Arabidopsis chromatin enriched for H3K4me was isolated by ChIP using antibodies that specifically recognize H3K4me1, H3K4me2 and H3K4me3 (Figure S1 in Additional data file 1). As a control, nucleosomal DNA was isolated by ChIP using an antibody against histone H3 regardless of its modifications. H3K4me ChIP samples were compared to the control nucleosomal DNA by hybridization to Affymetrix whole-genome tiling microarrays that represent approximately 97% of the sequenced Arabidopsis genome at 35-bp resolution.

H3K4me1, H3K4me2 and H3K4me3 regions identified here are highly consistent with results from recently published studies 38 (Figure S2 in Additional data file 1). In addition, real-time PCR validations were performed at a number of randomly chosen loci, all of which yielded results consistent with the ChIP-chip data (Figure S3 in Additional data file 1). Finally, only 0.10%, 0.66% and 0.57% of the chloroplast genome was falsely identified as containing H3K4me1, H3K4me2 and H3K4me3, respectively. Taken together, these results indicate that the ChIP-chip data here provide an accurate representation of the genome-wide distribution of H3K4me with a relatively low false positive rate.

A total of 15,475 (7.77 Mb) H3K4me1, 12,781 (7.17 Mb) H3K4me2 and 15,894 (14.48 Mb) H3K4me3 regions were identified as described above, representing 6.45%, 6.0% and 12.1% of the sequenced nuclear genome, respectively. All three types of H3K4me are highly enriched in the gene-rich euchromatin and absent from pericentromeric heterochromatin regions where transposons and other repetitive sequences cluster (Figure 1a). Such a euchromatic distribution may largely reflect the fact that H3K4me1, H3K4me2 and H3K4me3 localize almost exclusively in genes: 96.7%, 93.3% and 95.7% of all H3K4me1, H3K4me2 and H3K4me3 regions, respectively, are in or overlap with transcribed regions of genes or their promoters (defined as the 200-bp regions upstream of transcription start sites). Only a small fraction of the remaining H3K4me1, H3K4me2 and H3K4me3 regions (0.6%, 1.3% and 1.5% of total, respectively) overlap with intergenic repetitive sequences such as transposons. The distribution of HK4me in a representative eukaryotic region is shown in Figure 1b.

A total of 18,233 genes (approximately 68.0% of all annotated genes) were found to contain H3K4me in their promoters and/or transcribed regions, including 8,571 with H3K4me1, 10,396 with H3K4me2 and 14,712 with H3K4me3. The distribution patterns of H3K4me at the 5' regions of genes were determined by aligning genes by their transcription start sites, and the percentage of genes containing H3K4me in their promoters and the 5' transcribed regions was determined. Similarly, the distribution patterns of H3K4me at the 3' regions of genes were determined by aligning genes by the 3' end of their transcribed regions. These analyses were performed on a set of 5,809 genes that meet the following two criteria. First, they are located 1 kb or more away from the upstream and downstream genes such that ambiguity introduced by neighboring genes can be minimized. Second, they are longer than 1 kb so that there is sufficient gene space to determine the distribution of H3K4me. We further classified the 5,809 genes into four groups according to their length: long genes (>4 kb, 691 genes), intermediate genes (3 to 4 kb, 828 genes; 2 to 3 kb, 1,768 genes) and short genes (1 to 2 kb, 2,522 genes).

The distribution patterns of H3K4me on long genes are shown in Figure 2a. H3K4me1 is present at relatively low level at the 5' and 3' termini of transcribed regions, but is enriched in the internal regions with a slight 3' bias. In contrast, H3K4me2 and H3K4me3 are both enriched in the 5' end with H3K4me3 distributed slightly upstream of H3K4me2. Both H3K4me2 and H3K4me3 are also enriched in the promoters (200 bp upstream of transcription start sites) and 5' flanking regions (200 to approximately 400 bp upstream of transcription start sites), but are absent in the 3' half of the transcribed regions or the 3' flanking regions of the long genes.

