We initially searched for PPs that exhibit sequence similarity to any of the transcripts from the 21,921 genes annotated by the Ensembl project 29. The fact that PPs contain few if any introns enabled our search to generate 3,664 PP candidates (Table 1 and Additional data file 1; pseudogenes generated by DNA duplication contained many introns and were eliminated). These candidate PPs represented a minimum set because not all human genes have yet been annotated 30 and the search included only those PPs whose length is more than 90% of the respective mRNA. If the estimated 35,000 human genes 1428 had been used in the search and shorter PPs included in the analysis, over 7,000 PPs would have been expected.

Parental genes of human PPs are of various types, including those for enzymes, structural proteins and regulatory proteins such as ligand-binding proteins and transcription factors (Table 1). Of the total PPs analyzed, the relative frequency of those derived from genes encoding enzymes, structural proteins and ligand-binding proteins was 19%, 34%, and 9% (Figure 1b), respectively, whereas the PP parental genes for structural proteins constituted only 9% of the total parental genes (Figure 1a). Among 1,299 parental genes identified in this study, kinases, ribosomal proteins and ligand-binding proteins were predominant.

Table 2 shows a compilation of the abundant PPs in the human genome (see Additional data file 2). The three most abundant types of human PPs were derived from the genes for keratin 18, glyceraldehyde-3-phosphate dehydrogenase (GAPD) and ribosomal protein L21 (RP L21). These genes generated at least 52, 43 and 38 copies of PPs, respectively, in the genome. Keratin 18 is commonly expressed in internal epithelia and is one of the earliest intermediate filament proteins expressed during embryogenesis 31. The genes for GAPD and ribosomal proteins are housekeeping genes. These data suggest that mRNAs for keratin 18, GAPD and RPL21 were highly expressed or stable in either the germline cells or at an early stage of development, as heritable copies of these genes must have been reverse transcribed in one of those two instances.

As shown in Figure 1a and 1b, structural-protein PPs constitute the largest class (34%). The 50 most prolific PP parental genes include 25 ribosomal protein genes (Table 2) which contribute substantially to the high incidence of structural proteins among the total number of PPs presented in Figure 1b.

Human PPs are derived from mRNAs that exhibit a wide range of GC content. We examined the possible relationship between the number of PPs derived from a gene and the GC content of its mRNA (Figure 2). The rates of PP generation from parental genes within each GC group show no significant statistical difference except for genes of high GC content (> 0.62). This result differs from that of a previous study in which an inverse correlation between the number of ribosomal protein PPs and the GC content of the parental genes was observed 32. Because we analyzed a wide variety of PPs, including those of ribosomal protein genes, the correlation observed in this previous study probably reflects a specific correlation between GC content and either expression level or stability of ribosomal protein mRNAs.

The 3,664 PP candidates are distributed throughout all 24 chromosomes and were derived from genes on various chromosomes (Table 3 and Figure 3). No stringent bias of gene 'projections' (that is the insertion of the PP of a gene in a specific chromosomal location) toward specific chromosomes was observed. In some chromosomes, however, (for example chromosome 19), the ratios of self-projection are relatively high. Interestingly, the PP density within each chromosome roughly parallels gene density (Figure 4). For example, chromosomes that are gene-rich, such as 19, 17 and 11 1428, tend to be relatively PP-rich. On the other hand, chromosomes that are gene-poor, such as Y, 21, 13 and 4 1428, tend to also be poor in PPs. As human gene density shows strong positive correlation with local GC content 1428, this result suggests that the integration of PPs into chromosomes in general may be dependent on aspects of the genomic environment that are strictly related to chromosomal gene density 1428, such as local GC content and an open chromatin structure that facilitates transcription.

To approximate the age of each PP, we developed a method for estimating the level of nucleotide substitutions relative to the parental gene. Initially, this method corrected for the sequence divergence value (a consequence of nucleotide-substitution processes) by removing the contribution of mutations at CpG sites. The C-to-T transition rate in CpG pairs is around 12-fold higher than the rate for other transitions 33 and causes distortions when comparing different genomic elements of high (for example, Alus) or low (for example, L1s) CpG content. Assuming that CpG frequency (ι) in a genomic element that was generated by duplication of a functional gene of high CpG content decreases over time (t) and reaches a state of equilibrium (ε) (approximately 20% of the frequency 1433 expected from the local fraction of cytosines and guanosines 14), the time since the duplication (T) was calculated (see Materials and methods) from the given sequence divergence (D) and the neutral mutation rate (μ) of primates:

D = ∫₀^Tμ(1 + 11((ι - ε)/((0.01ι/(0.99ι-ε))t + 1) + ε))dt

Next, the quantity Σ (=μT) was corrected for multiple substitutions at the same site using the Jukes-Cantor model 34, giving the average number of substitutions per 100 base-pairs (bp), (K). For PPs, sequence divergences were defined as the mismatch rates of respective PPs relative to the current parental gene sequences. Finally, the levels of substitution that accumulated only in PPs were estimated (see Materials and methods). The estimated levels of substitutions in PPs (K(ψ)) were then calculated as K(ψ) = 0.705 K.

