The DNMT3A ADD domain is required for efficient de novo DNA methylation and maternal imprinting in mouse oocytes

Establishment of a proper DNA methylation landscape in mammalian oocytes is important for maternal imprinting and embryonic development. De novo DNA methylation in oocytes is mediated by the DNA methyltransferase DNMT3A, which has an ATRX-DNMT3-DNMT3L (ADD) domain that interacts with histone H3 tail unmethylated at lysine-4 (H3K4me0). The domain normally blocks the methyltransferase domain via intramolecular interaction and binding to histone H3K4me0 releases the autoinhibition. However, H3K4me0 is widespread in chromatin and the role of the ADD-histone interaction has not been studied in vivo. We herein show that amino-acid substitutions in the ADD domain of mouse DNMT3A cause dwarfism. Oocytes derived from homozygous females show mosaic loss of CG methylation and almost complete loss of non-CG methylation. Embryos derived from such oocytes die in mid-to-late gestation, with stochastic and often all-or-none-type CG-methylation loss at imprinting control regions and misexpression of the linked genes. The stochastic loss is a two-step process, with loss occurring in cleavage-stage embryos and regaining occurring after implantation. These results highlight an important role for the ADD domain in efficient, and likely processive, de novo CG methylation and pose a model for stochastic inheritance of epigenetic perturbations in germ cells to the next generation.

A previous study reported that the DNMT3A ADD domain normally blocks its own methyltransferase domain and, upon H3K4me0 binding, releases the autoinhibition in vitro [16].In contrast, the domain does not interact with H3K4me2/3, excluding chromatin regions with these marks, such as active enhancers and promoters, from targets of DNA methylation [12][13][14][15][16]. Two aspartic acids (D529 and D531) in the ADD domain of human DNMT3A are important for both autoinhibition and H3K4me0 binding [16].However, the exact role of the ADD-H3K4me0 interaction is not understood in vivo.
We herein report that substitutions of the aspartic acids in the mouse DNMT3A ADD domain cause dwarfism and female infertility.We further demonstrate that the DNMT3A ADD domain plays an important role in efficient, and maybe processive, de novo methylation and maternal imprinting in mouse oocytes.The stochastic loss of imprinting in derived embryos poses a model to study how partial changes in epigenetic modifications in germ cells can be inherited or eliminated during development.

Generation and phenotype of Dnmt3a ADA mice
Two aspartic acids (D529 and D531) in the ADD domain of human DNMT3A are respectively important for both autoinhibition and H3K4me0 binding [16].We generated mice carrying aspartic-acid-to-alanine substitutions at D525 and D527 (corresponding to human D529 and D531) using CRISPR/Cas9-mediated homology directed repair (see Materials and methods).We assumed that such DNMT3A would always adopt an active form but show reduced binding to H3K4me0 [16], the latter of which was confirmed by in vitro binding assay using wildtype and mutated recombinant ADD domains (S1A, and S1B Fig) .The mutated allele and protein are called Dnmt3a ADA and DNMT3A ADA , respectively (Fig 1A).We detected mutated DNMT3A2 protein in fully grown oocytes (FGOs), a major isoform expressed in this cell type, at a level comparable to wild-type protein (Fig 1B).Dnmt3a ADA/ADA homozygous mice were obtained at a near Mendelian ratio and showed dwarf phenotype (S1C, S1D and S1E Fig) , which contrasts with the overgrowth phenotype of Tatton-Brown-Rahman syndrome patients with mutations in this domain [22].
While homozygous males were fertile (litter size 7.1 ± 1.5 [n = 9 litters], versus 6.6 ± 1.4 for heterozygous control males [n = 9 litters]), homozygous females showed a low delivery rate per vaginal plug (4/7, versus 7/8 for heterozygous control females).Furthermore, the females suffered from prolonged pregnancy and gave no live pups with only a few stillborn pups (from 3 deliveries and one Caesarian section) (Fig 1C).To examine whether the phenotype is due to the mothers or the fetuses themselves, we performed in vitro fertilization (IVF) of Dnmt3a ADA/ ADA oocytes with wild-type sperm and transferred two-cell embryos to the oviducts of pseudopregnant females.The result showed a significant loss of embryos even in healthy mothers (S1 Table ).When a mixture of embryos from wild-type and homozygous oocytes was transferred, all derived pups were from wild-type oocytes.Importantly, two live Dnmt3a ADA/ADA oocytederived pups were obtained from a transferred female (S1 Table ), suggesting that homozygous oocytes can occasionally support full development and that homozygous females do have delivery problems.
We then examined the timing of embryonic loss by crossing homozygous females with wild-type JF1 males [23].While most embryos looked healthy at embryonic day 8.5-9.0 (E8.5-9.0), a small proportion was arrested at E10. 5 (Fig 1D and S2 Table).The actual sizes of the E10.5 embryos from homozygous females were significantly smaller than those from wild-type or heterozygous females (S1F Fig).At E14.5 and E18.5, we observed frequent fetal resorption, and only about half of the implanted embryos developed normally to these stages (S2 Table ).These results show that many embryos die during mid-to-late gestation in homozygous females.

