To identify transcribed pseudogenes, transcript sequences from the Unigene, RefSeq and H-InvDB databases were mapped onto the human genome and were examined for overlap with pseudogene annotations. These pseudogene annotations were taken from the VEGA http://vega.sanger.ac.uk/ and http://pseudogene.org websites (see Methods for details). We pooled these datasets with re-mappings of: (i) 'disrupted mRNAs' (dmRNAs) 16, and (ii) transcribed processed pseudogenes 14, from previous analyses 1416. Overlap of the transcripts with pseudogenes was verified through using the positions of mid-sequence disablements (frameshifts and premature stop codons) as 'anchors'. That is, at least one disablement position was required to occur in both the genomic DNA and the transcript sequence (see Methods for further details).

We found that ~11.5% (1750/15000) of human pseudogenes are transcribed (after correcting for pseudogene annotation overlaps within and between the various data sets) [see Additional file 1]. Table 1 summarises the numbers of transcribed pseudogene annotations (TPAs) in different categories and data sets. The number of processed pseudogenes that were identified to be transcribed is 3-4 times higher than in our previous analysis 14. Interestingly, in humans, regardless of category, only a small fraction of TPAs are transcribed completely in the antisense direction (~3-6%). Such a finding of significant avoidance of antisense transcription (Table 1) is surprising, especially for retropseudogenes. Retrotransposed sequences are inserted back into the genomic DNA irrespective of the position of existing local promoters. Thus, one would expect equal numbers of sense and antisense transcripts. However, the above finding indicates a general selection pressure against antisense transcribed pseudogenes, thus generally limiting the possibilities for complementary hybridization with transcripts and RNA elements from homologous genes.

We performed a similar survey in the mouse genome for TPAs. Surprisingly, in the mouse genome, we found a very low percentage of TPAs, in comparison to the human genome (<2%) (P << 0.001 for the likelihood of the number in mouse, given the human percentage as an expectation, using binomial statistics). This is despite these two animals having similar amounts of pseudogene annotation data (Table 1), and transcript data (203,785 transcript sequences in total for human, and 203,550 for mouse). This indicates that transcribed pseudogenes are rarer in mice than in humans.

Transcription of pseudogenes per se does not necessarily indicate functionality. It has been shown that transcriptional activation at a particular genomic locus has a ripple effect on the neighboring loci 21. It is therefore possible that many transcribed pseudogenes arise simply because of this. However, of the various identified human TPAs in our present study, those that are evolutionarily conserved across mammals due to natural selection are more likely to be biologically functional. Therefore, we set out to identify a list of such orthologous transcribed pseudogenes that have conserved in ≥2 of our panel of mammals.

Certain gene families are known to spawn large numbers of pseudogenes. Examples include olfactory receptors 22, ribosomal-protein genes 23, human thioredoxin and glutaredoxin 24, ABC transporters 25, and heat shock proteins 26. In such cases, identifying orthologs in other mammals using the standard bi-directional best-hit procedure is problematic, since the rates of sequence evolution may vary in different lineages and genomic regions. Furthermore, such a procedure does not work well for pseudogenes, since these sequences are not evolving like protein-coding genes, which are under strong purifying selection. Because of this, the best match obtained using blastp to a pseudogene query is expectedly the parent protein-coding gene or a pseudogene recently evolved from the parent gene. Thus, the standard bi-directional best-hit procedure alone is not sufficient. Therefore, here, we have used synteny information between two organisms to pin-point pseudogene orthologs. We have used synteny maps along with homology-based searches to identify conserved orthologs in five mammals (rhesus monkey, mouse, rat, dog and cow) (see Methods for details). We identified a set of 68 human TPAs that are conserved in at least two of these mammals, representing potentially functional candidates (see Additional file 2: Table S1). In general, although approximately half (742/1750) of the human TPAs are syntenically conserved in rhesus monkey, only 3% are syntenically conserved in mouse. This suggests that a large number of human transcribed pseudogenes are primate-specific.

A multiple sequence alignment of orthologous sequences for an example taken from Additional file 2: Table S2, is shown in Figure 1(a). The corresponding phylogenetic tree is drawn in Figure 1(b). This example is a human pseudogene named 'urn:lsid:pseudogene.org:9606.Pseudogene:4346', that is homologous to human ADP-ribose pyrophosphatase. In this case, one can see clearly that disablements at several positions in the alignment are conserved in divergent species, and parsimoniously would be assigned in the ancestral sequence. Also, in this phylogenetic tree, dog clusters closer to primates, than rodents do; this may be due to variance in local genomic mutation rates. Interestingly, we find that a significantly higher number of human transcribed pseudogenes were conserved in dog, compared to in mouse (Fisher's exact test, P-value: 0.0086) (Figure 2). There is some debate regarding whether human is phylogenetically closer to rodents than to dog, although most data analysis indicates a rodent-primate grouping 272829.

