Using the three hairpin assays described above, we obtained methylation-information for a total of 243 double-stranded DNA molecules, here termed “epialleles”, from males with FXS (Table 1). Of these molecules, 229 represented methylated epialleles, while only 14 represented unmethylated epialleles.

These double-stranded data are quantitatively consistent with our earlier single-stranded methylation data [3]. For the methylated epialleles, we observed extensive variation in methylation frequencies among CpG sites. Site-specific frequencies ranged from 0.5 (site 25) to 1.0 (site 10; Figure 4), similar to the variation described previously for some regions of FMR1 [3]. The mean frequency of methylation calculated here from both strands at CpG sites 1–22 in the males with FXS (0.91; Table 1) is also similar to our earlier estimates for the top-strand-only data for males with FXS (0.95) [3]. Our data for methylated, inactive-X alleles from normal females, reported here, (0.85; Table 1), are comparable to our previously published results (0.87) [3]. Thus, the double-stranded and single-stranded methylation data are in good agreement for these parameters.

To test for consistency of double-stranded methylation data collected using different hairpin protocols, we compared data at CpG sites 25 to 32, as ascertained using hairpins II and III, which overlapped for these sites. For this comparison, we focused on samples for which assays with both hairpin II and III were successful: non-expressing males #2, #23, and female #cF7 (Table 1). The two hairpin assays gave concordant results for all three individuals regarding inter-cell mosaicism. Male samples #2 and #23 showed only methylated epialleles, and the female control sample #cF7 showed a mixture of methylated and unmethylated epialleles, as would be predicted for a female having one active- and one inactive-X chromosome in each cell.

A comparison of site-specific methylation densities for male #2 and female #cF7 at these eight CpG sites yielded average densities of methylation that did not differ significantly (p>0.064, 1 d.f., χ₂ = 3.4; p>0.4, 1 d.f., χ₂ = 0.5, respectively). For the third individual, male #23, there was a significant difference in average methylation densities as estimated using hairpins II and III (p<0.0001, 1 d.f., χ₂ = 15.8). This p-value is more than ten times smaller than the significance limit of 0.0015 that is recommended after the conservative Bonferroni correction for multiple comparisons, in this case for the three individuals. We have no explanation for this significant difference in estimates. In all other respects, hairpins II and III gave concordant results for this individual. Unfortunately we were not successful in using both hairpins on samples from any other individual of this cohort.

Methylation densities obtained with the three different hairpins and averaged over the 52 analyzed CpG dyads on methylated epialleles were virtually indistinguishable between FMR1 expressing and non-expressing males (0.830 vs. 0.824, respectively). To further explore possible variation among sites, we next performed an analysis of variance (ANOVA) for data sets obtained with the three different hairpins. For hairpin I, there was no significant effect of FMR1 expression status (p>0.37), or individual (p>0.28). For hairpin III, the ANOVA indicated no significant effect of expression status (p>0.1), but a very significant effect of individual within the group of non-expressing males (p<0.00001). Thus, we conclude that this variation in methylation densities is not related to FMR1 expression but is instead due to variation among individuals with an inactive FMR1 allele. We have also seen variation in methylation densities in the control dataset from normal females, reflecting density of methylation on the inactive-X chromosome (Table 1).

The analysis for hairpin II was potentially interesting in that it revealed significant effects of expression group (p<0.01), and individual (p<0.03). We sought, but did not find, a subset of methylated sequences that had methylation densities low enough to permit transcription. The subset of methylated epialleles with the lowest densities of methylation in this region, identified as those with ≤13 of 20 CpGs (considering both strands), averaged 0.55 and 0.60 from expressing (34% of sequences) and non-expressing males (18% of sequences), respectively. There was no significant difference in methylation density between the two groups (p>0.5, 1 d.f., χ₂ = 0.39). Thus, there was no evidence for the existence of a subgroup of alleles with overall low-density methylation that might distinguish expressing from non-expressing males.

Of the ten CpG sites assayed by hairpin II (CpG sites 23–32), six are located within binding sites for transcription factors Sp1, Ap2 and USF1 (Figure 3), underscoring their potential relevance to FMR1 expression. None of these six CpG sites, as ascertained by hairpin II, had methylation densities that differed significantly between expressing and non-expressing groups.