A comparison of the distribution patterns of H3K4me on long genes and intermediate or short genes revealed several common features as well as some interesting differences. First, as gene length decreases, significantly smaller fractions of genes were found to contain H3K4me1, but the relative position of H3K4me1 in genes (that is, internal regions with a 3' bias) remains similar. Second, the distribution patterns of both H3K4me2 and H3K4me3 at the 5' ends of short or intermediate genes are largely similar to those on long genes, although the shortest genes seem to contain a lower level of H3K4me3 at the 5' end. Third, as gene length decreases, significantly more genes were found to contain H3K4me2 and H3K4me3 in their 3' regions. For example, in the last 200 bp, 10.8- and 13.3-fold more short genes contain H3K4me2 and H3K4me3 than long genes, respectively.

In order to obtain a more continuous view of the distribution of H3K4me, we analyzed the average distribution levels of H3K4me across entire genes. To do this, we divided the transcribed region of each gene into 20 bins (5% of the gene length per bin), and divided the 1-kb upstream and downstream flanking regions of each gene into 20 bins (50 bp per bin). The percentage of genes containing H3K4me in each bin was then determined (Figure 2b). Consistent with the results described above, H3K4me1 is highly enriched within the transcribed regions, but it is present at very low levels in promoters and 3' flanking regions. In addition, H3K4me1 is present at significantly higher levels and spans broader regions on longer genes. In contrast, H3K4me2 and H3K4me3 are enriched in promoters and the 5' half of transcribed regions, at comparable levels on genes with different lengths. Although H3K4me2 and H3K4me3 extend further towards the 3' end on shorter genes relative to gene length, the absolute positions remain virtually constant: regardless of gene length, the highest levels of H3K4me2 and H3K4me3 were found at approximately 600 to 800 bp and 400 to 600 bp downstream of transcription start sites, respectively (Figure 2a). In addition, for genes in all the length groups, H3K4me2 and H3K4me3 appear to be enriched (that is, present at the same or higher levels as they are at transcription start sites) downstream of transcription start sites for approximately 1.5 kb and 1 kb, respectively (Figure 2a).

The observation that H3K4me2 and H3K4me3 appear to cover the 5' regions of genes for a relatively constant length suggests that the length of a given gene may affect the association of this gene with different types of H3K4me, in particular H3K4me1. For example, while all three types of H3K4me are positively correlated with gene length (Figure 2b), such a relationship is significantly more pronounced for H3K4me1. To further study the relationship between gene length and H3K4me, we classified the 5,809 genes into 8 categories based on the 8 possible combinations of their associated H3K4me: H3K4me1 only (me1⁺me2^-me3^-), H3K4me2 only (me1^-me2⁺me3^-), H3K4me3 only (me1^-me2^-me3⁺), H3K4me1 and H3K4me2 but no H3K4me3 (me1⁺me2⁺me3^-), H3K4me1 and H3K4me3 but not H3K4me2 (me1⁺me2^-me3⁺), H3K4me2 and H3K4me3 but not H3K4me2 (me1^-me2⁺me3⁺), H3K4me1, H3K4me2 and H3K4me3 (me1⁺me2⁺me3⁺), and no H3K4me (me1^-me2^-me3^-). The frequencies of occurrences of these combinations within each length group were then determined. As shown in Figure 2c, all combinations that include H3K4me1 (regardless of H3K4me2 and H3K4me3) showed a strong positive correlation with gene length, and all combinations of H3K4me2 and H3K4me3 (in the absence of H3K4me1) showed a negative correlation with gene length. In addition, genes associated with H3K4me1 (me1⁺me2^-me3^-, me1⁺me2⁺me3^-, me1⁺me2^-me3⁺, me1⁺me2⁺me3⁺) are generally longer than average, with me1⁺me2^-me3⁺ and me1⁺me2⁺me3⁺ genes being significantly longer and including very few genes shorter than 2 kb (Figure 2d). In summary, by every measure, longer genes show higher levels of H3K4me1.

The distribution patterns of H3K4me2 and H3K4me3 described here are similar to results from analyzing genes on chromosomes 4 and 10 in rice 41. That is, in both species, H3K4me2 and H3K4me3 are enriched in the promoters and the 5' ends of transcribed regions, with H3K4me3 peaking slightly upstream of H3K4me2 (at approximately 400 to 600 bp and approximately 600 to 800 bp downstream of transcription start sites, respectively; Figure 2a). These results suggest that H3K4me2 and H3K4me3 may be involved in both transcription initiation and the early stages of transcription elongation. In contrast, the internal distribution of H3K4me1 observed here suggests that H3K4me1 might be primarily involved in the elongation step during the transcription of longer genes. Alternatively, the apparent preferential accumulation of H3K4me1 in the transcribed regions may be because this modification is reduced at gene ends (that is, H3K4 is preferentially di- or trimethylated at the 5' ends and unmethylated at the 3' ends).