Impact of Dnmt3a ADA mutation on CG and non-CG methylation in FGOs
We next examined the DNA methylation landscape of oocytes by whole-genome bisulfite sequencing (WGBS) with pooled FGOs (S3 Table ).Homozygous FGOs showed nearly 50% reduction in global CG methylation compared to wild-type and heterozygous FGOs and almost complete loss of non-CG methylation (Fig 2A) (non-CG methylation of homozygous FGOs was close to the bisulfite conversion error rate; S3 Table ).The methylation loss in homozygous FGOs occurred across the genome (Fig 2B and 2C).Among the different genomic annotations, intracisternal A particle elements, a class of endogenous retrovirus, were the least affected (Fig 2D ), perhaps consistent with their preferential methylation and resistance to demethylation [24][25][26][27][28][29].Notably, maternally methylated ICRs and promoter CpG islands that are fully CG methylated in wild-type FGOs showed 30%-60% reduction in CG methylation (Fig 2E and 2F).The examination of individual WGBS reads mapping to the ICRs revealed partial and mosaic CG methylation loss in each DNA template (S2 Fig) .These results suggest that the establishment of the normal CG and non-CG methylation landscape is impaired in homozygous FGOs.

Efficient, and likely processive, CG methylation is compromised in Dnmt3a ADA mutant FGOs
The mosaic loss of CG methylation in Dnmt3a ADA/ADA FGOs suggests that the ADD domain may play an important role in efficient, and possibly processive, methylation in oocytes.To gain insights into how this domain controls de novo methylation, we compared the distribution of CG methylation levels of individual WGBS reads at the maternally methylated ICRs in homozygous FGOs with that of wild-type growing oocytes (GOs), in which de novo methylation is still ongoing [4].As expected, we observed progressive CG methylation of the ICRs in GOs obtained at postnatal day 10 (P10) [30], P12, and P14 (diameter 60-70 μm) [31] (Fig 3A).Interestingly, the methylation levels of the individual WGBS reads often showed bimodal distribution at later stages (especially at P14; Fig 3A), suggesting that de novo ICR methylation tend to occur at multiple, proximal CG sites (processive methylation).By contrast, homozygous FGOs, which had CG methylation levels not very different from P12 and P14 GOs, showed a simple distribution suggestive of random methylation (Fig 3A).These results suggest that DNMT3A ADA likely have a problem in processive CG methylation.
To investigate this possibility further, we selected WGBS reads containing exactly the same number of CGs and the same number of methylated CGs (thus the same CG methylation ratio) from the entire genome of wild-type GOs at P10 and P12, and wild-type and homozygous FGOs.We then determined the longest stretches of methylated CGs (or the largest number of consecutively methylated CGs) in individual reads.This analysis showed that, in each WGBS read group with the same CG number and the same CG methylation ratio, the methylated CG stretches found in wild-type GOs and FGOs were clearly longer than those expected by mutually independent methylation of CG sites with the same methylation ratio (Fig 3B and S4 Table).This suggests the occurrence of processive CG methylation in wild-type oocytes.In contrast, homozygous FGOs had a stretch length distribution closest to that of independent methylation (Fig 3B ), which is supported by the Kullback-Leibler divergences (S5 Table ), suggesting that methylation mediated by DNMT3A ADA is more like distributive than processive in vivo.Although our analysis was limited by the number of CGs contained in individual reads generated by short-read WGBS, these results further support the idea that processive CG methylation is compromised in homozygous FGOs.

