Many of the advances in understanding gene silencing have been made in plants, largely using model reporter systems in which changes in silencing can be sensitively detected. For example, screens for genes that relieve the silencing of the Arabidopsis SUP and PAI loci that is induced by inverted repeats led to the discovery of three components of epigenetic processes: the CHROMOMETHYLASE3 (CMT3) DNA methyltransferase; an H3 K9 methyltransferase (KYP, also called SUVH4); and an Argonaute family member (AGO4) 6781314. These studies revealed that mobile elements are among the targets for DNA methylation by CMT3, an observation confirmed by microarray analysis of DNA methylation patterns in mutant cmt3^- plants 15. Recently, the correlations between DNA methylation, H3 K9 methylation and siRNAs were shown to extend over a large contiguous portion of the Arabidopsis genome 16. These features were primarily associated with mobile elements, suggesting that multiple silencing mechanisms are used for controlling genomic parasites. When a mutation in the ATP-dependent chromatin remodeling protein DDM1 was introduced, methylation of both DNA and H3 K9 was sharply decreased at mobile elements, with concomitant increases in transcription of these elements. Even the siRNAs were decreased in abundance in a ddm1^- mutant background. Interdependent mechanisms involving chromatin remodeling, DNA methylation, histone methylation and siRNAs therefore maintain mobile elements in a silent state in Arabidopsis.

What is responsible for this interdependence? One possibility is that DDM1 facilitates synthesis of siRNAs, which would trigger downstream silencing events. But the loss of siRNA-mediated silencing components has no obvious effects on the silencing of extensive genomic regions 1117. An alternative possibility is that multiple components of epigenetic silencing act at the same place at the same time. In this way, the elimination of the ATP-dependent remodeler that provides access to chromatin would have effects on multiple silencing components. For example, if DDM1 were to act at the replication fork to promote DNA methylation and chromatin assembly, then a concerted process of silencing would be disrupted in ddm1^- mutant plants. Indeed, there are intriguing connections between various epigenetic-silencing components and the machinery for DNA replication and replication-coupled chromatin assembly. Where silent regions are extensive, replication-coupled mechanisms might suffice, but where they are small, targeting by siRNAs would be required to reinforce silencing. Below, we explore these concepts in light of recent evidence.

Most DNA methylation in both plants and animals is on CG dinucleotides 1. A CG methylated on one strand but not the other (hemi-methylated), which results from replication of fully methylated DNA, serves as a substrate for a maintenance methyltransferase that restores the site to a fully methylated state. The enzyme responsible for maintenance of CG methylation, called DNA methyltransferase 1 (Dnmt1), was first cloned in mice and is associated with DNA replication foci during the S (synthesis) phase of the cell cycle 1. An orthologous Arabidopsis enzyme called MET1 is similarly required for maintenance of CG methylation 18.

The Dnmt1 subfamily of cytosine DNA methyltransferases should be capable of maintaining methylation passively, so that the only signal required for a methylation pattern to be propagated is the initial methylation itself. Some of the best evidence for this comes from experiments in plants after induction of RNA-dependent DNA methylation 19. CG methylation and transcriptional silencing can be maintained for generations after the RNA trigger has been eliminated from the plants, and MET1 function is required for silencing to be heritable in the absence of the RNA trigger 17202122. Also, mutations affecting H3 K9 methylation and RNA-dependent DNA methylation have little effect on CG methylation at most loci 6781123. CG methylation can thus be maintained passively, without the need for an active signal.

Non-CG methylation is generally found on CNG motifs or more rarely on asymmetrical motifs (CNN), where N is any nucleotide. Most CNG methylation in Arabidopsis is maintained by CHROMOMETHYLASE3 (CMT3) 1314, so named because of the presence of a chromodomain within the catalytic domain 24. CMT3 mutants lack virtually all CNG methylation in pericentric heterochromatin, and a number of transposable elements that reside there are reactivated 18. But at some silent regions that span only a few nucleosomes, loss of CMT3 function leads to only partial loss of methylation on CNG and asymmetric motifs 825. At these sequences, the remainder of non-CG methylation is catalyzed by the DRM family of methyltransferases. This effect is most pronounced at loci consisting of tandem direct repeats, such as FWA and MEA-ISR, where mutations in CMT3 have only a minor effect on non-CG methylation whereas DRM loss-of-function mutations eliminate this methylation completely 25.

The first clues about how CMT3 might be recruited came from the discovery of its partial dependence on the KYP/SUVH4 H3 K9 methyltransferase 67. Mutations in kyp/suvh4 mimic the cmt3 phenotype with respect to DNA methylation and transposon reactivation, although the effect is weaker than that of cmt3 mutants. A potential mechanism for how histone methylation may target CMT3 is suggested by the fact that CMT3 contains a chromodomain. CMT3 could therefore have the ability to directly interact with methylated histone H3.

