To profile methylation patterns, DNA samples are treated with a methylation-sensitive restriction endonuclease and are size-fractionated by sucrose gradient centrifugation 7. The low molecular weight fraction is collected and labeled with either of two fluorescent dyes, such that two samples can be compared by standard microarray analysis. If one sample is derived from a mutant in which methylation is reduced, then affected sites will be more frequently cleaved by the restriction endonuclease relative to wild-type. When cleavage results in an assayed fragment sedimenting faster than the 2.5 kb cutoff used in the fractionation, then there will be a stronger signal for mutant than wild-type. Conversely, if the mutant causes hypermethylation of a site, then the wild-type signal will be higher than the mutant signal. In this way, we can detect whether or not changes occur in methylation patterns from the ratio of the two dye signals, scoring as positive targets only those that are statistically significant based on repeated measurements from different biological samples 21. Positive targets can be scored as either hypomethylated or hypermethylated.

In our original methylation profiling study, we assayed PCR-amplified fragments from loci known or suspected to be targets of DNA methylation and also PCR-amplified single-copy approximately 700 base pair (bp) fragments chosen at random from the Arabidopsis genome sequence, 360 in all 7. In a subsequent study, we increased the size of this array to include 597 randomly chosen loci 8. In the present study, we have used this 'random-PCR' array, as well as a 'gene-oligo' array consisting of 26,090 oligonucleotides (70-mers) representing essentially all annotated Arabidopsis genes 22, of which 10% (2,633 of 26,090) are identified as containing transposons and repetitive elements detectable by RepBase 23 and RECON 24 analysis (unpublished observations). The gene-oligo array thus samples both distal and transposon-rich pericentric chromatin regions of the genome.

To detect CNG methylation, we used the MspI restriction endonuclease to digest DNA samples from five mutant lines: cmt3, kyp, ago4, drm1/2 (double mutant) and cmt3 drm1/2 (triple mutant), each paired with its wild-type parental line. Using the random-PCR array, we detected five loci as hypomethylated in the cmt3 mutant background; these represent single-copy targets of CMT3 that are methylated on the first C of one or more CCGG sites, a modification that blocks MspI digestion.

We then asked whether any of these methyl-CNG positive loci were also detected as targets of other proteins assayed in this way. KYP yielded four (hypomethylated) targets from among the 597 randomly chosen single-copy loci, of which three were the same as the CMT3 targets (Table 1). Using the gene-oligo array, we detected 536 CMT3 targets (498 hypomethylated and 38 hypermethylated in cmt3; Figure 1a,b), and 81 KYP targets (79 hypomethylated and 2 hypermethylated in kyp; Figure 1c), of which 79 were also CMT3 targets (Figure 2) 25. This nearly complete overlap shows that the interplay between CMT3 and KYP found for sensitive reporter loci 1617 is also true for at least a large fraction of CMT3 targets genome-wide. All CMT3 or KYP positives on the random-PCR array and >90% of the positives on the gene-oligo array were hypomethylated in the mutant, as would be expected if these enzymes work in tandem to maintain CNG methylation. The close agreement between two very different array platforms 26 provides a high degree of confidence in our conclusions.

AGO4 yielded one hypomethylated target locus that was also a CMT3 target in the gene-oligo array and one (hypomethylated) target in the random-PCR array (Figures 1d and 2). In addition, we detected two targets of DRM1/2 in the random-PCR array (Table 1) and ten DRM1/2 targets (nine hypomethylated and one hypermethylated in drm1/2; Figure 1f,g) in the gene-oligo array (Figure 2). The low number of AGO4 and DRM1/2 targets relative to the high number of CMT3 and KYP targets on the gene-oligo array suggests that AGO4 and DRM1/2 play only a minor role in maintaining CNG methylation throughout the genome. We also assayed a triple mutant combination of cmt3 drm1/2 and observed extensive overlap for both the random-PCR array (Table 1) and the gene-oligo array (313 hypomethylated and 33 hypermethylated in cmt3 drm1/2; Figures 1e and 2).

