Article

History of the rat

The rat, hated and loved at once, is both scourge and servant to mankind. The "Devil's Lapdog" is the first sign in the Chinese zodiac and traditionally carries the Hindu god Ganesh⁶. Rats are a reservoir of pathogens, known to carry over 70 diseases. They are involved in the transmission of infectious diseases to man, including cholera, bubonic plague, typhus, leptospirosis, cowpox and hantavirus infections. The rat remains a major pest, contributing to famine with other rodents by eating around one-fifth of the world's food harvest.

Paradoxically, the rat's contribution to human health cannot be overestimated, from testing new drugs, to understanding essential nutrients, to increasing knowledge of the pathobiology of human disease. In many parts of the world the rat remains a source of meat.

The laboratory rat (R. norvegicus) originated in central Asia and its success at spreading throughout the world can be directly attributed to its relationship with humans⁷. J. Berkenhout, in his 1769 treatise Outline of the Natural History of Great Britain, mistakenly took it to be from Norway and used R. norvegicus Berkenhout in the first formal Linnaean description of the species. Whereas the black rat (Rattus rattus) was part of the European landscape from at least the third century AD and is the species associated with the spread of bubonic plague, R. norvegicus probably originated in northern China and migrated to Europe somewhere around the eighteenth century⁸. They may have entered Europe after an earthquake in 1727 by swimming the Volga river.

The rat in research

R. norvegicus was the first mammalian species to be domesticated for scientific research, with work dating to before 1828 (ref. 9). The first recorded breeding colony for rats was established in 1856 (ref. 9). Rat genetics had a surprisingly early start. The first studies by Crampe from 1877 to 1885 focused on the inheritance of coat colour¹⁰. Following the rediscovery of Mendel's laws at the turn of the century, Bateson used these concepts in 1903 to demonstrate that rat coat colour is a mendelian trait¹⁰. The first inbred rat strain, PA, was established by King in 1909, the same year that systematic inbreeding began for the mouse¹⁰. Despite this, the mouse became the dominant model for mammalian geneticists, while the rat became the model of choice for physiologists, nutritionists and other biomedical researchers. Nevertheless, there are over 234 inbred strains of R. norvegicus developed by selective breeding, which 'fixes' natural disease alleles in particular strains or colonies¹¹.

Over the past century, the role of the rat in medicine has transformed from carrier of contagious diseases to indispensable tool in experimental medicine and drug development. Current examples of use of the rat in human medical research include surgery¹², transplantation^{13, 14, 15}, cancer^{16, 17}, diabetes^{18, 19}, psychiatric disorders²⁰ including behavioural intervention²¹ and addiction²², neural regeneration^{23, 24}, wound^{25, 26} and bone healing²⁷, space motion sickness²⁸, and cardiovascular disease^{29, 30, 31}. In drug development, the rat is routinely employed both to demonstrate therapeutic efficacy^{15, 32, 33} and to assess toxicity of novel therapeutic compounds before human clinical trials^{34, 35, 36, 37}.

The Rat Genome Project

Over the past decade, investigators and funding agencies have participated in rat genomics to develop valuable resources. Before the launch of the Rat Genome Sequencing Project (RGSP), there was much debate about the overall value of the rat genome sequence and its contribution to the utility of the rat as a model organism. The debate was fuelled by the naive belief that the rat and mouse were so similar morphologically and evolutionarily that the rat sequence would be redundant. Nevertheless, an effort spearheaded by two NIH agencies (NHGRI and NHLBI) culminated in the formation of the RGSP Consortium (RGSPC).

The RGSP was to generate a draft sequence of the rat genome, and, unlike the comparable human and mouse projects, errors would not ultimately be corrected in a finished sequence³⁸. Consequently, the draft quality was critical. Although it was expected to have gaps and areas of inaccuracy, the overall sequence quality had to be high enough to support detailed analyses.

The BN rat was selected as a sequencing target by the research community. An inbred animal (BN/SsNHsd) was obtained by the Medical College of Wisconsin (MCW) from Harlan Sprague Dawley. Microsatellite studies indicated heterozygosity, so over 13 generations of additional inbreeding were performed at the MCW, resulting in BN/SsNHsd/Mcwi animals. Most of the sequence data were from two females, with a small amount of whole genome shotgun (WGS) and flow-sorted Y chromosome sequencing from a male. The Y chromosome is not included in the current assembly.

A network of centres generated data and resources, led by the Baylor College of Medicine Human Genome Sequencing Center (BCM-HGSC) and including Celera Genomics, the Genome Therapeutics Corporation, the British Columbia Cancer Agency Genome Sciences Centre, The Institute for Genomic Research, the University of Utah, the Medical College of Wisconsin, The Children's Hospital of Oakland Research Institute, and the Max Delbrück Center for Molecular Medicine, Berlin. After assembly of the genome at the BCM-HGSC, analysis was performed by an international team, representing over 20 groups in six countries and relying largely on gene and protein predictions produced by Ensembl.

Atlas and the 'combined' sequencing strategy

Despite progress in assembling draft sequences^{2, 3, 39, 40, 41, 42, 43, 44} the question of which method produces the highest-quality products is unresolved. A significant issue is the choice between logistically simpler WGS approaches versus more complex strategies employing bacterial artificial chromosome (BAC) clones^{45, 46, 47, 48}. In the Public Human Genome Project² a BAC by BAC hierarchical approach was used and provided advantages in assembling difficult parts of the genome. The draft mouse sequence was a pure WGS approach using the ARACHNE assembler^{3, 49, 50} but underrepresented duplicated regions owing to 'collapses' in the assembly^{3, 51, 52, 53}. This limitation of the mouse draft sequence was tolerable owing to the planned full use of BAC clones in constructing the final finished sequence.

The RGSPC opted to develop a 'combined' approach using both WGS and BAC sequencing (Fig. 1). In the combined approach, WGS data are progressively melded with light sequence coverage of individual BACs (BAC skims) to yield intermediate products called 'enriched BACs' (eBACs). eBACs covering the whole genome are then joined into longer structures (bactigs). Bactigs are joined to form larger structures: superbactigs, then ultrabactigs. During this process other data are introduced, including BAC end sequences, DNA fingerprints and other long-range information (genetic markers, syntenic information), but the process is constrained by eBAC structures.

Figure 1: The new 'combined' sequence strategy and Atlas software.

a, Formation of 'eBACs'. The RGSP strategy combined the advantages of both BAC and WGS sequence data⁵⁴. Modest sequence coverage (1.8-fold) from a BAC is used as 'bait' to 'catch' WGS reads from the same region of the genome. These reads, and their mate pairs, are assembled using Phrap to form an eBAC. This stringent local assembly retains 95% of the 'catch'. b, Creation of higher-order structures. Multiple eBACs are assembled into bactigs based on sequence overlaps. The bactigs are joined into superbactigs by large clone mate-pair information (at least two links), extended into ultrabactigs using additional information (single links, FPC contigs, synteny, markers), and ultimately aligned to genome mapping data (radiation hybrid and physical maps) to form the complete assembly.

High resolution image and legend (80K)

To execute the combined strategy we developed the Atlas software package⁵⁴ (Fig. 1). The Atlas suite includes a 'BAC-Fisher' component that performs the functions needed to generate eBACs. WGS genome coverage was generated ahead of complete BAC coverage, so a BAC-Fisher web server was established at the BCM-HGSC to enable users to access the combined BAC and WGS reads as each BAC was processed (see Methods for data access). Each eBAC is assembled with high stringency to represent the local sequence accurately, and so provide a valuable intermediate product that assists all users of the genome data. Additional Atlas modules joined eBACs and linked bactigs to give the complete assembly (Fig. 1). Overall, the combined approach takes advantage of the strengths of both previous methods, with few of the disadvantages.

