International Human Genome Sequencing Consortium

Authors:
Addresses:
Correspondence:
Abstract:
Introduction:
Summary of Findings:
Methods:
Supplementary Information:
...
Repeat Content of the Human Genome:
    Transposon-Derived Repeats:
        Classes of Transposable Elements:
           Figure 17. Classes of interspersed repeat transposons in the human genome:
           LINEs:
           SINEs:
           LTR retroposons:
           DNA retroposons:
        Census of human repeats:
           Table 11. Number of copies and fraction of  genome for interspersed repeat:
        Age distribution:
           Figure 18. Age distribution of interspersed repeats in the human and mouse genomes:
        Comparisons with other organisms:
           Table 12. Number and nature of interspersed repeats in eukaryote genomes:
           Figure 20. Comparison of the age of interspersed repeats in eukaryote genomes:
         Variation in the distribution of repeats:
          ...
          Active transposons:
           Transposons as a creative force:
    Simple sequence repeats:
    ...
Gene Content of the Human Genome:
    Noncoding RNAs:
         Transfer RNA genes:
         Ribosomal RNA genes:
         Small nucleolar RNA genes:
         Spliceosomal RNAs and other ncRNA genes:
     Protein-coding genes:
          Exploring properties of known genes:
...
The next steps:
     Finishing the human sequence:
     Developing the IGI and IPI:
     Large-scale identification of regulatory regions:
     Sequencing of additional large genomes:
     Completing the catalogue of human variation:
     From sequence to function:
     Concluding thoughts:
DNA Sequence Databases:
Supplementary Information:
Acknowledgements:
References:
Background References:
Additional References:
Other Sites:
Feedback:

The human genome holds an extraordinary trove of information about human development, physiology, medicine and evolution. Here we report the results of an international collaboration to produce and make freely available a draft sequence of the human genome. We also present an initial analysis of the data, describing some of the insights that can be gleaned from the sequence.

The rediscovery of Mendel's laws of heredity in the opening weeks of the 20th century [1-3] sparked a scientific quest to understand the nature and content of genetic information that has propelled biology for the last hundred years. The scientific progress made falls naturally into four main phases, corresponding roughly to the four quarters of the century. The first established the cellular basis of heredity: the chromosomes. The second defined the molecular basis of heredity: the DNA double helix. The third unlocked the informational basis of heredity, with the discovery of the biological mechanism by which cells read the information contained in genes and with the invention of the recombinant DNA technologies of cloning and sequencing by which scientists can do the same.

The last quarter of a century has been marked by a relentless drive to decipher first genes and then entire genomes, spawning the field of genomics. The fruits of this work already include the genome sequences of 599 viruses and viroids, 205 naturally occurring plasmids, 185 organelles, 31 eubacteria, seven archaea, one fungus, two animals and one plant.

Here we report the results of a collaboration involving 20 groups from the United States, the United Kingdom, Japan, France, Germany and China to produce a draft sequence of the human genome. The draft genome sequence was generated from a physical map covering more than 96% of the euchromatic part of the human genome and, together with additional sequence in public databases, it covers about 94% of the human genome. The sequence was produced over a relatively short period, with coverage rising from about 10% to more than 90% over roughly fifteen months. The sequence data have been made available without restriction and updated daily throughout the project. The task ahead is to produce a finished sequence, by closing all gaps and resolving all ambiguities. Already about one billion bases are in final form and the task of bringing the vast majority of the sequence to this standard is now straightforward and should proceed rapidly.

The sequence of the human genome is of interest in several respects. It is the largest genome to be extensively sequenced so far, being 25 times as large as any previously sequenced genome and eight times as large as the sum of all such genomes. It is the first vertebrate genome to be extensively sequenced. And, uniquely, it is the genome of our own species.

Much work remains to be done to produce a complete finished sequence, but the vast trove of information that has become available through this collaborative effort allows a global perspective on the human genome. Although the details will change as the sequence is finished, many points are already clear.
 

In this paper, we start by presenting background information on the project and describing the generation, assembly and evaluation of the draft genome sequence. We then focus on an initial analysis of the sequence itself: the broad chromosomal landscape; the repeat elements and the rich palaeontological record of evolutionary and biological processes that they provide; the human genes and proteins and their differences and similarities with those of other organisms; and the history of genomic segments. (Comparisons are drawn throughout with the genomes of the budding yeast Saccharomyces cerevisiae, the nematode worm Caenorhabditis elegans, the fruitfly Drosophila melanogaster and the mustard weed  Arabidopsis thaliana; we refer to these for convenience simply as yeast, worm, fly and mustard weed.) Finally, we discuss applications of the sequence to biology and medicine and describe next steps in the project. (underlines by WebEditor).

A full description of the methods is provided as Supplementary Information on Nature's web site.

Repeat content of the human genome:

A puzzling observation in the early days of molecular biology was that genome size does not correlate well with organismal complexity. For example, Homo sapiens has a genome thais 200 times as large as that of the yeast S. cerevisiae, but 200 times as small as that of Amoeba dubia [139, 140]. This mystery (the C-value paradox) was largely resolved with the recognition that genomes can contain a large quantity of repetitive sequence, far in excess of that devoted to protein-coding genes (reviewed in refs [140, 141]).

In the human, coding sequences comprise less than 5% of the genome (see below), whereas repeat sequences account for at least 50% and probably much more. Broadly, the repeats fall into five classes: (1) transposon-derived repeats, often referred to as interspersed repeats; (2) inactive (partially) retroposed copies of cellular genes (including protein-coding genes and small structural RNAs), usually referred to as processed pseudogenes; (3) simple sequence repeats, consisting of direct repetitions of relatively short k-mers such as (A)_n, (CA)_nor (CGG)_n; (4) segmental duplications, consisting of blocks of around 10–300 kb that have been copied from one region of the genome into another region; and (5) blocks of tandemly repeated sequences, such as at centromeres, telomeres, the short arms of acrocentric chromosomes and ribosomal gene clusters. (These regions are intentionally under-represented in the draft genome sequence and are not discussed here.)

Repeats are often described as 'junk' and dismissed as uninteresting. However, they actually represent an extraordinary trove of information about biological processes. The repeats constitute a rich palaeontological record, holding crucial clues about evolutionary events and forces. As passive markers, they provide assays for studying processes of mutation and selection. It is possible to recognize cohorts of repeats 'born' at the same time and to follow their fates in different regions of the genome or in different species. As active agents, repeats have reshaped the genome by causing ectopic rearrangements, creating entirely new genes, modifying and reshuffling existing genes, and modulating overall GC content. They also shed light on chromosome structure and dynamics, and provide tools for medical genetic and population genetic studies.

The human is the first repeat-rich genome to be sequenced, and so we investigated what information could be gleaned from this majority component of the human genome. Although some of the general observations about repeats were suggested by previous studies, the draft genome sequence provides the first comprehensive view, allowing some questions to be resolved and new mysteries to emerge.

Transposon-derived repeats

Most human repeat sequence is derived from transposable elements [142, 143]. We can currently recognize about 45% of the genome as belonging to this class. Much of the remaining 'unique' DNA must also be derived from ancient transposable element copies that have diverged too far to be recognized as such. To describe our analyses of interspersed repeats, it is necessary briefly to review the relevant features of human transposable elements.

Classes of transposable elements.

In mammals, almost all transposable elements fall into one of four types (Fig. 17), of which three transpose through RNA intermediates and one transposes directly as DNA. These are long interspersed elements (LINEs), shortinterspersed elements (SINEs), LTR retrotransposons and DNA transposons.

Figure 17. Almost all transposable elements in mammals fall into one of four classes. (See text for details). 

SINEs are wildly successful freeloaders on the backs of LINE elements. They are short (about 100–400 bp), harbour an internal polymerase III promoter and encode no proteins.These non-autonomous transposons are thought to use the LINE machinery for transposition. Indeed, most SINEs 'live' by sharing the 3' end with a resident LINE element [144]. The promoter regions of all known SINEs are derived from tRNA sequences, with the exception of a single monophyletic family of SINEs derived from the signal recognition particle component 7SL. This family, which also does not share its 3' end with a LINE, includes the only active SINE in the human genome: the Alu element. By contrast, the mouse has both tRNA-derived and 7SL-derived SINEs. The human genome contains three distinct monophyletic families of SINEs: the active Alu, and the inactive MIR and Ther2/MIR3.

LTR retroposons are flanked by long terminal direct repeats that contain all of the necessary transcriptional regulatory elements. The autonomous elements (retrotransposons) contain gag and pol genes, which encode a protease, reverse transcriptase, RNAse H and integrase. Exogenous retroviruses seem to have arisen from endogenous retrotransposons by acquisition of a cellular envelope gene (env) [147]. Transposition occurs through the retroviral mechanism with reverse transcription occurring in a cytoplasmic virus-like particle, primed by a tRNA (in contrast to the nuclear location and chromosomal priming of LINEs). Although a variety of LTR retrotransposons exist, only the vertebrate-specific endogenous retroviruses (ERVs) appear to have been active in the mammalian genome. Mammalian retroviruses fall into three classes (I–III), each comprising many families with independent origins. Most (85%) of the LTR retroposon-derived 'fossils' consist only of an isolated LTR, with the internal sequence having been lost by homologous recombination between the flanking LTRs.

DNA transposons resemble bacterial transposons, having terminal inverted repeats and encoding a transposase that binds near the inverted repeats and mediates mobility through a 'cut-and-paste' mechanism. The human genome contains at least seven major classes of DNA transposon, which can be subdivided into many families with independent origins [148] (see RepBase, http://www.girinst.org/~server/repbase.html DNA transposons tend to have short life spans within a species. This can be explained by contrasting the modes of transposition of DNA transposons and LINE elements. LINE transposition tends to involve only functional elements, owing to the cis-preference by which LINE proteins assemble with the RNA from which they were translated. By contrast, DNA transposons cannot exercise a cis-preference: the encoded transposase is produced in the cytoplasm and, when it returns to the nucleus, it cannot distinguish active from inactive elements. As inactive copies accumulate in the genome, transposition becomes less efficient. This checks the expansion of any DNA transposon family and in due course causes it to die out. To survive, DNA transposons must eventually move by horizontal transfer to virgin genomes, and there is considerable evidence for such transfer [149-153].

Transposable elements employ different strategies to ensure their evolutionary survival. LINEs and SINEs rely almost exclusively on vertical transmission within the host genome [154] (but see refs 148, 155).DNA transposons are more promiscuous, requiring relatively frequent horizontal transfer. LTR retroposons use both strategies, with some being long-term active residents of the human genome (such as members of the ERVL family) and others having only short residence times.

Census of human repeats.