In our original study, we noticed that all four randomly chosen loci that were positive for CMT3 represent transposable elements 7. This correspondence was especially striking considering that the loci were chosen to be single-copy in the genome, so that these represent the rarest class of transposable elements. This conclusion is confirmed in the present study. For the purposes of our analysis, we considered only loci where methylation blockage of a single site could cause a fragment to sediment more rapidly than 2.5 kb, thus resulting in its exclusion from the DNA used as probe (see Materials and methods and 8). By this criterion, only a subset of restriction sites overlapped by a repeat or transposon would be scored as affected by the mutation. Nevertheless, we found a preponderance of transposable elements in this class for CMT3 (Table 2). Likewise, three of the four single-copy targets of KYP were scored as transposable elements. This preference for transposable elements in CMT3 targets was amply confirmed on the gene-oligo array, with 63% (104/164) of loci with sites falling within transposons, compared with only 13% (907/7,032) for all loci on the array. These 104 elements include long terminal repeat (LTR), long interspersed element (LINE) and short interspersed element (SINE) retrotransposons, DNA transposons and helitrons. KYP showed a similar preference to CMT3 with 68% (26/38) of loci falling within transposons. We conclude that transposable elements of all types are by far the predominant target of the CMT3-KYP system. We also found a preference of DRM1/2 for transposable elements (Table 2).

If the targets of CMT3 result from a general preference of these DNA methylation pathways for transposable elements, then we would predict that the distribution of targets would approximately correspond to that of transposable elements along the chromosome. We tested this by comparing the distribution of target distances for CMT3 from the centromeric gap to the locations of repetitive elements. We searched the Repbase library of consensus sequences 23 against the Arabidopsis genomic sequence to determine the distribution of repeats. Most transposable elements in Arabidopsis are located near the centromeric gap, gradually decreasing in abundance towards the telomere (Figure 3a). Similarly, CMT3 targets, whether repetitive or single-copy, are highly abundant close to the centromeric gap, decreasing as one moves distally along the chromosome arms.

We wondered whether the KYP and DRM1/2 targets also showed a transposon-like distribution. This would be expected if KYP and DRM1/2 targets mostly comprise a representative sample of CMT3 targets. Indeed, the KYP target distribution along the chromosome nearly superimposes over that for CMT3 (Figure 3b), which together with the nearly complete inclusion of KYP targets within the set of CMT3 targets, indicates that KYP and CMT3 have the same target preferences. In contrast, DRM1/2 targets are scattered throughout the chromosome arms, with only one of ten targets in the most proximal 2 Mb where about half of the CMT3 targets are found; this difference in the distribution of elements is statistically significant (p = 0.013, Fisher's exact test). We conclude that CMT3 and KYP target transposable elements in general, whereas DRM1/2 is required primarily at elements that are distally located along the chromosome.

Our conclusion that DRM1/2 targets are distinct from CMT3 targets is unlikely to have resulted from false positives in the DRM1/2 dataset. As previously mentioned, DRM1/2 targets are enriched in transposons. In addition, the close correspondence between CMT3 and KYP distributions, even considering just the 85% of CMT3 targets that are not KYP targets (Figure 3c), implies that the cutoff criteria used for target detection were very conservative. As indicated below, it appears that the large majority of KYP targets were simply too weak relative to CMT3 targets to be detected in the context of a whole-genome analysis. Furthermore, the CMT3 DRM1/2 dataset provides an independent test of the stringency of our cutoff criteria, because we would expect it to include all of the CMT3 targets; but it actually is a smaller set that only partially overlaps. This partial overlap is evidently not attributable to false positives in both datasets, because the distributions of CMT3 and CMT3 DRM1/2 targets essentially superimpose (Figure 3c), even considering just the 21% of CMT3 targets that are not CMT3 DRM1/2 targets. This indicates that the small number of DRM1/2 targets results from strict cutoff criteria that identify a subset of truly affected loci.

To confirm and quantify the array results, we performed bisulfite sequencing on a selection of target sites. We chose one positive example from the random-PCR array and five from the gene-oligo array. For locus 4:1813417-1814107 (Mu-PCR), detected as a target of CMT3, KYP and CMT3 DRM1/2 on the random-PCR array, wild-type methylation levels averaged 88% for the 11 CG sites and 47% for the 10 CNG sites assayed by bisulfite sequencing (Figure 4a; Table 3). In the cmt3 mutant background, the average level dropped to 63% for CG and to 1% for CNG methylation. This drastic decrease in CNG methylation is as expected considering that CMT3 is known to be responsible for nearly all of this modification at selected loci 2728. In a kyp mutant background, the average level dropped to 74% for CG and to 16% for CNG methylation. In this fragment, methylation of a single MspI site would account for a change in fragment size and its differential fractionation prior to microarray analysis. Remarkably, the kyp-induced decrease in methylation at the MspI site itself was only about one third (from 11 to 7 methyl-Cs of the 19 determined for this site; Table 3), confirming that methylation profiling on the random-PCR array is capable of detecting an intermediate drop in methylation levels.

