human

Background to the Human Genome Project

The Human Genome Project arose from two key insights that emerged in the early 1980s: that the ability to take global views of genomes could greatly accelerate biomedical research, by allowing researchers to attack problems in a comprehensive and unbiased fashion; and that the creation of such global views would require a communal effort in infrastructure building, unlike anything previously attempted in biomedical research. Several key projects helped to crystallize these insights, including:

(1) The sequencing of the bacterial viruses X174^{4, 5} and lambda⁶, the animal virus SV40⁷ and the human mitochondrion⁸ between 1977 and 1982. These projects proved the feasibility of assembling small sequence fragments into complete genomes, and showed the value of complete catalogues of genes and other functional elements.

(2) The programme to create a human genetic map to make it possible to locate disease genes of unknown function based solely on their inheritance patterns, launched by Botstein and colleagues in 1980 (ref. 9).

(3) The programmes to create physical maps of clones covering the yeast¹⁰ and worm¹¹ genomes to allow isolation of genes and regions based solely on their chromosomal position, launched by Olson and Sulston in the mid-1980s.

(4) The development of random shotgun sequencing of complementary DNA fragments for high-throughput gene discovery by Schimmel¹² and Schimmel and Sutcliffe¹³, later dubbed expressed sequence tags (ESTs) and pursued with automated sequencing by Venter and others^{14, 15, 16, 17, 18, 19, 20}.

The idea of sequencing the entire human genome was first proposed in discussions at scientific meetings organized by the US Department of Energy and others from 1984 to 1986 (refs 21, 22). A committee appointed by the US National Research Council endorsed the concept in its 1988 report²³, but recommended a broader programme, to include: the creation of genetic, physical and sequence maps of the human genome; parallel efforts in key model organisms such as bacteria, yeast, worms, flies and mice; the development of technology in support of these objectives; and research into the ethical, legal and social issues raised by human genome research. The programme was launched in the US as a joint effort of the Department of Energy and the National Institutes of Health. In other countries, the UK Medical Research Council and the Wellcome Trust supported genomic research in Britain; the Centre d'Etude du Polymorphisme Humain and the French Muscular Dystrophy Association launched mapping efforts in France; government agencies, including the Science and Technology Agency and the Ministry of Education, Science, Sports and Culture supported genomic research efforts in Japan; and the European Community helped to launch several international efforts, notably the programme to sequence the yeast genome. By late 1990, the Human Genome Project had been launched, with the creation of genome centres in these countries. Additional participants subsequently joined the effort, notably in Germany and China. In addition, the Human Genome Organization (HUGO) was founded to provide a forum for international coordination of genomic research. Several books^{24, 25, 26} provide a more comprehensive discussion of the genesis of the Human Genome Project.

Through 1995, work progressed rapidly on two fronts (Fig. 1). The first was construction of genetic and physical maps of the human and mouse genomes^{27, 28, 29, 30, 31}, providing key tools for identification of disease genes and anchoring points for genomic sequence. The second was sequencing of the yeast³² and worm³³ genomes, as well as targeted regions of mammalian genomes^{34, 35, 36, 37}. These projects showed that large-scale sequencing was feasible and developed the two-phase paradigm for genome sequencing. In the first, 'shotgun', phase, the genome is divided into appropriately sized segments and each segment is covered to a high degree of redundancy (typically, eight- to tenfold) through the sequencing of randomly selected subfragments. The second is a 'finishing' phase, in which sequence gaps are closed and remaining ambiguities are resolved through directed analysis. The results also showed that complete genomic sequence provided information about genes, regulatory regions and chromosome structure that was not readily obtainable from cDNA studies alone.

Figure 1: Timeline of large-scale genomic analyses.

Shown are selected components of work on several non-vertebrate model organisms (red), the mouse (blue) and the human (green) from 1990; earlier projects are described in the text. SNPs, single nucleotide polymorphisms; ESTs, expressed sequence tags.

High resolution image and legend (59K)

In 1995, genome scientists considered a proposal³⁸ that would have involved producing a draft genome sequence of the human genome in a first phase and then returning to finish the sequence in a second phase. After vigorous debate, it was decided that such a plan was premature for several reasons. These included the need first to prove that high-quality, long-range finished sequence could be produced from most parts of the complex, repeat-rich human genome; the sense that many aspects of the sequencing process were still rapidly evolving; and the desirability of further decreasing costs.

Instead, pilot projects were launched to demonstrate the feasibility of cost-effective, large-scale sequencing, with a target completion date of March 1999. The projects successfully produced finished sequence with 99.99% accuracy and no gaps³⁹. They also introduced bacterial artificial chromosomes (BACs)⁴⁰, a new large-insert cloning system that proved to be more stable than the cosmids and yeast artificial chromosomes (YACs)⁴¹ that had been used previously. The pilot projects drove the maturation and convergence of sequencing strategies, while producing 15% of the human genome sequence. With successful completion of this phase, the human genome sequencing effort moved into full-scale production in March 1999.

The idea of first producing a draft genome sequence was revived at this time, both because the ability to finish such a sequence was no longer in doubt and because there was great hunger in the scientific community for human sequence data. In addition, some scientists favoured prioritizing the production of a draft genome sequence over regional finished sequence because of concerns about commercial plans to generate proprietary databases of human sequence that might be subject to undesirable restrictions on use^{42, 43, 44}.

The consortium focused on an initial goal of producing, in a first production phase lasting until June 2000, a draft genome sequence covering most of the genome. Such a draft genome sequence, although not completely finished, would rapidly allow investigators to begin to extract most of the information in the human sequence. Experiments showed that sequencing clones covering about 90% of the human genome to a redundancy of about four- to fivefold ('half-shotgun' coverage; see Box 1) would accomplish this^{45, 46}. The draft genome sequence goal has been achieved, as described below.

The second sequence production phase is now under way. Its aims are to achieve full-shotgun coverage of the existing clones during 2001, to obtain clones to fill the remaining gaps in the physical map, and to produce a finished sequence (apart from regions that cannot be cloned or sequenced with currently available techniques) no later than 2003.

Strategic issues

Hierarchical shotgun sequencing

Soon after the invention of DNA sequencing methods^{47, 48}, the shotgun sequencing strategy was introduced^{49, 50, 51}; it has remained the fundamental method for large-scale genome sequencing^{52, 53, 54} for the past 20 years. The approach has been refined and extended to make it more efficient. For example, improved protocols for fragmenting and cloning DNA allowed construction of shotgun libraries with more uniform representation. The practice of sequencing from both ends of double-stranded clones ('double-barrelled' shotgun sequencing) was introduced by Ansorge and others³⁷ in 1990, allowing the use of 'linking information' between sequence fragments.