Sequence and genome data

Over 44 million DNA sequence reads were generated (Table 1; Methods). Following removal of low-quality reads and vector contaminants, 36 million reads were used for Atlas assembly, which retained 34 million reads. This was 7 sequence coverage with 60% provided by WGS and 40% from BACs. Slightly different estimates came from considering the entire 'trimmed' length of the sequence data (7.3), or only the portion of Phred20 quality or higher (6.9).

Table 1: Clones and reads used in the RGSP

Full table

The sequence data were end-reads from clones either derived directly from the genome (insert sizes of <10 kb, 10 kb, 50 kb and >150 kb) or from small insert plasmids subcloned from BACs. Overall, these provided 42-fold clone coverage, with 32-fold coverage having both paired ends represented. Approximately equal contributions of clone coverage were from the different categories.

Over 21,000 BACs were used for BAC skims (1.6 coverage) with an average sequence depth of 1.8 , giving an overall 2.8 genomic sequence coverage from BACs. This was slightly more than the most efficient procedure would require (1.2 each), because the genome size was not known at the project start.

Simultaneous with sequencing, 199,782 clones from the CHORI-230 BAC library⁵⁵ were fingerprinted by restriction enzyme digestion, representing 12-fold genomic coverage⁵⁶ (Methods). These were assembled into a 'fingerprint contig (FPC)' map (a contig is a set of overlapping segments of DNA) containing 11,274 FPCs. BAC selection for sequence skimming was based on overlaps between BACs using FPC mapping⁵⁶ (M.K. and C.F., unpublished work), ongoing BAC end sequencing (S.Z., unpublished work), and BAC sequence skimming⁵⁷. This strategy led to the sequence of a tiling path of BAC clones, covering the whole genome. In addition to the FPC map, a yeast artificial chromosome (YAC)-based physical map was constructed. 5,803 BAC and P1-derived artificial chromosome (PAC) clones from RPCI-32 and RPCI-31 libraries⁵⁵, respectively, were anchored to 51,323 YAC clones originating from two tenfold-coverage YAC libraries^{58, 225} assembled into 605 contigs⁵⁶. This map was subsequently integrated with the FPC map and the sequence assembly, reducing the total number of map contigs to 376 (minimum length of contig containing the 'typical' nucleotide, N₅₀ = 172 clones, 4.4 Mb; 358 anchored to the sequence assembly; Supplementary Information).

The combined strategy enabled development of resources such as the FPC map, BAC end sequences, and BAC skim sequences in parallel, rather than sequentially. In addition to allowing ongoing quality checking, this permitted the data-gathering phase of the project to be completed in less than two years.

Atlas assembly

Statistics for the Rnor3.1 assembly are in Table 2. Contigs within eBACs were ordered and oriented using read-pair information. Read-pair information was also used to add WGS reads to eBACs, even when sequence overlaps could not be reliably detected owing to repeated sequences. BAC skim reads with repeats were included in the assembly of eBACs because they clearly originated within BAC insert sequences. Over 19,000 eBACs were eventually generated.

Table 2: Statistics of the RGSP draft sequence assembly

Full table

More than 98% of eBACs were successfully merged to form bactigs (Fig. 1). Bactigs were subsequently reassembled to process all reads from overlapping BACs simultaneously, and then ordered and oriented with respect to each other using FPC map and BAC end sequence read-pair information. These superbactig and ultrabactig structures (see below) were aligned with chromosomes using external information, such as positions of genetic markers. Ultrabactigs represented the largest sequence units used to build chromosomes.

The current release of the rat genome assembly, version Rnor3.1, was generated using the data in Table 1. Earlier releases (Rnor2.0/2.1, Methods) were used for a substantial part of the annotation and analysis of genes and proteins, whereas the current release provided the genome description. Rnor3.1 has 128,000 contigs, with N₅₀ length 38 kb—larger than the expected genomic extent of a mammalian gene. These sequence contigs were linked into 783 superbactigs that were anchored to the radiation hybrid map⁵⁹. These larger units had N₅₀ length 5.4 Mb. Another 134 smaller superbactigs (N₅₀ length 1.2 Mb) could not be anchored, presumably because they fell into gaps between markers or because they were in repeated regions that could not be unambiguously placed. From placement on the radiation hybrid map, adjacent superbactigs were further linked to maximize continuity of sequence if appropriate read-pair mates existed or FPC suggested links. This reduced linked superbactigs to 419 pieces with 71 singletons. 291 ultrabactigs with N₅₀ length of nearly 19 Mb were placed on chromosomes. Orthology information with mouse and human sequences was also used to resolve conflicts and suggest placement of sequence units. Most of the 128 unplaced units were either singletons or small superbactigs that consisted of few clones. Thus, nearly the entire genome was represented in less than 300 large sequence units.

Quality assessment

Thirteen megabases of high-quality finished rat sequence from BACs were available for comparison with Rnor3.1 (Methods). This analysis showed that the majority of draft bases from within contigs were high quality (1.32 mismatches per 10 kb). This is essentially the accepted accuracy standard for finished sequence (1.0 errors per 10 kb)⁶⁰, so the overwhelming majority of contig bases are highly accurate. The highest frequency of mismatches occurred at the ends of contigs. We calculate the average size of these lower-accuracy regions to be 750 base pairs (bp) and they amount to less than 0.9% of the genome. These regions arise from misassembly of terminal reads due to repeated sequences.

Few mismatches were found within contigs. Six were found within contigs when compared with the 13 Mb of finished sequence, or one case per 2.2 Mb. All were insertions or deletions and may represent polymorphisms. Thus, at the fine structure level, the bulk of sequences that make up contigs is nearly the quality of finished sequence.

We judged accuracy of assembly at the chromosomal level by alignment with linkage maps⁶¹ and radiation hybrid map⁵⁹ (Fig. 2). Thirteen markers out of 3,824 from the SHRSP BN genetic map were placed on different chromosomes in the assembly and in the genetic map. Similarly, of the 20,490 sequence tagged sites placed on both the assembly and radiation hybrid (v3.4) map, 96.9% had consistent chromosome placement⁵⁹. Initial alignments identified regions of misassembly, and these were corrected, so that in Rnor3.1 the maps are congruent except for possible mismapped markers. The distribution of assembled sequence among the chromosomes and chromosome sizes in Rnor3.1 are in Supplementary Table SI-2.

Figure 2: Map correspondence.

Correspondence between positions of markers on two genetic maps of the rat (SHRSP BN intercross and FHH ACI intercross⁶¹), on the rat radiation hybrid map⁵⁹, and their position on the rat genome assembly (Rnor3.1).

High resolution image and legend (43K)

Genome size

Genomic assemblies are usually smaller than the actual genome size owing to under-representation of sequences affected by cloning bias, and sequencing and assembly difficulties. Simply equating the assembled genome size with the euchromatic, cloneable portion does not take into account heterochromatin that may be included⁶². We therefore estimated both an assembled genome size, scaled by the inverse of the fraction of features (genetic markers, expressed sequence tags (ESTs), and so on) found in the Rnor3.1 assembly, and a cloneable (or sampled) genome size, which was the part of the genome present in the WGS reads before assembly, as measured by analysing the distribution of short oligomers⁶³. The former may be an underestimate because non-repetitive, easily assembled regions can be enriched for known features. The latter should be an overestimate because there are likely to be regions (such as repeats) that can be cloned and sequenced, but not assembled.