We began by taking a census of the transposable elements in the draft genome sequence, using a recently updated version of the RepeatMasker program (version 09092000) run under sensitive settings (see
http://repeatmasker.genome.washington.edu). This program scans sequences to identify full-length and partial members of all known repeat families represented in RepBase Update (version 5.08; see http://www.girinst.org/~server/repbase.html and ref. [156]). Table 11 shows the number of copies and fraction of the draft genome sequence occupied by each of the four major classes and the main subclasses.
 

Table 11. Number of copies and fraction of genome for classes of interspersed repeat.

	Number of copies  x1,000	Total number of bases in the draft genome sequence  (Mb)	Fraction of the  draft genome  sequence (%)	Number of families (subfamilies)
SINEs	1,558	359.6	13.14
Alu	1,090	290.1	10.60	1(~20)
MIR	393	60.1	2.20	1(1)
MIR3	75	9.3	0.34	1(1)
LINEs	868	558.8	20.42
LINE1	516	462.1	16.89	1(~55)
LINE2	315	88.2	3.22	1(2)
LINE3	37	8.4	0.31	1(2)
LTR elements	443	227.0	8.29
ERV-class I	112	79.2	2.89	72(132)
ERV(K)-class II	8	8.5	0.31	10(20)
ERV(L)-class III	83	39.5	1.44	21(42)
MaLR	240	99.8	3.65	1(31)
DNA elements	294	77.6	2.84
hAT group
MER1-Charlie	182	38.1	1.39	25(50
Zaphod	13	4.3	0.16	4(10)
Tc-1 group
MER2-Tigger	57	28.0	1.02	12(28)
Tc2	4	0.9	0.03	1(5)
Mariner	14	2.6	0.10	4(5)
PiggyBac-like	2	0.5	0.02	10(20)
Unclassified	22	3.2	0.12	7(7)
Unclassified	3	3.8	0.14	3(4)
Total interspersed repeats		1,226.8	44.83

The number of copies and base pair contributions of the major classes and subclasses of transposable elements in the human genome. Data extracted from a RepeatMasker analysis of the draft genome sequence (RepeatMasker version 09092000, sensitive settings, using RepBase Update 5.08). In calculating percentages, RepeatMasker excluded the runs of Ns linking the contigs in the draft genome sequence. In the last column, separate consensus sequences in the repeat databases are considered subfamilies, rather than families, when the sequences are closely related or related through intermediate subfamilies.

Age distribution.

The age distribution of the repeats in the human genome provides a rich 'fossil record' stretching over several hundred million years. The ancestry and approximate age of each fossil can be inferred by exploiting the fact that each copy is derived from, and therefore initially carried the sequence of, a then-active transposon and, being generally under no functional constraint, has accumulated mutations randomly and independently of other copies. We can infer the sequence of the ancestral active elements by clustering the modern derivatives into phylogenetic trees and building a consensus based on the multiple sequence alignment of a cluster of copies. Using available consensus sequences for known repeat subfamilies, we calculated the per cent divergence from the inferred ancestral active transposon for each of three million interspersed repeats in the draft genome sequence.

The percentage of sequence divergence can be converted into an approximate age in millions of years (Myr) on the basis of evolutionary information. Care is required in calibrating the clock, because the rate of sequence divergence may not be constant over time or between lineages [139]. The relative-rate test [157] can be used to calculate the sequence divergence that accumulated in a lineage after a given timepoint, on the basis of comparison with a sibling species that diverged at that time and an outgroup species. For example, the substitution rate over roughly the last 25 Myr in the human lineage can be calculated by using old world monkeys (which diverged about 25 Myr ago) as a sibling species and new world monkeys as an outgroup. We have used currently available calibrations for the human lineage, but the issue should be revisited as sequence information becomes available fromdifferent mammals.

Figure 18a shows the representation of various classes of transposable elements in categories reflecting equal amounts of sequence divergence. In Fig. 18b the data are grouped into four bins corresponding to successive 25-Myr periods, on the basis of an approximate clock. Figure 19 shows the mean ages of various subfamilies of DNA transposons.

Figure 18. Age distribution of interspersed repeats in the human and mouse genomes. Bases covered by interspersed repeats were sorted by their divergence from their consensus sequence (which approximates the repeat's original sequence at the time of insertion). The average number of substitutions per 100 bp (substitution level, K ) was calculated from the mismatch level p assuming equal frequency of all substitutions (the one-parameter Jukes±Cantor model, K = -3/4ln(1- 4/3p)). This model tends to underestimate higher substitution levels. CpG dinucleotides in the consensus were excluded from the substitution level calculations because the C ! T transition rate in CpG pairs is about tenfold higher than other transitions and causes distortions in comparing transposable elements with high and low CpG content. a, The distribution, for the human genome, in bins corresponding to 1% increments in substitution levels. b, The data grouped into bins representing roughly equal time periods of 25 Myr. c,d, Equivalent data for available mouse genomic sequence. There is a different correspondence between substitution levels and time periods owing to different rates of nucleotide substitution in the two species. The correspondence between substitution levels and time periods was largely derived from three-way species comparisons (relative rate test [139, 157]) with the age
estimates based on fossil data. Human divergence from gibbon 20±30 Myr; old world monkey 25±35 Myr; prosimians 55±80 Myr; eutherian mammalian radiation, 100 Myr.

Several facts are apparent from these graphs. First, most interspersed repeats in the human genome predate the eutherian radiation. This is a testament to the extremely slow rate with which nonfunctional sequences are cleared from vertebrate genomes (see below concerning comparison with the fly).

Second, LINE and SINE elements have extremely long lives. The monophyletic LINE1 and Alu lineages are at least 150 and 80 Myr old, respectively. In earlier times, the reigning transposons were LINE2 and MIR [148, 158]. The SINE MIR was perfectly adapted for reverse transcription by LINE2, as it carried the same 50-base sequence at its 3' end. When LINE2 became extinct 80–100 Myr ago, it spelled the doom of MIR.

Third, there were two major peaks of DNA transposon activity (Fig. 19). The first involved Charlie elements and occurred long before the eutherian radiation; the second involved Tigger elements and occurred after this radiation. Because DNA transposons can produce large-scale chromosome rearrangements [159-162], it is possible that widespread activity could be involved in speciation events.

Fourth, there is no evidence for DNA transposon activity in the past 50 Myr in the human genome. The youngest two DNA transposon families that we can identify in the draft genome sequence (MER75 and MER85) show 6–7% divergence from their respective consensus sequences representing the ancestral element (Fig. 19), indicating that they were active before the divergence of humans and new world monkeys. Moreover, these elements were relatively unsuccessful, together contributing just 125 kb to the draft genome
sequence.

Finally, LTR retroposons appear to be teetering on the brink of extinction, if they have not already succumbed. For example, the most prolific elements (ERVL and MaLRs) flourished for more than 100 Myr but appear to have died out about 40 Myr ago [163, 164]. Only a single LTR retroposon family (HERVK10) is known to have transposed since our divergence from the chimpanzee 7 Myr ago, with only one known copy (in the HLA region) that is not shared between all humans [165]. In the draft genome sequence, we can identify only three full-length copies with all ORFs intact (the final total may be slightly higher owing to the imperfect state of the draft genome sequence).

More generally, the overall activity of all transposons has declined markedly over the past 35–50 Myr, with the possible exception of LINE1 (Fig. 18). Indeed, apart from an exceptional burst of activity of Alus peaking around 40 Myr ago, there would appear to have been a fairly steady decline in activity in the hominid lineage since the mammalian radiation. The extent of the decline must be even greater than it appears because old
repeats are gradually removed by random deletion and because old repeat families are harder to recognize and likely to be under-represented in the repeat databases. (We confirmed that the decline in transposition is not an artefact arising from errors in the draft genome sequence, which, in principle, could increase the divergence level in recent elements. First, the sequence error rate (Table 9) is far too low to have a significant effect on the apparent age of recent transposons; and second, the same result is seen if one considers only finished sequence.)

What explains the decline in transposon activity in the lineage leading to humans? We return to this question below, in the context of the observation that there is no similar decline in the mouse genome.

Comparison with other organisms

We compared the complement of transposable elements in the human genome with those of the other sequenced eukaryotic genomes. We analysed the fly, worm and mustard weed genomes for the number and nature of repeats (Table 12) and the age distribution (Fig. 20). (For the fly, we analysed the 114 Mb of
unfinished 'large' contigs produced by the whole-genome shotgun assembly [166], which are reported to represent euchromatic sequence. Similar results were obtained by analysing 30 Mb of finished euchromatic sequence.) The human genome stands in stark contrast to the genomes of the other organisms.

The complete genomes of fly, worm, and chromosomes 2 and 4 of mustard weed (as deposited at http://www.ncbi.nlm.nih.gov/genbank/genomes) were screened against the repeats in RepBase Update 5.02
(September 2000) with RepeatMasker at sensitive settings.

Figure 20. Comparison of the age of interspersed repeats in eukaryotic genomes. The copies of repeats were pooled by their nucleotide substitution level from the consensus.

(1) The euchromatic portion of the human genome has a much higher density of transposable element copies than the euchromatic DNA of the other three organisms. The repeats in the other organisms may have been slightly underestimated because the repeat databases for the other organisms are less complete than for the human, especially with regard to older elements; on the other hand, recent additions to these databases appear to increase the repeat content only marginally.

(2) The human genome is filled with copies of ancient transposons, whereas the transposons in the other genomes tend to be of more recent origin. The difference is most marked with the fly, but is clear for the other genomes as well. The accumulation of old repeats is likely to be determined by the rate at which organisms engage in 'housecleaning' through genomic deletion. Studies of pseudogenes have suggested that small deletions occur at a rate that is 75-fold higher in flies than in mammals; the half-life of such nonfunctional DNA is estimated at 12 Myr for flies and 800 Myr for mammals [167]. The rate of large deletions has not been systematically compared, but seems likely also to differ markedly.

(3) Whereas in the human two repeat families (LINE1 and Alu) account for 60% of all interspersed repeat sequence, the other organisms have no dominant families. Instead, the worm, fly and mustard weed genomes all contain many transposon families, each consisting of typically hundreds to thousands of elements. This difference may be explained by the observation that the vertically transmitted, long-term residential LINE and SINE elements represent 75% of interspersed repeats in the human genome, but only 5–25% in the other genomes. In contrast, the horizontally transmitted and shorter-lived DNA transposons represent only a small portion of all interspersed repeats in humans (6%) but a much larger fraction in fly, mustard weed and worm (25%, 49% and 87%, respectively). These features of the human genome are probably general to all mammals. The relative lack of horizontallytransmitted elements may have its origin in the well developed immune system of mammals, as horizontal transfer requires infectious vectors, such as viruses, against which the immune system guards.

