Pseudogenes are sequences of genomic DNA lacking the protein-coding capability of their paralogous counterpart [1,2]. A pseudogene can be arbitrarily similar to the original gene, but differ by the fact that it accumulates disablements (in-frame stop codons and sequence frameshifts), which protein-coding genes do not. Because pseudogenes are not protein-coding, they are often thought of as being without function and therefore released from selective pressure. The origin of a pseudogene is generally either a segmental duplication, or a retrotransposition where mature mRNA is reversely transcribed into cDNA and reinserted in a new genomic position. The resulting pseudogene is in the latter case called processed as compared with duplicated or nonprocessed.

Studies of pseudogene populations are often motivated by the dilemma that their similarity to ordinary genes constitutes for gene finders and hybridization experiments. Pseudogene sequences can, given their nonfunctionality, be viewed as a molecular fossil and have been used to measure background genomic substitution rates [3,4].

However, evidence has occasionally been found, in Drosophila and recently also in mouse [5], of pseudogene functionality, as well as of conservation (see [6] for a review). In [5], evidence is given for a regulatory role of the mouse Makorin1 pseudogene Makorin1-p1. It was proposed in [5] that the function of the transcribed pseudogene is to stabilize the Makorin1 mRNA. A follow-up study [7] established that Makorin1-p1 is in fact conserved across several mouse species, although it is not found in more distantly related species such as rat or human.

Several surveys [8–10] have located and annotated pseudogenes in the human and mouse genomes. Despite using slightly different pseudogene definitions and methodologies for finding them, they end up with similar numbers of human pseudogenes (altogether about 20,000 sequences out of which some 8,000 show evidence of processing). The authors of [11] used more restrictive criteria, and identified about 3,600 human processed pseudogenes. The main theme for these studies is that sequences sufficiently similar to a known protein sequence are considered potential pseudogenes. The final classification as pseudogene is based on proof of sequence disablements (primarily in-frame stop codons and sequence frameshifts), Ka/Ks values indicating neutral evolution, and, importantly, that the sequences are not overlapping any known gene.

In a recent article [12], the authors went further and looked specifically for human-transcribed processed pseudogenes. They found that some 4%–6% of all human processed pseudogenes could be confidently mapped to sequences in expression databases. The same group then continued with a more careful annotation of pseudogenes on Chromosome 22, utilising tiling microarray technology [13], concluding that this rate was probably an underestimate and that maybe as much as 1/5 of all pseudogenes are transcribed. Another investigation in the same spirit [14] found that the percentage of expressed pseudogenes differ significantly between human and mouse. They reported 2%–3% and 0.5%–1%, respectively, using the most restrictive criteria.

A human–mouse comparative study [12] concludes that the vast majority of transcribed human pseudogenes are lineage specific. Only some 5% were found to have potential orthologs in mouse.

That a pseudogene is transcribed is not sufficient evidence of biological function. To obtain functional candidates, we decided to look for conserved pseudogenes common to human and mouse, originating from one duplication predating the human–mouse species split and having evolved as pseudogenes since the species split. In cases where the species split occurred sufficiently early, strong conservation and ancient origin gives evidence of the potential functionality of the pseudogenes. We have developed a pairwise comparative genomics methodology based on an explicit evolutionary model, which focuses on pseudogenes common to the two lineages. We also test the potential functionality of the found pseudogenes using enrichment of transcription and synteny.

We describe our methodology using the example of a human–mouse comparison. Our procedure takes as input a quartet of sequences representing, respectively, a human gene, a corresponding human pseudogene, the orthologous mouse gene, and a corresponding mouse pseudogene, and analyzes how they have evolved. All four basic evolutionary scenarios that can occur with respect to duplication and gene-to-pseudogene transitions are described below. When analyzing how well a scenario describes the evolution of a quartet, different models of sequence evolution are used for gene and pseudogene lineages.

The first scenario, S1, is what we expect for conserved pseudogenes originating before the species split (see Figure 1).

