The genome sequences and annotations of seven organisms analyzed in this paper (human, dog, cow, mouse, rat, chicken and Tetraodon nigriviridis), were downloaded from the Ensembl Web site 15, in January 2005. Version 2.1 of the chimpanzee assembly (downloaded in April 2006) was also used. Putative retrogenes (PRs) were identified in the annotated proteomes of eight vertebrates using the pipeline in Figure 1. This procedure is described in detail below. In tandem, putative retropseudogenes (RψGs), and additional PRs outside of current protein annotations, were assigned using a modification of previous procedures (Figures 1 &2) 142. Genes from which PRs and RψGs are thought to have originated are called parent genes.

Retropseudogenes were annotated on the Ensembl genome versions used in this present analysis, using the rapid improvement of previous procedures to identify retropseudogenes described above (summarised in Figure 2) 1151423.

(1) Homology detection: Each proteome was compared against itself using BLAST to find similarities with e-value ≤ 10^-4 16. Any match to a potential pseudogene contaminant in the proteome annotations was removed (Figure 1).

(2) Exon seam analysis: Exon boundary information for each protein was extracted from the appropriate Ensembl genome annotation files. The positioning of exon boundaries in encoded protein sequences, i.e., 'exon seams', was then deduced 11. Using the positioning of exon seams, the BLAST matches between proteins were filtered to pick out alignments between a protein encoded by a multiple-exon gene, and a single exon of another gene. To define PRs, the length of the exon was required to be between 0.9 and 1.1 of the whole length of the multiple-exon protein. (This is stricter than the criterion of 0.7 of the length used for annotation of retropseudogenes 2).

(3) Assignment of parent genes: To assign parent genes to PRs, we calculated substitution rates at synonymous codon sites (i.e., Ks values) for all matches to PRs using the package PAML, for all instances where the amino-acid sequence identity for the pair of sequences is ≥ 70%. The sequence with the smallest Ks value was chosen as the 'parent gene'. For sequence identities <70%, saturation of substitutions is likely 17, and so the sequence with the highest BLAST bitscore in alignment with the PR was chosen as the putative parent.

(4) Additional filtering: In addition, PRs were discarded if they matched olfactory receptors (ORs) with BLAST e-value ≤ 10^-4 over ≥ 0.5 of the length of the OR, since recent olfactory receptors (ORs) have probably originated from different mechanism other than retrotransposition 2. Olfactory receptor sequences were taken from ORDB 18.

(5) Local gene order test for similarity of the chromosomal milieus for PRs and parent genes: To check that the PRs did not arise from local or segmental duplication, we derived a 'local gene order' test. For this test, we compared the chromosomal milieus of PRs and parents for significant similarity, as follows. Proteins encoded by genes adjacent to PRs in the chromosomes were BLASTed against the corresponding proteins from genes adjacent to putative parents, for a given window (w_genes) of number of genes in either direction (5' and 3'). A w_genes size of 7 (the PR or parent, plus 3 genes in either direction), with an allowance for one gap between the positions of matches within w_genes, was found to be suitable. The number of significant homologous matches N_homologs (BLAST e-value ≤ 10^-4, sequence identity >40%, and match ≥ 0.6 length of both PR and parent) between the milieus of PR and parent was tallied. An expected benchmark distribution for N_homologs was derived for the chromosomal milieus of 1,000 randomly-sampled pairs of proteins that have any significant BLAST match to each other (e-value ≤ 10^-4). From examination of this distribution, we found that 80% of such random pairs have N_homologs <1, and 87% have N_homologs ≤ 1. We thus chose N_homologs = 1, as a suitable threshold for local similarity arising from duplication of genomic DNA. However, the results differ little if a threshold of N_homologs = 0 is used. This procedure was applied to the genomes of the human and mouse. Interestingly, application of this criterion resulted in the exclusion of many sequences with large individual exons (PRs with FLE_parent ≥ 0.8; 36/86 = 42% of those excluded), that may be false positives in our data set of PRs. A large fraction of these sequences (68%) tend to have long, tandem arrays of Zn-finger domains covering more than a third of their sequences (Additional Figure 1).