For the rat genome, the assembled and cloneable genome sizes are very close. Considering the fraction of the marker set successfully mapped to Rnor3.1 (92%), or the fraction of sequence finished outside the BCM-HGSC (to reduce bias) present in Rnor3.1 (91%), together with the assembled bases in main scaffolds (2.533 Gb, Table 2), we suggest a genome size of 2.75 Gb. Alternatively, analysis of the WGS oligomers of length 24 to 32 predicted a genome size of between 2.76 and 2.81 billion bases. We have used the more conservative value of 2.75 Gb for the rat genome size, but this is still considerably higher (150 Mb) than the 2.6 Gb currently reported for the mouse draft genome sequence. A fraction of the size differences in these rodent genomes results from the different repeat content (see below); however, it is also recognized that segmental duplications may be under-represented in the mouse WGS draft sequence for technical reasons^{3, 51}.

Telomeres, centromeres and mitochondrial sequence

The rat has both metacentric and telocentric chromosomes, in contrast to the wholly telocentric mouse chromosomes. As expected from previous draft sequences, the rat draft does not contain complete telomeres or centromeres. Their physical location relative to the rat draft sequence can however be approximated; the centromeres of the telocentric rat chromosomes (2, 4–10 and X) must be positioned before nucleotide 1 of these assemblies, and those for the remaining chromosomes are estimated as indicated in Fig. 3. Several of these putative centromere positions coincide with both segmental duplication blocks (see below) and classical satellite clusters, consistent with enrichment of both of these sequence features in rat pericentromeric DNA. Human subtelomere regions are characterized by both an abundance of segmentally duplicated DNA and an enrichment of internal (TTAGGG)_n-like sequence islands⁶⁴. Approximately one-third of the euchromatic rat subtelomeric regions are similarly enriched, suggesting that Rnor3.1 might extend very close to the chromosome ends.

Figure 3: Distribution of segmental duplications in the rat genome.

Interchromosomal duplications (red) and intrachromosomal duplications (blue) are depicted for all duplications with 90% sequence identity and 20 kb length. The intrachromosomal duplications are drawn with connecting blue line segments; those with no apparent connectors are local duplications (spaced below the figure resolution limit). p arms are on the left and the q arms on the right. Chromosomes 2, 4–10, and X are telocentric; the assemblies begin with pericentric sequences of the q arms, and no centromeres are indicated. For the remaining chromosomes, the approximate centromere positions were estimated from the most proximal STS/gene marker to the p and q arm as determined by fluorescent in situ hybridization (FISH) (cyan vertical lines; no chromosome 3 data). The 'Chr Un' sequence consists of contigs not incorporated into any chromosomes. Green arrows indicate 1 Mb intervals with more than tenfold enrichment of classic rat satellite repeats within the assembly. Orange diamonds indicate 1 Mb intervals with more than tenfold enrichment of internal (TTAGGG)_n-like sequences. For more detail see http://ratparalogy.cwru.edu.

High resolution image and legend (112K)

Fragments of the rat mitochondrial genome were also propagated within the WGS libraries and subsequently sequenced, allowing the assembly of the complete 16,313 bp mitochondrial genome (Supplementary Information). Comparison with existing mitochondrial sequences in the public databases revealed variable positions totalling 95 bp (0.6%) between this strain and the wild brown rat. Considerably more variation (2.2%) was found when compared with the Wistar strain: 357 bp differences over the whole genome, including 78 positions that are conserved in the other mammalian sequences. Such variation has also been reported in mouse mitochondrial sequences and attributed to errors in previously sequenced genomes⁶⁵. The current sequence is very accurate, and we therefore favour the BN sequence as a reference for the rat mitochondrial genome.

Orthologous chromosomal segments and large-scale rearrangements

Multi-megabase segments of the chromosomes of the primate–rodent ancestor have been passed on to human and murid rodent descendants with minimal rearrangements of gene order^{66, 67, 68}. These intact regions, which are bounded by the breaks that occurred during ancient large-scale chromosomal rearrangements, are referred to as orthologous chromosomal segments. The same phenomenon has occurred in the descent of the rat and mouse from the genome of their common murid ancestor, and we were able to use the human genome, and in some cases other outgroup data, to tentatively reconstruct the sequence of many of these rearrangements in these lineages. To visualize the extent of orthologous chromosomal segments, each genome was 'painted' with the orthologous segments of the other two species (Fig. 4) using the Virtual Genome Painting method (M.L.G.-G. et al., unpublished work; http://www.genboree.org). Inspection shows the interleaving of events that both preceded and occurred subsequently to the rat–mouse divergence.

Figure 4: Map of conserved synteny between the human, mouse and rat genomes.

For each species, each chromosome (x axis) is a two-column boxed pane (p arm at the bottom) coloured according to conserved synteny to chromosomes of the other two species. The same chromosome colour code is used for all species (indicated below). For example, the first 30 Mb of mouse chromosome 15 is shown to be similar to part of human chromosome 5 (by the red in left column) and part of rat chromosome 2 (by the olive in right column). An interactive version is accessible (http://www.genboree.org).

High resolution image and legend (94K)

Comparing the three species at 1 Mb resolution, BLASTZ⁶⁹, PatternHunter/Grimm-Synteny^{70, 71}, Pash⁷², and associated merging algorithms^{66, 72, 73} produce virtually indistinguishable sets of orthologous chromosomal segments. PatternHunter and the GRIMM-Synteny algorithm⁷³ detect 278 orthologous segments between human and rat, and 280 between human and mouse. The mouse–rat comparison reveals a smaller number of segments (105) of larger average size. The larger number of breaks in orthologous segments between the human to the rodent pair is expected, because of the latter's closer evolutionary relationship.

Understanding the number and timing of rearrangement events that have occurred in each of the three individual lineages (see tree in Fig. 5a) since the common primate–rodent ancestor required a more detailed analysis. We initially focused on the X chromosome, because rearrangements between the X and the autosomes are rare⁷⁴ and its history is somewhat easier to trace completely. The X chromosome consists of 16 human–mouse–rat orthologous segments of at least 300 kb in size⁷³ (Fig. 6a). In the most parsimonious scenario (found with MGR and GRIMM⁷⁵), these were created by 15 inversions in the descent from the primate–rodent ancestor (Fig. 6b). Outgroup data from cat, cow⁷⁶ and dog⁷⁷ resolved the timing of these rearrangements more precisely. Most of these events occurred in the rodent lineage: five (or four) before the divergence of rat and mouse, five in the rat lineage, and five in the mouse lineage. At most one rearrangement occurred in the human lineage since divergence from the common ancestor with rodents. The timing of this one event was ambiguous, owing to the limited resolution of the outgroup data. Even given this uncertainty, it is clear that the large-scale architecture of the X chromosome in humans is largely unchanged since the primate–rodent ancestor⁷³, whereas there has been considerable activity in the rodents. The assignment of the accelerated activity to the rodent branch, following the primate–rodent divergence, is consistent with previous studies at significantly lower resolution (these showed complete conservation of marker order between the X chromosomes of human and cat⁷⁸, human and dog⁷⁷, and human and lemur⁷⁹, as well as similar karyotypes of the X chromosomes in human, chimpanzees, gorillas and orangutans⁸⁰).

Figure 5: Substitutions and microindels (1–10 bp) in the evolution of the human, mouse and rat genomes.

a, The lengths of the labelled branches in the tree are proportional to the number of substitutions per site inferred using the REV model²²² from all sites with aligned bases in all three genomes. b, The table shows the midpoint and variation in these branch-length estimates when estimated from different sequence alignment programs and different neutral sites, including sites from ancestral repeats³, fourfold degenerate sites in codons, and rodent-specific sites ('in neutral sites only' row; Supplementary Information). Other rows give midpoints and variation for micro-indels on each branch of the tree in a.

