An Arabidopsis jmjC domain protein protects transcribed genes from DNA methylation at CHG sites.

Differential cytosine methylation of genes and transposons is important for maintaining integrity of plant genomes. In Arabidopsis, transposons are heavily methylated at both CG and non-CG sites, whereas the non-CG methylation is rarely found in active genes. Our previous genetic analysis suggested that a jmjC domain-containing protein IBM1 (increase in BONSAI methylation 1) prevents ectopic deposition of non-CG methylation, and this process is necessary for normal Arabidopsis development. Here, we directly determined the genomic targets of IBM1 through high-resolution genome-wide analysis of DNA methylation. The ibm1 mutation induced extensive hyper-methylation in thousands of genes. Transposons were unaffected. Notably, long transcribed genes were most severely affected. Methylation of genes is limited to CG sites in wild type, but CHG sites were also methylated in the ibm1 mutant. The ibm1-induced hyper-methylation did not depend on previously characterized components of the RNAi-based DNA methylation machinery. Our results suggest novel transcription-coupled mechanisms to direct genic methylation not only at CG but also at CHG sites. IBM1 prevents the CHG methylation in genes, but not in transposons.

Introduction

Transposons and repeats, which are major deleterious constituents in genomes of vertebrates and plants, are epigenetically silenced by DNA methylation (Yoder ; Walsh ; Miura ; Singer ; Kato ; Bender, 2004; Chan ; Gehring and Henikoff, 2007). A central question in epigenetics is how the methylation machinery distinguishes between transposons and cellular genes.

The flowering plant Arabidopsis thaliana serves as a powerful model system to understand control of DNA methylation through genetic and genomic approaches. Genome-wide analyses of DNA methylation in Arabidopsis reveal that transposons are heavily methylated at both CG and non-CG sites, whereas the methylation in genes is at much lower level and limited to CG sites (Lippman ; Zhang ; Zilberman ; Cokus ; Lister ) (http://neomorph.salk.edu/epigenome.html). The methylation at CG sites is maintained by DNA methyltransferase MET1 (Finnegan ; Kankel ; Saze ). Methylation at non-CG sites depends on DNA methyltransferases CMT3 and DRM2; the former mainly methylates symmetrical CHG sites, whereas the latter can methylate asymmetrical CHH sites (H can be A, C, or T) (Bartee ; Lindroth ; Cao and Jacobsen, 2002). CMT3-dependent CHG methylation is guided by methylation of histone H3 at lysine 9 (H3K9me) (Jackson ; Malagnac ). H3K9me is an epigenetic mark of silent ‘heterochromatin' conserved from plants to fungi and animals (Bender, 2004; Chan ; Grewal and Elgin, 2007). DNA methylation at specific sequence can be induced by RNA of homologous sequence. This process, called RNA-directed DNA methylation (RdDM), requires de novo methylase DRM2, members of RNAi machinery, such as RDR2, DCL3, AGO4, RNA polymerase IV, as well as a putative chromatin remodeller (DRD1) and an SMC-domain containing protein (DMS3) (Chan ; Matzke ; Kanno ). The RdDM generally affects CG, CHG, and CHH sites, and the CHH methylation is normally associated with small RNA, consistent with the involvement of RNAi machinery.

In addition to these well-investigated positive regulators of DNA methylation, negative regulators of DNA methylation have been recently identified (Agius ; Gehring ; Saze ). Through a genetic screen for mutants with increased methylation of a gene called BONSAI, we have previously identified the jumonji-domain-containing protein IBM1 (increase in BONSAI methylation 1) (Saze and Kakutani, 2007; Saze ). IBM1 is a member of the JHDM2/KDM3 family, which is comprised of demethylases of H3K9me conserved from plants to mammals (Klose ; Lu ; Sun and Zhou, 2008). The ibm1 mutation induces a variety of developmental abnormalities, which are suppressed by mutants of the histone H3K9 methylase KYP/SUVH4 gene and CHG methylase CMT3 gene. The results suggest that ectopic deposition of the chromatin marks, H3K9me and CHG methylation, is responsible for the ibm1-induced developmental abnormalities (Saze ). However, actual genomic targets of the IBM1 protein have not been identified, with the exception of the BONSAI gene. The spectrum of genomic targets of IBM1 remained unknown.

Here, we directly determined genomic targets of IBM1 by high-resolution genome-wide analysis of DNA methylation. The ibm1 mutation induced extensive hyper-methylation in thousands of genes throughout the euchromatic chromosomal arm regions. The hyper-methylation was specific for CHG sites within genes. Unexpectedly, long transcribed genes were most severely affected. The results suggest that not only CG but also CHG methylation in the genes is directed by transcription-coupled mechanisms. We propose that IBM1 protects transcribed genes from the CHG methylation and maintains genome integrity by distinguishing between genes and transposons.

