The biochemical mechanism of short RNA-induced TGS (transcriptional gene silencing) in mammals is unknown. Two competing models exist; one suggesting that the short RNA interacts with a nascent transcribed RNA strand (RNA–RNA model) and the other implying that short RNA forms a heteroduplex with DNA from the unwound double helix, an R-loop structure (RNA–DNA model). Likewise, the requirement for DNA methylation to enact TGS is still controversial. In vitro assays using purified recombinant murine Dnmt (DNA methyltransferase) 1-dN (where dN indicates an N-terminal truncation), 3a and 3b enzymes and annealed oligonucleotides were designed to question whether Dnmts methylate DNA in a RNA–DNA heteroduplex context and whether a RNA–DNA heteroduplex R-loop is a good substrate for Dnmts. Specifically, model synthetic oligonucleotides were used to examine methylation of single-stranded oligonucleotides, annealed oligonucleotide duplexes, RNA–DNA heteroduplexes, DNA bubbles and R-loops. Dnmt methylation activity on the model substrates was quantified with initial velocity assays, novel ARORA (annealed RNA and DNA oligonucleotide-based methylation-sensitive restriction enzyme analysis), tBS (tagged-bisulfite sequencing) and the quantitative PCR-based method MethylQuant. We found that RNA–DNA heteroduplexes and R-loops are poor substrates for methylation by both the maintenance (Dnmt1) and de novo (Dnmt3a and Dnmt3b) Dnmts. These results suggest the proposed RNA/DNA model of TGS in mammals is unlikely. Analysis of tagged-bisulfite genomic sequencing led to the unexpected observation that Dnmt1-dN can methylate cytosines in a non-CpG context in DNA bubbles. This may have relevance in DNA replication and silencing of transcriptionally active loci in vivo.
Functionally, RNA falls into two broad categories, protein-coding RNA and ncRNA (non-coding RNA) . Some ncRNAs regulate the transcription and translation of protein-coding RNA targets through the process of RNAi (RNA interference). RNAi regulates biological processes, including cell growth and development through mRNA degradation, translational repression and heterochromatization. Dynamic regulation of heterochromatin facilitates antiviral defence, transposon silencing, imprinting and cell differentiation. Long ds (double-stranded) RNA processing into small staggered-ended dsRNA to meditate RNAi is a highly conserved process in eukaryotes . Small RNA species include siRNA (short interfering RNA) , miRNA (microRNA)  and piRNA (piwi-interacting RNA) [5,6]. In mammalian PTGS (post-transcriptional gene silencing), siRNAs are loaded into the RISC (RNA-induced silencing complex) which drives binding to cognate sequences of mRNA with subsequent down-regulation of translation and/or cleavage of the target mRNA . In addition to PTGS, appropriate siRNA can be directed towards a promoter region of an actively transcribed gene or genomic repeat element to promote TGS (transcriptional gene silencing) via an epigenetic mechanism probably enacted by DNA methylation, histone modification or both.
DNA methylation is involved in epigenetic regulation of gene expression . In mammals, cytosines within DNA are methylated via Dnmts (DNA methyltransferases). The methyltransferase Dnmt1 has high activity on hemimethylated dsDNA, so-called ‘maintenance’ methyltransferase activity, whereas the Dnmt3a and Dnmt3b enzymes exhibit de novo methylation activity by being most active on unmethylated dsDNA .
In mammals, despite implication of ncRNA such as Xist in imprinting heterochromatin of the X chromosome , evidence for TGS mediated by siRNA and analogous to that in plants  or Schizosaccharomyces pombe  was not forthcoming until more recently [13–18]. There are conflicting reports about the roles of histone and Dnmts in enacting TGS, including claims that DNA methylation is not required for effective TGS [13,16–18]. Efficacious siRNA reported to promote TGS have been mostly 21 nt, but have included those 18–28 nt in size [19,20]. Confusingly, transcriptional activation is also known to occur upon transfection of some promoter-directed siRNA species [21,22]. Evidence suggests that genes with higher levels of transcription are better targets for mammalian TGS , with a requirement for nuclear import of siRNA . A requirement exists for RNAPII (RNA polymerase II) [24,25], TRBP2 (transactivation-responsive RNA-binding protein 2)  and Ago1 (Argonaute 1) [23,25,26] with evidence that Ago1 specifically associates with the promoter of genes targeted for TGS . It also seems that HDAC1 (histone deacetylase 1) is recruited with apparent histone repositioning and induction of H3K9Me2 (histone 3 Lys9 dimethylation) . A recent paper reported transcriptional silencing of c-myc without changes in known epigenetic marks . This process was dependent upon Ago2 (Argonaute2) and a non-coding promoter-associated RNA initiated upstream and overlapping the transcription start site.
Although some studies have found little support for DNA methylation as the primary transcriptional silencing agent in mammalian TGS, there is evidence suggesting that both Dnmt3a  and Dnmt3b (, but see [28a]) are required for RNAi-directed TGS in human cells. There is also speculation that siRNA action is required over an extended period of time to establish DNA methylation . The contrary evidence supporting either chromatin changes or DNA methylation as the effectors of TGS does not discount the possibility that both processes are involved in TGS, suggesting a unified model of silencing. There may well be multiple modes of silencing. Interestingly, Dnmt3a has been shown to co-purify with HDAC1 and a histone H3 methyltransferase, whereas Dnmt1 and Dnmt3b co-fractionate .