High resolution image and legend (45K)

Figure 6: X chromosome in each pair of species.

a, GRIMM-Synteny⁷¹ computes 16 three-way orthologous segments ( 300 kb) on the X chromosome of human, mouse and rat, shown for each pair of species, using consistent colours. b, The arrangement (order and orientation) of the 16 blocks implies that at least 15 rearrangement events occurred during X chromosome evolution of these species. The program MGR (http://www.cs.ucsd.edu/groups/bioinformatics/MGR/) determined that evolutionary scenarios with 15 events are achievable and all have the same median ancestor (located at the last common mouse–rat ancestor). Shown is a possible (not unique) most parsimonious inversion scenario from each species to that ancestor. We note that the last common ancestor of human, mouse and rat should be on the evolutionary path between this median ancestor and human.

High resolution image and legend (114K)

Large-scale reconstruction of the entire ancestral murid genome suggests that it retained many previously postulated chromosome associations of the placental ancestor^{81, 82}. The most parsimonious scenario we found requires a total of 353 rearrangements: 247 between the murid ancestor and human, 50 from the murid ancestor to mouse and 56 from the murid ancestor to rat. A recent study⁸² implies that most of the 247 rearrangements between the murid ancestor and human occurred on the evolutionary subpath from the squirrel–mouse–rat ancestor to the murid ancestor. Our analyses confirm that the rate of rearrangements in murid rodents is much higher than in the human lineage⁷³.

Segmental duplications

Segmental duplications are defined here as regions of the genome that are repeated over at least 5 kb of length and >90% identity. The rat has approximately 2.9% of its bases in these duplicated regions (Fig. 3), whereas the human genome has 5–6%⁸³. In contrast to the greater rate of large-scale rearrangement, the mouse genome shows substantially fewer of these events³, with only 1.0–2.0%⁵¹ of its sequenced bases in duplicated regions. These duplicated structures are particularly challenging to assemble, and we attribute at least some of the mouse–rat differences to the BAC-based approach we used for Rnor3.1, compared with the WGS mouse approach. The vast majority of these sequences (73 of 82 Mb) were regions with <99.5% identity and thus were not simply overlapping sequences that had not been joined by the assembly program Phrap. The 'unplaced' chromosome in Rnor3.1 showed a marked enrichment for blocks of segmental duplication (nearly 44% of the total), which indicates problems with anchoring these elements to the genome.

Intrachromosomal duplications are represented at a three-to-one excess when compared with interchromosomal duplications, and are significantly enriched near the telomeres and in centromeric regions (Fig. 3). The pericentromeric accumulation of segmental duplications in the rat is reminiscent of that observed in human and mouse^{83, 84, 85, 86}, and seems to be a general property of mammalian chromosome architecture.

We observed considerable clustering of duplications⁸⁷, including 41 discrete genomic regions larger than 1 Mb in size in which duplications appear to be organized into groups with <100 kb between duplicated segments. For many of these clusters, the underlying sequence alignments showed a wide range in the degree of sequence identity, suggesting that these areas have been subject to duplication events more or less continuously over millions of years. In contrast, an analysis of the evolutionary distance between all duplicated regions showed an unusual bimodal distribution, particularly for intrachromosomal segmental duplications. Two peaks were observed at 0.045 substitutions per site and 0.075 substitutions per site. Given that the rat genome has accumulated 8–10% substitutions (see below) since the speciation from mouse 12–24 Myr ago, this bimodal distribution may correspond to bursts of segmental duplication that occurred approximately 5 and 8 Myr ago, respectively.

The segmental duplications in the rat genome were of considerable interest because they represent an important mechanism for the generation of new genes. We found that 63 NCBI reference sequence⁸⁸ (RefSeq; see http://www.ncbi.nih.gov/RefSeq/) genes were located completely or partially within rat duplicated regions, out of a genome total of 4,532 rat RefSeq genes. As discussed below, many of these genes are present in multiple copies and belong to gene familes that have been recently duplicated and contribute to distinctive elements of rat biology.

Gains and losses of DNA

In addition to large rearrangements and segmental duplications, genome architecture is strongly influenced by insertion and deletion events that add and remove DNA over evolutionary time. To characterize the origins and losses of sequence elements in the human, mouse and rat genomes, we categorized all the nucleotides in each of the three genomes, using our alignment data and RepeatMasker annotations of the insertions of repetitive elements (Fig. 7). The rodent repeat database used by RepeatMasker was greatly expanded by analysing the rat and mouse genomes⁸⁹, but it is clear that not all repeats are being recognized, especially the older ones. Thus, these estimates of the amount of rodent repeats represent lower bounds.

Figure 7: Aligning portions and origins of sequences in rat, mouse and human genomes.

Each outlined ellipse is a genome, and the overlapping areas indicate the amount of sequence that aligns in all three species (rat, mouse and human) or in only two species. Non-overlapping regions represent sequence that does not align. Types of repeats classified by ancestry: those that predate the human–rodent divergence (grey), those that arose on the rodent lineage before the rat–mouse divergence (lavender), species-specific (orange for rat, green for mouse, blue for human) and simple (yellow), placed to illustrate the approximate amount of each type in each alignment category. Uncoloured areas are non-repetitive DNA—the bulk is assumed to be ancestral to the human–rodent divergence. Numbers of nucleotides (in Mb) are given for each sector (type of sequence and alignment category). Detailed results are tabulated (Supplementary Table SI-1).

High resolution image and legend (42K)

About a billion nucleotides (39% of the euchromatic rat genome) align in all three species, constituting an 'ancestral core' that is retained in these genomes. This ancestral core contains 94–95% of the known coding exons and regulatory regions. Comparisons between the human and mouse genomes, using transposon relics retained in both species ('mammalian ancestral repeats') to model neutral evolution, have been used to estimate the fraction of the human genome that is accumulating substitutions more slowly than the neutral rate in both lineages since their divergence, and hence may be under some level of purifying selection³. Depending on details of methodology, such estimates have ranged between about 4% and 7%^{3, 90, 91}. The levels of three-way conservation observed here between the human, mouse and rat genomes in the ancestral core lend further support to these earlier estimates, giving values in the range of 5–6% when measured by two quite different methods (see Methods and ref. 92). In this constrained fraction, non-coding regions outnumber coding regions regardless of the strength of constraint⁹², an observation that supports recent comparative analyses limited to subsets of the genome^{93, 94}. The preponderance of non-coding elements in the most constrained fraction of the genome underscores the likelihood that they play critical roles in mammalian biology.

About 700 Mb (28%) of the rat euchromatic genome aligns only with the mouse. At least 40% of this comprises of rodent-specific repeats inserted on the branch from the primate–rodent ancestor to the murid ancestor, and some of the remainder can be recognized as mammalian ancestral repeats whose orthologues were deleted in the human lineage (Fig. 7). Another part is likely to consist of single-copy ancestral DNA deleted in the human lineage but retained in rodents. Although this 700 Mb of rodent-specific DNA is primarily neutral, it may also contain some functional elements lost in the human lineage in addition to sequences representing gains of rodent-specific functions, including some coding exons⁹⁵.

The remainder of the euchromatic rat genome (726 Mb, 29%) aligns with neither mouse nor human (Fig. 7). At least half of this (15% of the rat genome) consists of rat-specific repeats, and another large fraction (8% of the rat genome) consists of rodent-specific repeats whose orthologues are deleted in the mouse.

Substitution rates

The alignment data allow relatively precise estimates of the rates of neutral substitutions and microindel events ( 10 bp). Both synonymous fourfold degenerate ('4D') sites in protein-coding regions and sites in mammalian ancestral repeats were used in this analysis, as in previous studies comparing human and mouse^{3, 96}. We additionally used a class of primarily neutral sites whose identification is made uniquely possible by the addition of the rat genome sequence: namely, the rodent-specific sites discussed above, identified by their failure to align to human sequence.