Results

Genome-wide analysis of DNA methylation revealed targets of IBM1 protein

To understand the mode and significance of the control of DNA methylation by the IBM1 protein, we determined DNA methylation across the whole genome in the ibm1 mutant. For that purpose, we adapted a genomic tiling array method combined with immunoprecipitation of the genomic DNA using an anti-methylcytosine antibody. The methylation level was assessed by the hybridization signal ratio of immunoprecipitated DNA to the input DNA. Pericentromeric regions, which are rich in transposons and repeats, show more methylation signal than the gene-rich chromosome arms in the wild-type sample, as has been reported previously (Lippman ; Zhang ; Zilberman ; Cokus ; Lister ). Compared with the wild type, the ibm1 mutation caused a global increase in the signal intensity over the entire chromosome arms (Figure 1A and B).

Figure 1

Global view of DNA methylation in the ibm1 mutant and wild-type Columbia plants. (A) Comparison along the chromosomes. Vertical values represent log2 of the immunoprecipitated DNA signal divided by input control signal. Each point reflects sliding window of 300 kb (∼3000 probes). In each chromosome, position of centromere is indicated by black triangle on the x-axis. (B) The difference of the signal between the ibm1 mutant and wild type. (C, D) Methylation of transcription units compared between the ibm1 and wild-type plants. To evaluate significance of the methylation in each transposon or gene, we calculated ‘methylation significance index (MSI)' (see Materials and methods). Most of the transposons were unaffected (C), whereas a subset of genes showed increase in DNA methylation in ibm1 (D). (E) Distribution of difference of the MSI between ibm1 and wild type for genes (grey) and transposons (black). The x value is expressed using standard deviation of differences in the signals between wild type and ibm1 for every probes.

At higher resolution, the comparison revealed that the ibm1-induced hyper-methylation is not uniform within the arms. Interestingly, each region of increased DNA methylation coincides with a transcription unit (Figure 2A and D; Supplementary Figures S2–S13). We therefore calculated the methylation level for each transcription unit based on the recent annotation of the Arabidopsis genome (TAIR7: http://www.arabidopsis.org). When ibm1 and wild type were compared, thousands of genes showed a significant increase in DNA methylation signal (Figure 1D and E). On the other hand, transposons did not show comparable changes (Figure 1C and E). In independent experiments using different plants and different probe amplification methods, the wild-type and ibm1 samples showed consistent results for each gene (Supplementary Figure S1). For further analysis, we selected 3112 genes (651 species of class I genes and 2461 class II genes; with stronger effects in class I genes) that reproducibly showed the most dense increase in methylation signal induced by the ibm1 mutation (Supplementary Figure S1; Supplementary Table S1).

Figure 2

DNA methylation status of the ibm1 mutant and wild-type Columbia plants in 12 representative loci, which are analysed by bisulphite sequencing. For all the 12 loci, the array results with biological replica and full bisulphite sequencing results are shown in Supplementary Figures S2–S13. (A, D) DNA methylation analysed by tiling array for two loci (A, position 172 500–277 500; D, position 5 745 000–5 835 000). Blue boxes indicate genes, which oriented 5′–3′ on the top and the reverse in the bottom. Each vertical green bar represents the log2 signal of the immunoprecipitated DNA divided by input control. Magenta bars represent ibm1/WT ratios of signals. (C, F) Results of bisulphite sequencing in wild type and ibm1 mutants within two genes shown in (A, D). Results of sense strands are shown. Results for both sense and antisense strands are shown in Supplementary Figures S2–S4. The percentage of methylated cytosines at each site is indicated by vertical bar (red, CG; blue, CHG; black, asymmetric). Exons and introns are shown by black boxes and lines between the results of wild type and ibm1 mutant. (B, E, G–P) The percentage of methylated cytosines found in different contexts at genic sequences (B, E, G–M) and transposon sequences (N–P) in wild type and ibm1.

