The 5-methylcytosine DNA glycosylase ROS1 prevents paternal genome hypermethylation in Arabidopsis endosperm | Genome Biology

ROS1 contributes to endosperm DNA methylation patterning

Gene expression data indicate that all four 5-methylcytosine DNA glycosylases are expressed in endosperm [36] (Additional file 1: Fig. S1A). To determine whether 5-methylcytosine DNA glycosylases other than DME contribute to the endosperm DNA methylation landscape, we profiled DNA methylation in the absence of ROS1 and in the absence of ROS1, DML2, and DML3. We performed enzymatic-methyl sequencing (EM-seq) on three replicates each of wild-type Col-0, ros1, and ros1-3 dml2-1 dml3-1 (rdd) endosperm. For ros1 profiling, we used both ros1-3 mutants, which have a T-DNA insertion in exon 7, and ros1-7 mutants, which have a missense mutation such that a conserved glutamic acid residue in the DNA glycosylase catalytic domain is changed to a lysine residue [8, 37] (Additional file 1: Fig. S1B). We used two different mutant alleles of ros1 so we could confirm that observations were not line-specific. Based on previous studies, ros1−7 is expected to be a hypomorphic allele [37], whereas ros1−3 is a null allele [8]. For all samples, endosperm nuclei were isolated from whole seeds at 7 days after pollination (DAP) by fluorescence-activated nuclei sorting (FANS) based on their triploid DNA content (Additional file 1: Fig. S1C). We obtained methylomes of 98 to 194x genome coverage with high conversion rates (Additional file 2: Table S1). To facilitate direct comparison between endosperm and a vegetative tissue, we also profiled methylation in three replicates each of wild-type Col-0, ros1-3, and ros1-7 rosette leaves, isolating 2 C and 4 C nuclei by FANS (Additional file 1: Fig. S1D, Additional file 2: Table S1). Globally, the total fraction of endosperm methylated cytosines was greater in rdd than wild-type in the CG and CHG sequence contexts (Fig. 1A, p < 0.05, unpaired t-test, Bonferroni-corrected) but was not significantly different between ros1-3 and wild-type or ros1-7 and wild-type (Fig. 1A, p > 0.05, unpaired t-test, Bonferroni-corrected). To identify potential discrete regions of differential methylation we used Dispersion Shrinkage for Sequencing Data (DSS) to identify differentially methylated regions (DMRs) between the demethylase mutant endosperm and the wild type in the CG, CHG, or CHH methylation sequence contexts [38, 39] (Fig. 1B, C). The ros1-3 mutation is linked to genomic regions from the Ws ecotype on chromosome 2; these regions were removed from analysis in ros1-3, ros1-7, and Col-0 prior to identifying DMRs to avoid calling false-positive DMRs due to ecotype-specific methylation differences [8, 20]. For identifying DMRs between rdd and Col-0, we removed from consideration regions of the Ws genome on chromosomes 2 and 3 that are linked to ros1-3 and dml2-1, respectively [8, 20]. Using DSS, we identified 1,624 total DMRs in any sequence context between ros1-3 and Col-0, 913 total DMRs between ros1-7 and Col-0, and 1,319 total DMRs between rdd and Col-0 (Additional file 3: Table S2). We partially attribute the lower number of DMRs called in rdd relative to ros1-3 to greater variability between rdd biological replicates. Consistent with the molecular function of DNA demethylases, most DMRs were more highly methylated in demethylase mutant endosperm compared to wild-type endosperm; these are referred to as “hyperDMRs” (Fig. 1C, Table 1). The DMRs were short, with the most abundant fraction being 50–100 bp in length (Additional file 1: Fig. S2A). By genome browsing, we observed that DMRs identified by DSS were often surrounded by regions that also appeared differentially-methylated but were not called as DMRs (Fig. 1B). To further investigate, we calculated methylation levels in 50 bp windows 1 kb 5′ and 3′ of hyperDMRs (Additional file 1: Fig. S3). Regions flanking hyperDMRs were more methylated in ros1 mutant backgrounds than they were in wild-type, up to a few hundred base pairs from the center of the DMR (Additional file 1: Fig. S3).