We also looked for differences among mammals, by comparing the transposons in the human and mouse genomes. As with the human genome, care is required in calibrating the substitution clock for the mouse genome. There is considerable evidence that the rate of substitution per Myr is higher in rodent lineages than in the hominid lineages [139, 168, 169]. In fact, we found clear evidence for different rates of substitution by examining families of transposable elements whose insertions predate the divergence of the human and mouse lineages. In an analysis of 22 such families, we found that the substitution level was an average of 1.7-fold higher in mouse than human (not shown). (This is likely to be an underestimate because of an ascertainment bias against the most diverged copies.) The faster clock in mouse is also evident from the fact that the ancient LINE2 and MIR elements, which transposed before the mammalian radiation and are readily detectable in the human genome, cannot be readily identified in available mouse genomic sequence (Fig.18).

We used the best available estimates to calibrate substitution levels and time [169]. The ratio of substitution rates varied from about 1.7-fold higher over the past 100 Myr to about 2.6-fold higher over the past 25 Myr.

The analysis shows that, although the overall density of the four transposon types in human and mouse is similar, the age distribution is strikingly different (Fig. 18). Transposon activity in the mouse genome has not undergone the decline seen in humans and proceeds at a much higher rate. In contrast to their possible extinction in humans, LTR retroposons are alive and well in the mouse with such representatives as the active IAP family and putatively active members of the long-lived ERVL and MaLR families. LINE1 and a
variety of SINEs are quite active. These evolutionary findings are consistent with the empirical observations that new spontaneous mutations are 30 times more likely to be caused by LINE insertions in mouse than in human (3% versus 0.1%) [170] and 60 times more likely to be caused by transposable elements in general. It is estimated that around 1 in 600 mutations in human are due to transpositions, whereas 10% of mutations in mouse are due to transpositions (mostly IAP insertions).

The contrast between human and mouse suggests that the explanation for the decline of transposon activity in humans may lie in some fundamental difference between hominids and rodents. Population structure and dynamics would seem to be likely suspects. Rodents tend to have large populations, whereas hominid populations tend to be small and may undergo frequent bottlenecks. Evolutionary forces affected by such factors include inbreeding and genetic drift, which might affect the persistence of active transposable
elements [171]. Studies in additional mammalian lineages may shed light on the forces responsible for the differences in the activity of transposable elements [172].

Variation in the distribution of repeats.

We next explored variation in the distribution of repeats across the draft genome sequence, by calculating the repeat density in windows of various sizes across the genome. There is striking variation at smaller scales.

Some regions of the genome are extraordinarily dense in repeats. The prizewinner appears
to be a 525-kb region on chromosome Xp11, with an overall transposable element density of
89%. This region contains a 200-kb segment with 98% density, as well as a segment of
100 kb in which LINE1 sequences alone comprise 89% of the sequence. In addition, there
are regions of more than 100 kb with extremely high densities of Alu (> 56% at three loci,
including one on 7q11 with a 50-kb stretch of > 61% Alu) and the ancient transposons MIR
(> 15% on chromosome 1p36) and LINE2 (> 18% on chromosome 22q12).

In contrast, some genomic regions are nearly devoid of repeats. The absence of repeats
may be a sign of large-scale cis-regulatory elements that cannot tolerate being interrupted
by insertions. The four regions with the lowest density of interspersed repeats in the human
genome are the four homeobox gene clusters, HOXA, HOXB, HOXC and HOXD (Fig.
21). Each locus contains regions of around 100 kb containing less than 2% interspersed
repeats. Ongoing sequence analysis of the four HOX clusters in mouse, rat and baboon
shows a similar absence of transposable elements, and reveals a high density of conserved
noncoding elements (K. Dewar and B. Birren, manuscript in preparation). The presence of
a complex collection of regulatory regions may explain why individual HOX genes carried in
transgenic mice fail to show proper regulation.

It may be worth investigating other repeat-poor regions, such as a region on chromosome
8q21 (1.5% repeat over 63 kb) containing a gene encoding a homeodomain zinc-finger
protein (homologous to mouse pID 9663936), a region on chromosome 1p36 (5% repeat
over 100 kb) with no obvious genes and a region on chromosome 18q22 (4% over 100 kb)
containing three genes of unknown function (among which is KIAA0450). It will be
interesting to see whether the homologous regions in the mouse genome have similarly
resisted the insertion of transposable elements during rodent evolution.

Distribution by GC content.

We next focused on the correlation between the nature of
the transposons in a region and its GC content. We calculated the density of each repeat
type as a function of the GC content in 50-kb windows (Fig. 22). As has been reported [142,
173-176], LINE sequences occur at much higher density in AT-rich regions (roughly fourfold
enriched), whereas SINEs (MIR, Alu) show the opposite trend (for Alu, up to fivefold
lower in AT-rich DNA). LTR retroposons and DNA transposons show a more uniform
distribution, dipping only in the most GC-rich regions.

The preference of LINEs for AT-rich DNA seems like a reasonable way for a genomic
parasite to accommodate its host, by targeting gene-poor AT-rich DNA and thereby
imposing a lower mutational burden. Mechanistically, selective targeting is nicely explained
by the fact that the preferred cleavage site of the LINE endonuclease is TTTT/A (where
the slash indicates the point of cleavage), which is used to prime reverse transcription from
the poly(A) tail of LINE RNA [177].

The contrary behaviour of SINEs, however, is baffling. How do SINEs accumulate in
GC-rich DNA, particularly if they depend on the LINE transposition machinery [178]?
Notably, the same pattern is seen for the Alu-like B1 and the tRNA-derived SINEs in
mouse and for MIR in human [142]. One possibility is that SINEs somehow target GC-rich
DNA for insertion. The alternative is that SINEs initially insert with the same proclivity for
AT-rich DNA as LINEs, but that the distribution is subsequently reshaped by evolutionary
forces [142, 179].

We used the draft genome sequence to investigate this mystery by comparing the
proclivities of young, adolescent, middle-aged and old Alus (Fig. 23). Strikingly, recent Alus
show a preference for AT-rich DNA resembling that of LINEs, whereas progressively
older Alus show a progressively stronger bias towards GC-rich DNA. These results
indicate that the GC bias must result from strong pressure: Fig. 23 shows that a 13-fold
enrichment of Alus in GC-rich DNA has occurred within the last 30 Myr, and possibly more
recently.

These results raise a new mystery. What is the force that produces the great and rapid
enrichment of Alus in GC-rich DNA? One explanation may be that deletions are more
readily tolerated in gene-poor AT-rich regions than in gene-rich GC-rich regions, resulting in
older elements being enriched in GC-rich regions. Such an enrichment is seen for
transposable elements such as DNA transposons (Fig. 24). However, this effect seems too
slow and too small to account for the observed remodelling of the Alu distribution. This can
be seen by performing a similar analysis for LINE elements (Fig. 25). There is no
significant change in the LINE distribution over the past 100 Myr, in contrast to the rapid
change seen for Alu. There is an eventual shift after more than 100 Myr, although its
magnitude is still smaller than seen for Alus.

These observations indicate that there may be some force acting particularly on Alus. This
could be a higher rate of random loss of Alus in AT-rich DNA, negative selection against
Alus in AT-rich DNA or positive selection in favour of Alus in GC-rich DNA. The first two
possibilities seem unlikely because AT-rich DNA is gene-poor and tolerates the
accumulation of other transposable elements. The third seems more feasible, in that it
involves selecting in favour of the minority of Alus in GC-rich regions rather than against
the majority that lie in AT-rich regions. But positive selection for Alus in GC-rich regions
would imply that they benefit the organism.

Schmid [180] has proposed such a function for SINEs. This hypothesis is based on the
observation that in many species SINEs are transcribed under conditions of stress, and the
resulting RNAs specifically bind a particular protein kinase (PKR) and block its ability to
inhibit protein translation [181-183]. SINE RNAs would thus promote protein translation under
stress. SINE RNA may be well suited to such a role in regulating protein translation,
because it can be quickly transcribed in large quantities from thousands of elements and it
can function without protein translation. Under this theory, there could be positive selection
for SINEs in readily transcribed open chromatin such as is found near genes. This could
explain the retention of Alus in gene-rich GC-rich regions. It is also consistent with the
observation that SINE density in AT-rich DNA is higher near genes [142].

Further insight about Alus comes from the relationship between Alu density and GC content
on individual chromosomes (Fig. 26). There are two outliers. Chromosome 19 is even richer
in Alus than predicted by its (high) GC content; the chromosome comprises 2% of the
genome, but contains 5% of Alus. On the other hand, chromosome Y shows the lowest
density of Alus relative to its GC content, being higher than average for GC content less
than 40% and lower than average for GC content over 40%. Even in AT-rich DNA, Alus
are under-represented on chromosome Y compared with other young interspersed repeats
(see below). These phenomena may be related to an unusually high gene density on
chromosome 19 and an unusually low density of somatically active genes on chromosome Y
(both relative to GC content). This would be consistent with the idea that Alu correlates not
with GC content but with actively transcribed genes.

Our results may support the controversial idea that SINEs actually earn their keep in the
genome. Clearly, much additional work will be needed to prove or disprove the hypothesis
that SINEs are genomic symbionts.

Biases in human mutation.

Indirect studies have suggested that nucleotide substitution is not uniform across mammalian genomes [184-187]. By studying sets of repeat elements
belonging to a common cohort, one can directly measure nucleotide substitution rates in
different regions of the genome. We find strong evidence that the pattern of neutral
substitution differs as a function of local GC content (Fig. 27). Because the results are
observed in repetitive elements throughout the genome, the variation in the pattern of
nucleotide substitution seems likely to be due to differences in the underlying mutational
process rather than to selection.

The effect can be seen most clearly by focusing on the substitution process , where
   denotes GC or CG base pairs and  denotes AT or TA base pairs. If K is the equilibrium
constant in the direction of  base pairs (defined by the ratio of the forward and reverse
rates), then the equilibrium GC content should be 1/(1 + K). Two observations emerge.

First, there is a regional bias in substitution patterns. The equilibrium constant varies as a
function of local GC content:  base pairs are more likely to mutate towards  base pairs in
AT-rich regions than in GC-rich regions. For the analysis in Fig. 27, the equilibrium constant
K is 2.5, 1.9 and 1.2 when the draft genome sequence is partitioned into three bins with
average GC content of 37, 43 and 50%, respectively. This bias could be due to a reported
tendency for GC-rich regions to replicate earlier in the cell cycle than AT-rich regions and
for guanine pools, which are limiting for DNA replication, to become depleted late in the cell
cycle, thereby resulting in a small but significant shift in substitution towards  base pairs [186,
188]. Another theory proposes that many substitutions are due to differences in DNA repair
mechanisms, possibly related to transcriptional activity and thereby to gene density and GC
content [185, 189, 190].

