We first present our analysis of CGIs identified by Takai and Jones' algorithm 14 in ten mammalian genome sequences. The conclusions are essentially the same when we used the popular algorithm by Gardiner-Garden and Frommer 13 or the recently developed algorithm CpGcluster 15 (see Discussion). The species names and the sources of genome sequences are shown in the Materials and methods. Table 1 summarizes the genome information and statistics of CGIs. Except for the platypus, these genomes had similar sizes (2.0-3.3 Gb) and similar numbers of annotated genes (20,000-30,000; Additional data file 1). However, both the number of CGIs and the CGI density (measured by the average number of CGIs per Mb) vary greatly among genomes. The dog genome has the largest number of CGIs (58,327) and the platypus genome has the highest CGI density (35.9 CGIs/Mb). Remarkably, the number of CGIs in the dog genome is nearly three times that in the rat (19,568) or mouse (20,458) genome, even though the number of dog genes has been estimated to be smaller than those of human or mouse genes (dog, 19,300 18; human, 20,000-25,000 19; mouse, approximately 30,000 11). The CGI density (per Mb) ranges from 7.5 (opossum) to 35.9 (platypus) in the 10 genomes investigated. These results suggest that, although genes are often associated with CGIs, the extant CGIs are distributed very differently among genomic regions (for example, genes versus non-coding regions) in mammalian genomes.

We examined the correlations between CGI density and other genomic features. Because of incomplete genome sequence and lack of some chromosome data in platypus, we present the correlation results only for the other nine genomes; the conclusion will likely be the same when the platypus data become available (Additional data file 2). We found a highly significant positive correlation between CGI density and number of chromosome pairs in a genome (r = 0.88, P = 7.9 × 10^-4; Figure 1a) and a significant correlation between CGI density and number of chromosome arms (r = 0.62, P = 0.037). As expected, there was a significant positive correlation between CGI density and Obs_CpG/Exp_CpG (r = 0.63, P = 0.035). No significant correlation was found between CGI density and genome size (r = -0.53, P = 0.073) or genome GC content (r = 0.24, P = 0.27).

There were a total of 219 chromosomes available in these 9 genomes after excluding the Y chromosomes. We found a highly significant negative correlation between CGI density and log₁₀(chromosome size) (r = -0.51, P = 2.6 × 10^-16; Figure 1b), a highly significant positive correlation between CGI density and GC content of the chromosome (r = 0.65, P = 3.5 × 10^-28; Figure 1c), and a highly significant positive correlation between CGI density and Obs_CpG/Exp_CpG (r = 0.75, P = 2.8 × 10^-41; Figure 1d). We further separated the chromosomes into different groups by their sizes (<25, 25-50, 50-75, 75-100, 100-150, 150-200, and >200 Mb). Interestingly, as the average size of a chromosome group increases, the CGI density decreases (Table 2). Indeed, the CGI density in small mammalian chromosomes (size <25 Mb) is, on average, about three times that in large chromosomes (size >200 Mb). We noted that the platypus (2n = 52), which has six pairs of large chromosomes but many small chromosomes 20, has a much higher CGI density than the other nine mammalian genomes (Table 1). These results are consistent with the previous observation that CGIs are highly concentrated on the microchromosomes in chickens 21.

The dog has overall smaller chromosomes and high CGI density, while the opossum has a few large chromosomes and low CGI density. To check whether our correlation analysis was largely driven by these two species, we performed a similar analysis but excluded the dog and opossum data. The same conclusion still held. For example, we found a significant correlation between CGI density and number of chromosome pairs (r = 0.75, P = 0.026) and a significant correlation between CGI density and log₁₀(chromosome size) (r = -0.49, P = 5.9 × 10^-12).

CGIs are considered gene markers, so they are expected to highly correlate with gene density 222. It is interesting to investigate whether the above correlation results still hold when gene information is excluded. We identified CGIs in the intergenic regions of nine mammalian genomes and found significant correlations between intergenic CGI density and log₁₀(chromosome size) (r = -0.55, P = 7.3 × 10^-19), GC content of the chromosome (r = 0.39, P = 8.6 × 10^-10), and Obs_CpG/Exp_CpG (r = 0.67, P = 3.7 × 10^-30). Details are shown in Additional data file 3.