To validate the tiling array results, DNA methylation was determined at the nucleotide level by the bisulphite sequencing method for 12 loci (6 class I genes, 3 class II genes, and 3 transposons). For all the examined class I and II genes, the bisulphite sequencing confirmed the increase in DNA methylation detected by the tiling array (Figure 2; Supplementary Figures S2–S10); extensive hyper-methylation was noted. The bisulphite sequencing results also revealed that the increase in DNA methylation was mainly at CHG sites (Figure 2). This sequence specificity contrasts with genic methylation in wild type, which is almost exclusively at CG sites (Cokus ; Lister ) (http://neomorph.salk.edu/epigenome.html) (Figure 2B, G–M). Both sense and antisense strands of the genes were affected similarly (Figure 2B, E and G; Supplementary Figures S2–S4). On the other hand, transposons were methylated at CG, CHG, and CHH sites already in wild type, and this methylation was not affected by the ibm1 mutation (Figure 2N–P; Supplementary Figures S11–S13). The bisulphite sequencing data confirmed that the targets of the IBM1 function were not transposons but genes, as was suggested from the tiling array results (Figure 1).

In Arabidopsis, another class of negative regulators of DNA methylation has been identified; DEMETER, ROS1, DML1, and DML2 are structurally related DNA demethylases (Agius ; Gehring ). Triple mutation of ros1, dml1, and dml2 (rdd) causes increased methylation in ∼200 genes (Penterman ; Lister ). The rdd triple mutation can induce hyper-methylation in the 5′ and 3′ regions of the gene. In contrast, the ibm1 mutation most severely affected the central region of transcription units, whereas the 5′ and 3′ terminal regions were least affected (Figure 3). This bias is similar to the distribution of genic CG methylation in wild type (Zhang ; Zilberman ). In the ibm1 mutant, the bias was enhanced by including extensive CHG methylation.

Figure 3

Distribution of DNA methylation. Each gene was divided into five regions (5′ and 3′ 500-bp segments, and three internal regions with equal lengths) with 1000-bp flanking segments. (A) Average of signals in probes in each segment was plotted for ibm1 and wild type. In total, 862 genes hyper-methylated in ibm1 and longer than 2 kb were analysed. (B) Difference of ibm1 and wild type.

Long transcribed genes were most severely affected by the ibm1 mutation

Each transcription unit seems to be affected in a coordinated manner; some transcription units are affected over the entire gene, whereas others are not affected at all (Figures 1, 2A and D; Supplementary Figures S2–S13). This possibility was examined quantitatively by comparing the effects of the ibm1 mutation between the 5′ and 3′ halves of a transcription unit. This analysis revealed that they indeed behave in a coordinated manner (Supplementary Figure S14). The correlation between the 5′ and 3′ halves of a transcription unit was significant (r=0.67). In other words, when the ibm1 mutation affected the 5′ half of one gene, then the 3′ half of that gene tended to be affected as well. This behaviour was not simply due to the spread of hyper-methylation to nearby sequences, because the correlation with flanking regions outside the gene was much lower; r=0.29 for the 5′ regions and r=0.36 for the 3′ regions (Supplementary Figure S14).

In wild-type plants, moderately transcribed genes are most likely to be methylated, whereas genes at either extreme with regard to the expression level are less likely (Zhang ; Zilberman ). Notably, the ibm1 mutation affected moderately transcribed genes and constitutively expressed genes most severely (Figure 4A–D). The spectrum of affected genes was consistent with that seen for genic CG methylation in wild type, and the methylation was also found at CHG sites in the ibm1 mutant. In addition, some genes unmethylated in wild type also showed hyper-methylation in the ibm1 mutant (Figures 1D and 2D–F). Interestingly, genes unmethylated in wild type and hyper-methylated in ibm1 were also transcribed in wild type (Supplementary Figure S15). Overall, the methylation of transcribed genes was much enhanced by the ibm1 mutation (Figure 4).

Figure 4

Transcription and effect of ibm1 on DNA methylation. (A, B) Bias to moderately transcribed genes. All genes were rank-ordered and binned based on the sum of expression in all of Arabidopsis tissues (see Supplementary data). The vertical value shows mean methylation signal for genes in each percentile. (C, D) Bias to constitutively expressed genes. Same as (A, B), except that the genes were rank-ordered based on tissue specificity of expression measured by entropy level (see Supplementary data).

Another notable feature of the effect of the ibm1 mutation is that longer genes showed a much stronger response (Figure 5A and B). For genes longer than 5 kb, about two-thirds belong to ibm1-affected genes in class I or II, and the class I genes were enriched more than 10-fold compared with the total population of genes (Figure 5B). As short genes tend to have fewer introns, we examined whether the intron number mediates the effect of the gene length on ibm1-induced hyper-methylation. However, even within the genes of the same intron number, longer genes tend to be affected severely by the ibm1 mutation (Supplementary Figure S16).