An alternative scenario, S2, is expected if both pseudogenes originated independently of each other, after the species split (see Figure 2). In our human–mouse comparison, the evolution of most quartets are best described by S2. A likely explanation for this is that dead-on-arrival pseudogenes [15] originating before the human–mouse species split have most often diverged beyond the limit of recognition. With approximately 0.5 substitutions per site, fewer than 10% of neutrally evolving genomic elements can be found using BLAST [16].

The third scenario, S3 (see Figure 3), is similar to S1. The difference is that the transition from gene to pseudogene occurred subsequent to the species split. This could mean that a pair of pseudogenes was in fact functional genes prior to the transition, but has since then evolved without selective pressure.

A fourth scenario where the human gene has the mouse pseudogene as a sibling in the gene tree is conceivable. We have never observed this scenario.

We have applied our comparative methodology to human–mouse as well as to human–chimpanzee and found the first examples of human pseudogenes showing signs of functionality.

We started with the 12,687 presumably orthologous protein pairs retrieved (see Materials and Methods) from the Inparanoid web site [17] for which gene sequences could be found. We used BLAST to scan the human and mouse genomes for potential pseudogenic sequence pairs (see Materials and Methods). A pseudogene pair corresponding to a protein pair was then used together with the gene sequences to form the sequence quartets on which we base our analysis.

This initial search with subsequent refinement resulted in 168,855 such quartets. For the vast majority of these quartets, one or both pseudogene sequences overlap regions of known or predicted protein-coding genes. Gene position data from Ensembl were used to filter out known genes. Predicted genes are kept for further analysis, since it is known [8] that gene predictors sometimes mistake pseudogenes for protein-coding genes.

The set that remains after filtering constitutes 11,146 sequence quartets originating from 1,349 protein pairs. The distribution of quartets per protein pair is highly nonuniform. While many gene pairs lack corresponding pseudogene pairs, a handful (EF12, G3PT, LDHB, TSY1, UB46, and several ribosomal genes) are origins of large pseudogene families in both species. Using the mutual-best-hit filtering outlined in Figure 4, we, however, removed a large number of pairs likely to be insignificant; after this step 1,453 sequence quartets remained. We divide these into four classes according to the following: class 1—pairs that have detectable disablements and do not overlap any Ensembl gene prediction; class 2—pairs that have detectable disablements and overlap an Ensembl gene prediction; class 3—pairs that do not have detectable pseudogenic disablements and do not overlap any Ensembl gene prediction; and class 4—pairs that do not have detectable pseudogenic disablements and overlap an Ensembl gene prediction.

We used the partition of our data induced by this classification in combination with mutual-best-hit filtering. The number of sequence quartets belonging to the classes are: class 1—247 quartets; class 2—299; class 3—146; and class 4—761 (see Table 1).

Our aim is to find those quartets for which S1 is the most likely explanation. We use a probabilistic methodology to compare scenarios (see Materials and Methods), to obtain p-values for any possible alternative hypothesis with respect to the interesting one, namely, that S1 best describes a given quartet. For visualization purposes, we also consider quotients of type L₁/L₂, where L_i is the log-likelihood corresponding to scenario i; for a particular quartet, a value L₁/L₂ < 1 suggests that L₁ is preferable to L₂.

For the majority of our 1,453 quartets, data support S2 (Figure 5), the scenario where pseudogenes originated later than the species split. In 425 out of 1,453 cases (29%), the p-value for S1 being the scenario that best explains our data is less than 0.001.

Interestingly, we note a bimodal pattern with one large hump distributed around 1.1 and another one distributed around 0.9. That is, in a large majority of cases, data show clear preference for either S1 or S2; it is only for a comparatively small number of cases that the quotients are close to 1.

We now use the same technique to compare S1 and S3. Remember that S3 is the scenario where the transitions from genes to pseudogenes were independent of each other and occurred subsequently to the speciation. Hence it is only the models of sequence evolution used for genes and pseudogenes, respectively, that distinguish S1 from S3. The likelihood values are in this case much less varied, yielding many quotients close to one (Figure 6).