The mechanism of TGS is still not well defined and two alternative models speculate on the biochemical interactions [31,32] (Figure 1). The models both agree on the requirement of transcription and Watson–Crick siRNA antisense strand binding with the target nucleic acid. The ‘RNA–RNA’ model suggests the siRNA antisense strand contained by Ago1 or Dnmt3a binds directly to nascent ncRNA strands, localizing TGS to the chromatin neighbouring active transcription. Alternatively, the ‘RNA–DNA’ model predicts binding to the unwound ‘DNA bubble’ directly behind active transcription to create a RNA–DNA heteroduplex which directs TGS to the proximity of the resultant R-loop . Target structures containing ss (single-stranded) DNA to which RNA may bind could also be formed during replication, recombination or DNA repair, but the role of these processes in RNA-directed methylation has not been addressed.
Models for siRNA-directed TGS
In the present paper, we report that R-loops and heteroduplexes are poor substrates for all of the Dnmt enzymes examined, with CpG sites occluded by the binding of a complementary RNA strand showing little methylation. Dnmt3a was inhibited to the least degree, relative to unmethylated dsDNA. Free dsRNA or heteroduplex did not act as an inhibitor. These results suggest the proposed RNA–DNA model of TGS in mammals is unlikely. In addition, Dnmt1-dN (murine DNA methyltransferase 1 with an N-terminal truncation) was found to strongly methylate a non-CpG context cytosine within the ss region of the DNA bubble substrate. This is the first observation of efficient non-CpG methylation by Dnmt1.
MATERIALS AND METHODS
Design, labelling and annealing of DNA oligonucleotides and siRNA
Oligonucleotides of 76 nt and 30 nt with six CpG sites (S) and four CpG sites (S′) respectively, together with complementary strands (A, A′) and counterparts with all CpG sites methylated (mS, mA, mS′, mA′) were synthesized (Geneworks, Adelaide, Australia) (see Supplementary Table S1 at http://www.BiochemJ.org/bj/432/bj4320323add.htm). The CpG sites were flanked by the appropriate bases to create recognition sequences for the methylation-sensitive enzymes AciI, HpaII and HhaI. Non-annealed 28 nt siRNA complementary to the central section of these oligonucleotides (Q and R) was synthesized (Ambion, Austin, TX, U.S.A.). A further oligonucleotide complementary to the sense oligonucleotide but with mismatches around the RNA-complementary region was also designed (L), as was its complementary strand (K). In designing the oligonucleotides, careful considerations were given to avoiding high self-complementarity; however, the requirement for repeated palindromic enzyme recognition sequences meant not all self-complementarity could be eliminated.
S and L oligonucleotide (10 pmol) and 1.0 μl of 1:10 diluted 10 bp molecular mass ladder (Promega) were 5′-labelled using 6.72 pmol (40 μCi and 336 nM) of Redivue™ 5′-[γ-32P]ATP, triethylammonium salt (6000 Ci·mmol−1) (Amersham Biosciences) and 10 units of T4 polynucleotide kinase (NEB) according to the manufacturer's instructions. Labelling of 10 pmol of R siRNA was as above except with 1.68 pmol of 5′-[γ-32P]ATP. The reaction mixture was cleaned up with a Qiagen nucleotide removal kit according to the manufacturer's instructions and resuspended in 30 μl of TE (Tris/EDTA) buffer containing 1.0 mM DTT (dithiothreitol) with storage at −20 °C until required.
Oligonucleotides were kept ss or annealed. R and Q strands were annealed to create the duplex siRNA. S, mS, S′ and mS′ were annealed to their CpG methylated or unmethylated complements to create unmethylated, hemimethylated and fully methylated dsDNA. RNA–DNA heteroduplexes were formed by annealing S or S′ to R, whereas DNA bubbles were created by annealing S to L. Formation of R-loops required the annealing of S to R and L. Annealing of oligonucleotides was performed in non-stick microcentrifuge tubes (Scientific Specialties) with 3.33 pmol of each oligonucleotide (and 6.66 pmol of RNA if required), in 22.2 μl of annealing buffer (10 mM Tris/HCl, pH 8.0, 2.0 mM MgCl2 and 1.0 mM DTT); likewise, ds siRNAs were prepared by adding 6.66 pmol of R to equimolar Q in annealing buffer. All tubes were placed in a 95 °C heat block and left to cool to room temperature (21 °C) over approx. 90 min. Tubes were then briefly centrifuged at 16000 g for 30 s before storage at 4 °C until use. The clean up procedure after 32P labelling results in some loss of the labelled strand so duplex substrates will have the unlabelled strand present at slight excess. This small excess of a strand was found to be necessary to efficiently generate the R-loop and DNA bubble substrates. Annealing of products was confirmed by electrophoretic mobility-shift assay using Criterion™ 10% TBE (Tris/borate/EDTA) PAGE gels (Bio-Rad Laboratories) run at 150 V. For Dnmt initial velocity assays, ss substrates were heated to 95 °C and cooled rapidly on ice immediately before use.
Recombinant Dnmt enzymes
His-tagged mouse recombinant Dnmt3a and Dnmt3b enzymes and a truncated form of mouse Dnmt1-(291–1620) (Dnmt1-dN) enzyme were prepared and purified from Sf9 insect cells as described previously [33,34]. Dnmt1-dN was the only Dnmt1 preparation available with sufficient activity for these experiments so was used instead of commercial sources of Dnmt1. The Dnmt1 truncated N-terminus is involved in PCNA (proliferating-cell nuclear antigen) binding, allosteric activation by fully methylated DNA and is one of three DNA-binding regions in the enzyme [35,36]. The N-terminus truncation does not affect kinetics or substrate specificity .