Bisulfite sequencing of targets detected on the gene-oligo arrays confirmed that the positives detected on these arrays indeed reflect changes in the degree of methylation. For example, locus A000229 was detected as hypomethylated in both cmt3 and drm1/2 mutants, and bisulfite sequencing shows a reduction of methylation at one flanking MspI site in cmt3 and at the other flanking MspI site in drm1/2 (Figure 4b,c; Table 3). Interestingly, a reduction in methylation at the first MspI site was also seen in kyp. This partial loss of CNG methylation in kyp that was not detected on the gene-oligo array could account for the low fraction of CMT3 targets that are also targets of KYP on this array (Figure 2b,c). Of the three other loci examined in cmt3 and kyp mutants, all showed a major loss of CNG methylation in cmt3, one showed a major loss of CNG methylation in kyp, one showed a minor loss of CNG methylation, and one showed no loss (Figure 4c–e). This consistently strong effect of cmt3 and variable effect of kyp on CNG methylation at unselected sites is in agreement with studies of cmt3 and kyp/suvh4 mutants at particular loci 1617192728.

An unexpected finding was that CG methylation levels dropped five- to tenfold at two loci when both classes of de novo/CNG methyltransferases were absent (cmt3 drm1/2 in Figure 4c,e). This effect might be caused by an occasional failure of the MET1 CG maintenance methyltransferase, leading to a dependence on methyltransferases that do not require a hemi-methylated substrate 2930.

We wondered whether there is an inherent difference between transposons that require DRM1/2 for methylation and those that do not. Bisulfite sequencing revealed major losses of CNG methylation in drm1/2 mutants at three loci: A000229 and two loci corresponding to SINE3 elements (Figure 4c,e,f). The approximately 160 bp size of these SINE3 targets of DRM1/2 contrasts with the >5 kb size of the three loci that were not affected by drm1/2. Only one of these SINE3 elements is present in the parental strain of ago4 (Ler), and this showed a major drop in CNG methylation, whereas all three large elements that were unaffected by drm1/2 were also unaffected by ago4. Taken together, our results are consistent with the possibility that DRM1/2 and AGO4 are required to maintain DNA methylation at small, but not large transposable elements. The small size and low abundance of DRM1/2 and AGO4 targets might explain why so few of them were detected relative to CMT3 and KYP targets.

To determine the generality of our observations on DRM1/2 and AGO4 targets, we included bisulfite sequencing data from previous studies 1112 with our own in comparisons of levels of methylation reduction in mutants. drm1/2 and ago4 mutants show strongly correlated CNG methylation reductions (r = 0.91, p = 0.0002). In addition, methylation reduction shows a clear association with element size for both ago4 and drm1/2 (Figure 5a), with small elements preferentially affected. No significant associations with element size were seen for either cmt3 or kyp (Figure 5b). Methylation reduction is not attributable to differences in element type, because heterogeneous classes of sequences were found in both size classes, with DNA transposons and LINE elements in the large size class and SINEs, tandem and inverted repeats and genes in the small size class. Methylation reduction is also not attributable to differences in element abundance, because no association is seen between changes in methylation and the estimated copy number of elements in the genome (data not shown). It thus appears that AGO4 and DRM1/2 work together to maintain DNA methylation and silencing of small elements.

Arabidopsis thaliana mutants were previously described 1012. To control for background variability, lines were constructed by backcrossing parental mutant lines with either Ler (for cmt3, kyp, ago4 and drm1 drm2 cmt3) or Ws (for drm1 drm2), which served as the corresponding wild-type controls. Whole 5 week old plants were used to prepare genomic DNA using the CTAB extraction method 40. After ethanol precipitation, DNA samples were treated with DNase-free ribonuclease (Roche, Indianapolis, IN, USA) and precipitated by addition of 3 M sodium acetate and ethanol, then pelleted by centrifugation and air-dried. Bisulfite treatment of DNA, cloning into a Topo TA vector (Invitrogen, San Diego, CA, USA) and DNA sequencing were performed as described 27. Primer sets are listed in Table 4.