Stochastic CG methylation loss at maternally methylated ICRs in derived embryos
To examine the inheritance of decreased CG methylation, we examined three E10.5 embryos from Dnmt3a ADA/ADA females (mat-Dnmt3a ADA/ADA embryos #12, #13, and #16 in S3 Fig) and two embryos from wild-type females as controls (mat-Dnmt3a +/+ embryo #1 and #2).Single nucleotide polymorphisms (SNPs) between laboratory mice (C57BL/6J) and JF1 mice were  ).Published data were used for wild-type GOs collected at P10 [30] and P14 [31] (diameter 60-70 μm).The global methylation level of wild-type GOs at P10, P12, and P14 is 9.6%, 12.8%, and 22.9%, respectively.The global methylation level of wild-type and homozygous FGOs is 35.9% and 17.6%, respectively.Reads containing more than 4 CG sites were used.(B) Comparisons of the longest stretches of methylated CG in individual WGBS reads in wild-type GOs and wild-type and homozygous FGOs.Reads with the same number of CGs and methylated CGs (thus the same CG methylation ratio) were extracted from the WGBS reads.Proportions of WGBS reads with the indicated largest number of consecutive methylated CGs detected in each extracted read group are shown.Gray lines show the distribution of longest stretch lengths expected by mutually independent methylation of CG sites with the same used for allele-specific analyses [23].The overall CG methylation level was similar in all embryos regardless of the genotype (S4 Fig and S3 Table ), and a SNP-based analysis confirmed this for both parental genomes (Fig 4A and S3 Table).Various genomic annotations and specific CpG islands that showed reduced methylation in homozygous FGOs (Fig 2D and 2F) had normal methylation levels (Fig 4B and 4C), suggesting that the maternal genome regains the normal CG methylation level before this stage.We then examined the maternally methylated ICRs as their methylation loss could persist in embryos [1][2][3].Twelve maternally methylated and two paternally methylated ICRs had informative SNPs, and strikingly, the maternal allele of the maternally methylated ICRs showed "stochastic" loss of CG methylation in embryos derived from homozygous females (Fig 4D).There was embryo-to-embryo variation as to which ICR loses methylation, and in each embryo, some maternally methylated ICRs lost methylation while others regained full methylation.Partial methylation loss was also observed
A cluster analysis of the transcriptomes of the individual embryos revealed that the similarity of the transcriptomes depends on the severity of defect (see S3 Fig for morphology) rather than the genotype (Fig 5A).However, when we attempted to identify genes differentially expressed between mat-Dnmt3a +/+ and mat-Dnmt3a ADA/ADA embryos, eight genes were either upregulated or downregulated (FDR < 0.05), of which four were linked to the maternally methylated ICRs (Zdbf2, Blcap, Mest, and Nnat; maternally imprinted genes) (Fig 5B and S7 Table ).
We therefore examined the expression of a total of 45 maternally imprinted genes in individual embryos: while their expression level was relatively constant in mat-Dnmt3a +/+ embryos, it varied greatly in mat-Dnmt3a ADA/ADA embryos (S5A Fig and S8 Table ).The imprinted genes showed either upregulation or downregulation, depending on their functional relationship with the linked ICR [32], in some of the mat-Dnmt3a ADA/ADA embryos but not in others.A cluster analysis using the expression levels of the 45 imprinted genes clearly distinguished mat-Dnmt3a ADA/ADA embryos from the other genotypes (Fig 5C).However, the exact number of the affected imprinted genes and degrees of their changes were different from embryo to embryo (S5B Fig) .These results suggest that stochastic and varied misregulation of the maternally imprinted genes is a hallmark of embryos derived from homozygous females.Gm14296 was a strongly downregulated non-imprinted gene (Fig 5B and S7 Table), and it turned out that this gene was secondarily regulated by imprinting through Klf14, a maternally imprinted gene linked to the Mest ICR [33]  We then identified allelic SNPs between the JF1 and C57BL/6J genomes in 16 maternally and 2 paternally imprinted transcripts (genes) and used them to examine whether or not the stochastic methylation loss was linked to the stochastic misregulation of the maternally imprinted genes in cis.It was found that, in each embryo, there was indeed a perfect link between the loss of methylation at the ICRs and misregulation of the linked genes exclusively in the maternal genome (Fig 5D and S5C Fig).The aberrant upregulation of Zdbf2 was correlated with a gain of methylation at the secondarily differentially methylated region, not methylation loss at the ICR, consistent with the reported unique developmental regulation at this locus [34,35].Since imprinted genes are crucial for embryonic development [36], the stochastic loss of maternal imprinting (and hence the difference in affected gene set) likely explains the loss of embryos at various stages.To examine whether the conversion occurs before or after implantation, we derived diploid parthenogenetic blastocysts possessing only maternal genomes from ovulated wild-type and Dnmt3a ADA/ADA oocytes (see Materials and methods) and performed WGBS (S3 Table ).The advantage of using parthenogenetic blastocysts is that, unlike SNP-based analysis, all WGBS reads obtained from a small number of cells could be used to examine CG methylation of the maternal genome.While parthenogenetic blastocysts from wild-type oocytes showed 45%-80% CG methylation at the ICRs, recapitulating the methylation levels in normal blastocysts [37], those from Dnmt3a ADA/ADA oocytes showed less than 50% methylation, with some ICRs showing near complete methylation loss (Fig 6B).This suggests that the stochastic loss occurs, or at least initiates, before implantation but that the regaining occurs after implantation, most likely during the de novo methylation phase.