Dnmt initial velocity assays
Initial velocity assays based upon a method described previously  were used to examine the methylation of the unlabelled substrates in the presence of 1.62 pmol of Dnmt3a, 1.25 pmol of Dnmt3b or 0.33 pmol Dnmt1-dN. To establish substrate and enzyme concentrations, enzyme velocities on a number of the substrates were determined experimentally. In brief, 25 μl reaction volumes contained ‘M buffer’ (2.7 M glycerol, 5 mM EDTA, 0.2 mM DTT, 25 mM NaCl and 20 mM Tris/HCl, pH 7.4) with the above quantities of purified Dnmt, 10 pmol of oligonucleotide substrate or 0.1 μg of poly(dI-dC)·(dI-dC) or poly(dG-dC)·poly(dG-dC) and 67 pmol (1.0 μCi and 2.65 μM) of [3H]AdoMet (S-adenosylL-methionine) (15 Ci·mmol−1) (Amersham Biosciences). After 1 h of incubation, the reaction was terminated with 1.5 mM non-radioactive AdoMet. Subsequently, mixtures were incubated with 0.1 μg of proteinase K (Nakalai Tesque) at 50 °C for 20 min, then washing and scintillation counting was performed as described previously .
Sodium bisulfite conversion
Bisulfite conversion was performed using an EZ DNA Methylation-Gold Kit™ (Zymo Research), employing altered thermal cycling conditions, but otherwise according to the manufacturer's instructions. The altered conditions were 5 min at 99 °C, 25 min at 60 °C, 5 min at 99 °C, 85 min at 60 °C, 5 min at 99 °C and 175 min at 60 °C. After purification, samples were resuspended in 10 μl of elution buffer.
tBS (tagged-bisulfite sequencing)
Reaction volumes of 15 μl were composed of 394 fmol of substrate, 2.4 nmol of AdoMet (NEB), M buffer and Dnmt enzyme(s). Dnmt3a or Dnmt3b reactions used 3.15 pmol of enzyme, whereas Dnmt1-dN reactions used 0.418 pmol of enzyme with or without an additional 0.418 pmol of Dnmt3b. Dnmt reaction mixtures were incubated at 37 °C for 4 h before addition of 5 μl of 40 mM Tris/HCl (pH 8.0), 0.4% SDS, 0.025 μg·μl−1 proteinase K and 200 ng of phage λ carrier DNA. Tubes were incubated for 20 min at 50 °C before bisulfite conversion as described above. Then, 2 μl of the eluate was used per 25 μl PCR. Overhanging tailed primers (see Supplementary Table S1) were used to create an extended amplicon. BiS-F and BiS-R amplified the bisulfite-converted S oligonucleotide to give amplicon S, whereas BiA-F and BiA-R amplified the bisulfite-converted L oligonucleotide to give amplicon L. Control oligonucleotides (see Supplementary Table S1) designed to replicate post-bisulfite treatment S and L amplicons with CpG sites methylated (S-MAT, L-MAT) or unmethylated (S-MIS, L-MIS) were synthesized (Geneworks, Adelaide, Australia). PCR thermal cycling of controls or experimental DNAs used an initial cycle of denaturation at 95 °C for 8.25 min then a further two cycles of 90 °C for 15 s, 20 °C for 30 s and 72 °C for 30 s before 45 cycles of 90 °C for 15 s, 52 °C for 15 s and 72 °C for 30 s. PCRs used 0.5 unit of Platinum Taq (Invitrogen), 2 mM Mg2+, 200 nM primers and 200 μM of each dNTP. BiS-Gtag and BiS-Ctag or BiA-Gtag and BiA-Ctag primers (see Supplementary Table S1) based on the tag-modified bisulfite genomic DNA sequencing method  were used in a 30 μl touchdown PCR to amplify and add C-tag- and G-tag-compatible ends to 1 μl of amplicons S and L respectively. Thermal cycling conditions consisted of an initial 1 min 95 °C denaturation step then 20 cycles of denaturation at 95 °C for 15 s, annealing initially at 72 °C for 20 s and extension at 72 °C for 60 s. The annealing temperature was reduced by 1 °C per cycle until reaching 56 °C. PCR used 2 units of Platinum Taq, 1.5 mM Mg2+, 500 nM primers and 400 μM of each dNTP. PCR products were run on 4% Metaphor® (Lonza) agarose gels at 100 V; post-electrophoresis, GeneCatcher™ disposable band-picking tips (Gel Company) were used to remove the appropriately sized bands. DNA from agarose slabs was purified with a Wizard SV PCR and Gel Extract kit (Promega), according to the manufacturer's specifications. DNA was quantified using a NanoDrop ND-1000, and a sample of this DNA was subjected to forward and/or reverse DNA sequencing using an Applied Biosystems 3730xl DNA Analyzer using the C-tag forward primer and G-tag reverse primer (see Supplementary Table S1). At the six CpG sites, sequence chromatograms had C and T peak heights for forward reads, or G and A peak heights for reverse reads, determined manually within chromatogram-visualization software. Estimation of methylation was performed by dividing the peak height of the C (or antisense G) against the combined height of the C+T peaks and generating a C/C+T (or antisense A/A+G) ratio expressed as a percentage as described previously . Graphing was implemented using the ‘stars’ function within R version 2.6.1 .