Using the levels of nucleotide substitution in PPs estimated by K(ψ), we next evaluated the total number of PPs having the same substitution value, thus approximating the age distribution of PPs. We initially presumed that if PPs were generated at a roughly constant rate during primate evolution, their age distribution would be nearly flat. Surprisingly, a Poisson-like distribution was obtained (Figure 5a). This result indicates that PPs in general may have been generated at extraordinarily high rates during some periods. If the rate of nucleotide substitution is assumed to be 1.5 × 10^-9 per nucleotide per year 3435, then our data estimate that the peak of PP generation occurred approximately 40 million years ago, coincident with the onset of the radiation of the higher primates 3637.

The above results are reminiscent of the amplification profile of Alu repeats. Alu elements comprise approximately 10% of the human genome 14 and are restricted to primates 2223242526. It has been proposed that the average age of Alu repeats is around 40 million years and that the majority of Alus were generated around this time 14222326. We confirmed these previous results by re-estimating the age distribution of all human Alu repeats (Figure 5b). The Alus also showed a Poisson-like distribution with a sharp peak. Alus are classified into distinct subfamilies that can be identified on the basis of mutations shared among subfamily members 2326. Alu subfamilies were derived from a small number of source or master genes. Accordingly, a consensus sequence constructed from members of each subfamily represents each subfamily's source gene(s) 2326. To evaluate the contribution of each subfamily to the entire distribution of Alus, we estimated the age distribution of respective Alu subfamilies (Figure 5c). The peaks for respective subfamilies are grouped closely, and the subfamily Alu Sx strongly influences the overall distribution of Alus (compare with Figure 5b). Therefore, the Sx subfamily (and thus Alus in general) appears to have been amplified intensively over a relatively short period. To the best of our knowledge, many previous discussions of Alu amplification reflect this viewpoint of Alu evolution 14222326. However, our results show that the intensive generation of two distinct elements, PPs and Alus, occurred almost simultaneously suggesting that an unknown change in either the cellular environment or the proliferation mechanism itself enhanced the proliferation of such retroposons in ancestral primates 40-50 million years ago.

Recent progress in L1 biology shows that mammalian L1-encoded proteins are likely to have been involved in the reverse transcription of Alus and PPs 1718192021. To elucidate the cause of the elevated retrotransposition of PPs and Alus, we analyzed the age distribution of all human L1s (Figure 5d). Curiously, the rate of amplification (retrotransposition in cells and fixation within a population) of L1s does not peak around 7%, as was the case for PPs and Alus (compare with Figure 5a,b), raising the issue of how the rate of PP/Alu retrotransposition became elevated during a period of moderate change in L1 retrotransposition. To address this problem, L1s were divided into around 80 subfamilies 38, and age distributions for representative subfamilies are shown in Figure 5e. Although the distributions of respective subfamilies overlap, each subfamily has emerged successively during approximately 150 million years of mammalian evolution. Merging the distribution profiles of all the L1s yields a curve that is rather flat (almost equal to the curve that connects the apices of the respective bars in Figure 5d). Among a large number of L1 subfamilies, certain subfamilies, namely L1PA6, L1PA7 and L1PA8, were amplified intensively around 47 million years ago (the time corresponding to the 7% score). These data suggest that only one or a few L1 subfamilies may have contributed to the increased level of Alu and PP amplification (see Discussion).

Figure 6 shows phylogenetic relationships between L1 subfamilies. A considerable number of substitutions are evident that could explain a possible functional change in L1s between these subfamilies and the current L1 subfamily (L1Hs/L1PA1). There are several amino-acid substitutions within evolutionarily conserved domains (for example, 'C (cysteine)-rich domain') that result in altered residue polarity or charge (Figure 6). A key example is the highly variable amino-terminal half of the L1-encoded ORF1 protein, which contains residues that may be critical for interaction with other proteins 39. There is 41% (54/131) amino-acid divergence between the ORF1 amino-terminal halves of L1PA7 and L1Hs whereas the divergence is only 7% (14/207) in the carboxy-terminal half (data not shown).