The application of shotgun sequencing was also extended by applying it to larger and larger DNA molecules—from plasmids ( 4 kilobases (kb)) to cosmid clones³⁷ (40 kb), to artificial chromosomes cloned in bacteria and yeast⁵⁵ (100–500 kb) and bacterial genomes⁵⁶ (1–2 megabases (Mb)). In principle, a genome of arbitrary size may be directly sequenced by the shotgun method, provided that it contains no repeated sequence and can be uniformly sampled at random. The genome can then be assembled using the simple computer science technique of 'hashing' (in which one detects overlaps by consulting an alphabetized look-up table of all k-letter words in the data). Mathematical analysis of the expected number of gaps as a function of coverage is similarly straightforward⁵⁷.

Practical difficulties arise because of repeated sequences and cloning bias. Small amounts of repeated sequence pose little problem for shotgun sequencing. For example, one can readily assemble typical bacterial genomes (about 1.5% repeat) or the euchromatic portion of the fly genome (about 3% repeat). By contrast, the human genome is filled (> 50%) with repeated sequences, including interspersed repeats derived from transposable elements, and long genomic regions that have been duplicated in tandem, palindromic or dispersed fashion (see below). These include large duplicated segments (50–500 kb) with high sequence identity (98–99.9%), at which mispairing during recombination creates deletions responsible for genetic syndromes. Such features complicate the assembly of a correct and finished genome sequence.

There are two approaches for sequencing large repeat-rich genomes. The first is a whole-genome shotgun sequencing approach, as has been used for the repeat-poor genomes of viruses, bacteria and flies, using linking information and computational analysis to attempt to avoid misassemblies. The second is the 'hierarchical shotgun sequencing' approach (Fig. 2), also referred to as 'map-based', 'BAC-based' or 'clone-by-clone'. This approach involves generating and organizing a set of large-insert clones (typically 100–200 kb each) covering the genome and separately performing shotgun sequencing on appropriately chosen clones. Because the sequence information is local, the issue of long-range misassembly is eliminated and the risk of short-range misassembly is reduced. One caveat is that some large-insert clones may suffer rearrangement, although this risk can be reduced by appropriate quality-control measures involving clone fingerprints (see below).

Figure 2: Idealized representation of the hierarchical shotgun sequencing strategy.

A library is constructed by fragmenting the target genome and cloning it into a large-fragment cloning vector; here, BAC vectors are shown. The genomic DNA fragments represented in the library are then organized into a physical map and individual BAC clones are selected and sequenced by the random shotgun strategy. Finally, the clone sequences are assembled to reconstruct the sequence of the genome.

High resolution image and legend (49K)

The two methods are likely to entail similar costs for producing finished sequence of a mammalian genome. The hierarchical approach has a higher initial cost than the whole-genome approach, owing to the need to create a map of clones (about 1% of the total cost of sequencing) and to sequence overlaps between clones. On the other hand, the whole-genome approach is likely to require much greater work and expense in the final stage of producing a finished sequence, because of the challenge of resolving misassemblies. Both methods must also deal with cloning biases, resulting in under-representation of some regions in either large-insert or small-insert clone libraries.

There was lively scientific debate over whether the human genome sequencing effort should employ whole-genome or hierarchical shotgun sequencing. Weber and Myers⁵⁸ stimulated these discussions with a specific proposal for a whole-genome shotgun approach, together with an analysis suggesting that the method could work and be more efficient. Green⁵⁹ challenged these conclusions and argued that the potential benefits did not outweigh the likely risks.

In the end, we concluded that the human genome sequencing effort should employ the hierarchical approach for several reasons. First, it was prudent to use the approach for the first project to sequence a repeat-rich genome. With the hierarchical approach, the ultimate frequency of misassembly in the finished product would probably be lower than with the whole-genome approach, in which it would be more difficult to identify regions in which the assembly was incorrect.

Second, it was prudent to use the approach in dealing with an outbred organism, such as the human. In the whole-genome shotgun method, sequence would necessarily come from two different copies of the human genome. Accurate sequence assembly could be complicated by sequence variation between these two copies—both SNPs (which occur at a rate of 1 per 1,300 bases) and larger-scale structural heterozygosity (which has been documented in human chromosomes). In the hierarchical shotgun method, each large-insert clone is derived from a single haplotype.

Third, the hierarchical method would be better able to deal with inevitable cloning biases, because it would more readily allow targeting of additional sequencing to under-represented regions. And fourth, it was better suited to a project shared among members of a diverse international consortium, because it allowed work and responsibility to be easily distributed. As the ultimate goal has always been to create a high-quality, finished sequence to serve as a foundation for biomedical research, we reasoned that the advantages of this more conservative approach outweighed the additional cost, if any.

A biotechnology company, Celera Genomics, has chosen to incorporate the whole-genome shotgun approach into its own efforts to sequence the human genome. Their plan^{60, 61} uses a mixed strategy, involving combining some coverage with whole-genome shotgun data generated by the company together with the publicly available hierarchical shotgun data generated by the International Human Genome Sequencing Consortium. If the raw sequence reads from the whole-genome shotgun component are made available, it may be possible to evaluate the extent to which the sequence of the human genome can be assembled without the need for clone-based information. Such analysis may help to refine sequencing strategies for other large genomes.

Technology for large-scale sequencing

Sequencing the human genome depended on many technological improvements in the production and analysis of sequence data. Key innovations were developed both within and outside the Human Genome Project. Laboratory innovations included four-colour fluorescence-based sequence detection⁶², improved fluorescent dyes^{63, 64, 65, 66}, dye-labelled terminators⁶⁷, polymerases specifically designed for sequencing^{68, 69, 70}, cycle sequencing⁷¹ and capillary gel electrophoresis^{72, 73, 74}. These studies contributed to substantial improvements in the automation, quality and throughput of collecting raw DNA sequence^{75, 76}. There were also important advances in the development of software packages for the analysis of sequence data. The PHRED software package^{77, 78} introduced the concept of assigning a 'base-quality score' to each base, on the basis of the probability of an erroneous call. These quality scores make it possible to monitor raw data quality and also assist in determining whether two similar sequences truly overlap. The PHRAP computer package (http://bozeman.mbt.washington.edu/phrap.docs/phrap.html) then systematically assembles the sequence data using the base-quality scores. The program assigns 'assembly-quality scores' to each base in the assembled sequence, providing an objective criterion to guide sequence finishing. The quality scores were based on and validated by extensive experimental data.

Another key innovation for scaling up sequencing was the development by several centres of automated methods for sample preparation. This typically involved creating new biochemical protocols suitable for automation, followed by construction of appropriate robotic systems.