ARORA (annealed RNA and DNA oligonucleotide-based methylation-sensitive restriction enzyme analysis)
Reaction tubes were prepared similarly to those in the tBS assay except 15 μl reaction volumes contained 394 fmol of 5′-32P-labelled substrate and the reactions were allowed to continue overnight at 37 °C. To reaction mixtures containing 5′-[32P]S, the quantity of A was bought to 1.0 pmol in 20 μl. In reactions containing 5′-[32P]L, the quantity of K was bought to 2.0 pmol in 20 μl. The tubes were heated and left to cool as described above. After brief centrifugation at 16000 g for 30 s, 3.0 μl aliquots were digested in parallel with 10 units of HpaI, MspI, AciI or HhaI (NEB) in 15 μl volumes in non-stick tubes according to the manufacturer's instructions. Restriction enzyme digests were carried out overnight at 37 °C. Formamide loading buffer (30 μl) was added (95% deionized formamide, 0.025% Xylene Cyanol, 0.025% Bromophenol Blue, 24 mM EDTA and 0.025% SDS) and the opened tube was heated to 95 °C for 10 min and loaded immediately on to a Criterion™ 15% TBE/urea denaturing PAGE gel (Bio-Rad Laboratories). Gels were run at 250 V then dried, with subsequent exposure to a Molecular Dynamics storage phosphor screen and image capture via a Molecular Dynamics Typhoon 8600 imager. All image analysis was performed with ImageQuant TL software V2005 (GE Healthcare) and data were exported for processing and graphing within R. Gel-based analysis of RNase activity was performed by annealing 1.185 pmol of 5′-[32P]R with 394 fmol of S per tube and setting up ARORA reactions with this substrate. At time intervals of 0 h, 1 h, 4 h and overnight, 3.0 μl aliquots were removed and combined with 6 μl of formamide loading buffer with storage at 4 °C until being run on Criterion™ 10% TBE PAGE gels at 150 V.
DNA and RNA–DNA heteroduplex substrates
30-mer and 76-mer oligonucleotides were used to generate ds DNA–DNA and RNA–DNA duplexes. In order to form a stable R-loop structure, we used a 76-mer lower strand (L) that is pseudo-complementary to the S DNA strand such that a number of bases, particularly those near the termini of the 28-mer RNA strand, cannot form Watson–Crick base pairs. Following successive incubation of the RNA strand and the L strand with the S strand, it is possible to detect the formation of stable R-loop-like structures by electrophoretic mobility-shift assay. Figure 2 shows the migration profile of the R-loop structure, the DNA bubble structure lacking the RNA strand and other ss and ds substrates. The dual-S oligonucleotide bands (Figure 2) are a consequence of the annealing conditions and presumably represent a degree of intra- or inter-molecular hybridization. The S oligonucleotide was confirmed by PAGE under denaturing conditions to be the expected length (see Supplementary Figure S1 at http://www.BiochemJ.org/bj/432/bj4320323add.htm). The duplex, DNA bubble and R-loop were determined to be highly stable (see the Supplementary Online Data and Supplementary Figure S2 at http://www.BiochemJ.org/bj/432/bj4320323add.htm). In incubations with Dnmt enzymes we looked for evidence of DNase or RNase activity that might destabilize the annealed structures. No DNase activity was evident. Whereas free RNA was degraded in prolonged incubations, annealed RNA was stable (see Supplementary Figure S3 at http://www.BiochemJ.org/bj/432/bj4320323add.htm). The formation of stable synthetic R-loops and DNA bubbles provides a tool to study a proposed model of siRNA-directed DNA methylation mechanisms with purified enzymes in vitro.
Electrophoretic mobility-shift assay with annealed oligonucleotide products
Dnmt enzyme activity on control substrates
DNA and RNA–DNA substrates (Figure 3A) were incubated in vitro with purified recombinant Dnmt1-dN , Dnmt3a or Dnmt3b and activity was measured by incorporation of 3H-labelled methyl groups from AdoMet (Figure 3B). Reactions were stopped after 1 h, in the initial linear velocity phase. As expected, Dnmt1-dN had an activity much higher than that of Dnmt3a and Dnmt3b on hemimethylated DNA duplexes, consistent with its maintenance methylation function (Figure 3B). In the present study, the Dnmt1-dN enzyme was approx. 2.5-fold more active on hemimethylated dsDNA (mS–A or S–mA) than on unmethylated dsDNA (S–A) substrate. The favouring of hemimethylated over unmethylated substrate is relatively low compared with ratios of velocities in much of the literature for recombinant human and murine Dnmt1 purified from baculovirus-infected Sf9 cells ; however, it is in agreement with the relative reaction velocity of the study using the same Sf9-cell-derived Dnmt1-dN preparation . Little methylation was observed on fully CpG methylated duplexes or methylated ss oligonucleotide. Activities of Dnmt3a and Dnmt3b on poly(dI-dC)·(dI-dC) and dsDNA oligonucleotides are comparable with those of previous studies [33,34,40,41]. Unlike Dnmt1-dN, Dnmt3a displayed a strong preference for unmethylated DNA duplexes consistent with its role as a de novo Dnmt (Figure 3B). Dnmt3a methylation on hemimethylated substrates was about half the rate exhibited with the corresponding unmethylated substrate. The ss substrates were only methylated to around one-fifth of that demonstrated with the unmethylated dsDNA substrate. The remarkable consistency of reaction velocities across hemimethylated substrates and between all ssDNA substrates suggests that Dnmt3a does not show a distinct bias for sequences flanking the CpG site. Using an established initial velocity assay [33,40], we find no obvious flanking sequence preference by Dnmt3a.