A recent extensive survey of the human genome revealed a large number of ribosomal protein pseudogenes derived from the 79 functional ribosomal protein genes 32. The discussion of the ages of these pseudogenes is problematic, however, in that ages were calculated by simply dividing sequence divergences by mutation rate. As the sequence divergence of a PP relative to its parental gene (K) is dependent on substitutions in both the PP (K(ψ)) and the gene (K(f)) (see Materials and methods), the ages of the ribosomal protein pseudogenes were overestimated. For example, with respect to RPL21 mRNAs (the most predominant source of ribosomal protein pseudogenes in humans), the sequence divergence between human and mouse or rat is approximately 11% (NM_000982, NM_019647, NM_053330, and 40). The previous ribosomal protein pseudogene calculations dismissed sequence divergences between the present-day and primordial genes, probably overestimating the ages by around 10 million years (a few percent per 8-10% of divergence). Therefore, it is difficult to compare such values with the ages of Alus/L1s. Our method provides a clear solution to this matter, enabling us to compare the ages of different classes of retroposons. Hence, our method led us to the finding that there was a simultaneous burst of PPs and Alus - a 'retrotranspositional explosion' - in the primate genome.

Regarding the cause of the retrotranspositional explosion, it is worth considering the effect of a 'bottleneck' 3441 during primate evolution. Only individuals that experienced extensive genomic retrotransposition might have propagated to become a majority within a population of ancestral primates, via a mechanism involving a rapid reduction in the general population. Studies on the molecular phylogeny and demographic history of humans show, however, that the primate lineage leading to humans never experienced an extensive bottleneck, at least since its divergence from the prosimian lineage 42. Therefore, the effect of a bottleneck can be largely ignored.

The retrotranspositional explosion could be due to a change in the cellular environment of ancestral primates 40-50 million years ago, such as a higher transcriptional potential of parental (master) genes of PPs and Alus. A specific environment of the genome during the period of the retrotranspositional explosion, such as more available target sites of PPs and Alus, might have facilitated this event. Alternatively, a change in the proliferation mechanism of PPs and Alus, such as an increased amount of reverse transcriptase or an enhanced activity of enzymes for retrotransposition, might have promoted the explosion.

Recent studies on the L1 retrotransposons show that mammalian L1-encoded proteins may have been involved in the reverse transcription of Alus and PPs 1718192021. Here, we have shown that the intensive amplification of distinct genetic elements, namely PPs and Alus, seems to have occurred almost simultaneously around 40-50 million years ago, and suggests that only one or a few L1 subfamilies may have contributed to the observed high levels of Alu/PP retrotransposition.

How could a specific L1 subfamily (or subfamilies) have generated Alus and PPs at such an accelerated rate? We propose that L1s within specific subfamilies mobilized RNAs in trans at accelerated rates in ancestral primate genomes. Thus, a specific L1 subfamily may have mediated the Alu/PP retrotranspositional explosion. The age distributions estimated in this study allow the prediction of the most probable L1 subfamilies responsible for the explosion (care must be exercised when comparing ages between distinct genetic elements; see Materials and methods). The most probable candidate subfamilies are L1PA6, L1PA7 and L1PA8 (Figure 5e). As mentioned above, although the youngest L1 subfamily mobilizes cellular RNAs in trans at very low frequencies (0.01-0.05%) in HeLa cells, the frequency is not necessarily intrinsic to L1s. In fact, in cultured feline cells the frequency of L1-mediated PP formation in trans is 5% relative to that of L1 retrotransposition in cis 18. Moreover, an eel LINE family exhibits a high level of trans retrotransposition (up to 30% 43), and the frequency of L1-mediated Alu retrotransposition in HeLa cells is 100-1,000 times higher than control mRNAs 21. Although L1 subfamilies such as L1PA6, L1PA7 and L1PA8 appear to have been extinguished by cumulative mutations, the possibility that an ancient L1 subfamily exhibited an enhanced ability to mobilize RNAs in trans could be verified experimentally in HeLa cells using reconstructed L1 subfamilies 44 as sources for reverse transcription of trans RNAs.

Alu insertions mediate many genomic rearrangements, such as unequal crossing over, induction of alternative splicing, and the introduction of new promoters, poly(A) signals and even new exons 6. Inactivation of CMP-N-acetylneuraminic acid hydroxylase (around 2.8 million years ago) before brain expansion during human evolution occurred by an Alu-mediated inactivating mutation 45, representing yet another example of the impact of the Alu expansion. The current frequency of human endogenous insertional mutations caused by Alu retrotransposition is estimated at around 1 in every 16-200 individuals 2646. The frequency of Alu insertion at the time of the retrotranspositional explosion is estimated to have been 30-200 times higher than the frequency over the last 10 million years (26 and data not shown). This implies that at least one in seven individuals at the time carried new Alu insertions in their genomes (a maximum of 12 insertions per individual). This high Alu insertion rate may have had a much greater impact on ancestral primate genomes compared with the impact of present-day mutations.