It is also interesting to examine whether the correlations between CGI density and other genomic factors would hold in different genomic regions. We used human data because of their high quality annotations. According to gene annotations in the NCBI database, we identified 24,228 CGIs overlapped or within genes (gene-associated CGIs), 13,026 CGIs whose whole sequences were within intergenic regions (intergenic CGIs), 12,136 CGIs whose whole sequences were within gene regions (intragenic CGIs), and 11,192 CGIs overlapped with transcriptional start sites (TSS CGIs) in the human genome. Table 3 shows significant correlations between CGI density and genomic features (log₁₀(chromosome size), GC content, and Obs_CpG/Exp_CpG) in all genomic regions when we compare the data at the chromosome level.

Table 4 summarizes the correlations between CGIs and genomic features based on nine or ten genomes using three CGI identification algorithms.

Recombination rate correlates with both the number of chromosomes and the number of chromosome arms, and elevates the GC content, probably via biased gene conversion 2324. Fine-scale recombination rates vary extensively among populations 2526, genomic regions 27, or the homologous regions between two closely related organisms (human and chimpanzee) 2829, suggesting a rapid evolution of local pattern of recombination rates. Many genomic features, including CpG dinucleotide frequencies (but not CGIs or CGI density) in genomic sequences, have been employed to analyze the pattern of recombination rate. Here we examined specifically the relationship between CGI density and recombination rate at the genome level. We retrieved human recombination rate data (window size, 1 Mb, 2,772 windows) from the UCSC Genome Browser 30. We found a significant positive correlation between CGI density and recombination rate (r = 0.18, P = 1.1 × 10^-22).

We obtained another set of recombination rate data (in 5 Mb and 10 Mb windows) for the human, mouse and rat from Jensen-Seaman et al. 31. We discarded those regions that had more than 50% 'N's ('N' denotes an uncertain nucleotide in the sequence) or whose recombination rate was 0. In the latter case, it was likely due to insufficient available genetic markers or a small number of meioses used to construct the genetic maps 31. Again, we found a significant correlation between CGI density and recombination rate, regardless of window size (5 Mb or 10 Mb; Table 5 and Additional data file 4). For example, the correlation coefficient was 0.33 (P = 5.9 × 10^-16) for human recombination rates measured in a 5 Mb window (Figure 2). The correlation became stronger as the window size increased. Furthermore, the extent of the correlation was different among the three genomes. For example, the coefficients were 0.33 (human), 0.24 (mouse), and 0.17 (rat), respectively, when the 5 Mb window was used.

Recombination rates were found to increase from the centromeric towards telomeric regions 31. Interestingly, we observed a trend of higher CGI density in the telomeric regions (Figure 3) in many chromosomes. This feature supports a positive correlation between CGI density and recombination rate. However, this finding is opposite to a previous observation of no correlation between CGI features and chromosomal telomere position based on a small gene dataset 17.

To retrieve information on the CGIs in vertebrate genomes, we scanned CGIs in seven non-mammalian vertebrate genomes, including the chicken, lizard and five fish (tetraodon, medaka, zebrafish, stickleback and fugu) genomes. Except for lizard and fugu, all these genomes had assembled chromosomes.

Table 6 shows the CGIs and other genome information for the seven non-mammalian vertebrates. The CGI density had a much wider range (14.7-161.6 per Mb) among these genomes. The CGI densities in the chicken (23.0 per Mb) and green anole lizard (25.9 per Mb) were similar to that in the dog (25.3 per Mb), higher than that in the other eight therians, but lower than that (35.9 per Mb) in the platypus (prototherian) (Table 1). It is worth noting that both the chicken and platypus have many small chromosomes. The chicken karyotype consists of 39 chromosomes, of which 33 are classified as microchromosomes 32. At the DNA sequence level, chicken chromosomes were separated into three groups (large macrochromosomes, intermediate chromosomes and microchromosomes) by the International Chicken Genome Sequencing Consortium 33. Using this classification, we found that CGI density in the 20 chicken microchromosomes (51.7 per Mb) was much higher than that (15.0 per Mb) in the 6 large macrochromosomes (Table 6), consistent with an earlier report 21. We did not estimate the CGI density in the large or small chromosomes of platypus because the available assembled genome sequences (410 Mb) represent only a small portion of the genome, which is expected to be about the same size as the human genome 20.