Discussion
Our study revealed an important biological function of the ADD domain of DNMT3A in somatic cells and oocytes.The impact of the Dnmt3a ADA mutation in somatic cells was manifested as the dwarf phenotype of homozygous mice.This contrasts with the overgrowth phenotype of Tatton-Brown-Rahman syndrome patients with mutations in this domain [22].Although the cause of this discrepancy is currently unknown, it is noteworthy that mutations in the PWWP domain of human DNMT3A can lead to both overgrowth (Tatton-Brown-Rahman syndrome) and microcephalic dwarfism [8,22].The homozygous females also had delivery problems, but embryos from such females were lost at various stages even in healthy foster mothers, suggesting that homozygous oocytes, in most of the cases, cannot support normal development.This is most likely due to the partial loss of DNA methylation in such oocytes and stochastic loss of maternal imprinting in derived embryos (see below).
Second, our study suggested a new molecular function for the ADD domain: in addition to acting as an activation switch and excluding H3K4me2/3-marked regions from methylation, it likely contributes to processive CG methylation.Perhaps the ADD domain tethers DNMT3A to chromatin unmethylated at H3K4 and allows the enzyme to act on consecutive CG sites in a processive manner (S6 Fig).This explains why DNMT3A, which on its own does not show processivity [38], can perform processive CG methylation.The role of the ADD domain in non-CG methylation can also be explained in this context.While we note that the model needs to be tested preferably by an in vitro methylation assay with reconstituted nucleosomes, such an assay system is not yet in our hands: we therefore leave it for a future study.The fact that DNMT3L, an interacting partner of DNMT3A, also has an ADD domain [12] suggests the functional importance of this domain in the DNMT3A/DNMT3L complex.In this regard, it is interesting that the PWWP domain of DNMT3A, which recognizes histone H3K36me2/3 [5][6][7]10], only plays a role in limiting ectopic CG methylation [8,9,11].We thus speculate that, while both ADD and PWWP domains recognize specific modification states of the histone H3 tail, the former contributes to more efficient enzymatic activity and the latter to higher specificity.
Lastly, we found a stochastic loss of maternal imprinting in embryos derived from homozygous oocytes: the partial methylation loss at the maternally methylated ICRs in oocytes was frequently converted to either complete loss or full recovery of CG methylation in a stochastic manner during development.Our study on parthenogenetic embryos derived from homozygous oocytes revealed that the stochastic loss occurs before implantation and the regaining occurs after implantation.Thus, it appears that the conversion of the partial ICR methylation to an all-or-none pattern follows the global wave of demethylation in pre-implantation embryos and subsequent de novo methylation in early post-implantation embryos.Although the determinant for gain or loss is currently unknown, methylation states at some key CG sites, such as those recognized by ZFP57, could affect the entire ICR [39,40].ZFP57 is a nuclear factor that binds to its target sequences within the ICRs [40], only when the internal CG site is methylated, and maintains their methylated state against the demethylation wave in preimplantation embryos [39].Consistent with this idea, a recent study by epigenome editing showed that artificial gain and loss of CG methylation at ZFP57 binding sites respectively alter the imprinting state of the ICR and whole imprinted gene cluster in embryonic stem cells, which then persists through neuronal differentiation [41].The stochastic imprinting loss poses a model to study how partial epigenetic changes induced in germ cells by, for example, environmental factors [42,43] can be transmitted to the successive generations and affect their phenotype.
In conclusion, our study reveals that the DNMT3A ADD domain contributes to efficient, and likely processive, de novo DNA methylation, maternal imprinting, and normal embryonic development.Since mutations in this domain are found in congenital growth disorders and hematologic malignancy in humans [19][20][21][22], our findings provide a molecular basis for understanding these diseases.The study also identifies a case to study how stochastic epigenetic inheritance can occur in mammalian reproduction.