The fewer ros1-7 hyperDMRs identified compared to ros1-3 are consistent with the expectation of ros1-7 as a hypomorphic allele and ros1-3 as a null allele (Fig. 1C, Additional file 1: Fig. S1B). To assess the replicability of hypermethylation across the mutant genotypes, we calculated the weighted average of DNA methylation at ros1–3 and ros1–7 hyperDMRs in all genotypes (Fig. 1D, Additional file 1: Fig. S4). Across replicates and genotypes, ros1 mutants were more highly methylated than Col-0 at hyperDMRs identified in either ros1 mutant background, indicating a high degree of similarity between the two ros1 mutant genotypes. Finally, our analysis of methylation across genotypes showed that disruption of DML2 and DML3 caused only minor increases in endosperm methylation compared to loss of ROS1 alone (Fig. 1C, Additional file 1: Fig. S4). Thus, we focused our additional analyses and experiments on ROS1.

ROS1 prevents DNA methylation spreading from a subset of TEs in the endosperm

ROS1 is known to maintain DNA methylation boundaries at TEs, preventing aberrant DNA methylation “spread” [9, 10, 40]. The hypomethylation of the wild-type endosperm genome relative to leaf and seedling tissues prompted us to further investigate the impact of ROS1 at TE boundaries in endosperm. To identify features of endosperm ROS1 targets, we first merged ros1 hyperDMRs (Table 1) from all cytosine sequence contexts. Consistent with previous results, these regions were often found near TEs; over half were within 1 kb of or intersecting a TE (representing 1463 of the 34856 Araport11-annotated TE fragments for ros1-3, and 624 for ros1-7) (Fig. 1E). For subsequent analyses, we refer to the 1463 TEs within 1 kb of or intersecting a ros1-3 hyperDMR as ROS1 TEs (Additional file 4: Table S3). Seventy-nine percent of the DMRs associated with ROS1 TEs were in the 1 kb flanking regions, rather than in the TE body. In wild-type endosperm, ROS1 TEs were less methylated in flanking regions (~26% CG methylation) than non-ROS1 TEs (~40% CG methylation), with a sharper boundary between methylation in the flanking region and in the body of the TE (Fig. 2A, Additional file 1: Fig. S5, 6). Additionally, in ros1 mutant endosperm, we observed DNA methylation spreading up to ~1 kb beyond the ends of ROS1 TEs (Fig. 2A). Similar results were obtained for ROS1 TEs defined using the ros1-7 methylation data (Additional file 1: Figs. S5, S6). To investigate any endosperm-specific features of ROS1 TEs, we compared DNA methylation levels of the same TEs in leaves (Additional file 1: Figs. S5, S6). The total level of CG DNA methylation at both ROS1 TEs and non-ROS1 TEs was lower in the endosperm than in leaves (Additional file 1: Figs. S5, S6), consistent with endosperm DNA hypomethylation. However, the magnitude of increased CG methylation flanking ROS1 TEs was not different between endosperm and leaf. We note that bodies of both ROS1 and non-ROS1 TEs have higher levels of non-CG methylation in endosperm relative to leaf (Additional file 1: Figs. S5, S6). Overall, we conclude that ROS1 enforces sharp methylation boundaries at a subset of TEs in the endosperm, as it does in leaves.

Whereas the function of ROS1 in maintaining DNA methylation boundaries at the ends of TEs has been previously documented, we sought to further characterize the dynamics and mechanism of DNA methylation spreading. To address this, we utilized our high-coverage EM-seq data from Arabidopsis endosperm to investigate the dynamics of DNA methylation spread from individual TEs. We first identified ROS1 TEs that were methylated in wild-type Col-0 endosperm (those that are at least 10% methylated in all sequence contexts; n = 434), as these are TEs from which DNA methylation has the potential to spread. To visualize the molecular phenotype of DNA methylation spreading at individual TEs, we calculated average percent mCG in 100 bp windows of regions flanking TEs for each biological replicate and calculated the difference between the median ros1-3 replicate and the median Col-0 replicate at each window; differences were visualized using a clustered heatmap (Fig. 2B). We observed that DNA methylation spreading from individual TEs appeared to have a primary direction—either from the 5′ or 3′ end of the TE, which was masked in the meta profile (Fig. 2A). However, by definition, the detection of methylation spreading in ros1 is only possible if the proximal region of interest has low levels of methylation in wild-type. To test if bidirectional DNA methylation spread from TEs in a ros1 mutant was possible but did not occur, we separately examined the median wild-type Col-0 value (Fig. 2C, right) and ros1–3 value (Fig. 2C, left). TEs that were demethylated by ROS1 on both ends were not adjacent to a highly-methylated region on either end in wild-type. In contrast, TEs that appeared to be demethylated by ROS1 on only one end (Fig. 2B) were adjacent to a highly-methylated region on the non-spreading end in wild-type (Fig. 2C). Thus, these TEs lack the capacity for bidirectional spreading in the ros1 mutant. Therefore, spreading direction reflects the surrounding methylation state in wild-type, rather than being an inherently asymmetric process. One caveat to these analyses is the challenge of TE annotation; incorrect TE boundary annotations could obscure where a true TE “end” is. Future work investigating the nature and mechanism of DNA methylation spreading in a ros1 mutant will be valuable for understanding the nature and maintenance of epigenetic boundaries in Arabidopsis.