CGI densities in the five fish genomes varied to a much greater extent than in the mammalian genomes. The CGI densities in tetraodon (161.6 per Mb) and stickleback (157.8 per Mb) were about 11 times that in zebrafish (14.7 per Mb). The Obs_CpG/Exp_CpG ratios in the fish genomes (0.479-0.662) were also much higher than those (0.129-0.296) in the mammalian, the chicken (0.248) and the lizard (0.296) genomes. Fishes are cold-blooded vertebrates and lack GC-rich isochores 34. An early study found certain fish did not have elevated GC content in nonmethylated CGIs 35, so our comparison of CGIs in fishes should be taken with caution.

In contrast to the observation in mammalian genomes, the correlation between CGI density and number of chromosome pairs in the seven non-mammals was not significant (r = -0.42, P = 0.17). We further examined CGI density at the chromosome level in the five non-mammalian genomes (chicken, tetraodon, stickleback, medaka and zebrafish), whose assembled chromosomes are available, and compared it to the nine mammalian genomes. To distinguish the features of CGIs among different genomes, we separated them into different groups: primates (human, chimpanzee and macaque), rodents (mouse and rat), dog-horse-cow, opossum, chicken and fish (tetraodon, stickleback, medaka and zebrafish). Figure 4 shows the plots of CGI density over chromosome GC content. Although there is an overall trend of increasing CGI density with chromosome GC content in both the mammals and non-mammals, their distributions of CGI densities over the chromosome GC content are different. In mammals, CGI density is high in dog-horse-cow and low in rodents, but extensive overlaps are seen among different groups, especially between primates and other groups (Figure 4a). This pattern is more evident in the plots of CGI density versus log₁₀(chromosome size) or versus chromosome Obs_CpG/Exp_CpG ratios (Additional data file 5). Interestingly, we found an overall distinct distribution pattern among non-mammal genomes, especially among the fish genomes (Figure 4b). The chromosomes from each fish genome clustered but they were separated from other fish genomes (Figure 4b, Additional data file 5). Finally, when all species were plotted together, there were overlaps between mammals and non-mammals, but overall, fish chromosomes and chicken microchromosomes could be separated from the mammalian chromosomes (Figure 4b, Additional data file 5).

There are three major algorithms for identifying CGIs in a genomic sequence (reviewed in the Background). The major aim in this study is to investigate and compare the CGIs in today's mammalian genomes, rather than to identify CGIs in the mammalian ancestral sequences. Thus, our analysis may provide insights into how CGIs have evolved and their association with gene function and other genomic factors. Since CGIs have been widely documented to be approximately 1 kb long 26, Takai and Jones' stringent criteria seem to be the most appropriate for our analysis. To assure the reliability of our analysis, we performed similar analysis using Gardiner-Garden and Frommer's algorithm (only on the non-repeat portions of the genomes) and CpGcluster with the ten mammalian genomes and seven other vertebrate genomes under study. The conclusions were the same; see detailed results in Table 4 and Additional data files 6 and 7. For example, there was a significant positive correlation between CGI density and chromosome number, using Gardiner-Garden and Frommer's algorithm (r = 0.92, P = 2.0 × 10^-4; Additional data file 6) or CpGcluster (r = 0.81, P = 0.004; Additional data file 7).

However, we found that the number of CGIs identified by CpGcluster or Gardiner-Garden and Frommer's algorithm was remarkably larger than that identified by Takai and Jones' algorithm (Additional data file 8); for example, the numbers of CGIs identified in the human genome was 37,531 (Takai and Jones), 76,678 (Gardiner-Garden and Frommer), and 197,727 (CpGcluster). The number of genes was estimated to be approximately in the range 20,000-30,000 in mammalian genomes (Additional data file 1). Since CGIs have been widely considered as gene markers, both the Gardiner-Garden and Frommer algorithm and CpGcluster likely identified either many CGIs that are not associated with genes or multiple CGIs that share one gene. To address the latter case, we evaluated the length distribution of CGIs identified by the three algorithms. Among all these vertebrate genomes, the majority of CGIs identified by CpGcluster were shorter than 500 bp (Additional data file 8), which is the minimum length in Takai and Jones' algorithm. For example, the proportions of human CGIs identified by CpGcluster were 44.3% (<200 bp), 45.9% (200-500 bp), 7.3% (500-1,000 bp), 1.9% (1,000-1,500 bp), 0.4% (1,500-2,000 bp), and 0.2% (≥2,000 bp). For Gardiner-Garden and Frommer's algorithm, the proportion of CGIs shorter than 500 bp was also large, for example, 65.8% in the human CGIs and 64.8% in the opossum CGIs (Additional data file 8). Based on the evaluation above, we consider that our analysis using Takai and Jones' algorithm is the most reliable and appropriate, though further evaluation of species-specific algorithms may enhance our results.