To further test the relationship between H3K4me and transcription, we compared the expression level and tissue specificity of genes associated with different combinations of H3K4me, using a previously published expression profiling dataset 42. Of the 5,809 genes described above, 5,479 were analyzed here, as expression data were available for these genes. As shown in Figure 3a, me1⁺me2^-me3⁺, me1⁺me2⁺me3⁺ and me1^-me2^-me3⁺ genes are highly expressed, whereas me1⁺me2^-me3^-, me1^-me2⁺me3^- and me1⁺me2⁺me3^- genes are expressed at very low levels. The me1^-me2⁺me3⁺ group includes genes with a wide range of expression levels and seems to be enriched for moderately expressed genes. In addition, me1⁺me2^-me3⁺, me1⁺me2⁺me3⁺ and me1^-me2^-me3⁺ genes exhibit very low levels of tissue specificity, while me1⁺me2^-me3^-, me1^-me2⁺me3^- and me1⁺me2⁺me3^- genes are highly tissue specific (Figure 3b). Taken together, these results suggest that H3K4me3 is associated with and likely plays important roles in active transcription. H3K4me1 and H3K4me2, in the absence of H3K4me3, are preferentially associated with tissue-specific genes that are generally not expressed at the developmental stage assayed in this study. These results are consistent with previous reports that although H3K4me2 is generally associated with genes in Arabidopsis, its presence does not always correlate with active transcription 37.

In Drosophila, the Trithorax (TRX) family of H3K4 methyltransferases and the Enhancer of Zeste (E(z)) family of H3K27 methyltransferases function antagonistically to activate or repress a largely overlapping set of genes, respectively 4344. Interestingly, many genes are associated with both H3K4me and H3K27me3 in mammalian stem cells, and such a 'bivalent' histone modification has been suggested to play an important role in stem cell renewal and differentiation 45. Similarly, the co-existence and antagonistic functions of H3K4me3 and H3K27me3 have been described at the FLC and AGAMOUS genes in Arabidopsis 3839464748. We have indeed detected H3K4me2, H3K4me3 and H3K27me3 at the FLC gene. However, we found that AGAMOUS contains a low level of H3K4me2 but no significant level of H3K4me3. This apparent discrepancy is likely due to the different tissues used in the experiments: young seedlings were used in this studywhereas a previous study used mature rosettes (Z Avramova, personal communication).

We have previously found that H3K27me3 is associated with 4,000 to 5,000 tissue-specific genes in their repressed state in Arabidopsis 26. In order to test whether a preferential association of H3K4me with H3K27me3 exists that could indicate a functional connection, we first determined the fraction of genes with each combination of H3K4me that are also associated with H3K27me3. As shown in Table 1, we found that me1^-me2^-me3^- and me1^-me2⁺me3^- genes are associated with H3K27me3 more frequently than expected. In addition, the association frequencies of me1⁺me2^-me3^-, me1⁺me2⁺me3^- and me1^-me2⁺me3⁺ genes with H3K27me3 are all lower than expected. Finally, me1^-me2^-me3⁺, me1⁺me2^-me3⁺, and me1⁺me2⁺me3⁺ genes are even more depleted of H3K27me3. It should be noted that the differences in transcription levels cannot fully account for the differential association of H3K4me genes with H3K27me3. For example, the me1^-me2^-me3^- and me1^-me2⁺me3^- genes are significantly more frequently associated with H3K27me3 than me1⁺me2^-me3^- and me1⁺me2⁺me3^- genes, but these four categories of genes are expressed at very similar levels (Figure 3). The relationship between H3K4me and H3K27me3 was further examined by directly testing whether they co-localize to the same genomic regions. To do this, we determined the presence of each type of H3K4me in H3K27me3-containing genomic regions. As a control, we also determined the presence of H3K4me in a set of randomly chosen regions with the same length and genomic distributions of H3K27me3-containing regions. As shown in Table 2, whereas H3K4me1 and H3K4me3 are significantly depleted in H3K27me3-containing regions, H3K4me2 was found to overlap with H3K27me3 slightly more frequently than with random control regions.