Our estimates for the neutral substitution level between the two rodents range from 0.15 to 0.20 substitutions per site, while estimates for the entire tree of human, mouse and rat range from 0.52 to 0.65 substitutions per site (Fig. 5). This difference was predictable because of the evolutionary closeness of the two rodents. For all classes of neutral sites analysed, however, the branch connecting the rat to the common rodent ancestor is 5–10% longer than the mouse branch (Fig. 5a). Thus, for as yet unknown reasons, the rat lineage has accumulated substantially more point substitutions than the mouse lineage since their last common ancestor.

We also analysed four-way alignments including sequence from orthologous ancestral repeats in human, mouse and rat, along with the repeat consensus sequences, which approximate the sequence of the progenitor of the corresponding repeat family (Methods). These alignments allow us to distinguish substitutions on the branch from the primate–rodent ancestor to the rodent ancestor from substitutions on the branch descending to human⁷⁷. This revealed an overall speed-up in rodent substitution rates relative to human of about three-to-one, larger than estimated previously³, but consistent with other more recent studies which also use multiple sequence alignments^{77, 97, 98}.

Estimates for rates of microdeletion events are, for all branches, approximately twofold higher than rates of microinsertion (Fig. 5b), suggesting a fundamental difference in the mechanisms that generate these mutations. Furthermore, there are substantial rate differences for each class of event between the various lineages. In particular, the rat lineage has accumulated microdeletions more rapidly than the mouse, while the opposite holds true for microinsertions. As with substitutions, both microinsertion and microdeletion rates are substantially slower in the human lineage. The size distribution of microindels (1–10 bp) on the rat branch was heavily weighted towards the smallest indels: 45% of indels are single bases, 18% are 2 bp, 10% are 3 bp, 8% are 4 bp, and so on, monotonically decreasing. Separate distributions for insertions and for deletions were similar, as were distributions of indel sizes on the mouse branch.

Male mutation bias

As mouse and rat are similar in generation time and number of germline cell divisions^{99, 100}, we investigated a potential sex bias in different types of observed genome changes. We compared substitution and indel rates between the X chromosome and autosomes in ancestral repeat sites (5 Mb and 100 Mb in total for X and autosomes, respectively¹⁰¹). We discovered that in rodents, small indels (<50 bp) are male-biased, with a male-to-female rate ratio of 2.3. This is in contrast to a recent study in primates, based on a substantially smaller data set, that indicates no sex bias in small indels¹⁰². Our male-to-female nucleotide substitution rate ratio in rodents is 1.9, confirming earlier reports^{103, 104}. When substitution rates are compared for all sites aligned between mouse and rat (78 Mb and 1,691 Mb, respectively), we again observe an approximately twofold excess of small indels and nucleotide substitutions originating in males compared with females¹⁰¹. Interestingly, the ratio in the number of cell divisions between the male and female germlines is also about two^{99, 100}, suggesting that these substitutions may arise from mutations that occur primarily during DNA replication.

G + C content and CpG islands

The G + C content of the rat varies significantly across the genome (Fig. 8a), and the distribution more closely resembles that of mouse than human. The variation in G + C content is coupled with differences in the distribution of CpG islands—short regions that are associated with the 5' ends of genes and gene regulation^{2, 3, 105}, and that escape the depletion of CpG dinucleotides that occurs from deamination of methylated cytosine^{2, 105}. The 2.6 Gb rat genome assembly (including unmapped sequences) contains 15,975 CpG islands in non-repetitive sequences of the genome. This is similar to the 15,500 CpG islands reported in the 2.5 Gb mouse genome³, but far fewer than the 27,000 reported in the human genome^{2, 3, 105}.

Figure 8: Base composition distribution analysis.

a, The fraction of 20 kb non-overlapping windows³ with a given G + C content is shown for human, mouse and rat. b, The number of Ensembl-predicted genes per chromosome and the number of CpG islands per chromosome. The density of CpG islands averages 5.9 islands per Mb across chromosomes and 5.7 islands per Mb across the genome. Chromosome 1 has more CpG islands than other chromosomes, yet neither the island density nor ratio to predicted genes exceeds the normal distribution. The number of CpG islands per chromosome and the number of predicted genes are correlated (R² = 0.96).

High resolution image and legend (51K)

A summary of the CpG island distributions by chromosome is given in Fig. 8b. Chromosome X, with a low G + C content of 37.7%, has the fewest islands (362) and the lowest density of islands (2.6 per Mb). Chromosome 12 is at the other end of the range with a G + C content of 43.5% and the highest density of CpG islands (11.5 islands per Mb). This is similar to chromosome 10, with 11.3 islands per Mb. The average density of CpG islands is 5.7 islands per Mb over the whole genome and 5.9 CpG islands per Mb averaged by chromosome, which is similar to the distribution in mouse³. Neither rodent genome shows the extreme outliers in CpG island density that are seen for human chromosome 19 (ref. 2). The density of CpG islands in the rat genome correlates positively with the density of predicted genes (R of 0.96) (Fig. 8b).

These data show that the overall changes in CpG island content predate the rat–mouse split and are consistent with the accelerated loss of CpG dinucleotides in rodents compared with humans^{105, 106}. It remains possible, however, that occurrences such as the greater number of human regions with extremely high G + C content are due to distributional changes mostly in the primate, rather than in the rodent lineage.

Shift in substitution spectra between mouse and rat

The non-repetitive fraction of the rat genome is enriched for G + C content relative to the mouse genome, by 0.35% over 1.3 billion nucleotides. This is a subtle but substantial difference that may be explained, at least in part, by differences in the spectra of mutation events that have accumulated in the mouse and rat lineages. We analysed all alignment columns in which substitution events can be assigned to either the mouse or the rat lineage, by virtue of a nucleotide match between human and only one rodent⁹²; note that this is a small minority of substitutions. Of the 117 million alignment columns meeting this criteria, 60 million involve a change in the rat lineage versus 57 million in the mouse, reflecting the increase in rates of point substitution in the rat lineage (Fig. 5b). While 50% of these changes in rat involve a substitution from an A/T to a G/C, these events constitute only 47% of all mouse changes. The complementary change, G/C to A/T, exhibits relative excess in the mouse versus the rat lineage (38% versus 35%, respectively). No substantial difference between changes that do not alter G + C content is observed. In addition, this bias is not confined to particular transition or transversion events, nor can it be explained simply as a result of divergent substitution rates of CpG dinucleotides (data not shown). Thus, this shift appears to be a general change that results in an increase in G + C content in the rat genome. Biochemical changes in repair or replication enzymes might be responsible, and the observation that recombination rates are slightly higher in rat than in mouse¹⁰⁷ may suggest a role for G + C-biased mismatch repair^{108, 109}. However, population genetic factors, such as selection, cannot be ruled out.

Evolutionary hotspots

Comparison of the two rodent genomes, using human as outgroup, reveals regions that are conserved yet under different levels of constraint in mouse and rat. These regions may have distinct functional roles and contribute to species-specific differences. Analysis of the MAVID alignments¹¹⁰ revealed 5,055 regions 100 bp, in which there was at least a tenfold difference in the estimated number of substitutions per site on the mouse and rat branches. To avoid alignment problems and fast-evolving regions, the analysis was restricted to regions where the human branch had <0.25 substitutions per site¹¹¹. These regions are enriched twofold in transcribed regions: 39% of mouse hotspots were found in the 18% of the mouse genome covered by RefSeq genes; and 17% of the rat hotspots were found in the 8% of the rat genome covered by RefSeq genes. Similar numbers are observed when examining coding exon and EST regions (not shown). Half of all hotspots in the mouse genome lie totally in non-coding regions. Many hotspots are several hundred bases long, with average length 190 86 bp. Future work aimed at identifying the genomic differences that contribute to phenotypic evolution may benefit from analyses such as these, which will become more powerful as the repertoire of mammalian genome sequences expands.