It was hypothesized that CGIs arose once at the dawn of vertebrate evolution and vertebrate ancestral genes were embedded in entirely non-methylated DNA during the divergence of vertebrates 9. Genome-wide methylation has been found to be common in vertebrates (except for promoter-associated CGIs) and fractional methylation common in invertebrates. The transition from fractional to global methylation likely occurred around the origin of vertebrates 36. Many CGIs might have lost their typical features due to de novo methylation at their CpG sites and subsequent high deamination rates at the newly methylated CpG sites, leading to TpG and CpA dinucleotides. Excess of TpGs and CpAs as well as other vanishing CGI features (decreasing length, Obs_CpG/Exp_CpG ratio and GC content) has been found in the homologous gene regions, evidence of frequent CGI losses in mouse and human genes and a faster loss rate in mice 7917. Recent methylation studies revealed weak CGIs in promoter regions (promoters with intermediate CpG content, ICPs), most of which were not found in the CGI library, had a faster loss rate of CpGs than stronger CGIs (promoters with high CpG content, HCPs), suggesting that strong CGIs might be protected from methylation and are thus better conserved during evolution 223738. Using the data in Weber et al. 37 and Mikkelsen et al. 38, we found that HCP density has stronger correlations with genomic features than ICPs in both the human and mouse genomes. The CGIs identified by the Takai-Jones algorithm are different from HCPs or ICPs. However, when we separated the promoter-associated CGIs identified by the Takai-Jones algorithm into HCGIs (those that satisfied the HCP criteria) and non-HCGIs, we also found that HCGIs had stronger correlations with genomic features than non-HCGIs. This supports the observations from the methylation studies mentioned above. Although loss of CGIs is likely a major evolutionary scenario in mammals, little comparative analysis at the DNA sequence level has been performed yet, because CGIs have been thought to be poorly conserved between species 79. Our CGI analysis indicated that rodents have the lowest CGI density and most other eutherians have moderate CGI density when compared to platypus (Table 1). Platypus is one of the only three extant monotremes and has a fascinating mixture of features typical of mammals and of reptiles and birds. Monotremes (mammalian subclass Prototheria) are the oldest branch of the mammalian tree, diverging 210 million years ago from the therian mammals 20. Although the platypus genome is incomplete, its higher CGI density is likely true because high frequencies of GC and CG dinucleotides and high GC content have been reported 20. Further, our analysis of the chicken (bird) and green anole lizard genomic sequences, the only reptilian genome available at present, showed higher CGI density than most of the therians (except dogs) we examined. These data support an overall decrease in CGIs in mammalian genomes.

Below we discuss specific CGI features of a few species. The low number of CGIs in the rodent genome is likely due to a much higher rate of CGI loss and a weaker selective constraint in the rodent lineage 717. Interestingly, the dog has a notably large number of CGIs and high CGI density among the nine therians investigated. Our further analysis revealed that the difference is due to the substantial enrichment of CGIs in dog's intergenic and intronic regions, while the number of CGIs associated with the 5' end of genes is similar to the human and the mouse (data not shown). Whether and how CGIs have accumulated in dog requires further investigation. It is also worth noting that opossum, which belongs to metatheria, is another evolutionarily ancient lineage of mammals. The CGI density is very low (7.5 per Mb). This is likely attributed to its large chromosomes (Table 1), as large chromosomes are correlated with low CGI density (Figure 1). Large chromosomes reduce recombination rate, which has a positive correlation with CGI density (Figure 2).