To understand the mechanism of spreading in the endosperm, we examined whether methylation spreading was associated with 24-nucleotide small RNAs, which participate in RdDM, in wild-type Col-0 endosperm [41]. The proximal regions of ROS1 TEs produced on average more 24-nt sRNAs in wild-type Col-0 endosperm than in the embryo, whereas the opposite was true for non-ROS1 TEs (Additional file 1: Fig. S7A). However, the data indicated enrichment of endosperm 24-nt sRNAs was driven by only a few ROS1 TEs (Additional file 1: Fig. S7B). To quantify this, we summed the sRNA levels in 2 kb flanking regions of ROS1 TEs and identified TEs with a difference between endosperm and embryo that was 1 standard deviation or more from the mean level. This analysis identified 25 TEs driving the observed endosperm enrichment of sRNAs (Additional file 1: Fig. S7B, Additional file 4: Table S3).

To further clarify the relationship between ROS1 activity and endosperm 24-nt sRNAs, we quantified DNA methylation levels in ros1-3 and wild-type at regions previously determined to be enriched for 24-nt sRNA production in the endosperm relative to the embryo, referred to as endosperm differential sRNA regions (endosperm DSRs) [41]. Endosperm DSRs were lowly methylated in wild-type endosperm, consistent with previous results [41], and a subset gained DNA methylation in ros1-3 endosperm (Additional file 1: Fig. S7C). To quantify the fraction of endosperm DSRs which are demethylated by ROS1, we calculated a weighted average of CG methylation in endosperm DSRs in Col-0 and ros1-3, calculated the difference between median replicates of both, and filtered DSRs which were at least 30% more CG methylated in ros1-3 than in Col-0. We found 78 DSRs out of 2481 met this stringent threshold. Together, these results suggest that at regions enriched for sRNA production in the endosperm, the low level of DNA methylation observed in wild-type is at least partly a result of ROS1-mediated DNA demethylation, counteracting RdDM.

ROS1 targets have reduced capacity for hypermethylation in ros1 endosperm

To determine whether ROS1 acts at unique sites in the endosperm, we investigated the extent to which DNA hypermethylation was tissue-specific. We calculated the level of cytosine methylation at endosperm ros1-3 hyperDMRs (Table 1) in our Col-0 and ros1-3 leaf methylation data as well as in published data from wild-type Col-0 and ros1-3 sperm cells [27]. Sperm methylation was of particular interest because the endosperm is a product of fertilization between a sperm and the central cell. In wild-type plant tissues, ros1-3 CG hyperDMRs displayed DNA methylation features that have been observed at a genome-wide scale: wild-type endosperm has lower DNA methylation levels than does leaf (average 2.8% vs 7.8% in one replicate of each), and sperm has higher DNA methylation levels (~20%) than leaf or endosperm (Fig. 3A, Additional file 1: Fig. S8). Endosperm ros1-3 hyperDMRs were also hypermethylated in ros1 mutant leaf and sperm relative to the respective wild-type tissue (Fig. 3A, Supplemental Figures S4 and S8).

However, although ROS1 endosperm targets are also ROS1 targets in leaf and sperm, they displayed lower levels of CG DNA methylation in both Col-0 and in ros1 endosperm (Fig. 3A). What underlies the failure to reach the fully hypermethylated state in ros1 endosperm? We considered two non-mutually-exclusive possibilities: (1) these regions are variably methylated among nuclei of ros1 endosperm or (2) these regions are differently methylated between the maternal and paternal genomes in ros1 endosperm. In support of the second hypothesis, ROS1 has been shown to prevent hypermethylation of the paternal allele of the DOGL4 promoter in endosperm (35), and in our data, CG sites in the DOGL4 promoter were less methylated in ros1 endosperm compared to ros1 leaf (Additional file 1: Fig. S9).