The following additional analysis was performed on the PRs:

(i) Fraction of largest exon in parent: We calculated the fraction of the length any parent gene that is taken up by its largest exon. This is denoted FLE_parent. We found that there is no peculiar tendency for the parents of PRs to have a single large exon (which would yield a tendency for high FLE values) [Additional Figure 1].

(ii) Overlap with retropseudogenes annotations: Retropseudogenes were annotated on the Ensembl genome versions used in this present analysis, using the rapid improvement of previous procedures to identify retro(pseudo)genes described above. Any PR that overlapped one of these annotations was flagged (Table 1).

(iii) Filtering for potential transposable elements (TEs): Each proteome was compared using TBLASTN to libraries of transposable elements taken from the RepeatMasker distribution 19, using an e-value threshold of ≤ 10^-4. Any proteins containing SINEs, or near-complete matches to LINEs (≥ 0.8 of their lengths), were labeled as potentially TE-containing.

(iv) Whether single-exon gene or multiple-exon gene: The PRs were labeled as either single-exon genes or part of multiple-exon genes.

Orthologs of parent genes were identified using the bi-directional best hits method, using BLAST (e-value ≤ 10^-4, amino-acid sequence identity ≥ 40% and requiring the alignment to cover ≥ 0.6 of the lengths of each sequence. The bi-directional best hits method is a common procedure for guarding against considering paralogs.

The package PAML 20 was used to calculate maximum-likelihood Ka, Ks and Ka/Ks values for pairs of sequences (either PR versus computed ancestral sequences, or PR versus parent). In addition, branch-specific maximum-likelihood Ka/Ks values were calculated for three-way alignments of PR, parent and parent's orthologs from another close species.

We applied three different strategies based on analysis of Ks, to determine lists of genome-specific and lineage-specific PRs. For example, for the human genome, we calculated human-specific lists relative to the chimpanzee genome. Also, we calculated primate-specific lists for human plus chimpanzee, relative to a mammalian outgroup, such as dog or cow. To determine genome-specific lists of PRs, we investigated each of the following three methods ("parent's ortholog" refers to the ortholog of the parent in the most closely related genome):

(1) The distribution of Ks values for orthologous genes in the two organisms was calculated, and the median value m derived from this. If Ks [PR←→parent] <m and Ks [parent ←→parent's ortholog] >m, then a PR is labeled genome-specific;

(2) Secondly, a PR could be labeled genome-specific if Ks [PR ←→parent] < Ks [parent ←→parent's ortholog];

(3) Thirdly, a PR could be labeled genome-specific if, in a three-way tree of PR, parent and parent's ortholog, the branch-specific Ks [PR] is < (Ks [parent] + Ks [parent's ortholog])/2 ;

Lineage-specific lists were derived in a similar fashion. Additional Figure 2 shows how these three methods overlap which each other. Based on the overlaps observed, we used Method (3) for further analysis.

We assessed the reading frame conservation (RFC) in sequences using simulations of insertion and deletion governed by power-law insertion/deletion (indel) statistics 4. Power-law statistics for indels were extracted from recently-formed RΨGs having ≥ 85% amino-acid identity with their parent sequences. Power-law relationships were fitted, omitting points for any indels of size 3n, with n any positive integer. Expected ratios for insertions versus deletions were taken from this data; the expected number of indels per nucleotide substitution for several mammals was culled from the literature 2122. The program DNADIST 23 was used to calculate the nucleotide-level divergence of the PRs from ancestral sequences (calculated using PAML 20; see section on Ka/Ks analysis above). This divergence value is used as a target in simulations. For each PR, repeated simulations of the evolution of the ancestral sequence towards present-day, for 1000 iterations, was performed using a Kimura two-parameter model. In each case, the resulting simulated protein coding sequence was marked for frame disablements (stop codons and frameshifts). PR sequences whose simulations yielded frame-disrupted sequences ≥ 99% of the time were labeled as having significant RFC.

GO (Gene Ontology; 24) functional categories were taken from the annotation files provided on the Ensembl 15 and Gene Ontology websites 25. Further GO term annotations were derived by mapping functional GO annotations for the PDB (also downloaded from the GO website) onto Ensembl protein annotations, using 50% sequence identity and 0.8 fractional sequence coverage (for the protein domain) as thresholds, using alignment made by the program BLASTP (e-value ≤ 0.0001) 16. These thresholds were benchmarked on the complete SCOP protein domain sequence database 26, to give a 2% false positive rate for GO term transfer.