For 73 of the 425 quartets, S1 is the explanation favored by our method and for 30 of these 73 the p-value is lower than 0.1. For 352 sequence pairs, S3 is the most likely topology, and 262 of those clearly favor S3 (p-value again lower than 0.1).

To summarize, we have 30 quartets for which the sequences suggest that: 1) the pseudogenes are evolutionarily conserved since before the human and mouse speciation; 2) they have been pseudogenes since prior to the speciation.

Because we find 30 such quartets, and the number of quartets expected to pass our scenario test is 1453 * 0.001 * 0.1 ≈ 0.15, it is reasonable to conclude that a significant number of these 30 quartets are ancient pseudogenes, i.e., satisfying 1) and 2).

We are now going to investigate these 30 sequence quartets further, with the aim of testing their potential biological function. The criteria that will be our focus are synteny, expression evidence, and conservation.

We have presented and applied a semi-automated methodology to identify pseudogenes of potential biological function. To the best of our knowledge, functional pseudogenes have never been observed in human. Our method uses no prior knowledge other than publicly available data on orthologous relationships for proteins, gene sequences, gene positions, and synteny maps.

The term pseudogene is normally used for sequences derived from known proteins but with detectable disablements that make the translation to protein impossible. Detecting pseudogenes is complicated by the possibility that part of the copy can be disabled, while the rest is coding.

We use conserved ancient pseudogenes as candidates of potential function. A computational approach based on support for four different evolutionary scenarios is used to obtain putative ancient pseudogenes. The p-value thresholds used, as well as the tests applied, indicate that the set of putative ancient pseudogenes is significantly enriched for ancient pseudogenes. It is interesting to ask whether there are evolutionary mechanisms that could cause our scenario S1 to appear more likely than another correct scenario, i.e., mechanisms not taken into account in our approach. Notice that, for instance, homogenization, e.g., through gene conversion, cannot make S1 more likely, since S1 is supported by similarity between sequences in different species.

We test functionality of our candidates by means of enrichment of synteny as well as of transcriptional activity and degree of conservation. We see, as expected, a clear overrepresentation of synteny for human–mouse pseudogene pairs originating before the species split. Interestingly, we also see tendencies for those examples that have evolved as pseudogenes since the species split to be both more abundantly expressed and more often syntenic than those that have not evolved as pseudogenes. For the latter finding, we believe that enrichment of functionality among our pseudogenes is the most likely explanation.

Judging from what is known from earlier work, the number of detectable pseudogenes originating from before the human–mouse speciation is limited. In [12], the authors found 11 examples of potentially orthologous pseudogene pairs. Although there is considerable overlap between their analysis and ours, a numerical comparison is not straightforward. First, they have stricter criteria for classifying a sequence as a pseudogene (deploying a careful filtering to make sure that only sequences that are processed pseudogenes are investigated). Second, they compute human and mouse orthologs using reciprocal BLAST comparisons only, and investigate how many of the transcribed human genes have mouse orthologs.

Three out of four (PDZRN3 being the exception) of our S1 examples belonging to class 1 are also classified as pseudogenes in [8,12]. We use the same databases for expression analyses as do the authors of [12], and, as expected, the results are in agreement.

To determine functionality of a human pseudogene, it is probably not sufficient to use information about whether it has a mouse ortholog or not, because many young pseudogenes can be found among orthologous pairs. Instead, we select only those human pseudogenes with orthologous mouse pseudogenes that satisfy the additional constraint that the least common ancestor was a pseudogene.

The results we present suggest that while functional pseudogenes are relatively rare on a long evolutionary timescale, they nevertheless exist. Our findings include a handful of sequences that are conserved since before the split of primates and rodents. Some of these are sequences predicted by gene finders to be protein-coding. We have found examples with, as well as without, detectable in-frame disablements. Apart from their apparent functional conservation and sometimes extensive expression activity, all these are poorly characterized. This can be due to the fact that some of the originating proteins are themselves not very well known or to the common assumption that pseudogenes are nonfunctional. Further characterization of these genes, their respective pseudogenes, and the interactions between them are areas for further studies.