There is also an absolute bias in substitution patterns resulting in directional pressure
towards lower GC content throughout the human genome. The genome is not at equilibrium
with respect to the pattern of nucleotide substitution: the expected equilibrium GC content
corresponding to the values of K above is 29, 35 and 44% for regions with average GC
contents of 37, 43 and 50%, respectively. Recent observations on SNPs [190] confirm that the
mutation pattern in GC-rich DNA is biased towards  base pairs; it should be possible to
perform similar analyses throughout the genome with the availability of 1.4 million SNPs [97,
191]. On the basis solely of nucleotide substitution patterns, the GC content would be
expected to be about 7% lower throughout the genome.

What accounts for the higher GC content? One possible explanation is that in GC-rich
regions, a considerable fraction of the nucleotides is likely to be under functional constraint
owing to the high gene density. Selection on coding regions and regulatory CpG islands may
maintain the higher-than-predicted GC content. Another is that throughout the rest of the
genome, a constant influx of transposable elements tends to increase GC content (Fig. 28).
Young repeat elements clearly have a higher GC content than their surrounding regions,
except in extremely GC-rich regions. Moreover, repeat elements clearly shift with age
towards a lower GC content, closer to that of the neighbourhood in which they reside. Much
of the 'non-repeat' DNA in AT-rich regions probably consists of ancient repeats that are not
detectable by current methods and that have had more time to approach the local
equilibrium value.

The repeats can also be used to study how the mutation process is affected by the
immediately adjacent nucleotide. Such 'context effects' will be discussed elsewhere (A. Kas
and A. F. A. Smit, unpublished results).
 

Fast living on chromosome Y.

The pattern of interspersed repeats can be used to shed
light on the unusual evolutionary history of chromosome Y. Our analysis shows that the
genetic material on chromosome Y is unusually young, probably owing to a high tolerance
for gain of new material by insertion and loss of old material by deletion. Several lines of
evidence support this picture. For example, LINE elements on chromosome Y are on
average much younger than those on autosomes (not shown). Similarly, MaLR-family
retroposons on chromosome Y are younger than those on autosomes, with the
representation of subfamilies showing a strong inverse correlation with the age of the
subfamily. Moreover, chromosome Y has a relative over-representation of the younger
retroviral class II (ERVK) and a relative under-representation of the primarily older class
III (ERVL) compared with other chromosomes. Overall, chromosome Y seems to maintain
a youthful appearance by rapid turnover.

Interspersed repeats on chromosome Y can also be used to estimate the relative mutation
rates, m and f, in the male and female germlines. Chromosome Y always resides in males,
whereas chromosome X resides in females twice as often as in males. The substitution
rates, Y and X, on these two chromosomes should thus be in the ratio Y:X =
(m):(m + 2f)/3, provided that one considers equivalent neutral sequences. Several authors
have estimated the mutation rate in the male germline to be fivefold higher than in the
female germline, by comparing the rates of evolution of X- and Y-linked genes in humans
and primates. However, Page and colleagues [192] have challenged these estimates as too
high. They studied a 39-kb region that is apparently devoid of genes and resides within a
large segmental duplication from X to Y that occurred 3–4 Myr ago in the human lineage.
On the basis of phylogenetic analysis of the sequence on human Y and human, chimp and
gorilla X, they obtained a much lower estimate of Y:X = 1.36, corresponding to m:f =
1.7. They suggested that the other estimates may have been higher because they were
based on much longer evolutionary periods or because the genes studied may have been
under selection.

Our database of human repeats provides a powerful resource for addressing this question.
We identified the repeat elements from recent subfamilies (effectively, birth cohorts dating
from the past 50 Myr) and measured the substitution rates for subfamily members on
chromosomes X and Y (Fig. 29). There is a clear linear relationship with a slope of Y:X
= 1.57 corresponding to m:f = 2.1. The estimate is in reasonable agreement with that of
Page et al., although it is based on much more total sequence (360 kb on Y, 1.6 Mb on X)
and a much longer time period. In particular, the discrepancy with earlier reports is not
explained by recent changes in the human lineage. Various theories have been proposed for
the higher mutation rate in the male germline, including the greater number of cell divisions
in the formation of sperm than eggs and different repair mechanisms in sperm and eggs.

Active transposons.

We were interested in identifying the youngest retrotransposons in the draft genome sequence. This set should contain the currently active retrotransposons, as well as the insertion sites that are still polymorphic in the human population.

The youngest branch in the phylogenetic tree of human LINE1 elements is called L1Hs
(ref. 158); it differs in its 3' untranslated region (UTR) by 12 diagnostic substitutions from
the next oldest subfamily (L1PA2). Within the L1Hs family, there are two subsets referred
to as Ta and pre-Ta, defined by a diagnostic trinucleotide [193, 194]. All active L1 elements are
thought to belong to these two subsets, because they account for all 14 known cases of
human disease arising from new L1 transposition (with 13 belonging to the Ta subset and
one to the pre-Ta subset) [195, 196]. These subsets are also of great interest for population
genetics because at least 50% are still segregating as polymorphisms in the human
population [194, 197]; they provide powerful markers for tracing population history because
they represent unique (non-recurrent and non-revertible) genetic events that can be used
(along with similarly polymorphic Alus) for reconstructing human migrations.

LINE1 elements that are retrotransposition-competent should consist of a full-length
sequence and should have both ORFs intact. Eleven such elements from the Ta subset have
been identified, including the likely progenitors of mutagenic insertions into the factor VIII
and dystrophin genes [198-202]. A cultured cell retrotransposition assay has revealed that eight
of these elements remain retrotransposition-competent [200, 202, 203].

We searched the draft genome sequence and identified 535 LINEs belonging to the Ta
subset and 415 belonging to the pre-Ta subset. These elements provide a large collection of
tools for probing human population history. We also identified those consisting of full-length
elements with intact ORFs, which are candidate active LINEs. We found 39 such elements
belonging to the Ta subset and 22 belonging to the pre-Ta subset; this substantially increases
the number in the first category and provides the first known examples in the second
category. These elements can now be tested for retrotransposition competence in the cell
culture assay. Preliminary analysis resulted in the identification of two of these elements as
the likely progenitors of mutagenic insertions into the -globin and RP2 genes (R. Badge
and J. V. Moran, unpublished data). Similar analyses should allow the identification of the
progenitors of most, if not all, other known mutagenic L1 insertions.

L1 elements can carry extra DNA if transcription extends through the native transcriptional
termination site into flanking genomic DNA. This process, termed L1-mediated
transduction, provides a means for the mobilization of DNA sequences around the genome
and may be a mechanism for 'exon shuffling' [204]. Twenty-one per cent of the 71 full-length
L1s analysed contained non-L1-derived sequences before the 3' target-site duplication site,
in cases in which the site was unambiguously recognizable. The length of the transduced
sequence was 30–970 bp, supporting the suggestion that 0.5–1.0% of the human genome
may have arisen by LINE-based transduction of 3' flanking sequences [205, 206].

Our analysis also turned up two instances of 5' transduction (145 bp and 215 bp). Although
this possibility had been suggested on the basis of cell culture models [195, 203], these are the
first documented examples. Such events may arise from transcription initiating in a cellular
promoter upstream of the L1 elements. L1 transcription is generally confined to the
germline [207, 208], but transcription from other promoters could explain a somatic L1
retrotransposition event that resulted in colon cancer [206].

Transposons as a creative force.

The primary force for the origin and expansion of most transposons has been selection for their ability to create progeny, and not a selective advantage for the host. However, these selfish pieces of DNA have been responsible for important innovations in many genomes, for example by contributing regulatory elements
and even new genes.

Twenty human genes have been recognized as probably derived from transposons [142, 209].
These include the RAG1 and RAG2 recombinases and the major centromere-binding
protein CENPB. We scanned the draft genome sequence and identified another 27 cases,
bringing the total to 47 (Table 13; refs 142, 209). All but four are derived from DNA
transposons, which give rise to only a small proportion of the interspersed repeats in the
genome. Why there are so many DNA transposase-like genes, many of which still contain
the critical residues for transposase activity, is a mystery.

To illustrate this concept, we describe the discovery of one of the new examples. We
searched the draft genome sequence to identify the autonomous DNA transposon
responsible for the distribution of the non-autonomous MER85 element, one of the most
recently (40–50 Myr ago) active DNA transposons. Most non-autonomous elements are
internal deletion products of a DNA transposon. We identified one instance of a large
(1,782 bp) ORF flanked by the 5' and 3' halves of a MER85 element. The ORF encodes a
novel protein (partially published as pID 6453533) whose closest homologue is the
transposase of the piggyBac DNA transposon, which is found in insects and has the same
characteristic TTAA target-site duplications [210] as MER85. The ORF is actively transcribed
in fetal brain and in cancer cells. That it has not been lost to mutation in 40–50 Myr of
evolution (whereas the flanking, noncoding, MER85-like termini show the typical divergence
level of such elements) and is actively transcribed provides strong evidence that it has been
adopted by the human genome as a gene. Its function is unknown.

LINE1 activity clearly has also had fringe benefits. We mentioned above the possibility of
exon reshuffling by cotranscription of neighbouring DNA. The LINE1 machinery can also
cause reverse transcription of genic mRNAs, which typically results in nonfunctional
processed pseudogenes but can, occasionally, give rise to functional processed genes. There
are at least eight human and eight mouse genes for which evidence strongly supports such
an origin [211] (see http://www-ifi.uni-muenster.de/exapted-retrogenes/tables.html). Many
other intronless genes may have been created in the same way.

Transposons have made other creative contributions to the genome. A few hundred genes,
for example, use transcriptional terminators donated by LTR retroposons (data not shown).
Other genes employ regulatory elements derived from repeat elements [211].

Simple sequence repeats

Simple sequence repeats (SSRs) are a rather different type of repetitive structure that is
common in the human genome—perfect or slightly imperfect tandem repeats of a particular
k-mer. SSRs with a short repeat unit (n = 1–13 bases) are often termed microsatellites,
whereas those with longer repeat units (n = 14–500 bases) are often termed minisatellites.
With the exception of poly(A) tails from reverse transcribed messages, SSRs are thought to
arise by slippage during DNA replication212, 213.

We compiled a catalogue of all SSRs over a given length in the human draft genome
sequence, and studied their properties (Table 14). SSRs comprise about 3% of the human
genome, with the greatest single contribution coming from dinucleotide repeats (0.5%). (The
precise criteria for the number of repeat units and the extent of divergence allowed in an
SSR affect the exact census, but not the qualitative conclusions.)