Coordination and public data sharing

The Human Genome Project adopted two important principles with regard to human sequencing. The first was that the collaboration would be open to centres from any nation. Although potentially less efficient, in a narrow economic sense, than a centralized approach involving a few large factories, the inclusive approach was strongly favoured because we felt that the human genome sequence is the common heritage of all humanity and the work should transcend national boundaries, and we believed that scientific progress was best assured by a diversity of approaches. The collaboration was coordinated through periodic international meetings (referred to as 'Bermuda meetings' after the venue of the first three gatherings) and regular telephone conferences. Work was shared flexibly among the centres, with some groups focusing on particular chromosomes and others contributing in a genome-wide fashion.

The second principle was rapid and unrestricted data release. The centres adopted a policy that all genomic sequence data should be made publicly available without restriction within 24 hours of assembly^{79, 80}. Pre-publication data releases had been pioneered in mapping projects in the worm¹¹ and mouse genomes^{30, 81} and were prominently adopted in the sequencing of the worm, providing a direct model for the human sequencing efforts. We believed that scientific progress would be most rapidly advanced by immediate and free availability of the human genome sequence. The explosion of scientific work based on the publicly available sequence data in both academia and industry has confirmed this judgement.

Generating the draft genome sequence

Generating a draft sequence of the human genome involved three steps: selecting the BAC clones to be sequenced, sequencing them and assembling the individual sequenced clones into an overall draft genome sequence. A glossary of terms related to genome sequencing and assembly is provided in Box 1.

The draft genome sequence is a dynamic product, which is regularly updated as additional data accumulate en route to the ultimate goal of a completely finished sequence. The results below are based on the map and sequence data available on 7 October 2000, except as otherwise noted. At the end of this section, we provide a brief update of key data.

Clone selection

The hierarchical shotgun method involves the sequencing of overlapping large-insert clones spanning the genome. For the Human Genome Project, clones were largely chosen from eight large-insert libraries containing BAC or P1-derived artificial chromosome (PAC) clones (Table 1; refs 82,83,84,85,86,87,88). The libraries were made by partial digestion of genomic DNA with restriction enzymes. Together, they represent around 65-fold coverage (redundant sampling) of the genome. Libraries based on other vectors, such as cosmids, were also used in early stages of the project.

Table 1: Key large-insert genome-wide libraries

Full table

The libraries (Table 1) were prepared from DNA obtained from anonymous human donors in accordance with US Federal Regulations for the Protection of Human Subjects in Research (45CFR46) and following full review by an Institutional Review Board. Briefly, the opportunity to donate DNA for this purpose was broadly advertised near the two laboratories engaged in library construction. Volunteers of diverse backgrounds were accepted on a first-come, first-taken basis. Samples were obtained after discussion with a genetic counsellor and written informed consent. The samples were made anonymous as follows: the sampling laboratory stripped all identifiers from the samples, applied random numeric labels, and transferred them to the processing laboratory, which then removed all labels and relabelled the samples. All records of the labelling were destroyed. The processing laboratory chose samples at random from which to prepare DNA and immortalized cell lines. Around 5–10 samples were collected for every one that was eventually used. Because no link was retained between donor and DNA sample, the identity of the donors for the libraries is not known, even by the donors themselves. A more complete description can be found at http://www.nhgri.nih.gov/Grant_info/Funding/Statements/RFA/human_subjects.html.

During the pilot phase, centres showed that sequence-tagged sites (STSs) from previously constructed genetic and physical maps could be used to recover BACs from specific regions. As sequencing expanded, some centres continued this approach, augmented with additional probes from flow sorting of chromosomes to obtain long-range coverage of specific chromosomes or chromosomal regions^{89, 90, 91, 92, 93, 94}.

For the large-scale sequence production phase, a genome-wide physical map of overlapping clones was also constructed by systematic analysis of BAC clones representing 20-fold coverage of the human genome⁸⁶. Most clones came from the first three sections of the RPCI-11 library, supplemented with clones from sections of the RPCI-13 and CalTech D libraries (Table 1). DNA from each BAC clone was digested with the restriction enzyme HindIII, and the sizes of the resulting fragments were measured by agarose gel electrophoresis. The pattern of restriction fragments provides a 'fingerprint' for each BAC, which allows different BACs to be distinguished and the degree of overlaps to be assessed. We used these restriction-fragment fingerprints to determine clone overlaps, and thereby assembled the BACs into fingerprint clone contigs.

The fingerprint clone contigs were positioned along the chromosomes by anchoring them with STS markers from existing genetic and physical maps. Fingerprint clone contigs were tied to specific STSs initially by probe hybridization and later by direct search of the sequenced clones. To localize fingerprint clone contigs that did not contain known markers, new STSs were generated and placed onto chromosomes⁹⁵. Representative clones were also positioned by fluorescence in situ hybridization (FISH) (ref. 86 and C. McPherson, unpublished).

We selected clones from the fingerprint clone contigs for sequencing according to various criteria. Fingerprint data were reviewed^{86, 90} to evaluate overlaps and to assess clone fidelity (to bias against rearranged clones^{83, 96}). STS content information and BAC end sequence information were also used^{91, 92}. Where possible, we tried to select a minimally overlapping set spanning a region. However, because the genome-wide physical map was constructed concurrently with the sequencing, continuity in many regions was low in early stages. These small fingerprint clone contigs were nonetheless useful in identifying validated, nonredundant clones that were used to 'seed' the sequencing of new regions. The small fingerprint clone contigs were extended or merged with others as the map matured.

The clones that make up the draft genome sequence therefore do not constitute a minimally overlapping set—there is overlap and redundancy in places. The cost of using suboptimal overlaps was justified by the benefit of earlier availability of the draft genome sequence data. Minimizing the overlap between adjacent clones would have required completing the physical map before undertaking large-scale sequencing. In addition, the overlaps between BAC clones provide a rich collection of SNPs. More than 1.4 million SNPs have already been identified from clone overlaps and other sequence comparisons⁹⁷.

Because the sequencing project was shared among twenty centres in six countries, it was important to coordinate selection of clones across the centres. Most centres focused on particular chromosomes or, in some cases, larger regions of the genome. We also maintained a clone registry to track selected clones and their progress. In later phases, the global map provided an integrated view of the data from all centres, facilitating the distribution of effort to maximize coverage of the genome. Before performing extensive sequencing on a clone, several centres routinely examined an initial sample of 96 raw sequence reads from each subclone library to evaluate possible overlap with previously sequenced clones.

Sequencing

The selected clones were subjected to shotgun sequencing. Although the basic approach of shotgun sequencing is well established, the details of implementation varied among the centres. For example, there were differences in the average insert size of the shotgun libraries, in the use of single-stranded or double-stranded cloning vectors, and in sequencing from one end or both ends of each insert. Centres differed in the fluorescent labels employed and in the degree to which they used dye-primers or dye-terminators. The sequence detectors included both slab gel- and capillary-based devices. Detailed protocols are available on the web sites of many of the individual centres (URLs can be found at http://www.nhgri.nih.gov/genome_hub.html). The extent of automation also varied greatly among the centres, with the most aggressive automation efforts resulting in factory-style systems able to process more than 100,000 sequencing reactions in 12 hours (Fig. 3). In addition, centres differed in the amount of raw sequence data typically obtained for each clone (so-called half-shotgun, full shotgun and finished sequence). Sequence information from the different centres could be directly integrated despite this diversity, because the data were analysed by a common computational procedure. Raw sequence traces were processed and assembled with the PHRED and PHRAP software packages^{77, 78} (P. Green, unpublished). All assembled contigs of more than 2 kb were deposited in public databases within 24 hours of assembly.