Dnmt initial velocity assays
Although Dnmt3b also prefers unmethylated dsDNA substrates, it has different selectivity compared with Dnmt3a, with around half of the reaction velocity relative to unmethylated duplex, on ss substrates or hemimethylated duplexes. Also unlike Dnmt3a, Dnmt3b displays a bias towards methylating the S–mA 30-mer and 76-mer duplexes and the 30-S and A oligonucleotides, suggesting that Dnmt3b has some CpG site flanking base preference. Interestingly, Dnmt1-dN also shows a degree of strand bias, which, on 76-mer hemimethylated substrates, is the opposite of that seen with Dnmt3b. This opposing strand bias of Dnmt1 and Dnmt3b observed in initial velocity assays, together with a report of in vitro co-fractionation , led to later experiments to also considering Dnmt1 and Dnmt3b admixtures. Low methylation activity on the fully methylated substrates by all three Dnmts suggests that there is minimal non-CpG methylation.
Dnmt activity on RNA–DNA heteroduplex, R-loop and DNA bubble structures
Compared with the S–A duplex, all 76-mer ss DNAs, the DNA bubble or the R-loop exhibited around half the velocity. On the fully duplexed S–R′ 30-mer RNA–DNA heteroduplex, activity of all three Dnmts was reduced to less than 10% relative to the unmethylated dsDNA control (or by 3–5-fold relative to ssDNA) (Figure 3B). Dnmt3a retained the highest level of activity. Little activity was observed on the S–R RNA–DNA heteroduplex substrate. In this instance, Dnmt3b retained the highest level of activity, presumably from methylation of the ss ends not hybridized to RNA. Activity on the S–R heteroduplex was only ~30% of the activity recorded with the S–R–L R-loop. Interestingly, the sum of the L and S–R activities is similar to that seen with S–R–L. Compared with Dnmt3a, Dnmt3b had higher activity on the 76-mer S–R heteroduplex, the DNA bubble and R-loop substrates, consistent with its greater activity on ssDNA. It seems as though Dnmt3b is more capable than Dnmt3a of methylating non-duplex structures.
Further assays were run to elucidate whether RNA–DNA heteroduplexes act as inhibitors or are just poor substrates. A non-competitive inhibition mechanism was examined by assaying Dnmt3b methylation in the presence of 0.1 μg of poly(dG-dC)·poly(dG-dC) together with 10 pmol of S–R′ heteroduplex (see Supplementary Figure S4A at http://www.BiochemJ.org/bj/432/bj4320323add.htm). Lack of significant reduction of poly(dG-dC)·poly(dG-dC) methylation in the presence of the heteroduplex suggested that the enzyme was still completely active. Adding increasing amounts of heteroduplex to S–A′ duplex and assaying methylation potential of the mixture of substrates can reveal a competitive reaction mechanism if the initial velocity falls with increasing concentrations of inhibitor. Adding heteroduplex, even up to an equimolar amount with duplex, did not lower the initial velocity (Supplementary Figure S4B). Instead, the velocity was slightly increased, consistent with the heteroduplex not being an inhibitor but a poor substrate. If this is so, endonuclease cleavage of RNA bound to the DNA should restore methylation potential and the presence of RNA non-complementary to S should not significantly alter the methylation potential of S. Indeed, pre-incubation of heteroduplex substrates with 30 units of RNaseH restores methylation potential and the presence of RNA non-complementary to DNA only reduced methylation potential by approx. 25% (Supplementary Figure S4c). Although heteroduplexes are not inhibitors of Dnmt enzymes, they serve as poor substrates. In the presence of dsRNA, namely annealed Q–R siRNA, Dnmt3a and Dnmt3b exhibited an approx. 25% reduced rate of methylation relative to the S–A′ duplex or S′ ssDNAs alone (Figure 3B). The reduction in initial velocity observed in the presence of siRNA ds and ss non-complementary siRNA (Supplementary Figure S4C) could be due to inhibition by the binding of siRNA to an allosteric site on the enzyme or by acting as a ‘decoy’ nucleic acid in the active site. Alternatively, in the case of the duplex, melting of some of the siRNA and reannealing to the S′ strand to form the ‘poor’ heteroduplex substrate may explain this observation. Interestingly, incubation with RNaseH increases the initial velocity to that observed with S′ alone (Supplementary Figure S4C).