It is interesting to examine whether species traits such as lifespan, body temperature and body mass are related to CGI density. The small body size and short lifespan of mice were speculated to allow for their tolerance towards leaky control of gene activity, including erosion of CGIs 17. A previous study also revealed that methylation status is correlated with body temperatures in fish and affected by the local environment 39. It was also proposed that GC content of the isochores is driven by increasing body temperature, which has selective advantages because of being more thermally stable in higher GC-content regions 40. Our correlation analysis found a significant correlation between CGI density and body temperature in eight eutherians (r = 0.67, P = 0.035) and nine therians (r = 0.63, P = 0.034; Figure 5a). However, when platypus and/or chicken were added, the correlation became insignificant. Furthermore, we did not find a significant correlation between CGI density and lifespan in the eight eutherians (r = 0.14, P = 0.38) or nine therians (r = 0.26, P = 0.25; Figure 5b). Some factors might have affected the estimation of lifespan, making the analysis unreliable. First, living environments are much different between domesticated and wild animals; meanwhile, modern medical treatment has increased human longevity. Second, lifespan in the same species may differ according to factors such as sex 41 and hormonal regulation 4243. Third, the divergence among mammals is low when compared to other vertebrates. In summary, our analysis of these species traits should be considered preliminary.

We downloaded the assembled genome sequences (ten mammalian genomes and seven non-mammalian vertebrate genomes) from the National Center for Biotechnology Information (NCBI) 44 and the UCSC Genome Browser 30. The species names and data sources are provided in Table 7. The repeat-masked sequences of these genomes were downloaded from the UCSC Genome Browser 30. We used the EMBOSS package 45 to calculate the genome size, the GC content and the Obs_CpG/Exp_CpG ratios. Gene numbers were based on the annotations in Ensembl 46 and also in the literature (details are shown in Additional data file 1). At present, it remains a great challenge to obtain an accurate estimation of the gene number in a genome, but we suspect that the actual gene numbers in these genomes are likely in a smaller range than the range 20,000-30,000 in Additional data file 1.

We used three algorithms to identify CGIs. First, we used the stringent search criteria in the Takai and Jones algorithm 14: GC content ≥55%, Obs_CpG/Exp_CpG ≥0.65, and length ≥500 bp. Second, we used the algorithm originally developed by Gardiner-Garden and Frommer 13: GC content >50%, Obs_CpG/Exp_CpG >0.60, and length >200 bp. Because some repeats (for example, Alu) meet these criteria, we scanned CGIs in the non-repeat portions of these genomes only, as similarly done in other genome-wide identification studies 211. For both the Takai and Jones and the Gardiner-Garden and Frommer algorithms, we used the CpG island searcher program (CpGi130) available at 47. Third, we used CpGcluster developed by Hackenberg et al. 15 to scan CGIs in the whole genome.

We used the method of Jiang and Zhao 48 to identify CGIs in different genomic regions (genes, intergenic regions, intragenic regions, and TSS regions). Briefly, we compared the locations of CGIs with the coordinates of genic, intergenic, and intragenic regions and TSSs based on the human gene annotation information from the NCBI database (build 35.1) 4449. CGIs overlapped with any genes were classified as gene-associated CGIs; CGIs whose whole sequences were in intergenic regions were classified as intergenic CGIs; CGIs whose sequences were in gene regions were classified as intragenic CGIs; and CGIs overlapped with TSSs were classified as TSS CGIs.

We retrieved human recombination rate data based on the deCODE genetic map 50 from the UCSC Genome Browser 30. The recombination rates were measured in 1 Mb windows. We obtained another set of recombination rates from Jensen-Seaman et al. 31. These data were measured in 5 Mb and 10 Mb windows for the human, mouse and rat and are available in the supplementary material for Jensen-Seaman et al. 31. For both datasets, we discarded those regions having more than 50% 'N's 31. We also discarded those regions whose recombination rates were 0 because of too few genetic markers found in these regions 31.

Records of body temperature in a species may vary to some extent in the literature because they might be measured in different environments (for example, time of day, season, or geographical location) or different sites of the body. The body temperatures of ten mammals in this study were obtained from the literature (details are shown in Additional data file 9). When a species has a range of body temperatures in the literature, the average was used as the representative temperature. There are several measurements of lifespan, such as maximum lifespan, average lifespan, and lifespan of each sex. We used maximum lifespan, which was based on reports in the literature and from the AnAge database 51 (Additional data file 9).