Isoform-specific localization of DNMT3A regulates DNA methylation fidelity at bivalent CpG islands

Abstract

DNA methylation is a prevalent epigenetic modification involved in transcriptional regulation and essential for mammalian development. While the genome-wide distribution of this mark has been studied to great detail, the mechanisms responsible for its correct deposition, as well as the cause for its aberrant localization in cancers, have not been fully elucidated. Here, we have compared the activity of individual DNMT3A isoforms in mouse embryonic stem and neuronal progenitor cells and report that these isoforms differ in their genomic binding and DNA methylation activity at regulatory sites. We identify that the longer isoform DNMT3A1 preferentially localizes to the methylated shores of bivalent CpG island promoters in a tissue-specific manner. The isoform-specific targeting of DNMT3A1 coincides with elevated hydroxymethylcytosine (5-hmC) deposition, suggesting an involvement of this isoform in mediating turnover of DNA methylation at these sites. Through genetic deletion and rescue experiments, we demonstrate that this isoform-specific recruitment plays a role in de novo DNA methylation at CpG island shores, with potential implications on H3K27me3-mediated regulation of developmental genes.

Keywords: CpG islands; DNA methylation; DNMT3A; H3K27me3; Polycomb.

Figure 1. The DNMT3A isoforms DNMT3A1 and DNMT3A2 display differential localization along the mouse ES cell genome

1. Similarities in domain composition and amino acid sequence between murine DNMT3A and DNMT3B. Conserved amino acids are indicated as red bars. The longer and shorter DNMT3A isoforms DNMT3A1 and DNMT3A2 are shown.

2. Differential expression of DNMT3A isoforms measured by CAGE‐seq data obtained from the FANTOM5 consortium. Indicated are data points obtained from various tissues during early mouse development at the indicated embryonic days (E) and from adult animals. Shown are log₂‐transformed CAGE‐seq read counts overlapping the isoform‐specific promoters (see also Appendix Fig S1D and E).

3. Average profiles indicating DNMT3A1 binding around genomic regions containing unmethylated, low methylated, and fully methylated regions (UMR, LMR, and FMR, respectively, from Stadler et al, 2011). DNA methylation at these sites is indicated in gray (% m‐CpG); binding of DNMT3A1 is indicated in red (average read counts from ChIP‐seq data). At FMRs, DNMT3A1 binding is shown at three FMR groups stratified by methyl‐CpG‐density to indicate binding dependency on the density of methylated CpGs (see also Appendix Fig S3C and D).

4. Comparative binding analysis between DNMT3A1, DNMT3A2, and DNMT3B indicates lack of DNMT3A recruitment to transcribed gene bodies. Shown are average density plots for ChIP‐seq read counts around non‐overlapping genes scaled by gene length.

5. Representative genome browser view exemplifying differences in binding between the de novo methyltransferases. Shown are read counts per 100 bp for ChIP‐seq and input samples. Gene models, CpG islands, and repetitive elements from the UCSC genome browser are indicated below. Top track indicates differential binding between DNMT3A1 (red) and DNMT3A2 (orange).

6. Genome‐wide cross‐correlation analysis between de novo DNMT binding and various chromatin modifications. Pearson's correlations are calculated on log₂‐transformed read counts over 1‐kb‐sized tiles covering the mouse genome. Increased correlations between DNMT3A1 and H3K27me3, as well as for DNMT3B with transcriptionally active features, are indicated by red and blue boxes, respectively.

Figure EV1. DNMT3A1 localization around Polycomb CpG islands

1. Venn diagram indicating overlapping and individual binding sites for the de novo DNA methyltransferases identified from 1‐kb‐sized tiles covering the entire genome.

2. Averaged DNA methylation % around DNMT3A1‐bound sites (in gray) according to their position upstream (5′) or downstream (3′) of neighboring CpG island promoters. Blue line indicates DNMT3A1 sites downstream of the 3′‐end, and orange line upstream of the 5′‐end of CGI promoters.

3. Density plots indicating the distance between de novo DNA methyltransferase‐enriched sites (red: DNMT3A1, orange: DNMT3A2, blue: DNMT3B) and CpG island promoters separated by H3K27me3. Shaded area indicates a distance shorter than 5 kb, and the percentage of bound sites within that distance is indicated for all DNMT3 proteins in the corresponding color.

4. GREAT (McLean et al, 2010) analysis of DNMT3‐binding sites indicating the enriched Molecular Function Gene Ontology term of nearby genes. DNMT3A1 frequently associates with or near promoters of DNA sequence‐dependent transcriptional regulators.

Figure 2. The DNMT3A1 isoform preferentially localizes to H3K27me3‐bivalent CpG islands

1. Enrichment analysis of chromatin features under DNMT3‐protein‐binding sites reveals increased H3K27me3 localization at sites bound by DNMT3A1. Shown are average log₂‐enrichments of the indicated chromatin modifications and factors over input at sites exclusively bound by individual de novo DNMTs.

