To test whether human genes with CpG island promoters have a distinct chromatin organization, we analyzed the levels of histone modifications and other epigenetic modifications in 1,500 CpG and 1,500 non-CpG promoter genes with precisely matched expression levels (Materials and methods; Figure 1a; Additional files 1 and 2). Transcription is a strong influence on chromatin organization 24333536373839 and so it is necessary to control for expression level when examining additional potential influences. We first analyzed histone modifications previously associated with transcription elongation, that is, enriched within the bodies of expressed genes.

Trimethylation of the lysine 3 residue of histone H3 (H3K36me3) is catalyzed by Set2-type methyltransferases and is linked to transcription elongation 404142. As shown in Figure 1, CpG and non-CpG promoter genes with matched expression levels have very similar distributions of this modification, with a linear accumulation towards the 3' end of expressed genes, and levels that correlate with expression (Figure 1b). Only at the first nucleosome downstream of the site of initiation do highly expressed non-CpG promoters show a slightly higher level of this histone modification. The similar accumulation of H3K36me3 towards the end of the gene suggests a similar transcription elongation-coupled mechanism of H3K36me3 deposition in both gene classes.

In contrast to H3K36me3, however, our analysis revealed that other histone modifications enriched within the bodies of active genes have different distributions in CpG and non-CpG promoter genes. The modification H4K20me1 is catalyzed by SETD8/Pr-Set7 methyltransferases and has been variously linked to gene silencing, gene activation, transcription elongation, and to the early exons of highly expressed genes 4344454647. In expressed genes with CpG promoters, H4K20me1 levels increase rapidly within gene bodies, reaching a peak on nucleosomes approximately 1 to 1.5 kb after the start site (Figure 2a). In contrast, in non-CpG promoter genes, H4K20me1 levels are low on promoter-proximal nucleosomes, increasing only gradually to a plateau approximately 2 to 2.5 kb after the start site (Figure 2a). A distinct peak is not observed, and non-CpG promoter genes with medium expression only have very low levels of H4K20me1 (Figure 2a).

Levels of the transcription elongation-linked modifications H2BK5me1 and H3K79me1/2/3 also differ between the two gene classes, and in a manner that is similar to H4K20me1. Modification levels peak on nucleosomes 750 bp to 1 kb downstream of CpG promoter genes (Figure 2b, c). In contrast, they show only weak monotonic enrichment in the bodies of genes without CpG promoters (Figure 2b, c; Additional file 3).

The exon density (Additional file 4), nucleosome density (see below) and DNA methylation levels (Additional file 5) downstream of highly expressed CpG and non-CpG promoters are very similar, and so cannot account for these differences in the deposition of transcription elongation marks (Additional file 3). Also, the differences in the chromatin profiles remain largely unchanged when a different CpG island definition is used (Additional file 6 using CpG islands defined by Hackenberg et al. 48 and Additional file 7 using the promoter classification of Weber et al. 25), when we remove all genes with more than one transcription start site (Additional file 8), or even when we use an independent set of chromatin data from IMR90 cells (Additional file 9). Rather, genome-wide analysis suggests differences in the recruitment or activation of histone modifying enzymes during the early stages of transcription elongation in these two gene classes.

A second class of histone modifications revealed by our analysis is a set of transcription-associated histone methylations that are enriched in genes with CpG island promoters, but for which the location of the most modified nucleosomes differs depending upon the expression level (Figure 3). As for the previously described modifications, the levels of the methylations H3K9me1, H3K4me2, H3K4me1 and H3K27me1 increase with expression. However, in genes with higher expression, maximum modification levels are found on nucleosomes located further into the body of a gene (Figure 3a-h). Indeed, the +1 nucleosomes of highly expressed genes have lower levels of these histone methylations than the +1 nucleosomes of lowly expressed genes (Figure 3i-l). To our knowledge, this shifting of histone methylation peaks into gene bodies has not been previously noted.