CpG dyad 27, which does not overlap with a transcription-factor binding motif, was the only site assayed with hairpin II that showed a significant difference in methylation frequency between expressing and non-expressing males (0.52 and 0.76, respectively; p<0.02 after Bonferroni correction for multiple comparisons). However, the subset of low-methylation-density alleles, which would be most likely to express FMR1, showed no significant difference between expressing and non-expressing males in methylation frequency for CpG site 27 (0.29 versus 0.50, p<0.2); this p-value is well above the 0.005 required for significance with a Bonferroni correction for multiple comparisons. In addition, analysis of data obtained with hairpin III, which also assays methylation at CpG site 27, provides no evidence for a pivotal role of this site in FMR1 expression (p≤0.48).

In addition, the lack of methylation at site 27 is unlikely to be sufficient for transcription because an average of 20% of alleles from non-expressing males sampled in this study have this site unmethylated. Were this lack of methylation sufficient for transcription, males with FXS who exhibit this type of site-specific mosaicism would have FMR1 mRNA at levels greater than 1.4, assuming a seven-fold average level of elevated transcription for unmethylated, expanded alleles that approach the size of full mutation alleles [23]. The limit of detection of FMR1 mRNA – less than 0.01 of normal cells [16] – is substantially below such levels, establishing that an unmethylated CpG at site 27 of an individual allele is not sufficient for transcription.

Hairpin-bisulfite PCR enabled us to ask whether or not significant differences in DNA methylation densities exist between the top and bottom strands of individual DNA molecules (Figure 1C). For example, a marked difference in cytosine methylation could arise between the two strands of a DNA molecule if the maintenance DNA methyltransferase, DNMT1, were inhibited during, or immediately after DNA replication [24]. In such cases, an unmethylated bottom (template) strand could be transcribed and contribute to an unexpectedly high level of FMR1 mRNA. Such strand-specific methylation at transcribed loci has been described for a globin gene in chicken embryos [25], and has recently been reported to account for the transcriptional cycling of the pS2/TFF1 gene [26]. Hairpin-bisulfite PCR can readily detect differences in methylation densities between top and bottom strands of individual molecules [19], and thus enables us to examine the relationship between top- and bottom-strand methylation densities [27].

The coding and non-coding strands of double-stranded FMR1 molecules were highly correlated in their methylation densities for epialleles from both expressing (Figure 5a) and non-expressing (Figure 5b) males. Of particular relevance is our finding that none of the 111 methylated molecules from FMR1-expressing individuals had dense methylation on the non-coding, upper strand and low-density methylation on the coding, lower DNA strand (Figure 5a). Thus, we can exclude, with 95% confidence, the existence in FMR1-expressing males of a sub-population larger than 3.2% of cells that has markedly discordant methylation patterns between the two strands. This hypothetical 3.2% of cells would need to transcribe FMR1 to levels more than 30-fold greater than for normal FMR1 alleles to account for a relative mRNA level that averages 1.0 in these four expressing males. We conclude that methylation mosaicism between top and bottom strands of individual molecules is unlikely to account for FMR1 expression in these males with FXS.

Because each male cell carries only one X chromosome, the presence of both methylated and unmethylated FMR1 alleles in DNA isolated from an individual male is evidence of inter-cell mosaicism. For the detection of this type of mosaicism, hairpin I was the most reliable of the three assays. This hairpin assay yielded nearly unbiased amplification of methylated and unmethylated alleles from normal females (Table 1). For the five non-expressing males, all three hairpin assays yielded only methylated epialleles (total = 118). The recovery of 118 methylated epialleles and no unmethylated epialleles excludes with 95% confidence the existence of a subpopulation of unmethylated epialleles that constitutes more than 2.5% of cells in the non-expressing males.

Of the four DNA samples of FMR1-expressing males, two showed evidence of inter-cell methylation mosaicism (males #20 and #3; Table 1). DNA from male #3, previously found to have a level of FMR1 mRNA of 1.4 relative to the mean for normal males [16], had 31% (5/16) unmethylated FMR1 alleles. DNA from male #20, previously found to have a level of FMR1 mRNA of 1.0 [16], had 53% (9/17) unmethylated FMR1 epialleles (Table 1). Because we validated our hairpin-bisulfite PCR products with batch-stamps and barcodes [20], we can be confident that each FMR1 epiallele we analyzed was amplified from a different genomic template molecule. Likewise, none of our analyzed epialleles, methylated or unmethylated, arose through contamination or template redundancy. For these two males, the proportions of unmethylated epialleles reported here are sufficient to explain the observed levels of FMR1 mRNA. Cells carrying unmethylated FMR1 epialleles with expanded CGG repeats are known to transcribe FMR1 at an increased rate (up to 10-fold) relative to alleles with normal-sized repeats, thereby leading to unusually high expression levels detected by quantitative RT-PCR [23], [28], [29], [30]. It is thus clear that males #3 and #20 are inter-cell methylation mosaics whose levels of FMR1 mRNA are consistent with their degrees of mosaicism. Such a high level of mosaicism of unmethylated epialleles could have been missed by earlier Southern hybridization if the allele sizes were very heterogeneous.