Figure 3: The automated production line for sample preparation at the Whitehead Institute, Center for Genome Research.

The system consists of custom-designed factory-style conveyor belt robots that perform all functions from purifying DNA from bacterial cultures through setting up and purifying sequencing reactions.

High resolution image and legend (84K)

The overall sequencing output rose sharply during production (Fig. 4). Following installation of new sequence detectors beginning in June 1999, sequencing capacity and output rose approximately eightfold in eight months to nearly 7 million samples processed per month, with little or no drop in success rate (ratio of useable reads to attempted reads). By June 2000, the centres were producing raw sequence at a rate equivalent to onefold coverage of the entire human genome in less than six weeks. This corresponded to a continuous throughput exceeding 1,000 nucleotides per second, 24 hours per day, seven days per week. This scale-up resulted in a concomitant increase in the sequence available in the public databases (Fig. 4).

Figure 4: Total amount of human sequence in the High Throughput Genome Sequence (HTGS) division of GenBank.

The total is the sum of finished sequence (red) and unfinished (draft plus predraft) sequence (yellow).

High resolution image and legend (33K)

A version of the draft genome sequence was prepared on the basis of the map and sequence data available on 7 October 2000. For this version, the mapping effort had assembled the fingerprinted BACs into 1,246 fingerprint clone contigs. The sequencing effort had sequenced and assembled 29,298 overlapping BACs and other large-insert clones (Table 2), comprising a total length of 4.26 gigabases (Gb). This resulted from around 23 Gb of underlying raw shotgun sequence data, or about 7.5-fold coverage averaged across the genome (including both draft and finished sequence). The various contributions to the total amount of sequence deposited in the HTGS division of GenBank are given in Table 3.

Table 2: Total genome sequence from the collection of sequenced clones, by sequence status

Full table

Table 3: Total human sequence deposited in the HTGS division of GenBank

Full table

By agreement among the centres, the collection of draft clones produced by each centre was required to have fourfold average sequence coverage, with no clone below threefold. (For this purpose, sequence coverage was defined as the average number of times that each base was independently read with a base-quality score corresponding to at least 99% accuracy.) We attained an overall average of 4.5-fold coverage across the genome for draft clones. A few of the sequenced clones fell below the minimum of threefold sequence coverage or have not been formally designated by centres as meeting draft standards; these are referred to as predraft (Table 2). Some of these are clones that span remaining gaps in the draft genome sequence and were in the process of being sequenced on 7 October 2000; a few are old submissions from centres that are no longer active.

The lengths of the initial sequence contigs in the draft clones vary as a function of coverage, but half of all nucleotides reside in initial sequence contigs of at least 21.7 kb (see below). Various properties of the draft clones can be assessed from instances in which there was substantial overlap between a draft clone and a finished (or nearly finished) clone. By examining the sequence alignments in the overlap regions, we estimated that the initial sequence contigs in a draft sequence clone cover an average of about 96% of the clone and are separated by gaps with an average size of about 500 bp.

Although the main emphasis was on producing a draft genome sequence, the centres also maintained sequence finishing activities during this period, leading to a twofold increase in finished sequence from June 1999 to June 2000 (Fig. 4). The total amount of human sequence in this final form stood at more than 835 Mb on 7 October 2000, or more than 25% of the human genome. This includes the finished sequences of chromosomes 21 and 22 (refs 93, 94). As centres have begun to shift from draft to finished sequencing in the last quarter of 2000, the production of finished sequence has increased to an annualized rate of 1 Gb per year and is continuing to rise.

In addition to sequencing large-insert clones, three centres generated a large collection of random raw sequence reads from whole-genome shotgun libraries (Table 4; ref. 98). These 5.77 million successful sequences contained 2.4 Gb of high-quality bases; this corresponds to about 0.75-fold coverage and would be statistically expected to include about 50% of the nucleotides in the human genome (data available at http://snp.cshl.org/data). The primary objective of this work was to discover SNPs, by comparing these random raw sequences (which came from different individuals) with the draft genome sequence. However, many of these raw sequences were obtained from both ends of plasmid clones and thereby also provided valuable 'linking' information that was used in sequence assembly. In addition, the random raw sequences provide sequence coverage of about half of the nucleotides not yet represented in the sequenced large-insert clones; these can be used as probes for portions of the genome not yet recovered.

Table 4: Plasmid paired-end reads

Full table

Assembly of the draft genome sequence

We then set out to assemble the sequences from the individual large-insert clones into an integrated draft sequence of the human genome. The assembly process had to resolve problems arising from the draft nature of much of the sequence, from the variety of clone sources, and from the high fraction of repeated sequences in the human genome. This process involved three steps: filtering, layout and merging.

The entire data set was filtered uniformly to eliminate contamination from nonhuman sequences and other artefacts that had not already been removed by the individual centres. (Information about contamination was also sent back to the centres, which are updating the individual entries in the public databases.) We also identified instances in which the sequence data from one BAC clone was substantially contaminated with sequence data from another (human or nonhuman) clone. The problems were resolved in most instances; 231 clones remained unresolved, and these were eliminated from the assembly reported here. Instances of lower levels of cross-contamination (for example, a single 96-well microplate misassigned to the wrong BAC) are more difficult to detect; some undoubtedly remain and may give rise to small spurious sequence contigs in the draft genome sequence. Such issues are readily resolved as the clones progress towards finished sequence, but they necessitate some caution in certain applications of the current data.

The sequenced clones were then associated with specific clones on the physical map to produce a 'layout'. In principle, sequenced clones that correspond to fingerprinted BACs could be directly assigned by name to fingerprint clone contigs on the fingerprint-based physical map. In practice, however, laboratory mixups occasionally resulted in incorrect assignments. To eliminate such problems, sequenced clones were associated with the fingerprint clone contigs in the physical map by using the sequence data to calculate a partial list of restriction fragments in silico and comparing that list with the experimental database of BAC fingerprints. The comparison was feasible because the experimental sizing of restriction fragments was highly accurate (to within 0.5–1.5% of the true size, for 95% of fragments from 600 to 12,000 base pairs (bp))^{84, 85}. Reliable matching scores could be obtained for 16,193 of the clones. The remaining sequenced clones could not be placed on the map by this method because they were too short, or they contained too many small initial sequence contigs to yield enough restriction fragments, or possibly because their sequences were not represented in the fingerprint database.