In contrast to H3K36me3, which is located towards the 3' end of active gene bodies, monomethylation of histone H3 lysine 36 (H3K36me1) is detected in the 5' end of highly expressed genes 46. Surprisingly, our analysis suggests that H3K36 monomethylation is only enriched in the bodies of active genes that have CpG islands in their promoters (Figure 4a). Enrichment is detected on the fourth to eighth nucleosomes downstream of the start site, and then quickly returns to background levels within 2 kb (Figure 4a). In contrast, no enrichment is detected in highly expressed non-CpG promoter genes (Figure 4a). This suggests that the deposition of this modification may depend on the presence of a CpG island.

In contrast to the histone modifications described above that are shifted closer to the transcription start site in CpG promoter genes (H4K20me1, H2BK5me1, H3K79me1/2/3), show stronger peaks that shift in location in CpG promoter genes (H3K9me1, H3K4me1, H3K4me2, H3K27me1), or are only enriched in CpG promoter genes (H3K36me1), a fourth set of modifications consists of transcription-associated modifications that are actually depleted in the 5' end of CpG island promoter genes. In highly expressed genes with CpG promoters, the modifications H2AK5ac, H3K14ac and H3K23ac peak approximately 750 bp after the initiation site (Figure 4b-d). In contrast, in non-CpG genes, these modifications either peak immediately after the start site (Figure 4c, d) or do not show a clear peak (Figure 4b). Thus, some transcription activation-associated histone modifications are depleted proximal to the start site of CpG promoter genes.

In expressed genes with CpG islands in their promoters, there is therefore a series of characteristic transitions in transcription-associated histone modifications in the region between approximately 0.5 and 2 kb after the initiation site. In contrast, in non-CpG genes, the same modifications gradually increase more distally within the gene body (H4K20me1, H2BK5me1, H3K79me1/2/3), show weaker or no enrichment within the gene body (H3K9me1, H3K4me1, H3K4me2, H3K27me1, H3K36me1), or are enriched from immediately after the initiation site (H2AK5ac, H3K14ac, H3K23ac). Thus, the identity, positioning, shape and relative ordering of transcription-associated histone modifications all differ between genes with and without CpG island promoters (Figures 2 to 4).

We next compared overall levels of nucleosome occupancy within CpG and non-CpG promoter genes, as revealed by micrococcal nuclease digestion 49 (Figure 5). For highly expressed genes, patterns of occupancy are similar in the two gene classes: a characteristic nucleosome-depleted region is observed upstream of the transcription start site, and an array of well-positioned nucleosomes within the 5' end of the gene (Figure 5a). From an analysis of all genes it was previously proposed that nucleosome depletion upstream of transcription start sites is independent of gene expression level 49. Our analysis showed that this is indeed the case for CpG promoters (Figure 5a), but for non-CpG promoters there is a clear association between a nucleosome-depleted region and mRNA expression levels (Figure 5a). That is, CpG promoters have a constitutive nucleosome-depleted region, but non-CpG promoters do not.

Moreover, comparing the nucleosome occupancy detected in human promoters to that predicted by the intrinsic binding preferences of nucleosomes for DNA 50 revealed that only in non-expressed, non-CpG promoters is nucleosome occupancy actually reflecting the intrinsic binding preferences of nucleosomes for DNA (Figure 5a, b). Thus, in contrast to the situation in transcriptionally quiescent sperm 51, in somatic cells influences beyond the affinity of the DNA for nucleosomes must be important for determining nucleosome occupancy in most active and repressed human promoters.