Retrotransposition of PPs causes not only insertional mutations but also the propagation of new genes. These 'retrogenes' comprise PPs that inserted themselves next to resident promoter/enhancer elements and thereby escaped transcriptional silencing and PPs that were initially inactive but were reactivated at a later time when flanking regulatory elements became activated by mutation 6. Retrogenes are often observed in primate genomes 6, one example being the testis-specific human gene CDY (on the Y chromosome), which arose during primate evolution by retrotransposition of the ubiquitous mRNA of the gene CDYL located on chromosome 13 7. From the observed distribution of CDY homologs in primates, this event appears to have occurred in the simian lineage after its divergence from prosimians but before the split between Old and New World monkeys 7 during the period of the retrotranspositional explosion. We predict that further studies will demonstrate that many human retrogenes were generated during this period, and postulate that such retrogenes were involved in generating new characteristics that are specific to simian primates 847.

Over the course of eukaryotic evolution, the extensive propagation of new genes preceded apparent bursts of new organisms or the emergence of new hierarchies of morphological complexity 4849. The time of the retrotranspositional explosion can be estimated at 40-50 million years ago, assuming a nucleotide substitution rate of 1.5 × 10^-9 per nucleotide per year 3435. Fossil records show that before this period, simian ancestors (monkeys, apes and humans) diverged from prosimians (lemurs and lorises) with the divergence of New World monkeys and the radiation of the remaining primates proceeding immediately thereafter 3637 (Figure 7). The rapid amplification of Alus to the level of 10% of the primate genome and the creation of numerous replicas of various genes may have provided the molecular basis that led to the radiation of higher primates.

PPs were searched for in an assembled human genome sequence (Human Genome Project Working Draft, April 1 2001) 50 using BLAT 51. The BLAT setting was as follows: Assembly: April 1, 2001; Query type: DNA; Sort output: query, score; Output type: hyperlink. 'Confirmed cDNAs' (23,929 entries) in Ensembl DB (v1.1.0) 29 were used as queries. The subject with the highest score was regarded as the gene encoding the transcript. Multiple hits were subjected to analysis.

Subjects that contained over 90% of the query length were used. The number of aligning blocks, which usually corresponds to the number of exons, was compared between a gene and other subjects. If the number of aligning blocks was smaller than that of the gene, the subject was further analyzed, thus eliminating pseudogenes generated by DNA duplications. Subjects that were identified by intronless genes (single exon genes) were not included in the analysis to avoid confusing PPs and pseudogenes generated by DNA duplications. To avoid confusing phylogenetic relationships, loci (subjects) that were identified by multiple query hits were not included in the analysis. A series of Perl scripts were designed to analyze the BLAT search results.

To evaluate our annotation of PPs, our results for chromosomes 21 and 22 were compared with those from other studies. For chromosome 21, the PP total in this study was 34 whereas previous studies reported 41 5253 and 57 4. The number of annotations common to two studies totaled 18 (this study and 52), 14 (this study and 4) and 21 452. Annotations common to all studies totaled 10. For chromosome 22, the PP total in this study was 62, whereas previous studies reported 91 5455 and 73 4. The number of common annotations totaled 37 (this study and 54), 28 (this study and 4) and 52 454. Annotations common to all studies totaled 27. Differences between the numbers appear to derive mainly from differences in the gene sets used for the analyses 30.

For each Alu and L1 repeat, the genomic location and sequence divergence was obtained from the output file of the RepeatMasker program applied to the human genome draft sequence (22 December 2001 56). Sequence divergences were defined as the mismatch rates of respective repeats relative to the consensus sequence of respective subfamilies.

The level of substitutions that accumulated in a PP (K(ψ)) was estimated using the following method.