Further insight into the mechanism of CMT3 regulation was provided by the recovery of an allele of AGO4 from the screen for suppressors of silencing of the SUP locus 8. The ago4 mutant plants exhibited substantial loss of CNG methylation, dramatic loss of asymmetric methylation and decreased H3 K9 methylation at several silent regions that span only a few nucleosomes. Experiments with PAI silencing also provide evidence that siRNA-mediated silencing plays a role in targeting CMT3 26. An upstream promoter is responsible for making a transcript that reads through the inverted repeat responsible for silencing PAI, thus creating a long double-stranded RNA. Silencing of this promoter leads to loss of both non-CG methylation and silencing at PAI. Crosses that remove the inverted repeat produce similar results 20. Additionally, highly transcribed inverted-repeat loci designed to trigger RNA-dependent DNA methylation induce high levels of non-CG methylation, which is largely lost when the inverted repeat is removed by crossing. These observations suggest that CNG and asymmetric methylation, unlike CG methylation, need to be actively maintained.

So far we have described two general modes of maintenance DNA methylation: passive and active. Passive maintenance is self-perpetuating, and clearly distinct from de novo methylation on a naive template. Active maintenance, on the other hand, may be nothing more than recurring rounds of de novo methylation. Alternatively, the requirements for active maintenance methylation and de novo methylation may be different, despite the fact that they both require an active signal. At least two lines of evidence indicate that the latter is indeed the case. First, the Arabidopsis de novo methyltransferases of the DRM family are absolutely required for the establishment of the RNA-directed DNA methylation that is triggered by a number of loci involving inverted repeats, but DRM proteins have only a partial role in the active maintenance of asymmetric and CNG methylation at these loci 252728. The rest of the non-CG methylation is maintained by CMT3, which is not required for establishment of methylation. CMT3 is therefore capable, at least in some cases, of responding to an RNA signal in order to actively maintain, but not to establish, DNA methylation.

A second line of evidence comes from the effects of the ddm1 mutation on DNA methylation. After erasure of methylation by passage through a ddm1 mutant background, restoration of DDM1 activity by crossing into a wild-type background does not restore methylation 16. Restoration of DDM1 function is therefore not sufficient to regain methylation and silencing, despite the observation that many of the transposons in question are associated with siRNAs. Thus, in the same cell, the same silencing signal is sufficient to maintain DNA methylation and silencing of transposons on one set of chromosomes but is not sufficient to efficiently initiate DNA methylation and silencing of the same sequences on a different set of chromosomes.

How do these various processes fit together in order to maintain silent chromatin? Like Dnmt1, MET1 is thought to maintain CG methylation following DNA replication 18. Old histones are evenly distributed between the two products of replication, so each chromatid has a memory not only of the original DNA-methylation state but also of the original histone-modification state. New nucleosomes are deposited after replication by the Chromatin Assembly Factor 1 (CAF1) chaperone complex. The H3 K9 methylation state of old nucleosomes would provide cues for CAF1 to deposit methylated nucleosomes 29 (see Figure 1a). The modified regions recruit CMT3 in order to maintain CNG methylation. In support of a role for replication-coupled nucleosome assembly in helping to maintain silent chromatin, mutations in CAF1 components have recently been shown to destabilize heterochromatic silencing in Arabidopsis 30. Mutation of the DDM1 chromatin remodeler could thus simultaneously disrupt the maintenance of both DNA and histone methylation, leading to a profound loss of silencing.

Such replication-coupled maintenance may be all that is necessary to maintain silencing of large regions of chromatin. Regions that are only a few nucleosomes in length might be difficult to maintain, however, because the 'unit' of chromatin memory is a nucleosome and the distribution of old nucleosomes to daughter chromatids is random 31. For example, if two adjacent nucleosomes are distributed to one daughter chromatid, then all histone-associated information is lost from the corresponding region of the other chromatid (Figure 1a). Therefore, to maintain stable silencing, DNA and histone modifications that are limited to small regions may need to be occasionally reinforced by active targeting of siRNAs to the homologous DNA (Figure 1b). In accordance with this idea, mutations in genes affecting siRNA-mediated silencing and de novo methylation have the strongest effects on short stretches of silent chromatin interspersed in otherwise active regions 8112532. One of these genes encodes a putative ATP-dependent chromatin remodeling protein, thus providing a DDM1 counterpart where silent regions are of limited extent 32.

Thus, in this model, siRNAs would have a dual role in silencing: they would provide triggers for establishing silent regions, but when these regions are too small to maintain themselves siRNAs would provide active reinforcement for maintenance of silencing. This latter process would be especially important for the silencing of newly integrated transposable elements, where chromatin-based silencing alone may be unstable. Stable silencing of such elements requires reinforcement by siRNAs; it amounts to 'kicking them when they're down'.