Covariation of evolutionary and genomic features

To illustrate the genomic and evolutionary landscape of a single rat chromosome in depth, we characterized features for rat chromosome 10 at 1 Mb resolution (Fig. 9). This high-resolution analysis uncovered strong correlations between certain microevolutionary features^{89, 92, 98}. Particularly strongly correlated are the local rates of microdeletion (R² = 0.71; Fig. 9a), microinsertion (R² = 0.56; Fig. 9a), and point substitution (R² = 0.86; Fig. 9b) between the two independent lineages of mouse and rat. In addition, microinsertion rates are correlated with microdeletion rates (R² = 0.55; Fig. 9a). These strong correlations are also observed in an independent genome-wide analysis, both on the original data and after factoring out the effects of G + C content (not shown, see Supplementary Information).

Figure 9: Variability of several evolutionary and genomic features along rat chromosome 10.

a, Rates of microdeletion and microinsertion events (less than 11 bp) in the mouse and rat lineages since their last common ancestor, revealing regional correlations. b, Rates of point substitution in the mouse and rat lineages. Red and green lines represent rates of substitution within each lineage estimated from sites common to human, mouse and rat. Blue represents the neutral distance separating the rodents, as estimated from rodent-specific sites. Note the regional correlation among all three plots, despite being estimated in different lineages (mouse and rat) and from different sites (mammalian versus rodent-specific). c, Density of SINEs inserted independently into the rat or mouse genomes after their last common ancestor. d, A + T content of the rat, and density in the rat genome of LINEs and SINEs that originated since the last common ancestor of human, mouse and rat. Pink boxes highlight regions of the chromosome in which substitution rates, A + T content and LINE density are correlated. Blue boxes highlight regions in which SINE density is high but LINE density is low.

High resolution image and legend (107K)

Perhaps surprisingly, substantially less correlation is seen between microindel and point substitution rates (compare Fig. 9a and b). The amount of correlation varies among chromosomes (not shown), but is generally weaker than the relationships mentioned above. Further studies will be required to determine whether local evolutionary pressures, which must have remained stable since the separation of the mouse and rat lineages, differentially drive microindel and point substitution rates.

We also find that the local point substitution rate in sites common to human, mouse and rat strongly correlates with that in rodent-specific sites (R² = 0.57; Fig. 9b, blue line versus red/green). These two classes of sites, while interdigitated at the level of tens to thousands of bases, constitute sites that are otherwise evolutionarily independent. This result confirms that local rate variation is not solely determined by stochastic effects and extends, at high resolution, the previously documented regional correlation in rate between 4D sites and ancestral repeat sites^{3, 96}.

Construction of gene set and determination of orthology

The Ensembl gene prediction pipeline¹¹² predicted 20,973 genes with 28,516 transcripts and 205,623 exons (Methods). These genes contain an average of 9.7 exons, with a median exon number of 6.0. At least 20% of the genes are alternatively spliced, with an average of 1.3 transcripts predicted per gene. Of the 17% single exon transcripts, 1,355 contain frameshifts relative to the predicted protein and 1,176 are probably processed pseudogenes. Of the 28,516 transcripts, 48% have both 5' and 3' untranslated regions (UTRs) predicted and 60% have at least one UTR predicted.

These gene predictions considered homology to other sequences, including 26,949 rodent proteins, 4,861 non-rodent, vertebrate proteins, 7,121 rat complementary DNAs from RefSeq and EMBL, and 31,545 mouse cDNAs from Riken, RefSeq and EMBL. The majority (61%) of transcripts are supported by rodent transcript evidence. When combined with additional private EST data, the fraction of genes supported by transcript evidence could be increased to 72%¹¹³.

A number of other ab initio (GENSCAN¹¹⁴, GENEID¹¹⁵), similarity-based (FGENESH + + ; ref. 116) and comparative (SGP¹¹⁷, SLAM¹¹⁸, TWINSCAN1^{119, 120, 121}) gene-prediction programs were used to analyse the rat genome. The number of genes predicted by these programs ranged from 24,500 to 47,000, suggesting coding densities ranging from 1.2% to 2.2%. The coding fraction of RefSeq genes covered by these predictions ranged from 82% to 98%. Such comparative ab initio programs using the rat genome were successfully used to identify and experimentally verify genes missed by other methods in rat¹²¹ and human¹²². The predictions of these programs can be accessed through the UCSC genome browser and Ensembl websites.

RefSeq genes (20,091 human, 11,342 mouse and 4,488 rat) mapped onto genome assemblies with BLAT¹²³ and the UCSC browser revealed that the number of coding exons per gene and average exon length were similar in the three species. Differences were observed in intron length, with an average of 5,338 bp in human, 4,212 bp in mouse and 5,002 bp in rat. These differences were also found in a smaller collection of 6,352 confidently mapped orthologous intron triads (see 'Conservation of intronic splice signals' section below): average intron lengths in this collection were 4,240 bp in human, 3,565 bp in mouse and 3,638 bp in rat.

Properties of orthologous genes

Orthology relationships were predicted on the basis of BLASTp reciprocal best-hits between proteins of genome pairs (human–rat, rat–mouse and mouse–human)³ (Supplementary Information). Using these methods and the ENSEMBL prediction sets, 12,440 rat genes showed clear, unambiguous 1:1 correspondence with a gene in the mouse genome. This is an underestimate, because random sampling of different classes of rat genes with less stringent criteria for comparison to mouse always identified additional gene pairs. Errors arose from pseudogene misclassification, sequence loss, duplication or fragmentation in assemblies; and missing or inappropriate gene predictions, including coding-gene predictions from non-coding RNAs. Taking these errors into account, we estimate the true proportion of 1:1 orthologues in rat and mouse genomes to lie between 86 and 94% (Methods). The remaining genes were associated with lineage-specific gene family expansions or contractions. These overall observations are consistent with a careful analysis of rat proteases showing that 93% of these genes have 1:1 orthologues in mouse^{124, 125}.

Surprisingly, a similar proportion (89 to 90%) of rat genes possessed a single orthologue in the human genome. Because human represents an outgroup to the two rodents, it was expected that mouse and rat would share a higher fraction of orthologues. A close inspection of gene relationships indicates that these findings may suffer from incompleteness of rodent genome sequences, together with problems of misassembly and gene prediction within clusters of gene paralogues.

Further analysis of orthologous pairs considered the occurrence of nucleotide changes within protein-coding regions that reflected synonymous or non-synonymous substitutions. The majority of these studies measured evolutionary rates by determination of K_A (number of non-synonymous substitutions per non-synonymous site) and K_S (number of synonymous substitutions per synonymous site). K_A/K_S ratios of less than 0.25 indicate purifying selection, values of 1 suggest neutral evolution, and values greater than 1 indicate positive selection¹²⁶.

Evolutionary rates were first calculated from a reduced set of orthologue pairs that are embedded in orthologous genomic segments and are related by conservative values of K_S (Table 3) (Methods). A slight increase in median K_S values for rat–human as compared with mouse–human, was found, indicating that the rat lineage has more neutral substitutions in gene coding regions than the mouse lineage. Sequence conservation values were similar to those previously found using smaller data sets^{127, 128}, and the overall trend is consistent with results of other evolutionary rate analyses discussed above (Fig. 5).