ROS1 prevents hypermethylation on the paternal genome at target loci

To test if ROS1 targets are differentially methylated between paternal and maternal genomes we performed allele-specific whole-genome EM-seq using F1 endosperm isolated from reciprocal crosses between ros1 mutants in the Col-0 (ros1-3) and C24 (ros1-1) [1] backgrounds, along with appropriate controls. ros1-1 is a nonsense allele (Additional file 1: Fig. S1B). Endosperm nuclei were collected from three replicates each of Col-0 × C24, C24 × Col-0, ros1-3 × ros1-1, ros1-1 × ros1-3, C24 × C24, and ros1-1 × ros1-1 (female parent in cross written first). SNPs between C24 and Col-0 were used to assign reads to a parent-of-origin after sequencing [42] (Additional file 2: Table S1).

We compared the methylation of maternal and paternal alleles at ros1-3 hyperDMRs in wild-type and ros1 endosperm. In wild type, CG methylation of maternal and paternal alleles at ros1-3 CG hyperDMRs was highly correlated (Pearson’s r = 0.83) (Fig. 4A). In contrast, CG methylation of maternal and paternal alleles at the same regions in ros1 endosperm was not correlated (Pearson’s r = −0.08) (Fig. 4B). The lack of correlation was caused by gain of paternal allele methylation in ros1 endosperm (Fig. 4B). The phenomenon of paternal allele hypermethylation was replicable across ecotypes, as we observed a comparable paternal bias on the C24 genomes at CG hyperDMRs identified between ros1-1 and wild-type C24 endosperm (Fig. 4C–D, Additional file 2: Table S1,). Consistent with these findings, paternal alleles of ROS1 TEs gained more mCG in the ros1-3 mutant than did maternal alleles (Additional file 1: Fig. S10A). Paternal allele hypermethylation of ros1-3 CG hyperDMRs (defined in Col-0) was also observed when the ros1-1 allele was inherited paternally (ros1-3 × ros1-1), indicating overlap among regions where ROS1 prevents paternal hypermethylation in the Col-0 and C24 ecotypes (Additional file 1: Fig. S10B–C).

We also examined parent-of-origin specific methylation for non-CG hyperDMRs. Contrary to CG hyperDMRs, non-CG hyperDMRs were biallelically hypermethylated in their respective sequence contexts (Additional file 1: Fig. S11). Furthermore, non-CG methylation at ros1-3 CG hyperDMRs was not biased for the paternal allele, although the magnitude of non-CG methylation increase was much less than for CG methylation (Fig. 4E, Additional file 1: Fig. S12). Thus, the observed paternal bias of ROS1 is specific to the CG sequence context. Studies of ROS1 activity in vitro have not revealed a sequence context preference for ROS1 demethylase activity, and in vivo whole-genome sequencing data has repeatedly shown that ROS1 prevents hypermethylation of cytosines in all sequence contexts [2, 8, 43], including the results in the present study of the endosperm (Fig. 1B–D). We propose that it is more likely that the differences observed between sequence contexts indicate ROS1 target regions are differentially targeted by methylation-establishing and maintenance pathways on the maternal and paternal genomes in the endosperm. Overall, ROS1 prevents hypermethylation of the paternal allele more strongly than the maternal allele in the endosperm in the CG sequence context.

ROS1 promotes a bi-allelically demethylated state in endosperm together with DME

Greater paternal genome DNA methylation relative to maternal genome DNA methylation is a distinguishing feature of the endosperm epigenome [22, 23]. It is known that wild-type endosperm maternal allele hypomethylation depends, at least in part, on the activity of DME in the central cell before fertilization [3, 6, 22,23,24]. We have shown that ROS1 acts to restrict paternal genome hypermethylation in the endosperm. Considering previous work, we hypothesized that in wild-type endosperm the maternal allele of regions that become hypermethylated in ros1 are demethylated by DME. Using previously published data for maternal genome methylation in endosperm where either dme-2 or wild-type DME were inherited maternally [22], we quantified mCG at ros1-3 CG hyperDMRs and ros1-7 CG hyperDMRs (Additional file 1: Fig. S13). This analysis showed that the maternal alleles of ros1-3 and ros1-7 CG hyperDMRs were hypermethylated in dme heterozygous endosperm (Additional file 1: Fig. S13). As expected, the paternally-inherited allele of these regions, which were wild-type Ler in both cases, were not differentially methylated when dme was inherited maternally (Additional file 1: Fig. S13 A, B). This result suggests that a subset of ROS1 regions are shared demethylase targets, with DME acting on the maternal allele and ROS1 on the paternal allele.