We have noted with interest recent research activity concerning two of our top candidates, ATX1 and ATXN7L3. As is the case for the Ataxin gene family in general, these are associated with a number of neurodegenerative disorders primarily caused by expanded polyglutamine [24,25], but other than that their function is currently unknown [26]. Findings indicating that these genes have a regulating function [27,28] are of particular interest because it is reasonable to believe that their paralogs, if they are indeed functional pseudogenes, work on the RNA level. There are previously known examples of ncRNA genes in the Ataxin gene family. In [27] it is shown that Ataxin 8 regulates Kelch-like 1 by anti-sense regulation.

When extending the search to younger pseudogenes, i.e., applying our methodology to human–chimpanzee, the number of obtained pseudogenes is substantially larger than what was obtained in the human–mouse comparison. In this case, however, the assumption that nonfunctional pseudogenes originating before the speciation have diverged beyond recognition is not true. Consequently, filtering out nonfunctional pseudogenes is much harder than for the human–mouse case. Encouragingly, we found that the conservation of many pseudogenes is similar to that of nonsynonymous nucleotides in protein-coding genes (estimated to be 99.4% [23]).

There is an apparent tradeoff between the number of pseudogenes in the result set and the certainty with which we can state that they are functional. It is quite possible that both our choices of species pairs are in fact suboptimal, human–mouse being too evolutionarily distant and human–chimpanzee not distant enough. It will be interesting to apply our methodology on an intermediate timescale, and we plan to conduct a comparison between human and rhesus macaque.

Our methodology includes three main parts (see Figure 10), here termed: 1) pseudogene finding, 2) cross-species matching, and 3) pseudogene pair evaluation.

To locate pseudogenes, we adopted a large part of the methodology presented in [8]. When searching for pseudogenes, a hit was considered to be significant if it had an e-value < 10⁻¹⁰ from a six-frame TBLASTN search with the protein sequence (BLAST-package was downloaded from ftp://ftp.ncbi.nih.gov/blast). We also made use of TBLASTN capabilities to detect stop codons and, importantly, sequence frameshifts.

We used repeat-masked genomic sequence data NCBI35 (human) and NCBIm33 (mouse) downloaded in January 2005 from the Ensembl database version 27. The Expasy protein database (sprot44, human_trembl and rodent_trembl) was used to assemble protein sequence sets for the two species. We used the Inparanoid [17] database to retrieve orthologous protein pairs, by selecting from each set of inparalogs the sequence with highest score.

The mutual-best-hit filtering was performed for each pair of orthologous proteins, by aligning each pair of pseudogenes from the respective species, using bl2seq (from the BLAST package) and then selecting the pair with the best score. We aligned the thereby obtained quartets using the Dialign package [29], and we extracted from the Dialign output gap-free column triplets based on the reading frame induced by the genes. By using a local alignment program we reduced the risk of misalignments caused by introns in the pseudogenes.

To select the scenario (S1, S2, S3, S4) that best describes a given quartet, we adapted the method outlined in [30]. We work with a model describing the instantaneous substitution rate from codon i to codon j, q_ij, given that the equilibrium frequency of codon j is π_j. In their model, the substitution rates are specified by the instantaneous rate matrix Q = {q_ij} defined by: where μ is a normalizing rate factor, κ is the transition/transversion ratio, and ω is the nonsynonymous-to-synonymous ratio. In our model, we use one matrix, Q_g, for the parts of the tree where the sequences are supposed to evolve as a gene, and another matrix, Q_ψ, for the parts of the tree where the sequences are supposed to evolve as a pseudogene. The difference between the matrices is that different κ are used, for pseudogenes ω equals one, for genes we do not allow transitions to stop codons, and equilibrium frequencies are estimated from gene or pseudogene sequences, respectively.

We used the nonparametric version of the Kishino–Hasegawa bootstrap test with 1,000 bootstraps to obtain p-values for scenario support [31].