It is noteworthy that the highest number of conserved TPAs is on the human chromosome X (13 out of the total of 68; Figure 3 and Additional file 2: Table S2), followed by 11 cases on chromosome 6. There is also a significant over-representation of human conserved TPAs on these chromosomes after normalizing for the chromosome size and dosage in gametes (χ² test, d.f. = 1, P-value < 10^-3). Furthermore, it is chromosome X that is consistently and most significantly overrepresented in the whole population of TPAs, and in the datasets of TPAs that are conserved in monkey and in mouse (calculations not shown). This finding is in line with the observation that over the course of evolution there has been some extensive gene trafficking to/from the mammalian X chromosome 30.

Human TPAs from the VEGA data set, that have syntenically-conserved orthologs in rhesus monkey, were analyzed for significant selection pressure to maintain them. This was assessed through comparison of the nucleotide percentage sequence identity between orthologs, with the highest nucleotide percentage sequence identity for the immediately flanking genomic regions, as illustrated in Figure 4. We chose the VEGA data set for this analysis, since the genomic coordinates of this pseudogene annotation data set are more precisely annotated.

Our analysis indicated that TPA orthologs have higher sequence homology in comparison to their syntenic flanking regions. Average sequence identities among different syntenic regions in human and rhesus monkey are as follows: 75.0% (s.d. 12.6%) in the 5' areas, 81.0% in the 3' areas (s.d. 13.5%), and 86.7% (s.d. 9.6%) in the TPA sequences. The difference in the sequence identities between pseudogenes and the respective flanking regions is statistically significant (Wilcoxon signed rank test, P-value: < 5e^-50). In the majority of cases (~86%, 293/341), the percentage sequence identity between orthologous TPA sequences is greater than that of the flanking regions. This suggests that significant selection pressure exists to maintain them. We note that similar analysis comparing the human and chimpanzee genomes is not informative because the species are too similar. Also, comparisons of the human genome with the other mammals in our panel are not informative either, because the appropriate regions cannot be aligned accurately or significantly.

We examined whether there existed any sequence feature that distinguished conserved TPAs from the rest of the human pseudogene population. A positive finding would indicate that these pseudogenes are not evolving neutrally. It has been widely observed that genes tend to be GC-rich in comparison to non-transcribed genomic segments 31. Here, we examined whether TPAs and annotations of non-transcribed pseudogenes showed any difference in the GC contents relative to their flanking regions (GC_diff). Pseudogene populations from a variety of organisms have been shown to relax to the composition of intergenic DNA over evolutionary time 3233. Here, the GC content of neutrally-evolving pseudogenes would be expected to relax to that of the background genomic GC content. Interestingly, GC content calculations revealed that 84% (327/391) of the human TPAs derived from the VEGA pseudogene data set, that are conserved in rhesus monkey, have GC content greater than their flanking regions (Table 2). This compares to 74% (5289/7154) for the non-transcribed cases (Table 2). This difference is statistically significant (χ² test, d.f. = 1, P-value < 10^-4). A similar result is obtained if we examine the whole population of TPAs, or also if we just examine transcription of conserved processed pseudogenes. There is however no such significant differences for transcribed pseudogenes of the 'nonprocessed' and 'unclassified' pseudogene categories (Table 2). This shows that there is a greater tendency for the conserved TPAs to be GC-rich than for non-transcribed cases, and that this tendency arises primarily because of transcription of processed pseudogenes. This finding on GC content is further evidence that such transcribed pseudogenes are not evolving neutrally. Such GC trends can be explained by selection for transcriptional efficiency, as noted above 31.

We checked whether the age of pseudogenes could be causing the GC content differences noticed above. To do this, we looked at the GC_diff in the following (VEGA) subsets of TPAs, i.e.: (i) TPAs unique to humans; (ii) TPAs conserved in rhesus monkey only; (iii) TPAs conserved in more divergent mammals such as mouse, rat, dog and cow. We found that 82.5% (381/462) of set (i) have GC_diff > 0 (i.e., GC content of pseudogene greater than that of the flanking region). Similar percentages were observed in the other classes: 84.1% (317/377) in set (ii), and 77.78% (21/27) in set (iii). There was no statistically significant difference in the GC_diff between any two of the classes (χ² test, P-value > 0.55), suggesting that age of pseudogenes does not have any influence on the observed GC content differences.