There is approximately one SSR per 2 kb (the number of nonoverlapping tandem repeats is
437 per Mb). The catalogue confirms various properties of SSRs that have been inferred
from sampling approaches (Table 15). The most frequent dinucleotide repeats are AC and
AT (50 and 35% of dinucleotide repeats, respectively), whereas AG repeats (15%) are less
frequent and GC repeats (0.1%) are greatly under-represented. The most frequent
trinucleotides are AAT and AAC (33% and 21%, respectively), whereas ACC (4.0%),
AGC (2.2%), ACT (1.4%) and ACG (0.1%) are relatively rare. Overall, trinucleotide SSRs
are much less frequent than dinucleotide SSRs214.

SSRs have been extremely important in human genetic studies, because they show a high
degree of length polymorphism in the human population owing to frequent slippage by DNA
polymerase during replication. Genetic markers based on SSRs—particularly (CA)n
repeats—have been the workhorse of most human disease-mapping studies101, 102. The
availability of a comprehensive catalogue of SSRs is thus a boon for human genetic studies.

The SSR catalogue also allowed us to resolve a mystery regarding mammalian genetic
maps. Such genetic maps in rat, mouse and human have a deficit of polymorphic (CA)n
repeats on chromosome X30, 101. There are two possible explanations for this deficit. There
may simply be fewer (CA)n repeats on chromosome X; or (CA)n repeats may be as dense
on chromosome X but less polymorphic in the population. In fact, analysis of the draft
genome sequence shows that chromosome X has the same density of (CA)n repeats per
Mb as the autosomes (data not shown). Thus, the deficit of polymorphic markers relative
to autosomes results from population genetic forces. Possible explanations include that
chromosome X has a smaller effective population size, experiences more frequent selective
sweeps reducing diversity (owing to its hemizygosity in males), or has a lower mutation rate
(owing to its more frequent passage through the less mutagenic female germline). The
availability of the draft genome sequence should provide ways to test these alternative
explanations.

Segmental duplications

A remarkable feature of the human genome is the segmental duplication of portions of
genomic sequence [215-217]. Such duplications involve the transfer of 1–200-kb blocks of
genomic sequence to one or more locations in the genome. The locations of both donor and
recipient regions of the genome are often not tandemly arranged, suggesting mechanisms
other than unequal crossing-over for their origin. They are relatively recent, inasmuch as
strong sequence identity is seen in both exons and introns (in contrast to regions that are
considered to show evidence of ancient duplications, characterized by similarities only in
coding regions). Indeed, many such duplications appear to have arisen in very recent
evolutionary time, as judged by high sequence identity and by their absence in closely
related species.

Segmental duplications can be divided into two categories. First, interchromosomal
duplications are defined as segments that are duplicated among nonhomologous
chromosomes. For example, a 9.5-kb genomic segment of the adrenoleukodystrophy locus
from Xq28 has been duplicated to regions near the centromeres of chromosomes 2, 10, 16
and 22 (refs 218, 219). Anecdotal observations suggest that many interchromosomal
duplications map near the centromeric and telomeric regions of human
chromosomes218-233.

The second category is intrachromosomal duplications, which occur within a particular
chromosome or chromosomal arm. This category includes several duplicated segments, also
known as low copy repeat sequences, that mediate recurrent chromosomal structural
rearrangements associated with genetic disease215, 217. Examples on chromosome 17
include three copies of a roughly 200-kb repeat separated by around 5 Mb and two
copies of a roughly 24-kb repeat separated by 1.5 Mb. The copies are so similar (99%
identity) that paralogous recombination events can occur, giving rise to contiguous gene
syndromes: Smith–Magenis syndrome and Charcot–Marie–Tooth syndrome 1A,
respectively34, 234. Several other examples are known and are also suspected to be
responsible for recurrent microdeletion syndromes (for example, Prader–Willi/Angelman,
velocardiofacial/DiGeorge and Williams' syndromes [215, 235-240]).

Until now, the identification and characterization of segmental duplications have been based
on anecdotal reports—for example, finding that certain probes hybridize to multiple
chromosomal sites or noticing duplicated sequence at certain recurrent chromosomal
breakpoints. The availability of the entire genomic sequence will make it possible to explore
the nature of segmental duplications more systematically. This analysis can begin with the
current state of the draft genome sequence, although caution is required because some
apparent duplications may arise from a failure to merge sequence contigs from overlapping
clones. Alternatively, erroneous assembly of closely related sequences from nonoverlapping
clones may underestimate the true frequency of such features, particularly among those
segments with the highest sequence similarity. Accordingly, we adopted a conservative
approach for estimating such duplication from the available draft genome sequence.

Pericentromeres and subtelomeres.

We began by re-evaluating the finished sequences
of chromosomes 21 and 22. The initial papers on these chromosomes93, 94 noted some
instances of interchromosomal duplication near each centromere. With the ability now to
compare these chromosomes to the vast majority of the genome, it is apparent that the
regions near the centromeres consist almost entirely of interchromosomal duplicated
segments, with little or no unique sequence. Smaller regions of interchromosomal duplication
are also observed near the telomeres.

Chromosome 22 contains a region of 1.5 Mb adjacent to the centromere in which 90% of
sequence can now be recognized to consist of interchromosomal duplication (Fig. 30).
Conversely, 52% of the interchromosomal duplications on chromosome 22 were located in
this region, which comprises only 5% of the chromosome. Also, the subtelomeric end
consists of a 50-kb region consisting almost entirely of interchromosomal duplications.

Chromosome 21 presents a similar landscape (Fig. 31). The first 1 Mb after the centromere
is composed of interchromosomal repeats, as well as the largest (> 200 kb) block of
intrachromosomally duplicated material. Again, most interchromosomal duplications on the
chromosome map to this region and the most subtelomeric region (30 kb) shows extensive
duplication among nonhomologous chromosomes.

The pericentromeric regions are structurally very complex, as illustrated for chromosome 21
in Fig. 32a. The pericentromeric regions appear to have been bombarded by successive
insertions of duplications; the insertion events must be fairly recent because the degree of
sequence conservation with the genomic source loci is fairly high (90–100%, with an
apparent peak around 96%). Distinct insertions are typically separated by AT-rich or
GC-rich minisatellite-like repeats that have been hypothesized to have a functional role in
targeting duplications to these regions [233, 241].

A single genomic source locus often gives rise to pericentromeric copies on multiple
chromosomes, with each having essentially the same breakpoints and the same degree of
divergence. An example of such a source locus on Xq28 is shown in Fig. 32b. Phylogenetic
analysis has suggested a two-step mechanism for the origin and dispersal of these
segments, whereby an initial segmental duplication in the pericentromeric region of one
chromosome occurs and is then redistributed as part of a larger cassette to other such
regions [242].

A comprehensive analysis for all chromosomes will have to await complete sequencing of
the genome, but the evidence from the draft genome sequence indicates that the same
picture is likely to be seen throughout the genome. Several papers have analysed finished
segments within pericentromeric regions of chromosomes 2 (160 kb), 10 (400 kb) and 16
(300 kb), all of which show extensive interchromosomal segmental duplication215, 219, 232,
233. An example from another pericentromeric region on chromosome 11 is shown in Fig.
32c. Interchromosomal duplications in subtelomeric regions also appear to be a fairly
general phenomenon, as illustrated by a large tract (500 kb) of complex duplication on
chromosome 7 (Fig. 32d).

The explanation for the clustering of segmental duplications may be that the genome has a
damage-control mechanism whereby chromosomal breakage products are preferentially
inserted into pericentromeric and, to a lesser extent, subtelomeric regions. The possibility of
a specific mechanism for the insertion of these sequences has been suggested on the basis
of the unusual sequences found flanking the insertions. Although it is also possible that these
regions simply have greater tolerance for large insertions, many large gene-poor 'deserts'
have been identified93 and there is no accumulation of duplicated segments within these
regions. Along with the fact that transitions between duplicons (from different regions of the
genome) occur at specific sequences, this suggests that active recruitment of duplications to
such regions may occur. In any case, the duplicated regions are in general young (with
many duplications showing <6% nucleotide divergence from their source loci) and in
constant flux, both through additional duplications and by large-scale exchange among
similar chromosomal environments. There is evidence of structural polymorphism in the
human population, such as the presence or absence of olfactory receptor segments located
within the telomeric regions of several human chromosomes226, 227.

Genome-wide analysis of segmental duplications.

We also performed a global
genome-wide analysis to characterize the amount of segmental duplication in the genome.
We 'repeat-masked' the known interspersed repeats in the draft genome sequence and
compared the remaining draft genomic sequence with itself in a massive all-by-all BLASTN
similarity search. We excluded matches in which the sequence identity was so high that it
might reflect artefactual duplications resulting from a failure to overlap sequence contigs
correctly in assembling the draft genome sequence. Specifically, we considered only
matches with less than 99.5% identity for finished sequence and less than 98% identity for
unfinished sequence.

We took several approaches to avoid counting artefactual duplications in the sequence. In
the first approach, we studied only finished sequence. We compared the finished sequence
with itself, to identify segments of at least 1 kb and 90–99.5% sequence identity. This
analysis will underestimate the extent of segmental duplication, because it requires that at
least two copies of the segment are present in the finished sequence and because some true
duplications have over 99.5% identity.

The finished sequence consists of at least 3.3% segmental duplication (Table 16).
Interchromosomal duplication accounts for about 1.5% and intrachromosomal duplication
for about 2%, with some overlap (0.2%) between these categories. We analysed the
lengths and divergence of the segmental duplications (Fig. 33). The duplications tend to be
large (10–50 kb) and highly homologous, especially for the interchromosomal segments. The
sequence divergence for the interchromosomal duplications appears to peak between 96.5%
and 97.5%. This may indicate that interchromosomal duplications occurred in a punctuated
manner. It will be intriguing to investigate whether such genomic upheaval has a role in
speciation events.

In a second approach, we compared the entire human draft genome sequence (finished and
unfinished) with itself to identify duplications with 90–98% sequence identity (Table 17).
The draft genome sequence contains at least 3.6% segmental duplication. The actual
proportion will be significantly higher, because we excluded many true matches with more
than 98% sequence identity (at least 1.1% of the finished sequence). Although exact
measurement must await a finished sequence, the human genome seems likely to contain
about 5% segmental duplication, with most of this sequence in large blocks (> 10 kb). Such
a high proportion of large duplications clearly distinguishes the human genome from other
sequenced genomes, such as the fly and worm (Table 18).