It should be noted that the starting materials in this study (young seedlings) included many distinct cell types. It is likely that some genes are associated with H3K4me3 when they are expressed in some cell types, but are associated with H3K27m3 elsewhere when they are transcriptionally repressed. It is therefore possible that the low frequency of co-localization between H3K4me3 and H3K27me3 described here may still represent an overestimate. It is also possible, however, that co-localization of H3K4me3 and H3K27me3 at a given gene only occurs in specific cell types or during certain developmental stages. If this is the case, our results generated using mixed cell types from a single development stage could represent a gross underestimate of the prevalence of bivalent chromatin modification in plants. Future studies at cell-specific levels should more directly address the exact extent to which plant genes are bivalently modified. In any event, our results seem to indicate a mutually exclusive relationship between H3K4me3 and H3K27me3 at many genes in Arabidopsis seedlings. In animals, the H3K4 demethylase JARID1A (Jumonji, AT rich interactive domain 1A)/RBP2 (Retinol binding protein 2) is recruited to genomic targets through its interaction with the H3K27me3 methyltransferase complex Polycomb repressive complex (PRC) 2, where RBP2 mediates transcriptional repression by demethylating H3K4me3 to H3K4me2 (and to a lesser extent, H3K4me2 to H3K4me1) 4950. In addition, the H3K4me3-specific demethylase JARID1D interacts with Ring6a (Really interesting new gene 6a)/MBLR (Mel18 and Bmi1-like RING finger protein), which is closely related to the PRC1 components Bmi1 (B Lymphoma Mo-MLV insertion region 1) and Mel18 51. Interestingly, two Arabidopsis RING finger proteins, AtRING1a and AtRING1b, have been recently found to interact with the H3K27me3 methyltransferase CURLY LEAF and the H3K27me3-binding protein LIKE HETEROCHROMATIN PROTEIN1, and are required for the transcriptional repression of H3K27me3 target genes 52. The general mutual exclusion between H3K4me3 and H3K27me3 as well as the more frequent overlap of H3K4me2 and H3K27me3 suggest that similar mechanisms might also function in plants. That is, plant H3K4me3 demethylase(s) may function in transcriptional repression by interacting with PRC1 and/or PCR2. If this is the case, a fraction of the H3K4me2 in the Arabidopsis genome could be the demethylation product of H3K4me3.

We also observed that H3K4me1 tended not to co-localize with H3K27me3. One contributing factor could be the differential distribution patterns of these histone modifications along genes: H3K4me1 tends to be present at the 3' half of long genes, whereas H3K27me3 does not exhibit similar preferences for either location within genes or gene length (Figure S4 in Additional data file 1). Furthermore, H3K4me1 was present more frequently on ubiquitously expressed housekeeping genes, while H3K27m3 was more frequently present on tissue-specific genes.

Cytosine DNA methylation is an epigenetic silencing mechanism important for the developmental regulation of endogenous genes and the transcriptional silencing of transposons 53545556. A mechanistic relationship between DNA methylation and H3K4me has been described in mammals, where the DNA methyltransferase (DNMT) homolog DNMT3L specifically interacts with histone H3 containing unmethylated lysine 4 57. That DNMT3L also binds and stimulates the activity of the de novo DNA methyltransferase DNMT3A suggests that H3 with unmethylated K4 may play a role in targeting de novo DNA methylation in mammals 575859. However, a distinct small interfering RNA (siRNA)-directed pathway is responsible for de novo DNA methylation in plants 606162, and an interaction between DNA methyltransferase and histone has not been reported.

Three DNA methylation pathways have been described in plants: METHYLTRANSFERASE 1 (MET1) is a homolog of mammalian DNMT1 and primarily functions in maintaining DNA methylation in the CG sequence context ('CG methylation') 63646566. The DOMAIN REARRANGED METHYLASE (DRM) (homologous to mammalian DNMT3) interacts with the siRNA pathway and is required for de novo DNA methylation in all sequence contexts as well as the maintenance of DNA methylation in the CHH context (H = A, C or T; 'CHH methylation') 606162. The CHROMOMETHYLASE3 is specific to plant genomes and interacts with the H3K9me2 pathway to maintain DNA methylation in the CHG sequence context ('CHG methylation') 6768.