Table 3: One-to-one orthologous genes in human, mouse and rat genomes

Full table

Next, we investigated examples of rat genes shared with mouse, but with no counterparts in human. Such genes might be rapidly evolving so that homologues are not discernible in human, or they might have arisen from non-coding DNA, or their orthologues in the human lineage might have formed pseudogenes. Thirty-one Ensembl rat genes were collected that have no non-rodent homologues in current databases (Methods). These are twofold over-represented among genes in paralogous gene clusters, and threefold over-represented among genes whose proteins are likely to be secreted. This is consistent with observations³ that clusters of paralogous genes, and secreted proteins, evolve relatively rapidly. Detailed examination of the 31 genes using PSI-BLAST determined that ten genes cannot be assigned homology relationships to experimentally described mammalian genes. These ten rodent-specific genes may have evolved particularly rapidly, or have non-coding DNA homologues, or be erroneous predictions.

The paucity of rodent-specific genes indicates that de novo invention of complete genes in rodents is rare. This is not unexpected, because the majority of eukaryotic protein-coding genes are modular structures containing coding and non-coding exons, splicing signals and regulatory sequences, and the chances of independent evolution and successful assembly of these elements into a functional gene are small, given the relatively short evolutionary time available since the mouse–rat split. However, individual rodent-specific exons may arise more frequently, particularly if the exon is alternatively spliced¹²⁹. Applying a K_A/K_S ratio test^{130, 131} to sequences that align only between rat and mouse, we identified 2,302 potential novel rodent-specific exons, with EST support, in BLASTZ alignments of rat and mouse sequences. None of these individual exons matched human transcripts, but approximately half (1,116) appear to be present in alternative splice forms found in rodents. We speculate that these exons contain the few successful lineage-specific survivors of the constant process of gene evolution, by birth and death of individual exons.

Indels and repeats in protein-coding sequences

In contrast to small indels occurring in the bulk of the genome (above), indels within protein-coding regions are probably lethal, or deleterious and so are rapidly removed from the population by purifying selection. Indel rates within rat coding sequences were 50-fold lower than in bulk genomic DNA¹³². The whole genome excess of deletions compared with insertions (Fig. 5b) was also evident in coding sequences. The magnitude was less, with a genome-wide deletion-to-insertion ratio of 3.1:1 reducing to 1.7:1 in the rat. In mouse this value reduced from 2.5:1 to 1.1:1 (ref. 132). These data suggest that deletions are 16% more likely than insertions to be removed from coding sequences by selection.

Owing to the triplet nature of the genetic code, indels of multiples of three nucleotides in length (3_n indels) are less likely to be deleterious. Direct comparison of 3_n indel rates between bulk DNA (0.77 indels per kb for mouse, 0.83 indels per kb for rat) and coding sequence (0.087 indels per kb for mouse and 0.084 indel per kb for rat) showed that 3_n indels were ninefold under-represented in coding sequences. At least 44% of indels were duplicative insertion or deletion of a tandemly duplicated sequence, collectively termed sequence slippage¹³². Sequence slippage contributed approximately equally to observed insertions and deletions. The overall excess of deletions could be attributed specifically to an excess of non-slippage deletion over non-slippage insertion in both mouse and rat lineages¹³². Of the slippage indels, 13% were in the context of trinucleotide repeats (n > 2, excluding the inserted or deleted sequence) which are known to be particularly prone to sequence slippage and encode homopolymeric amino acid tracts^{133, 134}.

To gain better understanding of dynamic changes in the length of homopolymeric amino acid tracts on gene evolution and disease susceptibility, we searched for other characteristics of amino acid repeat variation by analysing all size-five or longer amino acid repeats in a data set of 7,039 rat, mouse and human orthologous protein sequences¹³⁵. Most species-specific amino acid repeats (80–90%) were found in indel regions, and regions encoding species-specific repeats were more likely to contain tandem trinucleotide repeats than those encoding conserved repeats. This was consistent with the involvement of slippage in the generation of novel repeats in proteins and extended previous observations for glutamine repeats in a more limited human–mouse data set¹³⁶.

The percentage of proteins containing amino acid repeats was 13.7% in rat, 14.9% in mouse and 17.6% in human¹³⁵. The most frequently occurring tandem amino acid repeats were glutamic acid, proline, alanine, leucine, serine, glycine, glutamine and lysine. Using the same threshold size cut-off, tandem trinucleotide repeats were significantly more abundant in human than in rodent coding sequences, in striking contrast to the frequencies observed in bulk genomic sequences (29 trinucleotide repeats per Mb in rat, 32 repeats per Mb in mouse and 13 repeats per Mb in human, see discussion of the general simple repeat structure below). The conservation of human repeats was higher in mouse (52%) than in rat (46.5%), suggesting a higher rate of repeat loss in the rat lineage than the mouse lineage.

Functional consequences of these in-frame changes in rat, mouse and human were investigated¹³² through clustering of proteins based on annotation of function and cellular localization¹¹², and mapping indels onto protein structural and sequence features. The rate that indels accumulated in secreted (3.9 10^-4 indels per amino acid) and nuclear (4.0 10^-4) proteins is approximately twice that of cytoplasmic (2.4 10^-4) and mitochondrial (1.4 10^-4) proteins. Likewise, ligand-binding proteins acquire indels (3.1 10^-4) at a higher rate than enzymes (2.1 10^-4)¹³². These trends exactly mirror those observed for amino acid substitution rates³, suggesting tight coupling of selective constraints between indels and substitutions. Transcription regulators showed the highest rate of indels (4.3 10^-4), a finding that may relate to the over-representation of homopolymorphic amino acid tracts in these proteins¹³⁵.

Known protein domains exhibited 3.3-fold fewer indels than expected by chance, again paralleling nucleotide substitution rate differences between domains and non-domain sequences³. Of the protein-sequence and structural categories considered (transmembrane, protein domain, signal peptide, coiled coil and low complexity), the transmembrane regions were the most refractory to accumulating indels, exhibiting a sixfold reduction compared with that expected by chance. Low-complexity regions were 3.1-fold enriched, reflecting their relatively unstructured nature and enrichment in indel-prone trinucleotide repeats. Mapping of indels onto groups of known structures revealed that indels are 21% more likely to be tolerated in loop regions than the structural core of the protein¹³².

We observed that indel frequency and amino acid repeat occurrence both correlated positively with the G + C coding sequence content of the local sequence environment^{132, 135}. This may be explained in part by the correlation of polymerase slippage-prone trinucleotide repeat sequences and G + C content¹³⁵. There is also a positive correlation between CpG dinucleotide frequency and coding sequence insertions, but not deletions. This effect diminishes rapidly with increasing distance from the site of the insertion¹³².

Transcription-associated substitution strand asymmetry

A recent study reported a significant strand asymmetry for neutral substitutions in transcribed regions¹³³. Within introns of nine genes, the higher rate of A G substitutions over that of T C substitutions, together with a smaller excess of G A over C T substitutions, leads to an excess of G + T over C + A on the coding strand (also verified on human chromosome 22). The authors¹³³ hypothesized that the asymmetries are a byproduct of transcription-coupled repair in germline cells. Examining the three-way alignments of rat, mouse and human, we verified that the strand asymmetries for neutral substitutions exist in introns across the genome (Table 4).

Table 4: Strand asymmetry of substitutions in introns of rat genes

Full table

Under the assumption of independence of sequence positions, large sample normal approximations to the binomial distribution allow us to test whether the fraction of G + T exceeds 0.5, and whether the rate at the numerator exceeds the rate at the denominator for each of the ratios in Table 4. With the large amount of data provided by pooling introns genome-wide, the tests are all highly significant (P values < 10^-4), except for the rate of G A in mouse, which does not significantly exceed that of C T (P value = 0.6369). These asymmetries are also seen if the study is limited to ancestral repeat sites, excludes ancestral repeat sites, excludes CpG dinucleotides, is limited to positions flanked by sites that are identical in the aligned sequences (in the case of observations 2 and 3 in Table 4), or considers introns of RefSeq genes for human or mouse. Thus it appears that strand asymmetry of substitution events within transcribed regions of the genome is a robust genome-wide phenomenon.

