitney U-test 44 we compared gene length difference between the first and the second sets, and between the first and the third sets to assess the statistical significance of the difference in gene structure.

We also analyzed the feature of the region (UTR or coding) involved in the overlap for all OGCs in the 3 sets, by counting the number of detectable overlaps after removing the UTRs. In this case, the statistical significance of the difference between the first and the second sets and the first and the third set were assessed by applying a chi-squared test (degree of freedom = 1) to the resulting 2 × 2 contingency matrix 44.

cDNA was obtained from polyadenylated breast cancer RNA (purity 85–90%). cDNA was normalized after reverse transcription to obtain a balanced mix of low and high abundance mRNA, as previously described 59. 2.1 micrograms of normalized, double-strand cDNA were then converted to a single strand library using the 454 protocol 46. Two independent cDNA libraries were generated with an average length per sequence read of 100 and 200 nt, respectively. A total of 198,658 non-redundant sequence reads, according to NCBI non-redundant database, were sequenced from each breast cancer cDNA library. The entire library was mapped against the 249,953 sequences of the human "all_mrna" transcript dataset from the UCSC human genome. A total of 37,774 reads corresponding to a specific cDNA and its related isoforms was identified (requiring blat perfect matches, 95% of the read covered by alignment). The reads were then aligned to the human RefSeq cDNA dataset from UCSC (25,922 sequences) requiring perfect coverage. 9,082 distinct matches were finally obtained, which were used for the subsequent calculations.

Reads-to-gene assignment was performed by blasting the nucleotide sequences of all OGs to the library. Only reads showing 100% identity with a transcript were used in the analyses. To ensure the 454 sequences were unambiguously matched to the assigned transcript, we removed reads mapped to more than one locus. Since the 454 sequencing process does not involve in-vivo cloning and the cDNA is subjected to nebulization, in the deriving library it is not possible to assign the strand when the two transcripts overlap. Thus, we removed all sequence reads mapping to more than one gene within the same cluster. In total, 36 out of 3701 reads were removed, corresponding to an estimated loss of 0.9%, which likely did not create a significant bias.

The statistical significance for the enrichment of co-expression in overlapping gene pairs was evaluated by an upper cumulative distribution function.