DME activity on the maternally-inherited genome in the central cell of the female gametophyte establishes parent-of-origin-specific DNA methylation in the endosperm. We propose that ROS1 antagonizes parent-of-origin-specific DNA methylation patterning in the endosperm, resulting in low methylation on both maternal and paternal alleles. To further investigate the role of ROS1 in the context of greater paternal genome methylation in the endosperm, we used our allele-specific methylation data to identify DMRs between maternal and paternal genomes within a wild-type background and within a ros1 mutant background (Table 1). For each genotype, DMRs were called in mCG, mCHG, and mCHH sequence contexts independently, but resulting maternally-hypomethylated DMRs were merged into one list of regions per genotype for subsequent analyses. Regions (n = 1586) that were demethylated on the maternal allele and methylated on the paternal allele in F1 endosperm of all four reciprocal crosses (WT × WT, ros1 × ros1) were likely regions where DME establishes parental DNA methylation asymmetry by acting on the maternally-inherited genome in the central cell before fertilization. This difference is maintained independently of ROS1, and we refer to these as “DME maternal regions” in the following text and figures (Fig. 5A, Table 1, Additional file 5: Table S4). Regions defined as maternally-hypomethylated in both ros1 × ros1 cross directions, but neither WT × WT cross direction were also identified, and based on subsequent analyses we refer to these as “ROS1 paternal, DME maternal regions” in the following text and figures (Fig. 5B, Table 1, Additional file 5: Table S4). These regions (n = 274) lacked DNA methylation on both alleles in wild-type endosperm, and gained DNA methylation relative to wild-type predominantly on the paternal allele in the absence of ROS1. Hypomethylation of the maternal allele relative to the paternal allele is observable in the ros1 mutant background at these regions (Fig. 5B). This suggests that in the wild type, the maternal allele of these regions lacks methylation due to a ROS1-independent mechanism, whereas the paternal allele is demethylated by ROS1 (Fig. 5B). The presence of ROS1 paternal, DME maternal regions, in addition to the paternal bias in CG hypermethylation at ros1 hyperDMRs (Fig. 4), indicates a role for ROS1 in preventing differential DNA methylation between maternal and paternal genomes in the endosperm, specifically differential methylation where the paternal allele is more highly methylated than the maternal allele.

We tested our assumption that maternally-demethylated regions in wild-type endosperm that were not dependent on ROS1 for proper methylation state (Fig. 5A, “DME maternal regions”) were demethylated by DME. Examination of methylation levels in and flanking these regions using published dme allele-specific endosperm data [22] indicated maternal allele hypermethylation compared to wild-type, confirming these regions as canonical DME targets (Fig. 6A, C, Additional file 1: Fig. S14A). We also examined whether maternal allele hypomethylation at ROS1 paternal, DME maternal regions was, as we hypothesized, dependent on DME. We observed maternal allele hypermethylation in dme endosperm at these regions but to a lesser extent than at DME maternal regions (Fig. 6B, D, Additional file 1: Fig. S14B). We repeated these analyses using an independent, non-allelic dme mutant endosperm dataset [24], which further confirmed that DME prevents hypermethylation at DME and ROS1 regions (Additional file 1: Fig. S15). Together, these results indicate that DME, in part, prevents hypermethylation of the maternal allele and ROS1 prevents hypermethylation of the paternal allele at some regions that are biallelically-demethylated in wild-type endosperm. Other factors are also likely involved in maternal allele hypomethylation of these regions. For example, it is possible that methyltransferases, such as MET1, do not maintain methylation of the maternal allele at these sites in the central cell and early endosperm, which could also prevent full maternal allele methylation irrespective of DNA demethylase activity.

Relationship between ROS1 and DME in the endosperm

We further investigated the relationship between ROS1 and DME function and targets in the endosperm. To identify the genome neighborhood of ROS1 and DME endosperm targets, we plotted the density of DMRs in 100 kb windows across chromosomes, plotting a rolling average of 10 windows (Fig. 7A). We found that ROS1 targets, both ros1-3 hyperDMRs and ROS1 paternal, DME maternal regions, were evenly distributed across chromosomes (Fig. 7A, green and pink lines). DME maternal regions, however, were denser in pericentromeric regions (Fig. 7A, gold line). This is consistent with previous results showing that endosperm vs embryo hypoDMRs, indicative of DME activity in the central cell, match the distribution of transposable elements in the genome [23]. In prior work, ROS1 and DME have been associated with distinct genomic features with regards to TEs. ROS1 targets TEs of any length, especially the ends of TEs near gene [8, 10]. In euchromatic regions, DME has been reported to demethylate the bodies of shorter TEs [22,23,24]. We quantified the number of DME maternal regions and ROS1 paternal, DME maternal regions within 1 kb or intersecting a gene or TE (Fig. 7B). More than 86% of DME maternal regions and 75% of ROS1 paternal, DME maternal regions were associated with TEs (Fig. 7B).