Refseq mRNAs and complete Unigene consensus sequences were downloaded from the NCBI website 27, for both human and mouse. These were mapped to the coding sequences of Ensembl gene annotations, using blastn (e-value ≤ 1 × 10^-10 for alignments ≥ 100 nucleotides) 16. All mappings that match with ≥ 99% sequence identity over ≥ 0.99 of the sequence length of the cDNA or mRNA, after removal of any polyadenylation, were retained. Further restriction of analysis of cDNA/mRNA mappings to those that do not match their putative parent sequences with >95% sequence identity, does not change the trends reported with regard to transcription evidence reported below.

Our analysis suggests that up to ~3% of the genes encoded in a vertebrate genome contain a PR (Table 1), with the smallest percentages in the chicken and puffer fish T. nigroviridis. By comparison, mammalian genomes have ~2,000–5,000 retropseudogenes (RψGs), that have at least 70% of the coding sequence of their parent genes, again with smaller numbers in non-mammal vertebrates (just 644 RψGs in T. nigroviridis) (Table 1). These results together indicate that there has been less, recent gene-retrotransposition activity in the two non-mammal vertebrates. These observations tally well with other evidence for less retrotransposition activity in chicken and Tetraodon. In chicken, there appear to be little or no SINEs 28, and only ~8% of the genomic DNA is comprised of the CR1 ('chicken repeat 1') LINE-1 2829, whose reverse transcriptase is thought not to copy polyadenylated mRNAs 29. In Tetraodon, <1% of the genome is comprised of retrotransposons, so gene retrotransposition should consequently be less likely 30. For the eight genomes studied here, there are no significant linear correlations between the number of genes or proteins from a genome, versus the number of PRs, or RψGs (data not shown). Small percentages of the PRs could be classified as homologs of retrotransposed transposable elements, such as LINEs (2–12%), or as overlapping pseudogene annotations (4–13%).

As described in detail in Methods, we applied a 'local gene order' test, to set aside any PRs that may have arisen through local or segmental duplication, specifically for the human and mouse genomes (Table 1). This filter allows for at most one homologous protein encoded within a window of +/-3 genes along the genomic DNA (i.e., N_homologs ≤ 1) (Table 1). The substantial majority of human and mouse PRs pass this filter (80–87%).

How old are these PRs? Is there any evidence for spurts of gene retrotransposition activity in mammalian evolution? To answer these questions, we examined the distribution of Ks values for PRs compared to their assigned parent genes, in the human and mouse genomes. (Only PRs passing the local gene order test, with threshold N_homologs ≤ 1 were analysed.) Ks is the rate of synonymous substitutions per synonymous site in codons, and has been generally used to age coding sequences. From comparing Ks values for PRs, their parents, and orthologs of their parents, we have also been able to derive lists of genome-specific and lineage-specific PRs (Figure 4; see Methods for details).

In human PRs, we see that there is a bimodal distribution of Ks (Figure 3A). The median Ks values for lists of PRs that are human-specific or that have otherwise been formed since divergence from dog, are labelled on the Ks histogram. The peak at Ks ~0.06–0.08 corresponds approximately to the median Ks value for PRs formed between human divergence from dog and from chimpanzee. This peak has been noted previously in analyses of retropseudogenes and total retrosequence populations 2312, peaking at approximately the point of human lineage divergence from the New World Monkeys, some ~40 million years ago 3. The peak at 0.0–0.02 (containing 21% of PRs) obviously corresponds to human-specific PRs. Some of these PRs may simply be too young to differentiate as a PR or retropseudogene (They have not been around long enough to acquire (apparent) reading-frame disablements.) Evidence for selection pressures on these sequences is discussed below.

By comparison, in the rodents, there is more, very recent gene retrotransposition activity. In mouse, we find proportionately more, genome-specific PRs (relative to rat), with 44% having Ks ≤ 0.02 (Figure 3B). In the two rodents, mouse and rat, there are >200 genome-specific PRs, compared to ~40 in each of the primates human and chimp. However, setting aside genome-specific examples, there are more gene retrotranspositions appearing in the primate lineage since its divergence from the dog or cow lineage (Figure 4).