Methylation of individual CpG sites across substrates
We initially used sensitivity to restriction digestion to estimate the level of methylation at individual CpG sites. After overnight incubation with Dnmt enzymes, radiolabelled substrates were denatured and cooled in the presence of excess oligonucleotide complementary to the labelled strand before cutting with methylation-sensitive restriction enzymes (Figure 4A). Using gel electrophoresis, the proportion of labelled material subject to cleavage, i.e. unmethylated, could be estimated (Figures 4B and 4C). Since the analysis of initial velocity data showed that Dnmt1-dN and Dnmt3b had complementary strand biases on ds substrates, we also examined methylation site specificities of Dnmt1-dN/Dnmt3b admixtures. In Dnmt negative controls, some labelled substrate was still resistant to cleavage. Despite technical limitations of this method (see the discussion in the Supplementary Online Data), a consistent increase in AciI and HhaI nuclease digestion of the S strand in the DNA bubble, R-loop, RNA–DNA heteroduplex or ssDNA was evident, implying reduced methylation potential of these substrates (Figure 4). The only exception was reduced AciI cutting of the S strand in the DNA bubble after Dnmt3b methylation. Methylation activity is greatest in Dnmt1-dN or Dnmt1-dN/Dnmt3b admixture preparations and less with Dnmt3a and Dnmt3b. Relative to the S–L DNA bubble, the S–R–L R-loop was more highly methylated in the L strand ss region. Comparison of the relative methylation of CpG Site 1 that is in dsDNA and the adjacent Site 2 in the bubble region of the S–L and S–R–L substrates shows the inhibitory effect of RNA–DNA heteroduplex formation. Also, Site 1 on the S strand and Site 6 on the L strand were slightly more methylated in the S–R–L substrate compared with S–L, suggestive of increased Dnmt occupancy of the duplexed strand with a free 5′ end.
ARORA methylation estimates
tBS provided a more sensitive method of measuring methylation at individual CpG sites. Here, Dnmts were incubated with substrates for 4 h before bisulfite treatment, then the upper (S) strand was amplified using tagged primers and sequenced, with methylation estimated. The DNA bubble and R-loop substrates also had the lower strand (L) separately amplified and sequenced. The four enzyme-exposed and one enzyme-negative control sets of tagged and amplified substrates as well as a pool of synthetic fully methylated and unmethylated tagged oligonucleotides were sequenced (Figure 5). Comparison of the forward and reverse sequencing estimates of methylation demonstrated good concordance across all six CpG sites (see Supplementary Figure S5 at http://www.BiochemJ.org/bj/432/bj4320323add.htm). The ratio-based method of methylation estimation used is prone to systematic error if chromatograms have a high degree of baseline noise. However, estimates of methylation in fully methylated and unmethylated controls were accurate, showing the systematic error to be low (see Supplementary Figure S6 at http://www.BiochemJ.org/bj/432/bj4320323add.htm). An established quantitative PCR assay, MethylQuant (see the Supplementary Online Data), was used as a further independent quantitative measure of the methylation of CpG Site 1, the site hardest to determine methylation accurately using sense strand tBS. There was relatively good concordance between the MethylQuant and Site 1 tBS data (see Supplementary Figure S7 at http://www.BiochemJ.org/bj/432/bj4320323add.htm). A consistency was observed across initial velocity data, ARORA enzyme cleavage data and tBS sequencing with respect to both relative enzyme activity and enzyme substrate preference. It is possible that the relative difference in observed methylation potential of the sense and antisense strand data is a technical artefact resulting from amplification bias in the tBS or MethylQuant procedures. ARORA data suggest that methylation on the S and L strands is similar. As such, comparisons of absolute tBS methylation estimates between the two partner strands must be interpreted with caution. Instead, the sequencing data should be considered as indicative of relative methylation potential at sites within the L strand.
tBS methylation estimates
In agreement with initial velocity results, the tBS sequencing data suggests that, on dsDNA, Dnmt3a shows remarkably consistent methylation rates across all CpG sites in the ds substrate, whereas Dnmt1-dN and Dnmt3b demonstrate significant site bias. The tBS data suggested that Dnmt3b strongly favours Site 3, whereas Dnmt1-dN methylated Sites 1, 3 and 4 more than Sites 2, 5 and 6. The site-specificity of all enzymes on the ssDNA was strikingly different from the same sequences in the context of dsDNA. For Dnmt3a and Dnmt3b, methylation occurred almost exclusively at CpG Site 3. Although this site was also strongly preferred by Dnmt1, there was additional methylation of other sites, particularly Sites 2, 5 and 6. Figure 4(C) shows that AciI cut very inefficiently across all substrates, consistent with the methylation preference of Site 3 observed in tBS data.
The presence of an R-loop (compare S–L and S–R–L) or RNA heteroduplex (compare S and S–R) strongly inhibits DNA methylation at CpG sites complementary to the RNA strand. Dnmt3a shows the most residual activity. The presence of RNase activity in the Dnmt3a preparation could account for this effect; however, incubation of labelled RNA–DNA heteroduplex with a Dnmt preparation suggested that RNase activity on annealed RNA was very low (see Supplementary Figure S3). When methylating the DNA bubble or ssDNA substrates, the de novo Dnmt enzymes show similar methylation biases. Dnmt1-dN shows results similar to the de novo enzymes, except that the DNA bubble is favoured for methylation over the ds part of the substrate.
If Dnmt3b and Dnmt1-dN are both present, enhanced methylation of the R-loop structure is observed, compared with activity of either enzyme alone. Compared with methylation of the DNA bubble structure, methylation of CpG Sites 2–5 in the RNA–DNA heteroduplex part of the R-loop S strand is reduced, whereas that of CpG Sites 1 and 6 in the ds flanks is enhanced. In contrast, there is a moderate increase in methylation of CpG sites in the ssDNA region. The low methylation activity of both the enzymes separately on R-loops and heteroduplex implies that RNase activity is low, so, upon combining preparations, the increased activity on the RNA-bound DNA strand is probably not due to RNase activity degrading the heteroduplex. Furthermore, the tBS methylation pattern on the DNA bubble by the admixture is different from that exhibited on the R-loop.