Figure 5

Long and expressed genes were affected by ibm1. (A) Relationship between the gene length and increase in methylation in ibm1. The methylation change shows mean of ibm1/WT signals in each gene. (B) ibm1 affects long genes. Genes were divided into five groups based on length. The proportions of class I–III genes are shown by different colours. These three classes were categorized on the basis of the mean of log2(ibm1/WT) signals in each gene, with class I being the most severely affected and class III the least affected by ibm1 (Supplementary Figure S1). (C, D) Genes shown in (A) were divided into high expressers (15 000 genes in (C)) and low expressers (10 620 genes in (D)) based on the transcript level in leaves. The effect of the length is less evident in the latter population.

The severe response of long genes was less evident in genes with low levels of expression (Figure 5D), suggesting that the effect of the gene length on the response to the ibm1 mutation also depends on the transcription.

The ibm1-induced CHG methylation did not depend on previously characterized RNAi-based components for DNA methylation

The results presented above suggest that the genic methylation of both CG and CHG sites can be induced by transcription-coupled mechanisms, and that the latter is masked by the function of IBM1 protein, which contributes to the differential methylation of genes and transposons. A question would be how transcription could affect genic methylation. A change in the chromatin state might be induced by passage of the transcription machinery itself, or by the action of the resultant transcript.

For possible effects of the RNA on chromatin states, the most extensively studied system in plants is RdDM, which is connected to small RNA. We therefore tested whether known components of the RdDM process, namely, DRM2, RDR2, RDR6, AGO4, and NRPD1a, are necessary for the ibm1-induced hyper-methylation or not. Methylation in the double mutants was examined for two target genes of IBM1, BONSAI and ERL2 (ERECTA-LIKE 2) (Figure 6; Supplementary Figure S17). Interestingly, in all the double mutants, the ibm1-induced hyper-methylation was detected. The results suggest that the ibm1-induced CHG methylation does not depend on any of these previously characterized RNAi-based components of the DNA methylation machinery. Taken together, these results suggest that novel transcription-coupled mechanisms direct genic methylation not only at CG but also at CHG sites.

Figure 6

Known components of RNAi-based DNA methylation are dispensable for the ibm1-induced genic hyper-methylation. DNA methylation status of the BONSAI gene was examined by methylation-sensitive restriction digestion and subsequent PCR. Detailed conditions were as described previously (Saze and Kakutani, 2007). The band can be detected when the restriction site is methylated and undigestible. Essentially the same results were obtained for another target gene, ERL2 (ERECTA-LIKE 2, At5g07180) (Supplementary Figure S17). (A) Four types of homozygotes segregating in self-pollinated progeny of a double heterozygote IBM/ibm1, DRM2/drm2. The double mutant plants also showed hyper-methylation. (B–E) Methylation status was examined in double mutants with RDR6, NRPD1a, RDR2, or AGO4 gene. Plants in F3 generation were examined in (C).

Discussion

In this study, we determined genomic targets of the jmjC domain protein IBM1 by genome-wide analysis of DNA methylation. The ibm1 loss-of-function mutation induced extensive hyper-methylation in thousands of genes. Long transcribed genes were most severely affected. Methylation of transcribed genes is limited to CG sites in wild type, but those genes were also methylated at CHG sites in the ibm1 mutant. Thus, the IBM1 protein prevents CHG methylation of transcribed genes. These observations broaden our understanding of the control of genic methylation, but important questions remain, as illustrated below.

Why are long and transcribed genes body-methylated?

It has been shown previously that transcribed genes tend to be body-methylated at CG sites (Zhang ; Zilberman ). Here, we report that, without the IBM1 function, transcribed genes are also body-methylated at CHG sites. For the body methylation of transcribed genes, Zilberman proposed that disruption of chromatin during the passage of transcription machinery allows activation of cryptic promoters, resulting in aberrant transcript formation, and subsequent DNA methylation. With regard to the ibm1-induced hyper-methylation of gene body, the entire transcription unit behaved in a coordinated manner (Supplementary Figure S14). If sporadic events, such as activation of cryptic promoters, affect the entire gene, a long gene would have more chance to initiate such an event. This prediction is consistent with our observation that long genes tend to be affected most severely by the ibm1 mutation (Figure 5).

On the other hand, our genetic analyses suggest that the ibm1-induced hyper-methylation does not depend on previously characterized components of RdDM, which include RNAi machinery and the de novo DNA methylase DRM2 (Figure 6; Supplementary Figure S17). Even if aberrant RNA is involved in genic methylation, the latter might use an as yet uncharacterized pathway.

Alternatively, the passage of transcription machinery itself, rather than the RNA produced, might mediate a change in the chromatin state and facilitate gene body methylation. In yeast, transcription machinery can be associated with histone methylases, such as SET1 and SET2 (Krogan ; Ng ). SET2-induced H3K36 methylation recruits histone deacetylase complex, which negatively regulates the transcription (Keogh ). Similar mechanisms might also be used for generating repressive histone marks such as H3K9me directly or indirectly. In addition, passage of transcription machinery can enhance histone replacement, which may also affect chromatin structure (Martin and Zhang, 2007).