Ethics statement
All animal experiments were performed under the ethical guidelines of Kyushu University and the protocols were approved by the Institutional Animal Care and Use Committee of Kyushu University (approval number: A22-087-1).Animals were group-housed in a specific-pathogen-free facility under the standard housing condition (12-h light/dark cycle with lights on at 8 am, temperature 20-22˚C with ad libitum access to water and food).

Mutant mice
Dnmt3a ADA mutant mice were generated using a CRISPR-Cas9 method previously described by Inui et al. [44].In brief, pX330 plasmid encoding Cas9 and guide RNA (gRNA) and singlestranded donor oligonucleotides (ssODN) were microinjected into mouse zygotes derived from (C57BL/6N× C3H/HeN) F1 females crossed with C57BL/6J males.The gRNA and ssODN sequences were described in S9 Table .The injected zygotes were transferred into the oviducts of pseudo-pregnant ICR females.Pups were genotyped by polymerase chain reaction (PCR) based Sanger sequencing.The primers used for genotyping are listed in S9 Table .A male carrying the expected Dnmt3a ADA mutation was crossed with C57BL/6J females and offspring carrying the mutation were backcrossed to C57BL/6J at least five times.

Fertility testing
Female mice at 8 to 24 weeks old were placed with a C57BL/6J or a JF1 male in the same cage, and vaginal plugs were checked the next morning.

Parthenogenetic blastocysts
Production of parthenogenetic blastocysts was performed as reported by Kishigami and Wakayama [45].To obtain MII oocytes, female mice (10 to 15 weeks old) were injected sequentially with pregnant mare serum gonadotropin and human chorionic gonadotropin.Cumulusoocyte complexes were collected from the ampulla of oviducts and treated with 10 μg/ml hyaluronidase in KSOM medium to remove cumulus cells.After at least 20 minutes culture at 37˚C with 5% CO 2 , the MII oocytes were activated by 5 mM SrCl 2 in 2 mM ethylene glycol-bis (β-aminoethylether)-N,N,N',N'-tetraacetic acid (EGTA)-containing KSOM medium in the presence of 5 μg/ml cytochalasin B for 6 h and cultured in KSOM medium at 37˚C and 5% CO 2 for 5 days (120 h).