Consistent with the overall nucleosome depletion, most histone modifications are also depleted at the transcription start sites of CpG promoter genes and at the start sites of highly expressed non-CpG promoter genes (Figures 3, 4 and 6). However, our analysis revealed one notable exception to this: the modification H4K20me3 is detected precisely at the transcription start site in active CpG promoters (Figure 5c). This signal is weak but detected across many initiation sites (Figure 5d). With the current data, we cannot exclude the possibility that the signal is coming from a subset of cells within the population. Nevertheless, a similar peak is not seen for other modifications (Additional file 3). Previously, H4K20me3 has been linked to transcriptional pausing 52, to heterochromatin 53, and to the body of KRAB and zinc finger genes 3946. Our analysis suggests that the modification is also specifically enriched on an unstable nucleosome precisely positioned at the transcription start site in active CpG island promoters. To our knowledge, this is the first evidence for a specific histone post-translational modification precisely marking transcription start sites.

Many epigenetic modifications, particularly histone acetylations, are enriched around the start sites of actively expressed genes (Additional file 3). One of these start site-proximal modifications, H3K4me3, was previously reported to show two distinct peaks in the forward and reverse directions, interpreted as corresponding to two divergent sites of RNA polymerase II initiation 16. Our expression-matched gene sets show that these dual peaks are observed for many promoter-associated modifications, and for both CpG and non-CpG promoter genes (Figure 6a, b). Interestingly, however, the symmetry of these dual peaks about the start site can differ between CpG and non-CpG promoters. For example, the modifications H2A.Z and H4K5ac (and to a lesser extent H4K8ac, H3K36ac, H3K27ac, H2BK5ac, H3K18ac, H3K9ac and H3K4me3) actually have higher levels upstream of the start site (in the antisense orientation) than downstream (in the sense orientation) specifically in promoters with CpG islands (Figure 6). This difference is characteristic comparing across promoters (Figure 6) and again points to a different transcription-associated chromatin organization in genes with CpG island promoters.

Finally, we examined the association of the chromosome organization CCTC-binding factor (CTCF) with CpG and non-CpG promoter genes, and how this association relates to gene expression. We chose to analyze CTCF because up to a quarter of all binding sites for CTCF have been reported to occur close to gene promoters 4654. Surprisingly, CTCF is detected at high levels immediately upstream of the start site in very many CpG island promoters (Figure 7a, b; Additional file 10), and at similar levels in genes with both medium and high expression (and medium and high levels of RNA polymerase II; Figure 7c, d). This suggests the intriguing possibility that CTCF binding might be part of the general architecture of an active CpG island promoter (see Discussion).

Human protein-coding genes were retrieved from Ensembl release 54 59. For each gene we considered only the most 5' transcription start site and we removed genes less than 3 kb long to avoid the inclusion of non-genic regions downstream of a gene. To avoid analyzing upstream regions that overlap another gene, we also removed all genes that have a promoter within 500 bp of another gene. We retrieved CpG islands from the UCSC genome browser 560. We considered a gene as having a CpG-island promoter when its first transcription start site overlaps a UCSC CpG island. Removing genes with more than one transcription start site did not change any of the observations (see Additional file 8). All chromatin profiles were repeated using the alternative CpG island definition from 48 (shown in Additional file 6). Further, all chromatin profiles were repeated using the promoter definition by 25 (shown in Additional file 7). Following the original definition, to annotate promoters as 'high CpG promoters' (HCPs) we scanned the region from -1,200 bp to +300 bp of the transcription start site for a 500-bp window with CpG observed over expected ratio of > 0.75 and GC content > 55. Promoters with all windows with CpG observed over expected ratio ≤ 0.48 were annotated as 'low CpG promoters' (LCPs). The rest of the promoters were annotated as 'intermediate CpG promoters' (ICPs).