An independent approach to placing sequenced clones on the physical map used the database of end sequences from fingerprinted BACs (Table 1). Sequenced clones could typically be reliably mapped if they contained multiple matches to BAC ends, with all corresponding to clones from a single genomic region (multiple matches were required as a safeguard against errors known to exist in the BAC end database and against repeated sequences). This approach provided useful placement information for 22,566 sequenced clones.

Altogether, we could assign 25,403 sequenced clones to fingerprint clone contigs by combining in silico digestion and BAC end sequence match data. To place most of the remaining sequenced clones, we exploited information about sequence overlap or BAC-end paired links of these clones with already positioned clones. This left only a few, mostly small, sequenced clones that could not be placed (152 sequenced clones containing 5.5 Mb of sequence out of 29,298 sequenced clones containing more than 4,260 Mb of sequence); these are being localized by radiation hybrid mapping of STSs derived from their sequences.

The fingerprint clone contigs were then mapped to chromosomal locations, using sequence matches to mapped STSs from four human radiation hybrid maps^{95, 99, 100}, one YAC and radiation hybrid map²⁹, and two genetic maps^{101, 102}, together with data from FISH^{86, 90, 103}. The mapping was iteratively refined by comparing the order and orientation of the STSs in the fingerprint clone contigs and the various STS-based maps, to identify and refine discrepancies (Fig. 5). Small fingerprint clone contigs (< 1 Mb) were difficult to orient and, sometimes, to order using these methods. In all, 942 fingerprint clone contigs contained sequenced clones. (An additional 304 of the 1,246 fingerprint clone contigs did not contain sequenced clones, but these tended to be extremely small and together contain less than 1% of the mapped clones. About one-third have been targeted for sequencing. A few derive from the Y chromosome, for which the map was constructed separately⁸⁹. Most of the remainder are fragments of other larger contigs or represent other artefacts. These are being eliminated in subsequent versions of the database.) Of these 942 contigs with sequenced clones, 852 (90%, containing 99.2% of the total sequence) were localized to specific chromosome locations in this way. An additional 51 fingerprint clone contigs, containing 0.5% of the sequence, could be assigned to a specific chromosome but not to a precise position. The remaining 39 contigs containing 0.3% of the sequence were not positioned at all.

Figure 5: Positions of markers on previous maps of the genome (the Genethon¹⁰¹ genetic map and Marshfield genetic map (http://research.marshfieldclinic.org/genetics/genotyping_service/mgsver2.htm ), the GeneMap99 radiation hybrid map¹⁰⁰, and the Whitehead YAC and radiation hybrid map²⁹) plotted against their derived position on the draft sequence for chromosome 2.

The horizontal units are Mb but the vertical units of each map vary (cM, cR and so on) and thus all were scaled so that the entire map spans the full vertical range. Markers that map to other chromosomes are shown in the chromosome lines at the top.The data sets generally follow the diagonal, indicating that order and orientation of the marker sets on the different maps largely agree (note that the two genetic maps are completely superimposed). In a, there are two segments (bars) that are inverted in an earlier version draft sequence relative to all the other maps. b, The same chromosome after the information was used to reorient those two segments.

High resolution image and legend (136K)

We then merged the sequences from overlapping sequenced clones (Fig. 6), using the computer program GigAssembler¹⁰⁴. The program considers nearby sequenced clones, detects overlaps between the initial sequence contigs in these clones, merges the overlapping sequences and attempts to order and orient the sequence contigs. It begins by aligning the initial sequence contigs from one clone with those from other clones in the same fingerprint clone contig on the basis of length of alignment, per cent identity of the alignment, position in the sequenced clone layout and other factors. Alignments are limited to one end of each initial sequence contig for partially overlapping contigs or to both ends of an initial sequence contig contained entirely within another; this eliminates internal alignments that may reflect repeated sequence or possible misassembly (Fig. 6b). Beginning with the highest scoring pairs, initial sequence contigs are then integrated to produce 'merged sequence contigs' (usually referred to simply as 'sequence contigs'). The program refines the arrangement of the clones within the fingerprint clone contig on the basis of the extent of sequence overlap between them and then rebuilds the sequence contigs. Next, the program selects a sequence path through the sequence contigs (Fig. 6c). It tries to use the highest quality data by preferring longer initial sequence contigs and avoiding the first and last 250 bases of initial sequence contigs where possible. Finally, it attempts to order and orient the sequence contigs by using additional information, including sequence data from paired-end plasmid and BAC reads, known messenger RNAs and ESTs, as well as additional linking information provided by centres. The sequence contigs are thereby linked together to create 'sequence-contig scaffolds' (Fig. 6d). The process also joins overlapping sequenced clones into sequenced-clone contigs and links sequenced-clone contigs to form sequenced-clone-contig scaffolds. A fingerprint clone contig may contain several sequenced-clone contigs, because bridging clones remain to be sequenced. The assembly contained 4,884 sequenced-clone contigs in 942 fingerprint clone contigs.

Figure 6: The key steps (a–d) in assembling individual sequenced clones into the draft genome sequence.

A1–A5 represent initial sequence contigs derived from shotgun sequencing of clone A, and B1–B6 are from clone B.

High resolution image and legend (28K)

The hierarchy of contigs is summarized in Fig. 7. Initial sequence contigs are integrated to create merged sequence contigs, which are then linked to form sequence-contig scaffolds. These scaffolds reside within sequenced-clone contigs, which in turn reside within fingerprint clone contigs.

Figure 7: Levels of clone and sequence coverage.

A 'fingerprint clone contig' is assembled by using the computer program FPC^84,451 to analyse the restriction enzyme digestion patterns of many large-insert clones. Clones are then selected for sequencing to minimize overlap between adjacent clones. For a clone to be selected, all of its restriction enzyme fragments (except the two vector-insert junction fragments) must be shared with at least one of its neighbours on each side in the contig. Once these overlapping clones have been sequenced, the set is a 'sequenced-clone contig'. When all selected clones from a fingerprint clone contig have been sequenced, the sequenced-clone contig will be the same as the fingerprint clone contig. Until then, a fingerprint clone contig may contain several sequenced-clone contigs. After individual clones (for example, A and B) have been sequenced to draft coverage and the clones have been mapped, the data are analysed by GigAssembler (Fig. 6), producing merged sequence contigs from initial sequence contigs, and linking these to form sequence-contig scaffolds (see Box 1).

High resolution image and legend (55K)

The draft genome sequence

The result of the assembly process is an integrated draft sequence of the human genome. Several features of the draft genome sequence are reported in Tables 5, 6 & 7, including the proportion represented by finished, draft and predraft categories. The Tables also show the numbers and lengths of different types of contig, for each chromosome and for the genome as a whole.