We observed that the flanking regions of DME maternal regions were more highly methylated in wild-type endosperm than were flanking regions of ROS1 paternal, DME maternal regions, most notably on the paternal allele (Fig. 6A, B, compare yellow lines). We also observed a slight increase in CG methylation on the paternal allele at regions flanking DME maternal regions in the ros1 mutant endosperm (Fig. 6A, green vs. yellow line). We hypothesized that this may be indicative of a relationship between regions targeted by DME alone and regions targeted by both DME and ROS1. We quantified the distance between each ROS1 paternal, DME maternal region to the nearest DME maternal region to test their proximity. The majority of regions were greater than 15 kb away from one another, but about 10% of ROS1 paternal, DME maternal regions were within 1 kb of a DME maternal region (Additional file 1: Fig. S16). Given the frequent proximity of both ROS1 paternal, DME maternal regions and DME maternal regions to TEs, we investigated if ROS1 might be preventing DNA methylation spread at the ends of TEs that are demethylated by DME. We identified 3263 TEs that were within 1 kb or intersecting a maternally-demethylated region (DME TEs) and calculated allele-specific DNA methylation levels in wild-type and ros1 (Fig. 7C). As expected, DME-targeted TE bodies and flanking regions were methylated on the paternal allele (50–70% mCG), but the maternal allele was less methylated (~30–60% mCG) in TE bodies and flanking regions (Fig. 7C). A small increase in paternal allele CG methylation at regions flanking DME TEs was observed in the ros1 mutant endosperm (Fig. 7C; green vs. yellow lines). To assess how widespread this phenomenon was, we calculated the difference between the median value of mCG in ros1-3 and Col-0 for DME TEs and plotted these using a clustered heatmap (Fig. 7D, Additional file 1: Fig. S17). This visualization shows that ROS1 prevents methylation spreading at a small subset of DME TE ends. To quantify this, we identified 653 TEs that met a threshold of one window or more with at least 30% higher methylation in ros1-3 than in Col-0. Using this definition, 20% of DME targeted-TEs showed some ROS1-dependency in their flanking regions in the endosperm. By genome browsing, we observed TEs that were maternally demethylated in a DME-dependent manner and where ROS1 prevented DNA methylation spreading into TE-flanking regions, predominantly on the paternal allele (Fig. 7E).

Although some DME maternal regions were proximal to ROS1 paternal, DME maternal regions (Fig. 7C–E), the majority were not (Additional file 1: Fig. S16). Thus we further investigated features that might distinguish ROS1 and DME target regions, examining the overlap with 9 chromatin states defined in seedling tissue [44]. Consistent with the distribution of DMRs (Fig. 7A), we found that in seedling tissue DME maternal regions were enriched in H3K9me2-marked heterochromatin in intergenic regions and TEs, with 66% of DME maternal regions found in chromatin state 8 (Fig. 7F). The largest single fraction of chromatin states represented in both ros1-3 and ros1-7 hyperDMRs was also state 8, but together there was more representation of states 4 and 5 (45.8% for ros1-3 hyperDMRs), which are marked by H3K27me3 and correspond to upstream regions of promoters and intergenic regions, respectively (Fig. 7F). The largest fraction of ROS1 paternal, DME maternal regions, 48%, were found in chromatin state 4, indicating these shared targets of ROS1 and DME are mostly found in upstream promoter regions marked by H3K27me3 (Fig. 7F). One limitation of this analysis is that it is unknown whether these same chromatin states are present in endosperm tissue. We thus further focused our investigation on H3K27me3 profiles in the endosperm. Using published data from H3K27me3 profiling in Col × Ler endosperm at 4 days after pollination, we evaluated if ROS1 and DME regions were coincident with H3K27me3 peaks in the endosperm [45]. We found that 32% and 64% of ros1-3 and ros1-7 hyperDMRs, respectively, were associated with an H3K27me3 peak in endosperm tissue. By contrast, ~20% of DME regions were associated with an endosperm H3K27me3 peak. ROS1 paternal, DME maternal regions were in between these values, with ~29% of regions associated with an H3K27me3 peak in endosperm. These results are consistent with the chromosomal localization and chromatin state of these regions, as H3K27me3 is relatively depleted from pericentromeric regions.