2. Box plots indicating log₂‐enrichment over input for H3K27me3 and H3K36me3 at DNMT3A1‐, DNMT3A2‐, and DNMT3B‐binding sites, and compared to average enrichment in the entire genome. Boxes denote the inter‐quartile range (IQR) and whiskers 1.5 × IQR.

3. Scatter plots indicating the specific binding preference of DNMT3A1 for H3K27me3 and DNMT3B for H3K36me3. Shown are genome‐wide enrichments over input for these histone modifications highlighting their mutual exclusiveness (gray) and the location of DNMT3A1 (red)‐, DNMT3A2 (orange)‐, and DNMT3B (blue)‐bound sites along these genomic distributions.

4. DNA methylation levels and CpG densities differ around de novo DNMT‐binding sites. Top: Shown are average methylation values for CpGs centered around DNMT‐binding sites (gray area). Bottom: Shown are CpG densities calculated as number of CpG dinucleotides per 100 bp at DNMT3‐binding sites. Average densities are calculated for exclusive binding sites of DNMT3A1 (red), DNMT3A2 (orange), and DNMT3B (blue).

5. DNMT3A1 preferentially localizes upstream and downstream of bivalent CpG island promoters. Shown are heat map density profiles for DNMT3 protein binding and chromatin features around all promoter‐associated CpG islands and CpG promoters separated by H3K27me3. CpG island promoters are oriented according to their downstream gene.

Figure 3. Dynamic colocalization of DNMT3A1 with H3K27me3 during neuronal differentiation

1. Both DNMT3A isoforms dynamically localize to de novo methylated regions during neuronal development. Increased DNMT3A binding in neuronal progenitors (NP) is shown as log₂‐fold difference between ES and NP cells (y‐axis) and is ranked based on de novo DNA methylation in NPs (x‐axis). Individual data points denote median log₂‐fold changes calculated from 500 consecutive windows ranked by DNA methylation.

2. Scatter plots indicate that genome‐wide dynamics in DNMT3A1 binding correlates with regions that change H3K27me3 during neuronal differentiation. Shown are log₂‐fold changes between ES and NP ChIP‐seq signals of DNMT3A isoforms and H3K27me3 calculated at 1‐kb windows positive for H3K27me3 in ES and/or NP cells.

3. Genome browser examples for regions that change DNMT3A1 binding and DNA methylation according to H3K27me3. Shown are read counts per 100 bp for ChIP‐seq datasets and percentage of DNA methylation per individual CpGs. Regions that show concomitant changes in DNMT3A1 and DNA methylation are highlighted in gray.

Figure 4. The N‐terminal part of DNMT3A1 is required for localization to H3K27me3 sites

1. Scatter plot indicating similarities in protein–protein interactions between DNMT3A1 and DNMT3A2. Shown are the unique spectral counts per protein obtained from co‐immunoprecipitation experiments performed in cell lines expressing biotin‐tagged DNMT3A1 or DNMT3A2. The Pearson's correlation coefficient is shown. Red dots indicate proteins listed in (B).

2. Table indicating top‐scoring proteins identified to interact with DNMT3A1 or DNMT3A2 in mouse ES cells. Shown are the results from (A) and an untagged cell line (empty) that serves as background control.

3. Genome browser example for regions indicating reduced binding of the chimeric N3B‐DNMT3A protein to H3K27me3 sites. Shown are read counts per 100 bp for ChIP‐seq datasets and percentage of DNA methylation per individual CpGs.

4. Average density plot around H3K27me3‐positive and H3K27me3‐negative CpG island promoters indicates reduced binding of the chimeric N3B‐DNMT3A protein to H3K27me3 sites (see Appendix Fig S6B).

Figure 5. Binding of DNMT3A1 to CpG island shores is required for DNA methylation regulation and promotes 5‐hmC deposition

1. Heat map profiles of DNMT protein binding and chromatin features around promoter‐associated unmethylated regions (UMR) oriented by gene direction and ranked by H3K27me3 show a preferential association of DNMT3A1 with H3K27me3‐decorated UMRs. Furthermore, H3K27me3‐positive UMRs are enriched for 5‐hmC signals flanking the unmethylated center.

2. Average density profiles for top 25% of UMRs ranked by H3K27me3 indicate a localized 5‐hmC signal at borders of UMRs that is localized between DNMT3A1‐ and TET1‐binding sites (top). Top 25% of UMRs ranked by H3K27ac lack DNMT3A1 enrichment and have a more uniform 5‐hmC signal (bottom). Shown are library‐normalized read counts.

3. Average density profiles for UMR regions clustered based on DNMT3A1 binding. Five clusters were defined by k‐means clustering that show distinct DNMT3A1 binding (d 1–5, see also Fig EV2 for more details). Clusters with highest DNMT3A1 binding are also enriched for H3K27me3 and also display elevated 5‐hmC signals at UMR borders.

4. Average density plots centered over UMR borders identify a sharp signal in 5‐hmC at the transition from FMR to UMR states in wild‐type cells. This signal is reduced in ES cells lacking DNMT3A and is absent in the Dnmt‐TKO cell line.