Embryo dissection and RNA/DNA extraction
Embryos and placentas were dissected from the uteri of female mice crossed with a JF1 male and stored in phosphate-buffered saline.DNA and total RNA were extracted from embryos using AllPrep DNA/RNA Mini kit (QIAGEN).

WGBS single read analysis
Individual WGBS reads containing the same number of CGs and the same number of methylated CGs (thus the same CG methylation ratio) were selected from the entire genome of wildtype GOs at P10 and P12, and wild-type and homozygous FGOs (6 methylation ratio, in the range of 41.7% (5/12) and 54.5% (6/11) CG methylation).We then determined the longest stretches of methylated CGs (or the largest number of consecutively methylated CGs) in individual reads and compared the mean maximum length between the different samples (genotype, developmental stage, expected) for each methylation ratio.Kullback-Leibler divergences of the observed distributions of the longest stretches of methylated CGs per read from the expected distribution were calculated for each methylation ratios using R.

Protein synthesis and purification
Complementary DNAs (cDNAs) were generated using 1 μg total RNA from the ovaries as templates using FastGene RNA Premium kit (NIPPON Genetics).cDNA fragments encoding either wild-type or mutated ADD domain (residues 472-605) were amplified by PCR and cloned into pGEX6p-1 vector.The primers used for cloning are listed in S9 Table .The glutathione S-transferase fusion proteins were expressed in Escherichia coli strain Rosetta.The cultures with optical density of 0.6-0.8 were added with 0.2 mM isopropyl-β-Dthiogalactopyranoside and incubated at 20˚C for 15 hours.After sonication, the supernatant of cell lysate was mixed with Glutathione Sepharose 4 Fast Flow (Cytiva) and applied onto columns for gravity-flow chromatography (Polypropylene Columns, QIAGEN).Then, the fusion proteins were recovered with elution buffer (50 mM Tris-HCl, 10 mM glutathione, pH 8.0).The eluted proteins were purified by Amicon Ultra Centrifugal Filters and subjected to SDS-PAGE followed by Coomassie blue staining for product size confirmation.

Fig 1 .
Fig 1.Generation and phenotype of Dnmt3a ADA mice.(A) Structure of mouse DNMT3A isoforms and positions of D525A and D527A substitutions.Known domains (PWWP and ADD) and motifs of the methyltransferase (MTase) domain are indicated by colored boxes.A major isoform expressed in oocytes is DNMT3A2.An example of genotyping by Sanger sequencing is shown.(B) Western blotting of DNMT3A2 in FGOs.Proteins prepared from 50 FGOs were loaded for each genotype.ACTB is a loading control.(C) Fertility of Dnmt3a ADA/ADA females crossed with a C57BL/6J male.Female #1 was subjected to Caesarian section, and two stillborn pups and six degenerated fetuses were obtained.(D) Representative images of embryos recovered at E10.5.Scale bar, 1 mm.https://doi.org/10.1371/journal.pgen.1010855.g001

Fig 2 .Fig 3 .
Fig 2. Impact of Dnmt3a ADA mutation on CG and non-CG methylation in FGOs.(A) Violin plots showing distributions of CG and non-CG methylation levels of 10-kb genomic bins in pooled FGOs of indicated genotypes.The number above each plot indicates the global methylation level.Data obtained from biological replicates (each comprising 100 to 250 FGOs) were combined after confirmation of reproducibility (S3 Table).(B) Genome browser view of CG and non-CG methylation levels of 10-kb bins in FGOs.(C) Scatter plots comparing the CG methylation levels of 10-kb bins between FGOs of indicated genotypes.(D) CG methylation levels of indicated genomic annotations including various repeats in FGOs.Gene bodies and intergenic regions include repeats.(E) CG methylation levels of maternally and paternally methylated ICRs in FGOs of indicated genotypes.(F) CG methylation levels of promoter CpG islands that are normally methylated in FGOs.Six representative islands among those that we identified to be methylated in wild-type FGOs (n = 251) are shown.https://doi.org/10.1371/journal.pgen.1010855.g002 (see Nespas-Gnasxl, Peg3 and Zac1 in Fig 4D).No change was observed in the paternally methylated ICRs.methylation ratio.The WGBS data from P14 GOs was not used due to the small data size.See above for the global methylation level of each sample.mCGs, methylated CGs.https://doi.org/10.1371/journal.pgen.1010855.g003