These observations are in keeping with the apparent maintenance of greater levels of LINE and SINE retrotransposition activity in the rodents 3132; also, they tally well with previous observations for a general fall-off in such retrotransposition activity in the primate lineage 233.

Focussing on human and mouse, we examined the proportion of PRs that could be mapped to a complete Unigene consensus cDNA or a complete Refseq mRNA from the NCBI (27; see Methods for details). For both organisms, we found that the PRs have significantly less mapping of this transcription evidence (P < 0.001, using the z-score for distribution of the sample mean). For human, only 23% of human PRs mapped to a Refseq mRNA or Unigene cDNA consensus sequence (compared to 41% for the whole proteome). This may be due to lower transcription levels, because they are novel gene sequences using co-opted promoter elements at a site distal to the genomic location of their parent genes 11. This general reduction in transcription is also to be expected, if some of the sequences are recent pseudogenes without disablements.

In addition, we examined how many PRs arise from alternatively-spliced genes. To do this, we cross-referenced the PR data with alternative splicings classified in the Alternative Splicing Database (ASD) at the EBI 34. We found that 24% of genes for human PRs arose from an alternatively spliced gene, compared to 59% of genes overall (significantly less, P < 0.001 using the z-score for distribution of the sample mean). A significant reduction in representation from alternatively-spliced genes was also observed in mouse (4%, compared to 29% overall).

We examined the divergence of PRs from their putative parents, in the context of transcription evidence. This is illustrated in Figure 5 for those PRs that pass the local gene order test for both mouse and human, with N_homologs ≤ 1. In human, there is a marked difference in the behaviour of transcribed PRs (purple bars in Figure 5A), compared to those without transcription evidence (blue bars in Figure 5A). There are relatively very few transcribed PRs with high sequence identities (i.e., that formed relatively recently). The bimodal character of these plots may arise because some of the PRs: (i) are in a younger population of PRs that are not under selection pressures, but which have not accumulated deleterious mutations, simply by chance; i.e., they are pseudogenes without disablements; or (ii) are in a state of relaxed selection, and thus concomitantly have low transcription levels. Similarly, only a small fraction of the PRs calculated to have formed since divergence from the dog lineage (Figure 4), in either human (15/207, 7%) or mouse (20/292, 7%), are transcribed (significantly less, P < 0.05 using χ² tests for both cases).

Is there any evidence for selection pressures on the PRs in human and mouse? We investigated this question for PRs that have been formed in human and mouse, since their divergences from dog. One standard indicator of selection pressures is the Ka/Ks ratio. This is the number of non-synonymous mutations per non-synonymous site, over the number of synonymous mutations per synonymous site, in codons. Negative (or 'purifying') selection in a specific lineage is indicated by a value significantly <1.0, where positive ('diversifying') selection is demonstrated by a value significantly >1.0. We calculated Ka/Ks values for PRs relative to ancestral sequences for the parents of the PRs (see Methods for details). We tested whether any of these Ka/Ks values were significantly < or >1.0 by generating 500 random pairs of sequences as diverged as the PR and ancestral sequence (to calculate expected means and standard deviations for the Ka/Ks values), and then deriving a P-value for the observed Ka/Ks.

Strikingly, when we correct for multiple hypothesis testing in the Ka/Ks calculations, we find only one PR sequence (formed since divergence from dog) that is under significant selection at the codon level in the human genome, and none in the mouse genome. (The one significant human example is a PR under purifying selection, from a family of proteins with the GTP-binding SAR1 domain.)