The ibm1 mutation induces genic CHG methylation, which is rarely found in wild type. Interestingly, the genic CHG methylation is also induced in the mutant background of CG methylase gene MET1 (Cokus ; Lister ). The CHG methylation in the met1 mutant might function to rescue the deleterious effects of the loss of the CG methylation (Mathieu ). We compared the spectrum of genes showing CHG methylation between the met1 and ibm1 mutants. The correlation was found to be significant between genes showing CHG hyper-methylation in ibm1 and genes with CG hypo-methylation and CHG hyper-methylation in the met1 mutant (Supplementary Figure S18), suggesting that IBM1 might be linked to a compensatory methylation pathway induced by CG hypo-methylation.

Is the CHG hyper-methylation sufficient for silencing the gene?

Considering the large impact of the ibm1 mutation on gene body methylation shown in this study, it may be surprising that effects of the ibm1 mutation on developmental phenotypes are not catastrophic (Saze ). Indeed, in our preliminary results on gene expression, the CHG hyper-methylation of gene body induced by the ibm1 mutation is not always associated with transcriptional silencing. Within the genes hyper-methylated by ibm1 mutation, some showed a decrease in the transcript level, whereas others were unaffected or even showed an increased expression (unpublished). On the other hand, in the wild-type background, non-CG methylation is not found in active genes, but found only in silent transposons and repeats. Of the transposon-specific non-CG methylation, CHG methylation is connected to H3K9me (Jackson ; Malagnac ), whereas CHH methylation is associated with small RNA (Matzke ). If CHG methylation is not sufficient for the gene silencing, interaction and/or collaboration with other layers of epigenetic modifications might be important for generating silent heterochromatin.

In this context, it may be interesting that double mutant of the IBM1 gene and a chromatin remodelling gene DDM1 (decrease in DNA methylation 1) shows much more severe developmental phenotypes than either single mutant (Saze ). As is the case for the ibm1 mutation, the ddm1 mutation induces methylation of the BONSAI gene. Unlike ibm1, however, ddm1 mutation causes BONSAI hyper-methylation in all the three contexts: CG, CHG, and CHH (Saze and Kakutani, 2007). In addition, although the ddm1-induced BONSAI methylation depends on the presence of a LINE retrotransposon nearby (Saze and Kakutani, 2007), our genome-wide analysis revealed that genes affected by the ibm1 mutation are not necessarily located near transposons (Supplementary Table S2). Targets of IBM1 are low copy genes, and repetitive sequences were unaffected (Figures 1 and 2). On the other hand, the direct targets of the DDM1 gene appears to be repeats (Vongs ; Lippman ), and the effect of ddm1 mutations on low copy sequences, such as BONSAI, might be indirect (Saze and Kakutani, 2007). Even if ibm1 and ddm1 mutations trigger the hyper-methylation by different mechanisms, these pathways may converge on overlapping targets, with possible interactions of different layers of epigenetic marks. The possible interactions may be clarified at the genome level by combination of genomics and genetics using these and other Arabidopsis mutations.

Materials and methods

Plant materials

Isolation and initial characterization of the ibm1 mutants were described previously (Saze ). A presumed null allele ibm1-4, which has T-DNA insertion in the coding region, was used throughout. Alleles or origins of the other mutants are rdr6-11, ago4-1, SALK059661 (RDR2), SALK128426 (NRPD1a), and cs6366 (DRM2).

Array design and data analysis

We used NimbleGen 3 × 385K array (Zilberman ), which covers the entire sequenced Arabidopsis genome with an interval of ∼100 bp. Details of the experimental procedures are described in the Supplementary data. The methylation signal for each probe was represented as log2 signal ratio of immunoprecipitated DNA to input DNA. Methylation of a gene (Figures 3, 4 and 5; Supplementary Figures S1 and S15) is represented by the average of the log2 signal for all probes covered by the gene. In the results shown in Figure 1C–E, we used ‘methylation significance index' of a gene, which is the sum of methylation signal for the probes covered by the gene divided by root of the probe number, because fluctuation of that value would be independent of the probe number per gene, assuming that the signal for each probe fluctuates independently.

Bisulphite sequencing

Bisulphite sequencing was performed as described previously (Saze ). In total, 38 amplicons were sequenced for the 12 loci examined. At least 12 clones were sequenced for each amplicon. Restriction enzymes and primer sequences used for the bisulphite sequencing are shown in Supplementary Table S3.