For two of the expressing males, #78 and #14, no unmethylated epialleles were observed (Table 1). One possible explanation is that for these males, transcription occurred from a densely methylated FMR1 promoter, as was suggested previously [16]. An alternate possibility is that unmethylated epialleles were present in these individuals, but were not detected. We used a statistical test to assess this possibility. The DNA sample from expressing male #78 gave only seven informative epialleles with hairpin II, all of which were methylated. As described in Materials and Methods, amplification using hairpin II is heavily biased in favor of methylated alleles, and is thus ineffective at revealing low levels of unmethylated alleles. With so few informative epialleles, and such biased amplification of methylated epialleles with hairpin II, we cannot exclude the possibility that the true frequency of unmethylated epialleles in DNA sample #78 exceeds 0.1, which, with overexpression, could account for the observed FMR1 mRNA level of 0.7 reported for this individual. Nonetheless, we cannot exclude the original proposal [16] that the observed level of FMR1 RNA in this individual arises from methylated epialleles.

For expressing male #14, the larger number of epialleles available for analysis (85; Table 1), allows a much smaller probability that there is significant but undetected inter-cell mosaicism. In order for the observed mRNA level (0.8) to result from transcription of unmethylated alleles, at least 8–10% of this individual's epialleles would have had to have been unmethylated. The probability that this level of mosaicism existed but was not detected in our study is less than 0.01. Thus, for this individual, inter-cell mosaicism is unlikely to account for the observed levels of FMR1 RNA.

Written consent was obtained from all participating patients and from female subjects, except for anonymous samples denoted ‘female cF#_’ in Table 1. The consent form was approved by the Institutional Review Board of the University of California at Davis. For “cF#” samples the Human Subjects Division at the University of Washington determined that use of these anonymous samples does not fall under the federal definition of “human subjects research.” Use of these samples is therefore not subject to 45 CFR 46 and does not require review by the Institutional Review Board.

The nine FXS DNA samples used in this study are a subset of the samples previously described by Tassone and colleagues [16]. We chose to analyze DNA samples from four of the males with FXS who showed significant levels of FMR1 mRNA expression. The other five FXS DNA samples used here were from males not expressing FMR1 mRNA (Table 1). DNAs used in control experiments were isolated from peripheral blood leukocytes of twelve females determined to have normal-sized FMR1 CGG repeats. Four of these DNA samples were provided by F.T. (the M.I.N.D. Institute), and the remaining eight, designated as ‘cF’, were from the C.D.L lab collection. It is not known whether or not skewed X inactivation occurred in any of the female DNA samples. None of the females examined had SNPs that would allow us directly to address the question of activation ratio. Indeed, it is precisely because the nucleotide sequence of the region examined was identical among the subjects examined that data from females can serve as a metric of the bias inherent to a given hairpin assay. Genetic identity at this locus ensures that any deviations from the 50∶50 methylated∶unmethylated ratio expected in data from females can be attributed to the assay itself, rather than to nucleotide-level differences between the two homologs. The recovery bias observed for females, for which the true methylated∶unmethylated ratio is known, can therefore be applied in the interpretation of data from males, for which the true methylated∶unmethylated ratio is unknown.

For each sample, double-stranded genomic DNA was digested with a restriction enzyme that leaves a single-stranded overhang near the region to be analyzed. A short hairpin oligonucleotide linker was ligated to this overhang at the FMR1 locus (Figure 2) [19]. This hairpin linker contains a defined “batchstamp” sequence that specifies the date of the reaction and the sample number; a barcode of eight random nucleotides, contained in the “loop” of the hairpin molecule, identifies the original genomic DNA molecule that served as the PCR template [20]. This information is indispensable in assembling validated sets of sequence data that are not corrupted by PCR template contamination or redundancy [20]. Because of the length of the region that we wished to study, three different hairpin linkers and primer sets were necessary to cover the 52 CpG dyads of the FMR1 promoter (Figure 3).