First, the sequence divergence value (D) was corrected by removing the contribution of mutations at CpG sites. Sequence divergence (δ) of a sequence (at a given time point) of length (N) including the number of CpG dinucleotides (n) is given as a function of the mutation rate at non-CpG dinucleotides (α) and CpG dinucleotides (β) as follows:

δ = α(1/2 - n/N) + βn/N     (1)

From the result of Sved and Bird 33, the ratio of β to α is ∼ 6.5. Therefore, designating α/2 = μ and n/N = ν in Equation 1 gives the following:

δ = μ(1 + 11ν)     (2)

Assuming that CpG frequency (ι) in a genomic element that was generated by duplication of a functional gene of high CpG content decreases over time (t) and reaches an equilibrium state (ε) (approximately 20% of the frequency 1433 expected from the local fraction of cytosines and guanosines 14), the CpG frequency (ν) at time (t) was calculated as follows:

ν = 1/(At + 1/(ι - ε)) + ε     (3)

If we accept the value of 1.5 × 10^-9 per nucleotide per year 3435 as the neutral mutation rate 41 and equate this to the mutation rate at non-CpG dinucleotides (μ) and use a time unit of 1 million years, then the mutation rate at CpG dinucleotides, β/2, will be around 1 per 100 nucleotides per million years (that is, ν will be reduced by 1% every million years). Therefore, ν(t = 1)/ν(t = 0) in Equation 3 gives:

(1/(A + 1/(ι - ε)) + ε)/ι ≈ 0.99

Solving for A gives:

A = 0.01ι/((0.99ι - ε)(ι - ε))     (4)

The sequence divergence value (D) is given as an integral of the sequence divergence (δ) from the present (t = 0) to the time of the duplication (t = T): D = ∫₀^Tδdt. From Equations 2, 3 and 4,

D = ∫₀^Tμ(1 + 11((ι - ε)/((0.01ι/(0.99ι - ε))t + 1) + ε))dt   (5)

Solving Equation 5 for T gives the time since the duplication. The following ι and ε values (ι, ε, respectively) were used for the retroposons shown 1457:

Alu (0.077, 0.020); L1 (0.012, 0.008); PPs (0.015, 0.010)

The substitution level (Σ) at sites other than CpG is given from the time since the duplication (T) and the neutral mutation rate (μ) of primates 41: Σ = μT. The quantity Σ was corrected for multiple substitutions at the same site using the Jukes-Cantor model 34, giving the average number of substitutions per 100 bp (K): K = - (3/4)ln(1 - (4/3)Σ).

For PPs, sequence divergences were defined as the mismatch rates of respective PPs relative to the current sequences of their parental genes. The mismatch rate of a PP relative to its parental gene (K) consists of the level of substitutions that accumulated only in the PP (K(ψ)) and the level of substitutions that accumulated only in the gene (K(f)): K = K(f) + K(ψ). K(f) and K(ψ) can be further subdivided into the number of synonymous (Ks) and nonsynonymous (Ka) substitutions 585960: K(f) = Ks(f) + Ka(f), K(ψ) = Ks(ψ) + Ka(ψ). Kuma and Miyata evaluated the average nucleotide substitution rates of 31 pairs of human PPs and their parental genes using homologs of other species as outgroups (K. Kuma and T. Miyata, personal communication). They used the following genes: ADP-ribosylation factor 1, aldolase A, aldose reductase, alpha-E-catenin, alpha-L-fucosidase, alpha-enolase, arylamine N-acetyltransferase, beta-tubulin, c-Raf protooncogene, cAMP-dependent protein kinase regulatory subunit, calmodulin, ceruloplasmin, creatine kinase, cyclophilin, cytochrome b5, cytochrome c, ferrochelatase, gamma-actin, glucocerebrosidase, glutamine synthetase, glyceraldehyde-3-phosphate dehydrogenase, histone H3.3, hsc70, hsp27, hsp60, lactate dehydrogenase-A, neurotrophin-4, phosphoglycerate kinase, prothymosine alpha, topoisomerase-I, triose phosphate isomerase. They calculated the following ratios: Rs(ψ), the synonymous substitutions in PPs to synonymous substitutions in their parental genes; Ra(f), the ratio of nonsynonymous substitutions in genes to synonymous substitutions in genes; and Ra(ψ), the ratio of nonsynonymous substitutions in PPs to synonymous substitutions in genes. The mean values of Rs(ψ), Ra(ψ) and Ra(f) were:

Rs(ψ) = Ks(ψ)/Ks(f) = 1.40     (6.1)

Ra(ψ) = Ka(ψ)/Ks(f) = 1.13     (6.2)

Ra(f) = Ka(f)/Ks(f) = 0.06     (6.3)

From Equations 6.1-6.3, and Equations K = K(f) + K(ψ), K(f) = Ks(f) + Ka(f), and K(ψ) = Ks(ψ) + Ka(ψ), the estimated level of substitutions in PPs (K(ψ)) is given by:

K(ψ) = 0.705K     (7)