We decided to assess the genomic evidence for a lack of protein-coding ability in the human TPAs that are syntenically conserved in rhesus monkey. We compared TPA characteristics to the characteristics of two other groupings: (i) known human protein-coding genes with orthologs in rhesus monkey; (ii) populations of simulated sequences that are randomly mutating without coding-sequence selection pressures. The human TPA sequences are used as starting sequences for these simulations. The protocol for these simulations is described in 'Methods: Ka/Ks ratio calculations'.

The ratio of non-synonymous to synonymous substitution rates (Ka/Ks) provides a measure of selection pressure for protein-coding ability on nucleotide sequences. Classically, values significantly <<1.0 indicate purifying selection, whereas sequences without coding ability theoretically yield values near 1.0. We examined the trends for Ka/Ks in the population of human TPAs that are conserved syntenically in the rhesus monkey. Ka/Ks was calculated using the PAML package (as described in Methods). The distribution of Ka/Ks is shown for TPAs, split into two groups, those that have a disrupted protein domain of known three-dimensional structure (TPA_DD, blue bar, Figure 5), and those that do not (TPA_NDD, red bar, Figure 5).

In addition, we calculated the distribution of Ka/Ks values for sequences that are randomly mutating without coding-sequence selection pressures. These sequences were generated using the simulation protocol described in 'Methods: Ka/Ks ratio calculations', using the human TPAs as starting sequences (green curve, Figure 5). The Ka/Ks distribution for these simulated sequences does not peak at ~1.0, as would be classically expected. This is likely due to some inaccuracy in modeling the expected frequency for the different possible nucleotide substitutions, which varies for different genomic areas 3. The distribution for TPA_DDs peaks in the range 0.6 to 1.0. This peak is similar for the randomly-mutating sequences (Figure 5). For the TPA_NDDs, the peak is at lower Ka/Ks values (0.4-0.6).

As a further comparison, we have calculated the Ka/Ks curve for orthologous pairs of protein-coding genes from the rhesus monkey and the human (blue curve, Figure 5). Clearly, these protein-coding sequences behave very differently from the TPAs, with a substantial mode in the range 0.0 to 0.2. In summary, these Ka/Ks trends indicate that the substitution patterns in the TPAs generally behave like non-protein-coding sequences, and not like protein-coding ones. This is despite the overall significant conservation relative to surrounding intergenic genomic DNA that was discussed in the previous section.

To gain a more complete picture, we also examined Ka/Ks values for TPAs that are conserved in two more divergent species, the dog and the mouse. We compared Ka/Ks values for orthologous TPA pairs (termed Ka/Ks_Ψ-ortho), with the corresponding Ka/Ks values for their parent genes (Ka/Ks_parent-ortho) (Figure 6). These were calculated for human/dog (Figure 6(a)), and human/mouse comparisons (Figure 6(b)). For human/dog comparisons, the substantial majority (83%) have Ka/Ks_Ψ-ortho > Ka/Ks_parent-ortho, whereas for human/mouse all of the pseudogene pairs have larger Ka/Ks values than their corresponding parent pairs.

The Ka/Ks results suggest that these transcribed pseudogenes are relaxing to higher Ka/Ks values, since origination from their parents. But why do they not have Ka/Ks values of ~1.0? We suggest that this is chiefly because: (i) there may be some inaccuracy in modelling the expected frequency for the different possible nucleotide substitutions, which varies for different genomic areas (as noted in the previous section); (ii) in some cases, the transcribed pseudogenes were originally protein-coding, and became disabled subsequently in multiple lineages; (iii) some of them maintain an imprint of the original coding sequence because of selection pressure for regulation of homologous genes via antisense interference (e.g., through genesis of siRNAs); (iv) selection pressures on non-synonymous codon substitution rates in protein-coding genes, may have relaxed in the pseudogenes, contributing to an apparent relative increase in Ks; (v) it is also possible that some of these sequences are currently protein-coding, and have evolved through multiple coding-sequence disablements, as discussed previously 4.

To examine these data more closely, we calculated whether the Ka/Ks_Ψ-ortho values are significantly less than would be expected for mutation without coding-sequence selection pressures (using the simulational analysis described in the Methods section). Several cases with such significant values (that may indicate purifying selection typical of protein-coding sequences), are observed (represented by circles in the Figure 6 plots). These Ka/Ks values (that apparently indicate protein-coding ability) may arise for the reasons listed in the preceding paragraph.