We also investigated ROS1 and DME target regions with regards to imprinting and gene expression in the endosperm. ROS1 and DME have been shown to act together to promote expression of genes in the vegetative nucleus of pollen, ultimately promoting proper pollen germination [27]. In the endosperm, DME maternal regions are associated with imprinted genes [31] (Additional file 6: Table S5). Although parent of origin-specific DNA methylation is not an absolute requirement for gene imprinting [30], the relatively demethylated status of both maternal and paternal alleles at ROS1 paternal, DME maternal regions suggests that they are likely not involved in regulation of gene imprinting. Consistent with this notion, no imprinted genes were within 1 kb of or intersecting ROS1 paternal, DME maternal regions. To investigate any differences in expression of DMR-associated genes in the seed, we utilized a single-nucleus RNA sequencing atlas of developing wild-type Arabidopsis seeds [36]. We found that the expression of genes near DME maternal regions was enriched in the endosperm compared to the embryo or seed coat, but that the expression of genes near ROS1 and ROS1/DME targets were not enriched in endosperm relative to embryo or seed coat. (Additional file 1: Fig. S18).

We conclude that both ROS1 and DME demethylate TEs, especially those near genes. A subset of TEs demethylated by DME are further demethylated by ROS1 in their flanking regions. However, ROS1 and DME targets are distinguished by their chromosomal distribution and their coincidence with different chromatin states. Furthermore, known imprinted genes are found near DME maternal regions, but not near ROS1 paternal, DME maternal regions. Overall, the functional consequence of co-targeting by DME and ROS1, either on gene expression or chromatin state, is unclear.

Inheritance of wild-type ROS1 does not rescue hypermethylation in F1 heterozygous endosperm

How and why does ROS1 primarily effect paternal allele methylation state in the endosperm? One explanation is that paternal allele-specific demethylation of these regions occurs in the endosperm after fertilization. This would mean that ROS1 selectively acts on paternal alleles despite the presence of maternal alleles. Another, not mutually-exclusive, possibility is that demethylation by ROS1 after fertilization is not allele-specific but that there is no methylation actively established or maintained on the maternal allele to be removed at this point in development. Finally, the paternally-biased effect of ROS1 in the endosperm could be a product of ROS1 activity pre-fertilization, with the consequent methylation state inherited and maintained after fertilization. Although not expressed in mature pollen or sperm, ROS1 is expressed in the microspore and bicellular pollen (Additional file 1: Fig. S19) [46, 47] and ROS1 endosperm targets are hypermethylated in ros1 sperm (Fig. 3). Thus, ROS1 could act prior to fertilization in the male germline, leading to inheritance of a demethylated paternal allele in the endosperm, without a requirement for active demethylation of these regions by ROS1 in the developing endosperm after fertilization.

Based on our comparisons between sperm and endosperm DNA methylation data, we sought to test if inheritance of a wild-type ROS1 allele is sufficient for paternal genome hypomethylation in the endosperm. More specifically, is the paternal allele hypermethylation that is observed in ros1 endosperm at ROS1 paternal, DME maternal regions rescued in the presence of a wild-type ROS1 allele that is inherited maternally? We reasoned that if ROS1 acts only through the male germline, then a wild-type paternal ROS1 allele would be necessary for demethylation of the paternal endosperm genome, and a wild-type maternal ROS1 allele would be insufficient. To test this hypothesis, we performed allele-specific methylation profiling on endosperm of F1 seed derived from reciprocal crosses between ros1-3 (in the Col-0 background) and wild-type C24 (Fig. 8A). In this design, the F1 endosperm is heterozygous for the ros1 mutation but either the maternal or paternal sporophytic tissues and female or male gametophytes are null for ROS1. If ROS1 activity is required before fertilization in the paternal sporophyte or male gametophyte to cause hypomethylation of paternal alleles in the endosperm after fertilization, then ROS1 paternal, DME maternal regions will be paternally hypermethylated in heterozygous ros1-3 endosperm when the mutation is inherited through the paternal parent. In contrast, under this model maternal inheritance of ros1 should not result in paternal allele hypermethylation in heterozygous endosperm. If ROS1 instead acts after fertilization to demethylate paternal alleles in endosperm, then paternal allele hypermethylation will not be observed in heterozygous ros1-3 endosperm when the mutation is inherited through the paternal parent.