The structure of large highly paralogous regions presents one of the 'serious and
unanticipated challenges' to producing a finished sequence of the genome [46]. The absence of
unique STS or fingerprint signatures over large genomic distances (1 Mb) and the high
degree of sequence similarity makes the distinction between paralogous sequence variation
and allelic polymorphism problematic. Furthermore, the fact that such regions frequently
harbour intron–exon structures of genuine unique sequence will complicate efforts to
generate a genome-wide SNP map. The data indicate that a modest portion of the human
genome may be relatively recalcitrant to genomic-based methods for SNP detection. Owing
to their repetitive nature and their location in the genome, segmental duplications may well
be underestimated by the current analysis. An understanding of the biology, pathology and
evolution of these duplications will require specialized efforts within these exceptional
regions of the human genome. The presence and distribution of such segments may provide
evolutionary fodder for processes of exon shuffling and a general increase in protein
diversity associated with domain accretion. It will be important to consider both
genome-wide duplication events and more restricted punctuated events of genome
duplication as forces in the evolution of vertebrate genomes.

...

Gene content of the human genome

Genes (or at least their coding regions) comprise only a tiny fraction of human DNA, but
they represent the major biological function of the genome and the main focus of interest by
biologists. They are also the most challenging feature to identify in the human genome
sequence.

The ultimate goal is to compile a complete list of all human genes and their encoded
proteins, to serve as a 'periodic table' for biomedical research [243]. But this is a difficult task.
In organisms with small genomes, it is straightforward to identify most genes by the
presence of long ORFs. In contrast, human genes tend to have small exons (encoding an
average of only 50 codons) separated by long introns (some exceeding 10 kb). This creates
a signal-to-noise problem, with the result that computer programs for direct gene prediction
have only limited accuracy. Instead, computational prediction of human genes must rely
largely on the availability of cDNA sequences or on sequence conservation with genes and
proteins from other organisms. This approach is adequate for strongly conserved genes
(such as histones or ubiquitin), but may be less sensitive to rapidly evolving genes (including
many crucial to speciation, sex determination and fertilization).

Here we describe our efforts to recognize both the RNA genes and protein-coding genes in
the human genome. We also study the properties of the predicted human protein set,
attempting to discern how the human proteome differs from those of invertebrates such as
worm and fly.

Noncoding RNAs

Although biologists often speak of a tight coupling between 'genes and their encoded protein
products', it is important to remember that thousands of human genes produce noncoding
RNAs (ncRNAs) as their ultimate product [244]. There are several major classes of ncRNA.
(1) Transfer RNAs (tRNAs) are the adapters that translate the triplet nucleic acid code of
RNA into the amino-acid sequence of proteins; (2) ribosomal RNAs (rRNAs) are also
central to the translational machinery, and recent X-ray crystallography results strongly
indicate that peptide bond formation is catalysed by rRNA, not protein [245, 246]; (3) small
nucleolar RNAs (snoRNAs) are required for rRNA processing and base modification in the
nucleolus [247, 248]; and (4) small nuclear RNAs (snRNAs) are critical components of
spliceosomes, the large ribonucleoprotein (RNP) complexes that splice introns out of
pre-mRNAs in the nucleus. Humans have both a major, U2 snRNA-dependent spliceosome
that splices most introns, and a minor, U12 snRNA-dependent spliceosome that splices a
rare class of introns that often have AT/AC dinucleotides at the splice sites instead of the
canonical GT/AG splice site consensus [249].

Other ncRNAs include both RNAs of known biochemical function (such as telomerase
RNA and the 7SL signal recognition particle RNA) and ncRNAs of enigmatic function
(such as the large Xist transcript implicated in X dosage compensation [250], or the small vault
RNAs found in the bizarre vault ribonucleoprotein complex [251], which is three times the
mass of the ribosome but has unknown function).

ncRNAs do not have translated ORFs, are often small and are not polyadenylated.
Accordingly, novel ncRNAs cannot readily be found by computational gene-finding
techniques (which search for features such as ORFs) or experimental sequencing of cDNA
or EST libraries (most of which are prepared by reverse transcription using a primer
complementary to a poly(A) tail). Even if the complete finished sequence of the human
genome were available, discovering novel ncRNAs would still be challenging. We can,
however, identify genomic sequences that are homologous to known ncRNA genes, using
BLASTN or, in some cases, more specialized methods.

It is sometimes difficult to tell whether such homologous genes are orthologues, paralogues
or closely related pseudogenes (because inactivating mutations are much less obvious than
for protein-coding genes). For tRNA, there is sufficiently detailed information about the
cloverleaf secondary structure to allow true genes and pseudogenes to be distinguished with
high sensitivity. For many other ncRNAs, there is much less structural information and so
we employ an operational criterion of high sequence similarity (> 95% sequence identity and
> 95% full length) to distinguish true genes from pseudogenes. These assignments will
eventually need to be reconciled with experimental data.

Transfer RNA genes.

The classical experimental estimate of the number of human tRNA
genes is 1,310 (ref. 252). In the draft genome sequence, we find only 497 human tRNA
genes (Tables 19, 20). How do we account for this discrepancy? We believe that the
original estimate is likely to have been inflated in two respects. First, it came from a
hybridization experiment that probably counted closely related pseudogenes; by analysis of
the draft genome sequence, there are in fact 324 tRNA-derived putative pseudogenes
(Table 20). Second, the earlier estimate assumed too high a value for the size of the human
genome; repeating the calculation using the correct value yields an estimate of about 890
tRNA-related loci, which is in reasonable accord with our count of 821 tRNA genes and
pseudogenes in the draft genome sequence.

The human tRNA gene set predicted from the draft genome sequence appears to include
most of the known human tRNA species. The draft genome sequence contains 37 of 38
human tRNA species listed in a tRNA database [253], allowing for up to one mismatch. This
includes one copy of the known gene for a specialized selenocysteine tRNA, one of several
components of a baroque translational mechanism that reads UGA as a selenocysteine
codon in certain rare mRNAs that carry a specific cis-acting RNA regulatory site (a
so-called SECIS element) in their 3' UTRs. The one tRNA gene in the database not found
in the draft genome sequence is DE9990, a tRNAGlu species, which differs in two positions
from the most related tRNA gene in the human genome. Possible explanations are that the
database version of this tRNA contains two errors, the gene is polymorphic or this is a
genuine functional tRNA that is missing from the draft genome sequence. (The database
also lists one additional tRNA gene (DS9994), but this is apparently a contaminant, most
similar to bacterial tRNAs; the parent entry (Z13399) was withdrawn from the DNA
database, but the tRNA entry has not yet been removed from the tRNA database.)
Although the human set appears substantially complete by this test, the tRNA gene numbers
in Table 19 should be considered tentative and used with caution. The human and fly (but
not the worm) are known to be missing significant amounts of heterochromatic DNA, and
additional tRNA genes could be located there.

With this caveat, the results indicate that the human has fewer tRNA genes than the worm,
but more than the fly. This may seem surprising, but tRNA gene number in metazoans is
thought to be related not to organismal complexity, but more to idiosyncrasies of the demand
for tRNA abundance in certain tissues or stages of embryonic development. For example,
the frog Xenopus laevis, which must load each oocyte with a remarkable 40 ng of tRNA,
has thousands of tRNA genes [254].

The degeneracy of the genetic code has allowed an inspired economy of tRNA anticodon
usage. Although 61 sense codons need to be decoded, not all 61 different anticodons are
present in tRNAs. Rather, tRNAs generally follow stereotyped and conserved wobble
rules [255-257]. Wobble reduces the number of required anticodons substantially, and provides
a connection between the genetic code and the hybridization stability of modified and
unmodified RNA bases. In eukaryotes, the rules proposed by Guthrie and Abelson [256]
predict that about 46 tRNA species will be sufficient to read the 61 sense codons (counting
the initiator and elongator methionine tRNAs as two species). According to these rules, in
the codon's third (wobble) position, U and C are generally decoded by a single tRNA
species, whereas A and G are decoded by two separate tRNA species.

In 'two-codon boxes' of the genetic code (where codons ending with U/C encode a
different amino acid from those ending with A/G), the U/C wobble position should be
decoded by a G at position 34 in the tRNA anticodon. Thus, in the top left of Fig. 34, there
is no tRNA with an AAA anticodon for Phe, but the GAA anticodon can recognize both
UUU and UUC codons in the mRNA. In 'four-codon boxes' of the genetic code (where U,
C, A and G in the wobble position all encode the same amino acid), the U/C wobble position
is almost always decoded by I34 (inosine) in the tRNA, where the inosine is produced by
post-transcriptional modification of an adenine (A). In the bottom left of Fig. 34, for
example, the GUU and GUC codons of the four-codon Val box are decoded by a tRNA
with an anticodon of AAC, which is no doubt modified to IAC. Presumably this pattern,
which is strikingly conserved in eukaryotes, has to do with the fact that IA base pairs are
also possible; thus the IAC anticodon for a Val tRNA could recognize GUU, GUC and
even GUA codons. Were this same I34 to be utilized in two-codon boxes, however,
misreading of the NNA codon would occur, resulting in translational havoc. Eukaryotic
glycine tRNAs represent a conserved exception to this last rule; they use a GCC anticodon
to decode GGU and GGC, rather than the expected ICC anticodon.

Satisfyingly, the human tRNA set follows these wobble rules almost perfectly (Fig. 34).
Only three unexpected tRNA species are found: single genes for a tRNATyr-AUA,
tRNAIle-GAU, and tRNAAsn-AUU. Perhaps these are pseudogenes, but they appear to
be plausible tRNAs. We also checked the possibility of sequencing errors in their
anticodons, but each of these three genes is in a region of high sequence accuracy, with
PHRAP quality scores higher than 70 for every base in their anticodons.

As in all other organisms, human protein-coding genes show codon bias—preferential use of
one synonymous codon over another [258] (Fig. 34). In less complex organisms, such as yeast
or bacteria, highly expressed genes show the strongest codon bias. Cytoplasmic abundance
of tRNA species is correlated with both codon bias and overall amino-acid frequency (for
example, tRNAs for preferred codons and for more common amino acids are more
abundant). This is presumably driven by selective pressure for efficient or accurate
translation [259]. In many organisms, tRNA abundance in turn appears to be roughly
correlated with tRNA gene copy number, so tRNA gene copy number has been used as a
proxy for tRNA abundance [260]. In vertebrates, however, codon bias is not so obviously
correlated with gene expression level. Differing codon biases between human genes is more
a function of their location in regions of different GC composition [261]. In agreement with the
literature, we see only a very rough correlation of human tRNA gene number with either
amino-acid frequency or codon bias (Fig. 34). The most obvious outliers in these weak
correlations are the strongly preferred CUG leucine codon, with a mere six tRNALeu-CAG
genes producing a tRNA to decode it, and the relatively rare cysteine UGU and UGC
codons, with 30 tRNA genes to decode them.