Protocols were designed to ensure that the CGG repeats of FMR1 were removed prior to hairpin-linker ligation. The expanded and sometimes hypermethylated repeats in DNAs from patients could affect the denaturability of DNAs during PCR amplification of the FMR1 promoter region, exacerbating any inherent amplification bias. To eliminate this potential source of bias, our protocol removes the CGG repeat with appropriate restriction enzymes as specified below. Methylated and unmethylated FMR1 epialleles from patients and control females are thus rendered equivalent at the nucleotide-sequence level for the region analyzed. Hence, it is reasonable to assume, as we do here, that estimates of amplification bias from normal female FMR1 epialleles can be used to correct for bias in the recovery of expanded FMR1 epialleles (Text SI).

Following linker ligation, genomic DNA was treated with bisulfite under conditions that convert unmethylated cytosines to uracil with very high efficiency, while leaving methylated cytosines unchanged. After bisulfite conversion, the two strands of linked genomic DNAs are no longer strongly complementary. This reduction in strand complementarity allows conventional PCR amplification of the two now-linked strands of individual DNA molecules, yielding both top- and bottom-strand methylation information in a single sequence read [19].

PCR products were subcloned using the TOPO-TA cloning kit (Invitrogen), and sequenced on an ABI 3100 in the Comparative Genomics Center of the Department of Biology, the University of Washington. Resulting sequences were obtained and scored for CpG-cytosine methylation as described previously [20]. Only molecules with no more than one failure of conversion for non-CpG cytosines were used for this analysis, corresponding to a conversion efficiency exceeding 99%.

The FMR1 promoter is especially difficult to denature in DNA of individuals with FXS because of augmented levels of DNA methylation [3]. We used ten denaturation steps during the bisulfite conversion reaction to achieve good conversion of this CG-rich promoter. Our protocol for hairpin I has been described previously [19]; in brief, ∼5 ug of DNA was digested with restriction endonucleases DraIII and AluI prior to ligation of the hairpin linkers (5′PAGCGTAGCDDDDDDDDGCTACGCTTGA, where D represents A, G or T, thus generating random barcodes that are not altered by bisulfite conversion). Specified differences in the nonrandom sequences of the hairpin stems, the batch-stamps, create distinct versions of hairpin I for each individual examined. Primers used for hairpin I PCR amplification were: 5′- CCTCTCTCTTCAAATAACCTAAAAAC-3′ (primer 1), and 5′- GTTGYGGGTGTAAATATTGAAATTA-3′ (primer 2), where Y represents C or T. Restriction endonuclease PspGI was used for hairpin II, to cleave the DNA prior to ligation of the hairpin linkers (5′P- CCTGGTGCACGTTDDDDDDDDAACGTGCA-3′, with variations in the hairpin stem as described above. Primers used for PCR amplification with hairpin II were: 5′- CCTAACTAAAACCRAACCCCC-3′, where R represents G or A (primer 1), and 5′- GGAGTTGAGYGTTTGATTAGG-3′ (primer 2).

PCR for hairpins I and II was performed using HotStarTaq MasterMix (Qiagen) and 2% DMSO, with denaturation at 95°C for 15 min., followed by 41 cycles of denaturation at 95°C for 30 sec., annealing at 56°C for 30 sec, and extension at 72°C for 50 sec, followed by a final extension at 72°C for 5 min. For hairpin III, restriction endonuclease PspGI was used to cleave the DNA prior to ligation of the hairpin linkers (5′P- CCAGGAGCGATGCDDDDDDDDGCATCGCT-3′, with variations in the hairpin stem as described above). Primers used for PCR amplification with hairpin III were: 5′- AAAAACRAAACCRAAAAACTAAACCC-3′ (primer 1), and 5′- ATTTGAAGAGAGAGGGYGG-3′ (primer 2). PCR conditions were as described above with annealing at 54°C for 30 sec, and extension at 72°C for 50 sec, followed by a final extension at 72°C for 5 min.

Common sources of error in data from bisulfite-treated DNA include biased PCR amplification of individual genomic templates, and biased cloning of individual PCR products [3]. Both types of bias can lead to inaccurate estimates of the proportions of hypomethylated and hypermethylated epialleles. Our method for estimating the extent of bias in PCR amplification is included in Text SI.