Primer selection, amplification and spotting in duplicate onto glass slides have been described in our original methylation profiling study 7. The random-PCR array in the present study consisted of 960 loci of which 597 were randomly chosen approximately 700 bp single-copy loci and 363 were selected control loci of different lengths 8. The 597 loci were selected as random non-overlapping 1 kb single-copy fragments from a non-redundant database consisting of contigs representing A. thaliana chromosomes 2 and 4 taken from a December 1999 version of the A. thaliana TIGR assembly 41. Chromosomes 1, 3 and 5 were pieced together from the A. thaliana genome project clone table from an August 2000 version of the A. thaliana TIGR assembly. Most of the selected loci were amplified from segments of known targets of the gene products under study 1227. Primers were designed and checked by BLAST searching to avoid redundancy as described 7. The gene-oligo array consisted of 26,090 70-mer oligonucleotides from the Arabidopsis genome oligo set Version 1.0 41, arrayed and hybridized as previously described 7.

Samples were dissolved in Tris-EDTA (TE) buffer and 50 to 60 μg aliquots were subjected to digestion by addition of 200 units of restriction endonuclease for 3 to 4 h. MspI endonuclease was obtained from New England Biolabs (New England Biolabs, Ipswich, MA, USA). Digested DNAs were size-fractionated on 5% to 30% sucrose gradients as described 42. Aliquots of DNA fractions were examined by agarose gel electrophoresis to verify DNA fragment size and concentration. Fractions in the <2.5 kb range were pooled, precipitated by addition of ethanol, and fluorescently labeled with either Cy3 or Cy5-dCTP (Amersham, Piscataway, NJ, USA) by random priming (Invitrogen) as described 42. Oppositely labeled samples from mutant and wild-type were mixed together and hybridized to microarrays on glass slides and processed as described 42.

Slides were scanned using a GenePix 4000 fluorescent scanner (Axon Instruments Inc, Union City, CA, USA). For each mutant comparison, three to four biological replicates were performed, all with dye-swaps. A lowess normalization was applied to the gene-oligo array to correct for non-linearity in this dataset. Methylation profiles were analyzed and p values assigned using Cyber-T microarray analysis software, which applies a Bayesian T-statistic method 21. The data-versus-model weighting factor was adjusted to 8 for the random-PCR array and to 6 for the gene-oligo array. A window size of 161 was used for the random-PCR array and 201 for the gene-oligo array. Bayesian-derived p values were adjusted for multiple hypotheses testing using a Bonferroni correction (p = 0.05) for the random-PCR array and a false discovery rate of p = 0.05 for the gene-oligo array. Note that the use of statistical criteria to delineate targets results in greatly reduced sensitivity of the gene-oligo array relative to the much smaller random-PCR array. An additional criterion for significance was implemented using 'self versus self' control experiments to assess experimental variation within the system. Accordingly, a lower-bound threshold for the log₂ methylation ratios (cy3/cy5) was defined as 3 standard deviations for the random-PCR array (4 for the gene-oligo array) from the corrected mean of the distributions of log₂-transformed ratios.

To facilitate comparison of datasets, we implemented a relational database (mysql) with a web browser display (Methprof 25). Methprof has utilities for processing raw data and for statistical analysis by CyberT 21. Methprof displays positive hits based on CyberT analysis for individual and combined datasets, together with a graphical chromosomal map of all the loci. Each hit in a Methprof table links to annotation data and displays user-provided descriptions, the number and identity of datasets in which it is positive, and whether the hit is hypo- or hyper-methylated.

In addition, a Javascript program (Region Viewer, developed by us) was implemented to display annotation and restriction site data for loci on the PCR-based array, and Methprof was adapted to display similar information for the oligo-based array. For each locus, a 'neighborhood' centered on a locus was defined such that blockage of a methyl-sensitive restriction site anywhere in the region could increase a fragment from less than to greater than the 2.5 kb cutoff. The blocked site was inferred as that most likely to have caused the depletion from the <2.5 kb fraction, ignoring ambiguous cases. Gene information was parsed from Genbank entries. Repeat information was generated using the program Censor4.1 43 on the A. thaliana repeat library athrep.ref 23. Repeat information was also obtained by BLASTN searching of an A. thaliana library of consensus sequences generated by the RECON program (Zhirong Bao, personal communication) 24. Data and maps used in this study are available for querying, browsing or downloading using Methprof 25, and all raw data can be downloaded from the GEO database under Accession number GSE3109 44.