5. DNA methylation is reduced around bivalent CpG islands in DNMT3A‐KO ES cells. Shown is the average DNA methylation measured by RRBS around UMR borders.

Figure EV2. Preferential DNMT3A1 binding to CpG island shores coincides with elevated Polycomb activity and TET‐mediated oxidation

1. Heat map profiles for all promoter‐associated UMRs clustered by DNMT3A1 binding and ranked by size. Shown are DNMT3A1 and various chromatin features clustered and ordered according to DNMT3A1 binding into five clusters. Note that clusters 1 and 4 contain the highest enrichment for DNMT3A1, whereas DNMT3A1 binding in cluster 2 is moderate and occurs upstream of the UMR.

2. Box plots indicating enrichment of Polycomb marks and proteins within UMRs at clusters preferentially bound by DNMT3A1. Shown are log₂‐enrichments over input. Boxes denote the inter‐quartile range (IQR) and whiskers 1.5 × IQR. Notches display the confidence interval around the median.

3. Average density plots indicate elevated transcriptional activity at clusters 2, 3, and 5. Shown are binding profiles for RNA polymerase II phosphorylated on serine 2 and H3K36me3 as markers of transcriptional activity. In addition, H3K4me1 that scales similarly to 5‐hmC at DNMT3A1‐enriched clusters is shown (compare to Fig 5C).

4. Average density plots for DNA methylation in wild‐type cells measured by WGBS around UMRs clustered based on DNMT3A1 binding.

Figure 6. De novo DNA methylation activity of DNMT3A1 is directed to bivalent CpG island shores

1. A

   Analysis of de novo DNA methylation in Dnmt‐TKO cells re‐expressing DNMT3A1 shows preferential DNA methylation activity at DNMT3A1‐binding sites. Exclusive binding sites from DNMT3A1, DNMT3A2, and DNMT3B are shown in red, orange, and blue, respectively.

2. B

   Expression of DNMT3A1, but not DNMT3A2, in Dnmt‐TKO cells results in preferential de novo DNA methylation around bivalent CpG islands. Density plots show average methyl‐CpG percent calculated from WGBS data around UMRs binned based on H3K27me3 or H3K27ac enrichment (top 25% ranked UMRs).

3. C, D

   Re‐expression of DNMT3A1 but not DNMT3A2 results in 5‐hmC accumulation at bivalent CpG island shores. (C) Average density plots indicating 5‐hmC read counts at UMR borders separated by H3K27me3 or H3K27ac from two independent replicates. (D) Representative genome browser views.

4. E

   Percentage of genes with significant increase in gene expression in Dnmt‐TKO over wild‐type cells (FDR < 0.001 & log₂‐FC > 0.5) according to their promoter classification. Promoters with more than 80% DNA methylation in wild‐type cells (black); promoters with < 20% DNA methylation in wild‐type cells (gray); or unmethylated CpG island promoters falling into one of five clusters based on DNMT3A1 binding (see Figs 5C and EV2A). P‐values calculated based on hypergeometric tests are shown only for fractions with P‐value < 0.05.

5. F

   DNMT3A1 is preferentially recruited to the shores of bivalent CpG islands to maintain DNA methylation in the presence of TET‐mediated oxidation of methyl‐cytosines to 5‐hmC.

Figure EV3. Polycomb‐regulated CpG island promoters bound by DNMT3A1 are frequently deregulated in the absence of DNA methylation

1. Box plots showing reduction in H3K27me3 in Dnmt‐TKO and Dnmt3a/Dnmt3b‐DKO cells at UMR promoters clustered based on DNMT3A1 enrichments (clustering based on Fig EV2A). Boxes denote the IQR and whiskers 1.5 × IQR. Notches display the confidence interval around the median.

2. H3K27me3 reduction and spreading in Dnmt3a/3b‐DKO cells is partially re‐established upon re‐expression of DNMT3A1. Genome browser examples for regions with dynamic H3K27me3 in Dnmt‐TKO cells and Dnmt‐DKO cells expressing DNMT3A1 or DNMT3A2. Shown are read counts per 100 bp.

3. Average density plots showing library‐normalized H3K27me3 ChIP‐seq reads around UMR promoters bound by DNMT3A1 (cluster 1 from Fig EV2A) in Dnmt‐TKO cells and Dnmt‐TKO cells expressing individual de novo DNMTs.

4. Box plot showing increased DNaseI hypersensitivity (DHS) signals in Dnmt‐TKO cells at UMR promoters enriched by DNMT3A1 (same clusters as in Fig EV2A). Boxes denote the IQR and whiskers 1.5 × IQR. Notches display the confidence interval around the median.

5. Scatter plots showing correlated increase in DHS or H3K27ac with loss of H3K27me3 at UMR promoters in Dnmt‐TKO cells, suggesting that reduction in H3K27me3 leads to increased promoter accessibility and activity. Red dots denote the UMR promoters within cluster 1 bound by DNMT3A1.

Data information: Datasets used in (A–C) were obtained from King et al (2016). Datasets used in (D and E) were obtained from Domcke et al (2015).