The tRNA genes are dispersed throughout the human genome. However, this dispersal is
nonrandom. tRNA genes have sometimes been seen in clusters at small scales [262, 263] but
we can now see striking clustering on a genome-wide scale. More than 25% of the tRNA
genes (140) are found in a region of only about 4 Mb on chromosome 6. This small region,
only about 0.1% of the genome, contains an almost sufficient set of tRNA genes all by
itself. The 140 tRNA genes contain a representative for 36 of the 49 anticodons found in
the complete set; and of the 21 isoacceptor types, only tRNAs to decode Asn, Cys, Glu and
selenocysteine are missing. Many of these tRNA genes, meanwhile, are clustered
elsewhere; 18 of the 30 Cys tRNAs are found in a 0.5-Mb stretch of chromosome 7 and
many of the Asn and Glu tRNA genes are loosely clustered on chromosome 1. More than
half of the tRNA genes (280 out of 497) reside on either chromosome 1 or chromosome 6.
Chromosomes 3, 4, 8, 9, 10, 12, 18, 20, 21 and X appear to have fewer than 10 tRNA genes
each; and chromosomes 22 and Y have none at all (each has a single pseudogene).

Ribosomal RNA genes.

The ribosome, the protein synthetic machine of the cell, is made
up of two subunits and contains four rRNA species and many proteins. The large ribosomal
subunit contains 28S and 5.8S rRNAs (collectively called 'large subunit' (LSU) rRNA) and
also a 5S rRNA. The small ribosomal subunit contains 18S rRNA ('small subunit' (SSU)
rRNA). The genes for LSU and SSU rRNA occur in the human genome as a 44-kb tandem
repeat unit [264]. There are thought to be about 150–200 copies of this repeat unit arrayed on
the short arms of acrocentric chromosomes 13, 14, 15, 21 and 22 (refs 254, 264). There are
no true complete copies of the rDNA tandem repeats in the draft genome sequence, owing
to the deliberate bias in the initial phase of the sequencing effort against sequencing BAC
clones whose restriction fragment fingerprints showed them to contain primarily tandemly
repeated sequence. Sequence similarity analysis with the BLASTN computer program
does, however, detect hundreds of rDNA-derived sequence fragments dispersed throughout
the complete genome, including one 'full-length' copy of an individual 5.8S rRNA gene not
associated with a true tandem repeat unit (Table 20).

The 5S rDNA genes also occur in tandem arrays, the largest of which is on chromosome 1
between 1q41.11 and 1q42.13, close to the telomere [265, 266]. There are 200–300 true 5S
genes in these arrays265, 267. The number of 5S-related sequences in the genome, including
numerous dispersed pseudogenes, is classically cited as 2,000 (refs 252, 254). The long
tandem array on chromosome 1 is not yet present in the draft genome sequence because
there are no EcoRI or HindIII sites present, and thus it was not cloned in the most heavily
utilized BAC libraries (Table 1). We expect to recover it during the finishing stage. We do
detect four individual copies of 5S rDNA by our search criteria ( 95% identity and  95%
full length). We also find many more distantly related dispersed sequences (520 at
P  0.001), which we interpret as probable pseudogenes (Table 20).

Small nucleolar RNA genes.

Eukaryotic rRNA is extensively processed and modified in
the nucleolus. Much of this activity is directed by numerous snoRNAs. These come in two
families: C/D box snoRNAs (mostly involved in guiding site-specific 2'-O-ribose
methylations of other RNAs) and H/ACA snoRNAs (mostly involved in guiding
site-specific pseudouridylations) [247, 248]. We compiled a set of 97 known human snoRNA
gene sequences; 84 of these (87%) have at least one copy in the draft genome sequence
(Table 20), almost all as single-copy genes.

It is thought that all 2'-O-ribose methylations and pseudouridylations in eukaryotic rRNA are
guided by snoRNAs. There are 105–107 methylations and around 95 pseudouridylations in
human rRNA [268]. Only about half of these have been tentatively assigned to known guide
snoRNAs. There are also snoRNA-directed modifications on other stable RNAs, such as
U6 (ref. 269), and the extent of this is just beginning to be explored. Sequence similarity has
so far proven insufficient to recognize all snoRNA genes. We therefore expect that there
are many unrecognized snoRNA genes that are not detected by BLAST queries.

Spliceosomal RNAs and other ncRNA genes.

We also looked for copies of other known ncRNA genes. We found at least one copy of 21 (95%) of 22 known ncRNAs, including the spliceosomal snRNAs. There were multiple copies for several ncRNAs, as expected; for example, we find 44 dispersed genes for U6 snRNA, and 16 for U1 snRNA (Table 20).

For some of these RNA genes, homogeneous multigene families that occur in tandem
arrays are again under-represented owing to the restriction enzymes used in constructing
the BAC libraries and, in some instances, the decision to delay the sequencing of BAC
clones with low complexity fingerprints indicative of tandemly repeated DNA. The U2
RNA genes are located at the RNU2 locus, a tandem array of 10–20 copies of nearly
identical 6.1-kb units at 17q21–q22 (refs 270, 271, 272). Similarly, the U3 snoRNA genes
(included in the aggregate count of C/D snoRNAs in Table 20) are clustered at the RNU3
locus at 17p11.2, not in a tandem array, but in a complex inverted repeat structure of about
5–10 copies per haploid genome [273]. The U1 RNA genes are clustered with about 30 copies
at the RNU1 locus at 1p36.1, but this cluster is thought to be loose and irregularly organized;
no two U1 genes have been cloned on the same cosmid [271]. In the draft genome sequence,
we see six copies of U2 RNA that meet our criteria for true genes, three of which appear
to be in the expected position on chromosome 17. For U3, so far we see one true copy at
the correct place on chromosome 17p11.2. For U1, we see 16 true genes, 6 of which are
loosely clustered within 0.6 Mb at 1p36.1 and another 6 are elsewhere on chromosome 1.
Again, these and other clusters will be a matter for the finishing process.

Our observations also confirm the striking proliferation of ncRNA-derived pseudogenes
(Table 20). There are hundreds or thousands of sequences in the draft genome sequence
related to some of the ncRNA genes. The most prolific pseudogene counts generally come
from RNA genes transcribed by RNA polymerase III promoters, including U6, the hY
RNAs and SRP-RNA. These ncRNA pseudogenes presumably arise through reverse
transcription. The frequency of such events gives insight into how ncRNA genes can evolve
into SINE retroposons, such as the tRNA-derived SINEs found in many vertebrates and the
SRP-RNA-derived Alu elements found in humans.

Protein-coding genes

Identifying the protein-coding genes in the human genome is one of the most important
applications of the sequence data, but also one of the most difficult challenges. We describe
below our efforts to create an initial human gene and protein index.

Exploring properties of known genes.

Before attempting to identify new genes, we explored what could be learned by aligning the cDNA sequences of known genes to the draft genome sequence. Genomic alignments allow one to study exon–intron structure and local GC content, and are valuable for biomedical studies because they connect genes with the genetic and cytogenetic map, link them with regulatory sequences and facilitate the development of polymerase chain reaction (PCR) primers to amplify exons. Until now, genomic alignment was available for only about a quarter of known genes.

The 'known' genes studied were those in the RefSeq database [110], a manually curated
collection designed to contain nonredundant representatives of most full-length human
mRNA sequences in GenBank (RefSeq intentionally contains some alternative splice forms
of the same genes). The version of RefSeq used contained 10,272 mRNAs.

The RefSeq genes were aligned with the draft genome sequence, using both the Spidey (S.
Wheelan, personal communication) and Acembly (D. Thierry-Mieg and J. Thierry-Mieg,
unpublished; http://www.acedb.org) computer programs. Because this sequence is
incomplete and contains errors, not all genes could be fully aligned and some may have been
incorrectly aligned. More than 92% of the RefSeq entries could be aligned at high
stringency over at least part of their length, and 85% could be aligned over more than half
of their length. Some genes (16%) had high stringency alignments to more than one location
in the draft genome sequence owing, for example, to paralogues or pseudogenes. In such
cases, we considered only the best match. In a few of these cases, the assignment may not
be correct because the true matching region has not yet been sequenced. Three per cent of
entries appeared to be alternative splice products of the same gene, on the basis of their
alignment to the same location in the draft genome sequence. In all, we obtained at least
partial genomic alignments for 9,212 distinct known genes and essentially complete
alignment for 5,364 of them.

Previous efforts to study human gene structure [116, 274, 275] have been hampered by limited
sample sizes and strong biases in favour of compact genes. Table 21 gives the mean and
median values of some basic characteristics of gene structures. Some of the values may be
underestimates. In particular, the UTRs given in the RefSeq database are likely to be
incomplete; they are considerably shorter, for example, than those derived from careful
reconstructions on chromosome 22. Intron sizes were measured only for genes in finished
genomic sequence, to mitigate the bias arising from the fact that long introns are more likely
than short introns to be interrupted by gaps in the draft genome sequence. Nonetheless,
there may be some residual bias against long genes and long introns.

There is considerable variation in overall gene size and intron size, with both distributions
having very long tails. Many genes are over 100 kb long, the largest known example being
the dystrophin gene (DMD) at 2.4 Mb. The variation in the size distribution of coding
sequences and exons is less extreme, although there are still some remarkable outliers. The
titin gene276 has the longest currently known coding sequence at 80,780 bp; it also has the
largest number of exons (178) and longest single exon (17,106 bp).

It is instructive to compare the properties of human genes with those from worm and fly.
For all three organisms, the typical length of a coding sequence is similar (1,311 bp for
worm, 1,497 bp for fly and 1,340 bp for human), and most internal exons fall within a
common peak between 50 and 200 bp (Fig. 35a). However, the worm and fly exon
distributions have a fatter tail, resulting in a larger mean size for internal exons (218 bp for
worm versus 145 bp for human). The conservation of preferred exon size across all three
species supports suggestions of a conserved exon-based component of the splicing
machinery [277]. Intriguingly, the few extremely short human exons show an unusual base
composition. In 42 detected human exons of less than 19 bp, the nucleotide frequencies of
A, G, T and C are 39, 33, 15 and 12%, respectively, showing a strong purine bias.
Purine-rich sequences may enhance splicing [278, 279], and it is possible that such sequences
are required or strongly selected for to ensure correct splicing of very short exons. Previous
studies have shown that short exons require intronic, but not exonic, splicing enhancers [280].

In contrast to the exons, the intron size distributions differ substantially among the three
species (Fig. 35b, c). The worm and fly each have a reasonably tight distribution, with most
introns near the preferred minimum intron length (47 bp for worm, 59 bp for fly) and an
extended tail (overall average length of 267 bp for worm and 487 bp for fly). Intron size is
much more variable in humans, with a peak at 87 bp but a very long tail resulting in a mean
of more than 3,300 bp. The variation in intron size results in great variation in gene size.