In addition, we examined whether the TPAs contain a protein domain of known three-dimensional structure, that is disabled by a frameshift or a premature stop codon (denoted 'TPA_DDs'; see Methods for details of annotation of such domains). The TPA_DDs are indicated by unfilled symbols in parts (a) and (b) of Figure 6. Interestingly, in the human-dog comparisons, there are three cases of TPA orthologous pairs that have such a disabled protein domain, despite Ka/Ks values that indicate apparent purifying selection. These sequences are thus of 'intermediate' character, i.e., they have evidence of both protein-coding ability and pseudogenicity.

Transcribed pseudogenes can regulate the expression of other genes by RNA interference mechanisms. For example, antisense transcribed RNA from a NOS pseudogene regulates the expression of neural nitric oxide synthase (nNOS) protein through formation of RNA duplex 3435. Therefore, we investigated how many of the TPAs have antisense homology to the annotated full-length human cDNAs (E-value < 1e^-10 and alignment length > = 100 nucleotides). A small proportion (8.3%, 69/828) of the human (VEGA) pseudogenic transcripts have either complete or partial, but significant, reverse complement (antisense) homology to human cDNAs. Of these, 63 have short length strong antisense homology to human cDNAs (alignment length > = 20, mismatches < = 2). However, there is no significant association of such antisense homologies to pseudogene transcription, since non-transcribed pseudogenes have similar levels of antisense homology (7.65%, χ² test P-value = 0.5).

Out of the identified 68 human conserved TPAs, 3 have antisense homology to human cDNAs (E-value < 1e^-10, 5 if alignment length > = 50 nucleotides is considered) (Table 3). These are cases that may generate small interfering RNAs (siRNAs) that could potentially regulate the expression levels of their homologous genes. Pseudogenes have been implicated in the negative regulation of parental genes (for a review, see ref. 2) and in the Dicer-mediated generation of small RNAs 1819. It would be interesting to verify experimentally whether these pseudogene transcripts can indeed generate small interfering RNAs, through the action of Dicer.

Transcribed pseudogenes can also regulate the transcription of genes by producing siRNAs that have antisense homology 1819. Due to unavailability of genome-wide human siRNA data, we used the siRNA data for the mouse genome from Tam et al. 19 and Watanabe et al. 18 to check how many of the small RNAs mapped to mouse transcribed pseudogenes that we identified. Interestingly, 24 out of 136 (17.6%) mouse TPAs had siRNA mappings compared to ~1% (178/18168) of the total mouse pseudogenes. The above difference is statistically significant (P-value < 0.05, using normal statistics for the distribution of the mean number of transcribed pseudogenes in a sample of 136 cases). This demonstrates that transcribed pseudogenes are significantly likely to generate siRNAs in mouse. For comparison, in Arabidopsis thaliana, ~40% of 572 pseudogenes have small RNA mappings 36.

Complete genome sequences of mammals were obtained from http://www.ensembl.org (Ensembl release 47 for human genome; Ensembl release 48 for other mammals, namely, rhesus monkey, mouse, rat, cow and dog). Pseudogene annotations for both processed and nonprocessed categories, were obtained from [http://www.pseudogene.org; 3738] and for VEGA pseudogenes from http://vega.sanger.ac.uk/, for disrupted mRNAs (dmRNAs) from Harrison and Yu 16 and for other transcribed processed pseudogenes from Harrison et al. 14. The Blastx program 39 was used to determine the parent protein coding genes for VEGA pseudogenes (using E-value < 1e^-09 as significance threshold), whereas for other datasets the annotations were readily available at the respective websites mentioned above.

Transcription data for human and mouse were taken from RefSeq database 40, Unigene database at the NCBI http://www.ncbi.nlm.nih.gov, H-InvDB database http://www.h-invitational.jp/ and Fantom3 database http://fantom3.gsc.riken.jp/. To identify putative transcribed pseudogenes, individual transcript sequences were mapped onto the respective genomes using GMAP software 41 with match criteria of >99% sequence identity and >99% sequence coverage. Transcript sequences that mapped to pseudogenes were aligned to parent protein sequences of respective pseudogenes to identify disablements such as frame shift or premature stop codon using the 'GeneWise' program (Wise2 - version 2.1.20 package downloaded from the European Bioinformatics Institute, http://www.ebi.ac.uk/Tools/Wise2/index.html) 42. The disablement positions in pseudogenes and transcript sequences were then used as 'anchors' to confirm the transcription of pseudogenes as in previous analyses 141543. Additional data file (file 1) contains the list of transcribed human pseudogenes. For a schematic representation of the annotation pipeline, see Figure. 7.