An interesting and unexpected result was observed in sequencing of the Dnmt1-dN products on the DNA bubble structure. Two ‘S’ strand cytosines, A and B, proximal to each end of the ss region and not within CpG sites were significantly methylated, both in the presence of Dnmt1-dN alone and to a lesser degree in its admixture with Dnmt3b (Figure 6A and see Supplementary Figure S8A at http://www.BiochemJ.org/bj/432/bj4320323add.htm). Similarly, two non-CpG cytosines on the opposite ‘L’ strand, C and D, also showed methylation levels similar to that of CpG sites in the presence of Dnmt1-dN (Figure 6B and Supplementary Figure S8B). In the ss bubble region on the S and L strands, cytosines in a non-CpG context were mostly positioned immediately upstream of a CpG site or a thymine and, in one instance, an adenine, so ‘CpCpG’, ‘CpT’ and ‘CpA’ sites respectively. All observable non-CpG methylation was at CpT sites. No CpT methylation was observed with either of the de novo enzymes alone, or on the DNA duplex, the R-loop or RNA–DNA heteroduplex. No methylation of Site A, the site most strongly methylated in the bubble structure occurred on ssDNA, indicating that such methylation by Dnmt1-dN is not a general property of its action on ssDNA. However, some methylation was evident at CpT Site B on ssDNA. The possible formation of secondary structure or inter-molecule hybdridization of the ss ‘S’ (Figure 2), may provide a suitable substrate for such methylation.
Non-CpG cytosine methylation within the DNA bubble substrate
Dnmt activity on DNA structures
The three Dnmt enzymes showed clear preferences in methylation of unmethylated and hemimethylated DNAs. Dnmt1-dN showed an approx. 2.5-fold higher relative activity on hemimethylated compared with unmethylated DNA, whereas Dnmt3a and Dnmt3b had significantly reduced activity (approx. 25–50%) on hemimethylated DNA. Dnmt3a has been shown previously to exhibit 1.2–3-fold higher activity on unmethylated dsDNA than hemimethylated dsDNA [41,42]. As Dnmt3a and Dnmt3b are known to be activated by Dnmt3L and various chromatin-modifying proteins , the higher specific activity of Dnmt1-dN on unmethylated dsDNA, relative to the de novo enzymes, does not immediately suggest a de novo function of Dnmt1 in vivo. All three methyltransferases showed significantly less activity (20–60%) on ssDNA compared with dsDNA, with Dnmt3a showing the lowest activity and Dnmt3b the highest. The lower activity on ssDNA was also reflected in lower activity on the DNA bubble structure.
At the level of individual CpG sites, Dnmt1-dN and Dnmt3b both showed significant bias in methylation on an unmethylated dsDNA substrate, compared with Dnmt3a where all six sites were methylated to a similar level. In contrast, others have seen profound Dnmt3a flanking sequence methylation bias both in vitro [42,44] and in vivo [45,46]. The distinct site preference towards Sites 1, 3, 4 and 6 by Dnmt1-dN and Dnmt3b cannot easily be explained by phenomena such as CpG spacing [47,48], instead flanking sequence preference by the enzymes is more likely. Features of site preference for methylation of individual CpG sites are discussed in more detail in the Supplementary Online Data.
Bisulfite sequencing analysis revealed distinct differences in the methylation profile by the Dnmts on ssDNA, with a strong preference of all three enzymes for methylation of CpG Site 3, almost exclusively in the case of Dnmt3a and Dnmt3b. For Dnmt1-dN and its admixture with Dnmt3b, the relative levels of methylation of Sites 5 and 6 relative to Sites 1 and 4 in particular, were reversed. It is possible the ss site preference data is obfuscated due to intra- and inter-molecular secondary structure. In the context of the DNA bubble structure, all enzymes unexpectedly showed a preference for methylation of CpG sites within the ss region, again with methylation of CpG Site 3 predominating. As kinetic data showed reduced activity on the bubble structure in comparison with dsDNA, we had expected that any remaining activity would be toward the ds parts of the structure. One interpretation for this preference is that the activity on ss substrates is limited by initial DNA binding. Once bound to the duplex part of the bubble structure, CpGs in the ss region could be available for methylation without the necessary conformational flip required to make cytosines in dsDNA available to the active site of the enzyme . Regardless of mechanism, this difference in site preference for CpG site methylation in ssDNA compared with dsDNA may have significance in the context of DNA replication or repair.