The variation in gene size and intron size can partly be explained by the fact that GC-rich
regions tend to be gene-dense with many compact genes, whereas AT-rich regions tend to
be gene-poor with many sprawling genes containing large introns. The correlation of gene
density with GC content is shown in Fig. 36a, b; the relative density increases more than
tenfold as GC content increases from 30% to 50%. The correlation appears to be due
primarily to intron size, which drops markedly with increasing GC content (Fig. 36c). In
contrast, coding properties such as exon length (Fig. 36c) or exon number (data not shown)
vary little. Intergenic distance is also probably lower in high-GC areas, although this is hard
to prove directly until all genes have been identified.

The large number of confirmed human introns allows us to analyse variant splice sites,
confirming and extending recent reports [281]. Intron positions were confirmed by applying a
stringent criterion that EST or mRNA sequence show an exact match of 8 bp in the
flanking exonic sequence on each side. Of 53,295 confirmed introns, 98.12% use the
canonical dinucleotides GT at the 5' splice site and AG at the 3' site (GT–AG pattern).
Another 0.76% use the related GC–AG. About 0.10% use AT–AC, which is a rare
alternative pattern primarily recognized by the variant U12 splicing machinery [282]. The
remaining 1% belong to 177 types, some of which undoubtedly reflect sequencing or
alignment errors.

Finally, we looked at alternative splicing of human genes. Alternative splicing can allow
many proteins to be produced from a single gene and can be used for complex gene
regulation. It appears to be prevalent in humans, with lower estimates of about 35% of
human genes being subject to alternative splicing [283-285]. These studies may have
underestimated the prevalence of alternative splicing, because they examined only EST
alignments covering only a portion of a gene.

To investigate the prevalence of alternative splicing, we analysed reconstructed mRNA
transcripts covering the entire coding regions of genes on chromosome 22 (omitting small
genes with coding regions of less than 240 bp). Potential transcripts identified by alignments
of ESTs and cDNAs to genomic sequence were verified by human inspection. We found
642 transcripts, covering 245 genes (average of 2.6 distinct transcripts per gene). Two or
more alternatively spliced transcripts were found for 145 (59%) of these genes. A similar
analysis for the gene-rich chromosome 19 gave 1,859 transcripts, corresponding to 544
genes (average 3.2 distinct transcripts per gene). Because we are sampling only a subset of
all transcripts, the true extent of alternative splicing is likely to be greater. These figures are
considerably higher than those for worm, in which analysis reveals alternative splicing for
22% of genes for which ESTs have been found, with an average of 1.34 (12,816/9,516)
splice variants per gene. (The apparently higher extent of alternative splicing seen in human
than in worm was not an artefact resulting from much deeper coverage of human genes by
ESTs and mRNAs. Although there are many times more ESTs available for human than
worm, these ESTs tend to have shorter average length (because many were the product of
early sequencing efforts) and many match no human genes. We calculated the actual
coverage per bp used in the analysis of the human and worm genes; the coverage is only
modestly higher (about 50%) for the human, with a strong bias towards 3' UTRs which tend
to show much less alternative splicing. We also repeated the analysis using equal coverage
for the two organisms and confirmed that higher levels of alternative splicing were still seen
in human.)

Seventy per cent of alternative splice forms found in the genes on chromosomes 19 and 22
affect the coding sequence, rather than merely changing the 3' or 5' UTR. (This estimate
may be affected by the incomplete representation of UTRs in the RefSeq database and in
the transcripts studied.) Alternative splicing of the terminal exon was seen for 20% of 6,105
mRNAs that were aligned to the draft genome sequence and correspond to confirmed 3'
EST clusters. In addition to alternative splicing, we found evidence of the terminal exon
employing alternative polyadenylation sites (separated by > 100 bp) in 24% of cases.

...

Segmental history of the human genome

In bacteria, genomic segments often convey important information about function: genes
located close to one another often encode proteins in a common pathway and are regulated
in a common operon. In mammals, genes found close to each other only rarely have
common functions, but they are still interesting because they have a common history. In
fact, the study of genomic segments can shed light on biological events as long as 500 Myr
ago and as recently as 20,000 years ago.

Conserved segments between human and mouse

Humans and mice shared a common ancestor about 100 Myr ago. Despite the 200 Myr of
evolutionary distance between the species, a significant fraction of genes show synteny
between the two, being preserved within conserved segments. Genes tightly linked in one
mammalian species tend to be linked in others. In fact, conserved segments have been
observed in even more distant species: humans show conserved segments with fish350, 351
and even with invertebrates such as fly and worm352. In general, the likelihood that a
syntenic relationship will be disrupted correlates with the physical distance between the loci
and the evolutionary distance between the species.

Studying conserved segments between human and mouse has several uses. First,
conservation of gene order has been used to identify likely orthologues between the species,
particularly when investigating disease phenotypes. Second, the study of conserved
segments among genomes helps us to deduce evolutionary ancestry. And third, detailed
comparative maps may assist in the assembly of the mouse sequence, using the human
sequence as a scaffold.

Two types of linkage conservation are commonly described353. 'Conserved synteny'
indicates that at least two genes that reside on a common chromosome in one species are
also located on a common chromosome in the other species. Syntenic loci are said to lie in a
'conserved segment' when not only the chromosomal position but the linear order of the loci
has been preserved, without interruption by other chromosomal rearrangements.

An initial survey of homologous loci in human and mouse354 suggested that the total number
of conserved segments would be about 180. Subsequent estimates based on increasingly
detailed comparative maps have remained close to this projection353, 355, 356
(http://www.informatics.jax.org). The distribution of segment lengths has corresponded
reasonably well to the truncated negative exponential curve predicted by the random
breakage model357.

The availability of a draft human genome sequence allows the first global human–mouse
comparison in which human physical distances can be measured in Mb, rather than cM or
orthologous gene counts. We identified likely orthologues by reciprocal comparison of the
human and mouse mRNAs in the LocusLink database, using megaBLAST. For each
orthologous pair, we mapped the location of the human gene in the draft genome sequence
and then checked the location of the mouse gene in the Mouse Genome Informatics
database (http://www.informatics.jax.org). Using a conservative threshold, we identified
3,920 orthologous pairs in which the human gene could be mapped on the draft genome
sequence with high confidence. Of these, 2,998 corresponding mouse genes had a known
position in the mouse genome. We then searched for definitive conserved segments, defined
as human regions containing orthologues of at least two genes from the same mouse
chromosome region (< 15 cM) without interruption by segments from other chromosomes.

We identified 183 definitive conserved segments (Fig. 46). The average segment length was
15.4 Mb, with the largest segment being 90.5 Mb and the smallest 24 kb. There were also
141 'singletons', segments that contained only a single locus; these are not counted in the
statistics. Although some of these could be short conserved segments, they could also
reflect incorrect choices of orthologues or problems with the human or mouse maps.
Because of this conservative approach, the observed number of definitive segments is likely
be lower than the correct total. One piece of evidence for this conclusion comes from a
more detailed analysis on human chromosome 7 (ref. 358), which identified 20 conserved
segments, of which three were singletons. Our analysis revealed only 13 definitive segments
on this chromosome, with nine singletons.

The frequency of observing a particular gene count in a conserved segment is plotted on a
logarithmic scale in Fig. 47. If chromosomal breaks occur in a random fashion (as has been
proposed) and differences in gene density are ignored, a roughly straight line should result.
There is a clear excess for n = 1, suggesting that 50% or more of the singletons are indeed
artefactual. Thus, we estimate that true number of conserved segments is around 190–230,
in good agreement with the original Nadeau–Taylor prediction354.

Figure 48 shows a plot of the frequency of lengths of conserved segments, where the x-axis
scale is shown in Mb. As before, there is a fair amount of scatter in the data for the larger
segments (where the numbers are small), but the trend appears to be consistent with a
random breakage model.

We attempted to ascertain whether the breakpoint regions have any special characteristics.
This analysis was complicated by imprecision in the positioning of these breaks, which will
tend to blur any relationships. With 2,998 orthologues, the average interval within which a
break is known to have occurred is about 1.1 Mb. We compared the aggregate features of
these breakpoint intervals with the genome as a whole. The mean gene density was lower
in breakpoint regions than in the conserved segments (13.8 versus 18.6 per Mb). This
suggests that breakpoints may be more likely to occur or to undergo fixation in gene-poor
intervals than in gene-rich intervals. The occurrence of breakpoints may be promoted by
homologous recombination among repeated sequences359. When the sequence of the
mouse genome is finished, this analysis can be revisited more precisely.

A number of examples of extended conserved segments and syntenies are apparent in Fig.
46. As has been noted, almost all human genes on chromosome 17 are found on mouse
chromosome 11, with two members of the placental lactogen family from mouse 13
inserted. Apart from two singleton loci, human chromosome 20 appears to be entirely
orthologous to mouse chromosome 2, apparently in a single segment. The largest apparently
contiguous conserved segment in the human genome is on chromosome 4, including roughly
90.5 Mb of human DNA that is orthologous to mouse chromosome 5. This analysis also
allows us to infer the likely location of thousands of mouse genes for which the human
orthologue has been located in the draft genome sequence but the mouse locus has not yet
been mapped.

With about 200 conserved segments between mouse and human and about 100 Myr of
evolution from their common ancestor360, we obtain an estimated rate of about 1.0
chromosomal rearrangement being fixed per Myr. However, there is good evidence that the
rate of chromosomal rearrangement (like the rate of nucleotide substitutions; see above)
differs between the two species. Among mammals, rodents may show unusually rapid
chromosome alteration. By comparison, very few rearrangements have been observed
among primates, and studies of a broader array of mammalian orders, including cats, cows,
sheep and pigs, suggest an average rate of chromosome alteration of only about 0.2
rearrangements per Myr in these lineages361. Additional evidence that rodents are outliers
comes from a recent analysis of synteny between the human and zebrafish genomes. From
a study of 523 orthologues, it was possible to project 418 conserved segments350. Assuming
400 Myr since a common vertebrate ancestor of zebrafish and humans362, we obtain an
estimate of 0.52 rearrangements per Myr. Recent estimates of rearrangement rates in
plants have suggested bimodality, with some pairs showing rates of 0.15–0.41
rearrangements per Myr, and others showing higher rates of 1.1–1.3 rearrangements per
Myr363. With additional detailed genome maps of multiple species, it should be possible to
determine whether this particular molecular clock is truly operating at a different rate in
various branches of the evolutionary tree, and whether variations in that rate are bimodal or
continuous. It should also be possible to reconstruct the karyotypes of common ancestors.

Ancient duplicated segments in the human genome

Another approach to genomic history is to study segmental duplications within the human
genome. Earlier, we discussed examples of recent duplications of genomic segments to
pericentromeric and subtelomeric regions. Most of these events appear to be evolutionary
dead-ends resulting in nonfunctional pseudogenes; however, segmental duplication