Orthologous counterparts to a human pseudogene are detected by the presence of a homologous sequence at the syntenic position in the other mammalian genome. Based on this criterion, a search was carried out within 100 kb nucleotides distance of the exact syntenic coordinate (because genes can shuffle locally) in the target mammal as indicated in the synteny maps, to locate the orthologous pseudogenes. 'GeneWise' tool 42 was used, to align the above-obtained genomic DNA sequence and the human parental protein sequence, and to detect disablements in the alignment. The following mammals were included in the analysis: monkey, mouse, rat, cow and dog. The pair wise synteny map data for the various mammals were obtained from http://genome.ucsc.edu/.

Flanking sequences 5' and 3' of human pseudogenes were individually obtained, of length equal to the length of the human pseudogene, and were each globally aligned using 'needle' module of EMBOSS package http://www.ebi.ac.uk to the corresponding flanking region sequences (10000 nucleotides 5' and 3') of monkey in a sliding window of size also equal to the length of human pseudogene. The window in which best identity score was obtained was considered as the most optimum alignment between the flanking regions, representing syntenic regions. The Wilcoxon signed rank test was used for assessing the statistical significance of the difference between the degrees of homology calculated between two orthologous pseudogenes and that between the respective (orthologous) flanking regions. Cases with pair wise sequence identities <40% were excluded.

For sequence length and GC percentage calculations, only the exonic segments of pseudogenes were considered. One thousand nucleotides upstream and downstream of a pseudogenes were considered as flanking regions. GC content is calculated as the sum of guanine and cytosine nucleotides divided by the total number of nucleotides represented in terms of percentage.

'PAL2NAL' 44 was used to construct codon alignments between protein sequences (conceptual amino acid translation sequences in the case of pseudogenes) and corresponding DNA sequences, separately, for orthologous pseudogenes and parental protein coding genes. 'PAML 4' package 45 was used to calculate Ka/Ks ratios. Orthologs of human parental protein coding genes were identified using a similar approach as that for pseudogene orthologs discussed above, and also obtained from Ensembl database.

We derived a simulation protocol to calculate Ka/Ks values for evolution without coding-sequence selection pressures. This simulation protocol is as follows: (i) the nucleotide distance (D_nt) between a sequence and its ortholog was calculated, using the program DNADIST 46; (ii) for each sequence, samples of 500 simulated sequences were generated, by randomly mutating the human sequence until the D_nt value was reached; (iii) Ka/Ks was calculated using PAML 45, for each simulated sequence compared to the original human sequence; (iv) those original human sequences that have Ka/Ks values < 95% of simulated Ka/Ks values were labeled as potentially under significant purifying selection. For these simulations, all Ka/Ks calculations are performed on the longest ORF in the sequence.

We also analysed simulated distributions of Ka/Ks for populations of sequences mutating without coding-sequence selection pressures, starting from the human TPA sequences. These were derived simply by merging the simulated distributions of Ka/Ks for each individual TPA.

Protein domains were assigned to the TPAs, using protein structure domain sequences downloaded from the ASTRALSCOP database http://astral.berkeley.edu, as described previously 4. Protein domains sequences were aligned to the TPA nucleotide sequences to assess for disablement by a frameshift or premature stop codon at least 15 amino acids from the end of the aligned subsequence. Disablements were required to be detected both by blast/bl2seq and by the TFASTX program 439.

Transcribed human pseudogenes were aligned to full-length annotated human cDNA to examine for any antisense homology by using the sequence-searching program BlastN from the BLAST package (E-value < 1e^-10).

siRNAs have been previously determined in the mouse genome 1819. Using this data we mapped the siRNA sequences onto the mouse genome using GMAP software 41, and checked how many of these overlap with the annotations of transcribed mouse pseudogenes.

Ortholog sequences to the human transcribed ADP-ribose pyrophosphatase pseudogene (urn:lsid:pseudogene.org:9606.Pseudogene:4346; see Table 1), were obtained from the various studied mammals and were aligned using the online ClustalW tool http://www.ebi.ac.uk/clustalw/. The most conserved segment representing 257-396 positions of the human pseudogene was considered for the phylogenetic analysis. Phylogenetic tree was constructed using 'PHYLIP' software 46. The tree was evaluated statistically using 1000 bootstrap iterations and was visualized using the 'NJplot' tool 47.