In addition, we calculated the distribution of Ka/Ks values from directly comparing the PRs versus their parents. From this specific sort of comparison, the neutral expectation for Ka/Ks is not ~1, because of non-synonymous mutations accumulating in the parent genes 2. A significant excess of Ka/Ks values <0.5, however, may be indicative of purifying selection in the data set. For comparison, we also similarly calculated a Ka/Ks distribution for RψGs versus their respective parents, carefully parsing out disablements (frameshifts and premature stop codons) from the RψG sequences. This Ka/Ks distributional analysis is performed for both human and mouse (Figure 6). For human (Figure 6A), we find no significant excess of PRs with Ka/Ks values <0.5 relative to RψGs, either for the whole data set of PRs, or for the subset formed in the primate lineage, contrary to a previous report 12 (χ² test or Fisher's exact test). This distribution is thus consistent with a set of largely neutral retrotranspositions, behaving like RψGs. However, for mouse-specific PRs, there is a significant excess with Ka/Ks <0.5 (P ≤ 0.001, χ² test and Fisher's exact test) (Figure 6B), indicating that some of these mouse PR sequences are under purifying selection.

Conservation of open reading frames without disablements (frameshifts or stop codons), can also be an indicator of coding ability 123536. We derived a method for assessing significant conservation of open reading frames, using simulation with power-law insertion/deletion (indel) statistics 4. Using simulations with calculated neutral rates of substitution, insertion and deletion, the likelihood of conservation of an open reading frame without interruption by frameshifts and stop codons, can be determined (see Methods for details). To give sufficient power, a P-value threshold of ≤ 0.01 was used as an indicator for significant reading-frame conservation (RFC). This calculation is complementary to the Ka/Ks analysis.

The results are listed Table 2. Even though there were no significant Ka/Ks values in mouse, we find over 30 mouse-specific PRs that have significant reading frame conservation, and a further 17 that were formed since divergence from dog (Table 2). In human, we find in total, 59 PRs with significant RFC, that have arisen since divergence from dog (Table 2). A phylogenetic tree for an example of one of the mouse PRs with significant RFC, which is homologous to citrate synthase, is depicted in Additional Figure 3, with a depiction of the chromosomal milieu of this PR and of its parent in Additional Figure 4. Two further representative examples of human PRs are also shown in Additional Figure 4 (one with significant RFC and the other without), with varying degrees of age and transcription evidence.

These results are evidence for conservation of protein open reading frames, even though we found no evidence for purifying selection from examination of the sequences individually for Ka/Ks. This would arise if the PRs were generally under relaxed or positive selection pressures at the codon level. The existence of relaxed selection is consistent with the markedly low numbers of PRs found to be transcribed in both human and mouse, particularly those that were formed since divergence from dog.

Out of those with significant RFC, is there any evidence for non-neutral Ka/Ks trends? We checked for significant excess of PRs with Ka/Ks values <0.5, > 0.5, <1.0 or >1.0 for each of the subsets listed (Table 2). In the human lineage, we find a significant excess of PRs with Ka/Ks >0.5 (40/59, P < 0.05, χ² test and Fisher exact test, compared to an expectation from RψG sequences) formed between divergence from dog and from chimp. This non-neutral trend is suggestive of positive selection distributed throughout this specific subset population of RψG sequences. Out of the other subsets listed in Table 2, the only other significant non-neutral Ka/Ks tendency is for an excess of mouse-specific PRs to be under purifying selection (26/35 having Ka/Ks <0.5, compared to an expectation from RψG sequences, P < 0.05, χ² test and Fisher exact test).

To assess whether the PRs and RψGs have any unusual functional associations, we assigned functional categories using the Gene Ontology (GO) functional classification (Table 3). As noted previously, 'Structural constituent of ribosome' is a prevalent functional category for RψGs 2314. Noticeably, for mouse, there are more retropseudogenes in metabolic categories, than in human (Table 3). A notable prevalence indicative of origin in retrotransposition, 'structural constituent of ribosome', occurs in the top ten of all PR (sub)sets, and is ranked number one for PRs formed since divergence from dog, for both mouse and human (Table 3). 'Structural constituent of ribosome' is also the only Gene Ontology term that is statistically overrepresented in all of the retrotransposed gene sets listed (Table 3 legend; P' <0.05, using binomial statistics and a Bonferroni correction for multiple hypothesis testing 37). The functional category preferences are not caused by over-representation of any one parent, since when representations of PRs on a parent-by-parent basis are tallied up, we find only a very small number of parents giving rise to five or more PRs (Suppl. Table 1); the substantial majority of parents have only one PR offspring (286/353 [81%] for mouse, and 279/347 [80%] for human).