The genome-wide distribution of DNA methylation in Arabidopsis has been determined by a number of studies using microarray analyses or ultra-high-throughput deep sequencing of bisulfite treated DNA 2225697071727374757677. Results from these studies are largely consistent: CG, CHG and CHH methylation is highly enriched in transposons and other repetitive sequences, suggesting that the RNA interference, H3K9me2 and DNA methylation pathways function together in the transcriptional repression at these loci. DNA methylation is generally depleted in the promoters and 5' ends of endogenous genes. However, over one-third of Arabidopsis genes contain DNA methylation exclusively in the CG sequence context that is enriched in the 3' half of their transcribed regions (termed 'body-methylation'). Most body-methylated genes are expressed at moderate to high levels, and it is therefore unclear whether CG methylation alone in the transcribed regions of genes plays a direct and significant repressive role in transcription.

In order to determine the relationship between DNA methylation and H3K4me in Arabidopsis, we compared DNA methylation levels in genomic regions containing H3K4me to the whole-genome average of DNA methylation. As shown in Table 3, CHG and CHH methylation is significantly depleted in genomic regions containing H3K4me1, H3K4me2 or H3K4me3. CG methylation is also significantly depleted in H3K4me2- and H3K4me3-containing regions. In stark contrast, we found that CG methylation is highly enriched in H3K4me1-containing regions (Table 3). In addition, nearly two-thirds of H3K4me1-containing regions (8,841 of 14,599, approximately 60.6%) with two or more CG dinucleotides are methylated at two or more CG sites, compared to approximately 7.0% (842 of 12,100) and approximately 11.7% (1,750 of 14,918) for H3K4me2- and H3K4me3-containing regions, respectively.

The low level of CHG and CHH methylation in H3K4me-containing regions can be explained by the virtual absence of siRNAs and H3K9me2 within actively transcribed endogenous genes. The lack of CG methylation in H3K4me2- and H3K4me3-containing regions could be due to an active mutual exclusion mechanism (for example, MET1 may be discouraged from chromatin containing H3K4m2 or H3K4me3) similar to what was recently described between DNA methylation and the deposition of the histone variant H2A.Z 78, or simply the differential localization of DNA methylation and H3K4me2/H3K4me3 relative to genes (a 5' bias for H3K4me2/H3K4me3 and a 3' bias for DNA methylation). The high level of CG methylation in H3K4me1-containing regions was unexpected. It is possible that CG methylation and H3K4me1 interact with each other and therefore co-localize at the 3' transcribed regions of genes. It is also possible that the overlap of these two epigenetic marks merely reflects their preferential localization in the similar regions of highly expressed genes. In either case, these results indicate that CG methylation per se and H3K4me1 do not appear to interfere with each other. Finally, genomic regions free of H3K4me frequently lack DNA methylation, suggesting that the absence of H3K4me alone is insufficient to trigger DNA methylation.

In order to test whether direct mechanistic links exist between DNA methylation and H3K4me (that is, whether DNA methylation per se excludes H3K4me2/H3K4me3 and whether gene body methylation facilitates H3K4me1), we determined the genome-wide distribution of H3K4me in the met1 mutant by ChIP-chip. Previous studies have shown that loss of MET1 eliminates CG methylation as well as substantial fractions of CHG and CHH methylation, resulting in massive transcriptional reactivation of transposons 7172747677.

All three types of H3K4me were found to be present at much higher levels in the pericentromeric heterochromatin regions in met1 (Figure 4). A closer examination revealed that hyper-H3K4me in met1 is almost always associated with ectopic over-expression of transposons or pseudogenes (Figure 4). However, the loss of DNA methylation does not appear to directly trigger hyper-H3K4me. In contrast to the transcription-independent ectopic accumulation of H2A.Z in DNA hypomethylated regions in the met1 mutant 78, no major change in H3K4me was observed in genomic regions that are DNA-hypomethylated but not transcribed. This suggests that the ectopic transcriptional activity resulting from the loss of DNA methylation, but not the loss of DNA methylation per se, is associated with hyper-H3K4me. In addition, at nearly all genes, a complete loss of gene body methylation in met1 had no significant effect on H3K4me1, H3K4me2 or H3K4me3 (Figure 4), suggesting that CG methylation in genes is dispensable for the normal accumulation of H3K4me1.