We retrieved MAS5 normalized mRNA expression data for CD4+ T cells from Schones et al. 49 (Gene Expression Omnibus (GEO) accession GSE10437), mapping U133-PLUS-2 probes to genes using Ensembl. Probes matching multiple genes were discarded. In total, 16,781 protein-coding genes had annotated expression levels. Genes were ranked according to their expression level, using the (replicate-averaged) value of the most sensitive probe. We then split all genes into three equally sized groups of genes according to their expression level (low, middle and high expression). From these we randomly sampled 500 low, 500 intermediate and 500 high expression non-CpG promoter genes. For each non-CpG promoter gene we then identified a CpG promoter gene with a very similar expression level (absolute difference between log_e expression of non-CpG promoter gene and log_e expression of CpG promoter gene ≤ 0.1). These three expression-matched sets of CpG and non-CpG promoter genes were used to generate average chromatin profile plots. The entire sets of high, intermediate and low expression genes were used to generate chromatin profile heatmaps (the number of genes in each class are included in Additional file 1). The expression-matched sets of genes and their annotations are included in Additional file 2. A small number of genes with extreme expression levels for which we could not find a similarly expressed gene with a different promoter type were removed. We repeated the analysis in exactly the same way for IMR90 cells using microarray gene expression data from Kim et al. 36 (GEO accession GSE2672).

We retrieved the mapped sequenced reads of nucleosome fragments 49 and ChIP-Seq mapped read data for H2A.Z, 20 histone methylations 46 and 18 histone acetylations 38. Similarly, we analyzed DNA methylation data from the same cell-type 61. These datasets contained all reads that match the genome in a unique position with up to two mismatches. To minimize sequence amplification bias, we removed identical reads. We shifted the start position of the reads by 75 bp in the direction of sequencing (75 bp is approximately half of the length of the isolated DNA fragments), this way transforming the read start positions to nucleosome dyad positions. All datasets were rescaled to 10 million uniquely mapped nucleosome fragments. To generate the average chromatin profiles shown in the figures, we counted the number of dyads that fall at each position along the region surrounding the gene start site. Smoothed lines were generated based on the per-base-pair averaged position-shifted read count using the loess regression function in R (with 180 bp span) 62. The predict.loess R function was used for the calculation of 95% confidence intervals. For the background subtracted chromatin profiles included in Additional file 3 we used a 75-bp window sliding by 1 bp and calculated the difference between the number of shifted reads from the histone modification (or H2A.Z) and the number of shifted reads from the nucleosome occupancy. At each position with respect to the transcription start site we then calculated the mean and the standard error of the background-subtracted values assuming a Normal distribution. Heatmaps were generated using Java Tree View 1.1.5r2 63. We repeated all chromatin profiles using data from a fetal lung fibroblast cell line (IMR90) generated by the NIH Roadmap Epigenomics Project 6465. We downloaded the mapped reads provided as BED files. Because these reads were mapped to human genome version hg19, we converted all gene promoters from hg18 to hg19 using the LiftOver tool. These profiles are shown in Additional file 9. The accession identifiers of the samples used for these profiles are included in the figure legend. Regions of statistically significant CTCF binding in CD4+ T cells (used in Additional file 10), based on the data from Barski et al. 46, were retrieved from Ensembl (regulatory build of Ensembl release 68). We defined distal CTCF binding sites as those not overlapping any annotated Ensembl gene. The coordinates of CTCF peaks were converted from human genome assembly hg19 to hg18 using the liftOver tool 60.

Intrinsic nucleosome binding preference calculations along the human genome were predicted by Kaplan et al. and downloaded from the authors' website 50. The average nucleosome affinity model score around the transcription start site of each gene set is shown in Figure 5b. We also carried out the same analysis using the predicted probability that a nucleosome will be formed at each genomic position 50, which gave very similar results (data not shown). Similarly, as a control, we also calculated the G+C content profile at ± 3 kb around the gene start of each group (Additional file 11).

To compare the asymmetry of a histone modification around the start of CpG promoter and non-CpG promoter genes, for each highly expressed gene we calculated the number of position-shifted reads that map to the 500 bp upstream of the gene start and divided by the total number of position-shifted reads that map within 500 bp both upstream and downstream of the gene start. We then compared the distributions of these 'asymmetry scores' at highly expressed CpG and non-CpG promoter genes using the Wilcoxon rank sum test.