Table 5: The draft genome sequence

Full table

Table 6: Clone level contiguity of the draft genome sequence

Full table

Table 7: Sequence level contiguity of the draft genome sequence

Full table

The contiguity of the draft genome sequence at each level is an important feature. Two commonly used statistics have significant drawbacks for describing contiguity. The 'average length' of a contig is deflated by the presence of many small contigs comprising only a small proportion of the genome, whereas the 'length-weighted average length' is inflated by the presence of large segments of finished sequence. Instead, we chose to describe the contiguity as a property of the 'typical' nucleotide. We used a statistic called the 'N50 length', defined as the largest length L such that 50% of all nucleotides are contained in contigs of size at least L.

The continuity of the draft genome sequence reported here and the effectiveness of assembly can be readily seen from the following: half of all nucleotides reside within an initial sequence contig of at least 21.7 kb, a sequence contig of at least 82 kb, a sequence-contig scaffold of at least 274 kb, a sequenced-clone contig of at least 826 kb and a fingerprint clone contig of at least 8.4 Mb (Tables 6, 7). The cumulative distributions for each of these measures of contiguity are shown in Fig. 8, in which the N50 values for each measure can be seen as the value at which the cumulative distributions cross 50%. We have also estimated the size of each chromosome, by estimating the gap sizes (see below) and the extent of missing heterochromatic sequence^{93, 94, 105, 106, 107, 108} (Table 8). This is undoubtedly an oversimplification and does not adequately take into account the sequence status of each chromosome. Nonetheless, it provides a useful way to relate the draft sequence to the chromosomes.

Figure 8: Cumulative distributions of several measures of clone level contiguity and sequence contiguity.

The figures represent the proportion of the draft genome sequence contained in contigs of at most the indicated size. a, Clone level contiguity. The clones have a tight size distribution with an N50 of  160 kb (corresponding to 50% on the cumulative distribution). Sequenced-clone contigs represent the next level of continuity, and are linked by mRNA sequences or pairs of BAC end sequences to yield the sequenced-clone-contig scaffolds. The underlying contiguity of the layout of sequenced clones against the fingerprinted clone contigs is only partially shown at this scale. b, Sequence contiguity. The input fragments have low continuity (N50 = 21.7 kb). After merging, the sequence contigs grow to an N50 length of about 82 kb. After linking, sequence-contig scaffolds with an N50 length of about 274 kb are created.

High resolution image and legend (61K)

Table 8: Chromosome size estimates

Full table

Quality assessment

The draft genome sequence already covers the vast majority of the genome, but it remains an incomplete, intermediate product that is regularly updated as we work towards a complete finished sequence. The current version contains many gaps and errors. We therefore sought to evaluate the quality of various aspects of the current draft genome sequence, including the sequenced clones themselves, their assignment to a position in the fingerprint clone contigs, and the assembly of initial sequence contigs from the individual clones into sequence-contig scaffolds.

Nucleotide accuracy is reflected in a PHRAP score assigned to each base in the draft genome sequence and available to users through the Genome Browsers (see below) and public database entries. A summary of these scores for the unfinished portion of the genome is shown in Table 9. About 91% of the unfinished draft genome sequence has an error rate of less than 1 per 10,000 bases (PHRAP score > 40), and about 96% has an error rate of less than 1 in 1,000 bases (PHRAP > 30). These values are based only on the quality scores for the bases in the sequenced clones; they do not reflect additional confidence in the sequences that are represented in overlapping clones. The finished portion of the draft genome sequence has an error rate of less than 1 per 10,000 bases.

Table 9: Distribution of PHRAP scores in the draft genome sequence

Full table

Broad genomic landscape

What biological insights can be gleaned from the draft sequence? In this section, we consider very large-scale features of the draft genome sequence: the distribution of GC content, CpG islands and recombination rates, and the repeat content and gene content of the human genome. The draft genome sequence makes it possible to integrate these features and others at scales ranging from individual nucleotides to collections of chromosomes. Unless noted, all analyses were conducted on the assembled draft genome sequence described above.

Figure 9 provides a high-level view of the contents of the draft genome sequence, at a scale of about 3.8 Mb per centimetre. Of course, navigating information spanning nearly ten orders of magnitude requires computational tools to extract the full value. We have created and made freely available various 'Genome Browsers'. Browsers were developed and are maintained by the University of California at Santa Cruz (Fig. 10) and the EnsEMBL project of the European Bioinformatics Institute and the Sanger Centre (Fig. 11). Additional browsers have been created; URLs are listed at www.nhgri.nih.gov/genome_hub. These web-based computer tools allow users to view an annotated display of the draft genome sequence, with the ability to scroll along the chromosomes and zoom in or out to different scales. They include: the nucleotide sequence, sequence contigs, clone contigs, sequence coverage and finishing status, local GC content, CpG islands, known STS markers from previous genetic and physical maps, families of repeat sequences, known genes, ESTs and mRNAs, predicted genes, SNPs and sequence similarities with other organisms (currently the pufferfish Tetraodon nigroviridis). These browsers will be updated as the draft genome sequence is refined and corrected as additional annotations are developed.

Figure 9: Overview of features of draft human genome.

Please note that this figure is too large to display in image form. Instead it has been split into four PDFs. PDF 1 (3265K) shows chromosomes 1 - 3 and 20 - 22, PDF 2 (3049K) shows chromosomes 4 - 6 and 17 - 19, PDF 3 (2287K) shows chromosomes 7 - 9 and 20 - 22 and PDF 4 (2737K) shows chromosomes 10 - 11, X, Y, and 12 - 13.

The Figure shows the occurrences of twelve important types of feature across the human genome. Large grey blocks represent centromeres and centromeric heterochromatin (size not precisely to scale). Each of the feature types is depicted in a track, from top to bottom as follows. (1) Chromosome position in Mb. (2) The approximate positions of Giemsa-stained chromosome bands at the 800 band resolution. (3) Level of coverage in the draft genome sequence. Red, areas covered by finished clones; yellow, areas covered by predraft sequence. Regions covered by draft sequenced clones are in orange, with darker shades reflecting increasing shotgun sequence coverage. (4) GC content. Percentage of bases in a 20,000 base window that are C or G. (5) Repeat density. Red line, density of SINE class repeats in a 100,000-base window; blue line, density of LINE class repeats in a 100,000-base window. (6) Density of SNPs in a 50,000-base window. The SNPs were detected by sequencing and alignments of random genomic reads. Some of the heterogeneity in SNP density reflects the methods used for SNP discovery. Rigorous analysis of SNP density requires comparing the number of SNPs identified to the precise number of bases surveyed. (7) Non-coding RNA genes. Brown, functional RNA genes such as tRNAs, snoRNAs and rRNAs; light orange, RNA pseudogenes. (8) CpG islands. Green ticks represent regions of  200 bases with CpG levels significantly higher than in the genome as a whole, and GC ratios of at least 50%. (9) Exofish ecores. Regions of homology with the pufferfish T. nigroviridis ²⁹² are blue. (10) ESTs with at least one intron when aligned against genomic DNA are shown as black tick marks. (11) The starts of genes predicted by Genie or Ensembl are shown as red ticks. The starts of known genes from the RefSeq database¹¹⁰ are shown in blue. (12) The names of genes that have been uniquely located in the draft genome sequence, characterized and named by the HGM Nomenclature Committee. Known disease genes from the OMIM database are red, other genes blue. This Figure is based on an earlier version of the draft genome sequence than analysed in the text, owing to production constraints. We are aware of various errors in the Figure, including omissions of some known genes and misplacements of others. Some genes are mapped to more than one location, owing to errors in assembly, close paralogues or pseudogenes. Manual review was performed to select the most likely location in these cases and to correct other regions. For updated information, see http://genome.ucsc.edu/ and http://www.ensembl.org/.