We compared the paternal genome average CG methylation level in WT, ros1, and ros1/+ endosperm at ROS1 paternal, DME maternal regions (Fig. 8B). Paternal allele methylation was correlated between the heterozygous and homozygous ros1-3 endosperm, when ros1-3 was inherited paternally (Fig. 8B). We observed that paternal alleles were highly methylated in the C24 × ros1-3 F1 heterozygous endosperm, like in ros1 endosperm, relative to wild-type paternal alleles (Fig. 8C, Additional file 1: Fig. S20, “ros1-3 paternal”). Thus, inheritance of a wild-type maternal ROS1 allele is not sufficient for wild-type mCG levels at ROS1 paternal, DME maternal regions. We also observed little difference in methylation on maternally-inherited alleles between heterozygous and homozygous ros1 mutants relative to wild-type; maternal allele methylation was low in these regions regardless of the ROS1 genotype (Additional file 1: Fig. S20 “ros1-3 maternal”). We conclude that the majority of paternal allele hypermethylation in ros1 mutant endosperm is inherited through the male germline. Thus, in wild-type, the majority of regions are demethylated by ROS1 prior to fertilization. However, we observed a slight, non-significant, reduction in paternal mCG at ROS1 and DME regions in the ros1/+ F1 heterozygous endosperm relative to the ros1 homozygous endosperm (Fig. 8B–C, Additional file 1: Fig. S20 “ros1-3 paternal”) indicative of active maternal ROS1 in the endosperm after fertilization that is able to partially rescue paternal CG hypermethylation. However, in the same C24 × ros1-3 F1 heterozygous endosperm, methylation of maternal alleles (inherited from C24) of ROS1 paternal, DME maternal regions was also decreased (Additional file 1: Fig. S20, “C24 maternal”), so this reduction is likely not indicative of specific ROS1 activity on paternal alleles only in the endosperm post-fertilization.

We also investigated the role of maternally and paternally-inherited ROS1 in the endosperm at the larger set of ros1-3 CG hyperDMRs. Paternal allele methylation was also correlated between C24 × ros1-3 F1 heterozygous endosperm and ros1-1 × ros1-3 F1 endosperm at ros1-3 CG hyperDMRs (Additional file 1: Fig. S21). Maternally-inherited wild-type ROS1 was not sufficient for wild-type methylation levels on paternal alleles of ros1-3 CG hyperDMRs (Additional file 1: Fig. S21B, “ros1-3 paternal”). We again observed a slight decrease in paternal allele hypermethylation at some ros1-3 CG hyperDMRs in the ros1/+ F1 heterozygous endosperm (Additional file 1: Fig. S21A, 21B “ros1-3 paternal”). Additionally, we observed a non-significant increase in paternal allele mCG in the ros1/+ F1 heterozygous endosperm when ros1-3 was inherited maternally (Additional file 1: Fig. S21B, “C24 paternal”), further implicating maternally-inherited ROS1 in preventing hypermethylation at a subset of ROS1 target regions in the endosperm after fertilization. Overall, we conclude that paternal allele CG methylation patterning at ROS1 regions is largely inherited through the male germline, but some ROS1 regions are actively demethylated by ROS1 in the endosperm on both maternal and paternal alleles, post-fertilization.

Finally, as it has been shown that ROS1 and DME act together at some regions in pollen [27], we investigated whether DME had any role in the male germline at ROS1 target regions. As expected, ROS1 paternal, DME maternal regions were hypermethylated in ros1-3 mutant sperm relative to wild-type sperm (Additional file 1: Fig. S22A). In sperm collected from dme-2 heterozygous plants, ROS1 paternal, DME maternal regions were not hypermethylated (Additional file 1: Fig. S22A), although sufficient sperm methylation data was only available for 103 regions. To evaluate more ROS1 target regions, we performed the same analysis using ros1-3 and ros1-7 CG hyperDMRs. There was a significant increase in CG methylation at ros1-3 and ros1-7 CG hyperDMRs in dme-2 heterozygous sperm compared to wild-type, although of a lower magnitude than that observed in ros1 sperm, suggesting DME is partially capable of demethylating ROS1 target regions earlier in development, which could further impact paternal allele methylation in the endosperm (Additional file 1: Fig. S22B–C). Overall, these results are consistent with previous reports that DME contributes to demethylation in somatic tissues and in the pollen vegetative nucleus, but not to the same extent as ROS1 [20, 48].