A further unexpected finding from our sequence analysis was that CpT sites near the termini of the ss regions within the DNA bubble were methylated by Dnmt1-dN and to a lesser extent by the admixture of Dnmt1-dN and Dnmt3b. This non-CpG methylation was only observed to a significant degree on DNA bubble substrates when Dnmt1-dN was present. This observation of non-CpG methylation by Dnmt1-dN was unexpected and is the first account of such a high degree of non-CpG methylation by a mammalian Dnmt. This finding generates several questions. It remains to be seen whether the methylation of cytosines internal to a DNA bubble is reliant upon CpT sites, or whether Dnmt1 can methylate a cytosine in any context (a CpN site). Interestingly, murine Dnmt1 has previously been shown in vitro to methylate cytosine at CpT sites, but not at CpC or CpA sites . If the strong non-CpG cytosine methylation activity of Dnmt1-dN on the DNA bubble is indicative of Dnmt1 activity in vivo, it is easy to speculate that this property is associated with DNA replication and/or RNA transcription. If it is the latter, then Dnmt1 might well have a role in TGS. Recruitment of Dnmt1 to the unwound DNA trailing an active RNAPII complex may cause high levels of CpN methylation leading to long-term silencing. Indeed, high levels of CpN methylation have been documented when siRNA is expressed continuously from an integrated shRNA (short hairpin RNA) construct . Alternatively, this finding may be an in vitro artefact. The duplex/DNA bubble boundary may provide a structure analogous to a Dnmt1 reaction intermediate, where the enzyme is bound to dsDNA, but nearby cytosine bases are unpaired and can interact with the enzyme active site independently of their sequence context. Since the de novo Dnmt enzymes have a similar C-terminal active site to Dnmt1 and do not exhibit observable non-CpG cytosine methylation on DNA bubbles, this lends stronger support to the idea that non-CpG methylation is the product of an underlying Dnmt1 function. It is unlikely that non-CpG methylation can be attributed to the truncation of Dnmt1-dN, as the truncated N-terminus is involved in DNA binding, not in catalysis. Non-CpG methylation at significant levels has recently been reported in human embryonic stem cells  and the observations of the present study may have relevance to mechanisms that underlie its generation.
Dnmt activity on RNA heteroduplex and R-loop structures
All of the Dnmt enzyme preparations had significantly reduced activity on substrates containing annealed RNA as determined by enzyme initial velocities, bisulfite sequencing and ARORA methods, particularly on CpG sites opposite the RNA strand. Whereas the inhibition of enzyme activity on DNA by annealed RNA was profound with Dnmt1-dN and Dnmt3b, bisulfite sequencing data suggest that the inhibition was less so with Dnmt3a. Dnmt3a has been shown previously to be associated with RNA-mediated TGS  and to bind ss or ds siRNA [24,51]. It is possible that Dnmt3a binds siRNA and DNA substrate in a mutually exclusive fashion similar to methyl-binding domain proteins . If so, this may explain the reduction in initial velocities observed on S–A′ and S′ substrates in the presence of siRNA duplex. Activity on the L strand within the R-loop by the Dnmt1-dN/Dnmt3b admixture was enhanced by the annealing of RNA to the opposite S strand. Interestingly, when Dnmt1-dN and Dnmt3b were combined in an admixture, they exhibited a synergistic effect in methylating R-loops and the RNA–DNA heteroduplex. However, activity (particularly at CpG Site 3) within the heteroduplex was reduced considerably relative to the DNA bubble. We cannot exclude that the full-length enzyme may have activity on a RNA–DNA heteroduplex not observed with Dnmt1-dN.
One model for targeted transcriptional silencing by RNA involves formation of sequence-specific RNA–DNA heteroduplexes that trigger localized silencing by either or both induction of repressive chromatin modifications and DNA methylation. We set out to examine whether the Dnmts themselves were able to act on such RNA–DNA heteroduplex structures and so could be directly involved in TGS. Our data clearly demonstrate that the formation of an RNA–DNA heteroduplex is strongly inhibitory of de novo methylation in comparison with the equivalent ssDNA. The admixture of Dnmt1-dN and Dnmt3a was most able to methylate the RNA–DNA heteroduplex (S–R′) or R-loop structure (S–R–L), but, even in this instance, methylation of the hybridized DNA strand was less than in the equivalent ss (S or S′) or bubble (S–L) DNA structure.
A variation of the RNA–DNA TGS model speculates that RNA binding to a single strand of an unwound DNA duplex drives methylation of the DNA opposite the RNA, and such methylation is seen for the Dnmt1-dN/Dnmt3a admixture. From our combined data, we conclude that Dnmts are unlikely to be involved in TGS through direct methylation of DNA stands in RNA–DNA heteroduplex structures. However, we cannot discount all variations of the RNA–DNA model as our research leaves open the possibility that binding of an in vivo cofactor to a RNA–DNA heteroduplex structure may recruit chromatin-modifying and/or Dnmt enzymes to effect localized transcriptional repression. Similarly, it does not exclude the possibility that a Dnmt enzyme may bind to a heteroduplex in a non-methylating context and in doing so recruit chromatin-remodelling factors. Notwithstanding, Dnmts are not the primary effectors of TGS; perhaps instead they are recruited to an established TGS complex. DNA methylation as a downstream secondary characteristic of TGS fits observations that effective TGS can be observed without DNA methylation [13,16–18]. Although there is likely to be more than one pathway to TGS, previous publications provide stronger support for an RNA–RNA-mediated process .
We kindly thank Dr Keith Rand and Dr Lloyd Graham for critically reviewing the paper before submission.
annealed RNA and DNA oligonucleotide-based methylation-sensitive restriction enzyme analysis
murine DNA methyltransferase 1 with an N-terminal truncation
histone deacetylase 1
post-transcriptional gene silencing
RNA polymerase II
short interfering RNA
transcriptional gene silencing
The study was conceived by Peter Molloy. Jason Ross and Peter Molloy designed the study. Jason Ross performed the experiments and analysed the data. Isao Suetake assisted in performing the initial velocity experiments. Isao Suetake and Shoji Tajima helped in interpretation of the initial velocity experimental data. Jason Ross and Peter Molloy wrote the paper with assistance from Shoji Tajima.
This work was performed under the Commonwealth Scientific and Industrial Research Organization (CSIRO) Emerging Science Initiative and CSIRO Preventative Health Flagship and the International Collaborative Research Program of the Institute for Protein Research, Osaka University.