High resolution image and legend (7K)

Figure 10: Screen shot from UCSC Draft Human Genome Browser.

See http://genome.ucsc.edu/.

High resolution image and legend (52K)

Figure 11: Screen shot from the Genome Browser of Project Ensembl.

See http://www.ensembl.org/.

High resolution image and legend (56K)

In addition to using the Genome Browsers, one can download from these sites the entire draft genome sequence together with the annotations in a computer-readable format. The sequences of the underlying sequenced clones are all available through the public sequence databases. URLs for these and other genome websites are listed in Box 2. A larger list of useful URLs can be found at http://www.nhgri.nih.gov/genome_hub. An introduction to using the draft genome sequence, as well as associated databases and analytical tools, is provided in an accompanying paper¹¹¹.

In addition, the human cytogenetic map has been integrated with the draft genome sequence as part of a related project. The BAC Resource Consortium ¹⁰³ established dense connections between the maps using more than 7,500 sequenced large-insert clones that had been cytogenetically mapped by FISH; the average density of the map is 2.3 clones per Mb. Although the precision of the integration is limited by the resolution of FISH, the links provide a powerful tool for the analysis of cytogenetic aberrations in inherited diseases and cancer. These cytogenetic links can also be accessed through the Genome Browsers.

Long-range variation in GC content

The existence of GC-rich and GC-poor regions in the human genome was first revealed by experimental studies involving density gradient separation, which indicated substantial variation in average GC content among large fragments. Subsequent studies have indicated that these GC-rich and GC-poor regions may have different biological properties, such as gene density, composition of repeat sequences, correspondence with cytogenetic bands and recombination rate^{112, 113, 114, 115, 116, 117}. Many of these studies were indirect, owing to the lack of sufficient sequence data.

The draft genome sequence makes it possible to explore the variation in GC content in a direct and global manner. Visual inspection (Fig. 9) confirms that local GC content undergoes substantial long-range excursions from its genome-wide average of 41%. If the genome were drawn from a uniform distribution of GC content, the local GC content in a window of size n bp should be 41  ((41)(59)/n)%. Fluctuations would be modest, with the standard deviation being halved as the window size is quadrupled—for example, 0.70%, 0.35%, 0.17% and 0.09% for windows of size 5, 20, 80 and 320 kb.

The draft genome sequence, however, contains many regions with much more extreme variation. There are huge regions (> 10 Mb) with GC content far from the average. For example, the most distal 48 Mb of chromosome 1p (from the telomere to about STS marker D1S3279) has an average GC content of 47.1%, and chromosome 13 has a 40-Mb region (roughly between STS marker A005X38 and stsG30423) with only 36% GC content. There are also examples of large shifts in GC content between adjacent multimegabase regions. For example, the average GC content on chromosome 17q is 50% for the distal 10.3 Mb but drops to 38% for the adjacent 3.9 Mb. There are regions of less than 300 kb with even wider swings in GC content, for example, from 33.1% to 59.3%.

Long-range variation in GC content is evident not just from extreme outliers, but throughout the genome. The distribution of average GC content in 20-kb windows across the draft genome sequence is shown in Fig. 12. The spread is 15-fold larger than predicted by a uniform process. Moreover, the standard deviation barely decreases as window size increases by successive factors of four—5.9%, 5.2%, 4.9% and 4.6% for windows of size 5, 20, 80 and 320 kb. The distribution is also notably skewed, with 58% below the average and 42% above the average of 41%, with a long tail of GC-rich regions.

Figure 12: Histogram of GC content of 20-kb windows in the draft genome sequence.

High resolution image and legend (36K)

Bernardi and colleagues^{118, 119} proposed that the long-range variation in GC content may reflect that the genome is composed of a mosaic of compositionally homogeneous regions that they dubbed 'isochores'. They suggested that the skewed distribution is composed of five normal distributions, corresponding to five distinct types of isochore (L1, L2, H1, H2 and H3, with GC contents of < 38%, 38–42%, 42–47%, 47–52% and > 52%, respectively).

We studied the draft genome sequence to see whether strict isochores could be identified. For example, the sequence was divided into 300-kb windows, and each window was subdivided into 20-kb subwindows. We calculated the average GC content for each window and subwindow, and investigated how much of the variance in the GC content of subwindows across the genome can be statistically 'explained' by the average GC content in each window. About three-quarters of the genome-wide variance among 20-kb windows can be statistically explained by the average GC content of 300-kb windows that contain them, but the residual variance among subwindows (standard deviation, 2.4%) is still far too large to be consistent with a homogeneous distribution. In fact, the hypothesis of homogeneity could be rejected for each 300-kb window in the draft genome sequence.

Similar results were obtained with other window and subwindow sizes. Some of the local heterogeneity in GC content is attributable to transposable element insertions (see below). Such repeat elements typically have a higher GC content than the surrounding sequence, with the effect being strongest for the most recent insertions.

These results rule out a strict notion of isochores as compositionally homogeneous. Instead, there is substantial variation at many different scales, as illustrated in Fig. 13. Although isochores do not appear to merit the prefix 'iso', the genome clearly does contain large regions of distinctive GC content and it is likely to be worth redefining the concept so that it becomes possible rigorously to partition the genome into regions. In the absence of a precise definition, we will loosely refer to such regions as 'GC content domains' in the context of the discussion below.

Figure 13: Variation in GC content at various scales.

The GC content in subregions of a 100-Mb region of chromosome 1 is plotted, starting at about 83 Mb from the beginning of the draft genome sequence. This region is AT-rich overall. Top, the GC content of the entire 100-Mb region analysed in non-overlapping 20-kb windows. Middle, GC content of the first 10 Mb, analysed in 2-kb windows. Bottom, GC content of the first 1 Mb, analysed in 200-bp windows. At this scale, gaps in the sequence can be seen.