Conservation of intronic splice signals

Using 6,352 human–mouse–rat orthologous introns from 976 genes (Methods), we examined the dynamics of evolution of consensus splice signals in mammalian genes. We found that intron class¹³⁷ is extremely well conserved: we did not observe any U2 to U12 intron conversion, or vice versa, nor within U12 introns did we find any switching between the major AT–AC and GT–AG subtypes, although such events are documented at larger evolutionary distances¹³⁷. In contrast, conversions between canonical GT–AG and non-canonical GC–AG subtypes of U2 introns are not uncommon. Only 70% of GC–AG introns are conserved between human and mouse/rat, and only 90% are conserved between mouse and rat. Using human as the outgroup, we detected nine GT to GC conversions after divergence of mouse and rat (from 6,282 introns that were likely to have been GT–AG before human and rodents split), and two GC to GT conversions (from 34 GC–AG introns that probably predated the human and rodent split). These results give some indication of the degree to which mutation from T to C is tolerated in donor sites. The GC donor site appears to be better tolerated in introns with very strong donor sites, because in these introns the proportion of GC donor sites is 11%, much higher than the 0.7% overall frequency of GC donor sites in U2 introns. Although we found a variety of other non-canonical configurations in U2 introns, very few are conserved, which suggests that most correspond to transient, evolutionarily unstable states, pseudogenes, or mis-annotations.

Gene duplications

Duplication of genomic segments represents a frequent and robust mechanism for generating new genes¹³⁸. Because there were no compelling data showing rat-specific genes arising directly from non-coding sequences, we examined gene duplications to measure their potential contribution to rat-specific biology. A previous study showed that gene clusters in mouse without counterparts in human are subject to rapid, adaptive evolution^{3, 139}. We used two methods to identify recent gene duplications: methods that directly identified paralogous clusters, and methods that analysed genomic segmental duplications (see above).

Using the first approach, we found 784 rat paralogue clusters containing 3,089 genes (Methods). This was lower than in mouse (910 clusters/3,784 genes), but the difference probably reflects the larger number of gene predictions from the mouse assembly.

To investigate the timing of expansion of these individual families, we measured rates of local gene duplication and retention within clusters. BLAST is not suited to this^{140, 141} and so we instead calculated the number of synonymous substitutions per synonymous site (K_S) between all pairs of homologous genes; constructed K_S-derived phylogenetic trees; and predicted orthology or paralogy gene duplication events automatically from their topologies (Supplementary Information). The results showed that the neutral substitution rate varies among orthologues by approximately twofold (Fig. 10). This is similar to chromosomal variation shown previously by a study of mouse and human ancestral repeats³. Rates of change among ancestral gene duplications (those that predate the mouse–rat split) were relatively constant. Mouse-specific and rat-specific duplications occurred at similar rates, except for those with K_S < 0.04, which are reduced in mouse-specific duplications (Fig. 10). More data are required to determine whether this reduction is a biological effect, as it might be accounted for by different protocols for assembling mouse and rat genomes, which differentially collapse areas of nearly identical sequence.

Figure 10: Variation in the frequency of gene duplications during the evolutionary histories of the rat and mouse.

The sequence of gene duplication events was inferred from phylogenetic trees determined from pairwise estimates of genetic divergence under neutral selection (K_S, Methods). The median K_S value for mouse:rat 1:1 orthologues is 0.19. This value corresponds to the divergence time of mouse and rat lineages.

High resolution image and legend (41K)

The rat paralogue pairs that probably arose after the rat–mouse split (12–24 Myr ago) have K_S values of 0.2 (Table 3). We found 649 K_S < 0.2 gene duplication events in rat, a lower number than is found in mouse (755). For both rodents, this represents a likelihood of a gene duplicating of between 1.3 10^-3 and 2.6 10^-3 every Myr. These are necessarily estimates, because gene deletions, conversions and pseudogene formation are not considered. Interestingly, the data are consistent with a previous estimate for Drosophila genes, but are an order of magnitude lower than an estimate for Caenorhabditis elegans genes¹⁴⁰.

A subset of clusters have at least three gene duplications with K_S < 0.2 (Table 5). These are expected to be enriched in genes whose duplications persist as a consequence of positive selection. The group is dominated by genes involved in adaptive immune response and chemosensation⁸⁷. Inspection of the K_S-derived trees allowed us to infer the gene numbers in these clusters for the common ancestor of rat and mouse (that is, at K_S = 0.2), assuming no gene deletions or pseudogene generation (Table 5). Immunoglobulin, T-cell receptor -chain, and _2u-globulin genes appear to be duplicating at the fastest rates in the rat genome (Table 5). Since divergence with mouse, these rat clusters have increased gene content several-fold. This recapitulates previous observations that rapidly evolving and duplicating genes are over-represented in olfaction and odorant detection, antigen recognition and reproduction¹⁴².

Table 5: Recent gene duplications (K_S < 0.2) in the rat lineage

Full table

An examination of duplicated genomic segments showed this enrichment for most of the same genes and also elements involved in foreign compound detoxification (cytochrome P450 and carboxylesterase genes)⁸⁷. Together, these are exciting findings because each of these categories can easily be associated with a familiar feature of rat-specific biology, and further investigation could explain some differences between rats and their evolutionary neighbours.

Conservation of gene regulatory regions

As the third mammal to be fully sequenced, the rat can add significantly to the utility of nucleotide alignments for identifying conserved non-coding sequences^{143, 144, 145, 146, 147}. This power increases roughly as a function of the total amount of neutral substitution represented in the alignment^{97, 98}, and rat adds about 15% to the human–mouse comparison (Fig. 5). Many conserved mammalian non-coding sequences are expected to have regulatory function, and can be predicted using further analyses based upon these alignments^{93, 148, 149, 150}.

We applied such methods for detecting significantly conserved elements^{97, 151} and scoring regulatory potential^{148, 152} to the genome-wide human–mouse–rat alignments. Typical results show strong conservation for a coding exon, as well as for several non-coding regions (Fig. 11). For example, the intronic region in Fig. 11 contains 504 bp that are highly conserved in human, mouse and rat. The last 100 bp of this alignment block are identical in all three species. Peaks in regulatory potential score are correlated with conservation score, and in the highly conserved intronic segment, they are higher for the three-way regulatory potential score than for the two-way scores using human and just one rodent¹⁵². These data are illustrative, but form the foundation of ongoing efforts to identify genome sequences involved in gene regulation.

Figure 11: Close-up of PEX14 (peroxisomal membrane protein) locus on human chromosome 1 (with homologous mouse chromosome 4 and rat chromosome 5).

Conservation score computed on three-way human–mouse–rat alignments (parsimony P values¹⁵¹) presents a clear coding exon peak (grey bar) and very high values in a 504 bp non-coding, intronic segment (right; last 100 bp of alignment are identical in all three organisms). The latter segment showed a striking difference between the inferred mouse and rat branch lengths^110,111,222: the grey bracket corresponds to a phylogenetic tree where the logarithm of mouse to rat branch-length ratio is -6. Regulatory potential scores^148,152 that discriminate between conserved regulatory elements and neutrally evolving DNA are calculated from three-way (human–mouse–rat) and two-way (human–rodent) alignments. Here the three-way regulatory potential scores are enhanced over the two-way scores.

High resolution image and legend (60K)

Requiring conservation among mammalian genomes greatly increases the specificity of predictions of transcription factor binding sites. Transcription factor databases such as TRANSFAC¹⁵³ contain known transcription factor binding sites and some knowledge of their di