Fig 4 .
Fig 4. Stochastic CG methylation loss at maternally methylated ICRs in embryos.(A) Violin plots showing the distributions of CG methylation levels of 10-kb genomic bins in the maternal and paternal genomes of E10.5 embryos.The maternal genotype and embryo ID are indicated.Only SNP-containing reads were used for this analysis.Data from embryos of the same genotype were combined.The number above each plot indicates the global CG methylation level.(B) CG methylation levels of indicated genomic annotations including various repeats in individual embryos.(C) CG methylation levels of promoter CpG islands that are normally methylated in FGOs in individual embryos.The islands are the same ones as in Fig 2F (D) CG methylation levels at the maternally and paternally methylated ICRs in the maternal and paternal genomes of individual embryos.ICRs with more than 10 cumulative CG counts (redundant) in SNP-containing reads were analyzed.Red triangles indicate stochastic CG methylation loss.https://doi.org/10.1371/journal.pgen.1010855.g004 (Fig 5D), further corroborating the imprinting-related transcriptomic changes.Lastly, three mat-Dnmt3a ADA/ADA embryos showing largest changes in the imprinted gene expression (#16, #18, and #20) (S5B Fig) were clearly underdeveloped (S3 Fig), suggesting a close link between the imprinted gene misregulation and phenotype.
The frequent occurrence of almost fully methylated and almost fully unmethylated ICRs in E10.5 embryos (Fig 4D) can in principle originate from oocytes or arise during development after fertilization.However, considering the methylation distribution of the WGBS reads in oocytes (Fig 3A), their mosaic partial methylation patterns (S2 Fig), and the size of the maternally methylated ICRs (0.7-6.7 kb, mean 3.0 kb), the occurrence of ICRs with such extreme methylation levels is unlikely in oocytes.Our data thus suggest that the partial CG methylation at each ICR is frequently converted to either full or no methylation during development in a stochastic manner (Fig 6A).

Fig 5 .
Fig 5.The altered expression of imprinted genes is correlated with loss of ICR methylation.(A) A cluster analysis of the transcriptomes from E10.5 embryos.The maternal genotype of the mother and embryo ID (#1-20) are indicated.The most retarded embryos in S3 Fig are indicated in red.(B) Volcano plot showing gene expression changes in embryos of indicated maternal genotypes.Red dots show the 45 maternally imprinted genes (S8 Table).(C) A cluster analysis using the expression profiles of the maternally imprinted genes.(D) Allelic CG methylation states of maternally methylated ICRs and allelic expression states of linked genes in representative imprinted gene clusters.SNPbased allele-specific analyses were conducted.Red triangles show stochastic CG methylation loss and the altered expression of the linked genes.sDMR, secondarily differentially methylated region.CPM, count per million.https://doi.org/10.1371/journal.pgen.1010855.g005

Fig 6 .
Fig 6.Stochastic loss of ICR methylation occurs during development.(A) A model showing the conversion of a partial loss of CG methylation in oocytes to either full or no methylation in embryos.A hypothetical maternally methylated ICR containing five CG sites is shown.Filled and open circles indicate methylated and unmethylated CG sites, respectively.Each ID indicates a single oocyte or embryo.(B) CG methylation levels at maternally and paternally methylated ICRs in individual parthenogenetic blastocysts derived from oocytes of indicated genotypes.https://doi.org/10.1371/journal.pgen.1010855.g006