High resolution image and legend (42K)

Fickett et al.¹²⁰ have explored a model in which the underlying preference for a particular GC content drifts continuously throughout the genome, an approach that bears further examination. Churchill¹²¹ has proposed that the boundaries between GC content domains can in some cases be predicted by a hidden Markov model, with one state representing a GC-rich region and one representing an AT-rich region. We found that this approach tended to identify only very short domains of less than a kilobase (data not shown), but variants of this approach deserve further attention.

The correlation between GC content domains and various biological properties is of great interest, and this is likely to be the most fruitful route to understanding the basis of variation in GC content. As described below, we confirm the existence of strong correlations with both repeat content and gene density. Using the integration between the draft genome sequence and the cytogenetic map described above, it is possible to confirm a statistically significant correlation between GC content and Giemsa bands (G-bands). For example, 98% of large-insert clones mapping to the darkest G-bands are in 200-kb regions of low GC content (average 37%), whereas more than 80% of clones mapping to the lightest G-bands are in regions of high GC content (average 45%)¹⁰³. Estimated band locations can be seen in Fig. 9 and viewed in the context of other genome annotation at http://genome.ucsc.edu/goldenPath/mapPlots/ and http://genome.ucsc.edu/goldenPath/hgTracks.html.

CpG islands

A related topic is the distribution of so-called CpG islands across the genome. The dinucleotide CpG is notable because it is greatly under-represented in human DNA, occurring at only about one-fifth of the roughly 4% frequency that would be expected by simply multiplying the typical fraction of Cs and Gs (0.21  0.21). The deficit occurs because most CpG dinucleotides are methylated on the cytosine base, and spontaneous deamination of methyl-C residues gives rise to T residues. (Spontaneous deamination of ordinary cytosine residues gives rise to uracil residues that are readily recognized and repaired by the cell.) As a result, methyl-CpG dinucleotides steadily mutate to TpG dinucleotides. However, the genome contains many 'CpG islands' in which CpG dinucleotides are not methylated and occur at a frequency closer to that predicted by the local GC content. CpG islands are of particular interest because many are associated with the 5' ends of genes^{122, 123, 124, 125, 126, 127}.

We searched the draft genome sequence for CpG islands. Ideally, they should be defined by directly testing for the absence of cytosine methylation, but that was not practical for this report. There are various computer programs that attempt to identify CpG islands on the basis of primary sequence alone. These programs differ in some important respects (such as how aggressively they subdivide long CpG-containing regions), and the precise correspondence with experimentally undermethylated islands has not been validated. Nevertheless, there is a good correlation, and computational analysis thus provides a reasonable picture of the distribution of CpG islands in the genome.

To identify CpG islands, we used the definition proposed by Gardiner-Garden and Frommer¹²⁸ and embodied in a computer program. We searched the draft genome sequence for CpG islands, using both the full sequence and the sequence masked to eliminate repeat sequences. The number of regions satisfying the definition of a CpG island was 50,267 in the full sequence and 28,890 in the repeat-masked sequence. The difference reflects the fact that some repeat elements (notably Alu) are GC-rich. Although some of these repeat elements may function as control regions, it seems unlikely that most of the apparent CpG islands in repeat sequences are functional. Accordingly, we focused on those in the non-repeated sequence. The count of 28,890 CpG islands is reasonably close to the previous estimate of about 35,000 (ref. 129, as modified by ref. 130). Most of the islands are short, with 60–70% GC content (Table 10). More than 95% of the islands are less than 1,800 bp long, and more than 75% are less than 850 bp. The longest CpG island (on chromosome 10) is 36,619 bp long, and 322 are longer than 3,000 bp. Some of the larger islands contain ribosomal pseudogenes, although RNA genes and pseudogenes account for only a small proportion of all islands (< 0.5%). The smaller islands are consistent with their previously hypothesized function, but the role of these larger islands is uncertain.

Table 10: Number of CpG islands by GC content

Full table

The density of CpG islands varies substantially among some of the chromosomes. Most chromosomes have 5–15 islands per Mb, with a mean of 10.5 islands per Mb. However, chromosome Y has an unusually low 2.9 islands per Mb, and chromosomes 16, 17 and 22 have 19–22 islands per Mb. The extreme outlier is chromosome 19, with 43 islands per Mb. Similar trends are seen when considering the percentage of bases contained in CpG islands. The relative density of CpG islands correlates reasonably well with estimates of relative gene density on these chromosomes, based both on previous mapping studies involving ESTs (Fig. 14) and on the distribution of gene predictions discussed below.

Figure 14: Number of CpG islands per Mb for each chromosome, plotted against the number of genes per Mb (the number of genes was taken from GeneMap98 (ref.100)).

Chromosomes 16, 17, 22 and particularly 19 are clear outliers, with a density of CpG islands that is even greater than would be expected from the high gene counts for these four chromosomes.

High resolution image and legend (20K)

Comparison of genetic and physical distance

The draft genome sequence makes it possible to compare genetic and physical distances and thereby to explore variation in the rate of recombination across the human chromosomes. We focus here on large-scale variation. Finer variation is examined in an accompanying paper¹³¹.

The genetic and physical maps are integrated by 5,282 polymorphic loci from the Marshfield genetic map¹⁰², whose positions are known in terms of centimorgans (cM) and Mb along the chromosomes. Figure 15 shows the comparison of the draft genome sequence for chromosome 12 with the male, female and sex-averaged maps. One can calculate the approximate ratio of cM per Mb across a chromosome (reflected in the slopes in Fig. 15) and the average recombination rate for each chromosome arm.

Figure 15: Distance in cM along the genetic map of chromosome 12 plotted against position in Mb in the draft genome sequence.

Female, male and sex-averaged maps are shown. Female recombination rates are much higher than male recombination rates. The increased slopes at either end of the chromosome reflect the increased rates of recombination per Mb near the telomeres. Conversely, the flatter slope near the centromere shows decreased recombination there, especially in male meiosis. This is typical of the other chromosomes as well (see http://genome.ucsc.edu/goldenPath/mapPlots). Discordant markers may be map, marker placement or assembly errors.

High resolution image and legend (50K)

Two striking features emerge from analysis of these data. First, the average recombination rate increases as the length of the chromosome arm decreases (Fig. 16). Long chromosome arms have an average recombination rate of about 1 cM per Mb, whereas the shortest arms are in the range of 2 cM per Mb. A similar trend has been seen in the yeast genome^{132, 133}, despite the fact that the physical scale is nearly 200 times as small. Moreover, experimental studies have shown that lengthening or shortening yeast chromosomes results in a compensatory change in recombination rate¹³².

Figure 16: Rate of recombination averaged across the euchromatic portion of each chromosome arm plotted against the length of the chromosome arm in Mb.

For large chromosomes, the average recombination rates are very similar, but as chromosome arm length decreases, average recombination rates rise markedly.

High resolution image and legend (26K)

The second observation i