We generated HELP tagging libraries with HpaII- or MspI-digested DNA derived from human ES cells using the experimental approach shown in Figure 1. The assay differs from MSCC 15 by using EcoP15I instead of MmeI, generating longer flanking sequences (27 as opposed to 18 to 19 bp) and the addition of a T7 polymerase and reverse transcription step to allow the generation of the library without contaminating single-adapter products, while in addition obviating the need for gel extraction. After the library preparation, a single band of 125 bp in length was generated, as expected. Libraries were sequenced using an Illumina Genome Analyzer (36 bp single end reads) and the sequences were analyzed and aligned using Illumina pipeline software version 1.3 or 1.4. A summary of the Illumina analysis results for each replicate is shown in Table S1 in Additional file 1.

Based on our experimental design, successfully generated products would be expected to possess a 5'-CGCTGCTG sequence at the 3' end of the read, the first two nucleotides (CG) representing the cohesive end for ligation of HpaII/MspI digestion products, the remaining six nucleotides the EcoP15I restriction enzyme recognition site. In order to evaluate the yield of desired products, we counted the number of reads containing this sequence and plotted the starting positions of this sequence within the reads obtained. We observed that approximately two-thirds of the reads contained the expected sequence, and found that the majority was located at base positions 25 and 26, consistent with the known digestion properties of the restriction enzyme 19. Removal of the approximately 30% of reads lacking the CG-EcoP15I sequence was performed to eliminate spurious sequences. In order to investigate sequence quality further, we also determined the number and relative position of Ns (ambiguous base calls) within the reads obtained. Overall, few reads were found to contain Ns, and where they were present, they were found to be evenly distributed by position within the sequence. To test data reproducibility, we compared the results of three experimental replicates against each other using the Pearson correlation coefficient metric. The results of this study showed that all replicates were highly correlated (all the r values exceed 0.9), which confirmed that the technical reproducibility of this assay was excellent (Table S2 in Additional file 1).

We merged three lanes of MspI data and observed that approximately 80% of the 2,292,198 annotated HpaII sites in the human genome (hg18) were represented by at least one read, for a total of over 1.8 million loci throughout the genome. The mean numbers of reads per locus for MspI and HpaII were 3.94 and 1.82, respectively, and MspI counts were distributed evenly across all genomic compartments examined (Table S3 in Additional file 1). We hypothesize that a combination of incomplete genomic coverage and polymorphisms within some CCGG sites (as we have previously observed 10) accounts for the 20% of HpaII sites that were not represented by any reads.

When we plot the MspI count on the x-axis and HpaII count on the y-axis for each HpaII site, we can see two major groups of values in the plot (Figure 2a), separated into loci with high or with minimal HpaII counts. This plot helped us to develop a new method for normalizing HpaII counts in terms of variability of the MspI representation. We recognize that hypomethylated loci are associated with relatively greater HpaII counts and a larger angle B (Figure 2b, left) whereas methylated loci will be defined by smaller angle values (Figure 2b, middle). Furthermore, some loci will tend to be sequenced more readily than others, and may have identical B values but differing distances from the origin (c distance), allowing a confidence score for identical methylation values (B) in terms of the c distance values (Figure 2b, right). To test this model, we used bisulfite MassArray to test quantitatively the cytosine methylation values for 61 HpaII sites (Tables S4, S5 and S6 in Additional file 1), choosing loci representing all components of the B angle spectrum of values. In Figure 2c, d we show the correlations between these gold standard cytosine methylation values and raw HpaII counts or B angle values. We find that there is the same negative correlation (R² = 0.502) between HpaII counts and cytosine methylation values as demonstrated in the MSCC technique 15, and that the angular transformation of the data incorporating the MspI normalization substantially improves this correlation (R² = 0.826), defining the optimal approach for processing of these data. We represent the data for University of California Santa Cruz (UCSC) genome browser visualization as wiggle tracks, with higher B angle values defining less methylated loci. Methylated loci with zero values that would be otherwise difficult to visualize as having been tested are represented as small negative values. We show the details of the analytical workflow in Figure 3 and an example of a UCSC genome browser representation of HELP-tagging data in Figure S1 in Additional file 1. All data are available through the Gene Expression Omnibus database (accession number [GEO:GSE19937]) and as UCSC genome browser tracks 20.

As the number of reads at CCGG sites following MspI digestion should not be influenced by methylation, the representation obtained from MspI digestion allowed us to look for systematic sources of bias inherent to the assay. A major concern was that base composition could be a source of such bias, as it has been reported that Illumina sequencing can be influenced by GC composition 21, possibly because of the gel extraction step 22. Our protocol does not require gel extraction and only begins to show an under-representation of sequences when the (G+C) content exceeds approximately 80% (Figure 4a). We also tested to see whether the sizes of the MspI fragments generated influenced the counts obtained, as the digestion by type III endonucleases like EcoP15I is most efficient when a pair of enzymes is present in convergent orientation on the same DNA molecule 19. We find that there is indeed an over-representation for shorter (≤300 bp) and a corresponding modest under-representation for larger MspI fragments (Figure 4b).

Whereas MSCC used MmeI and generates an 18- to 19-bp sequence flanking the HpaII site 15, our use of EcoP15I generates a 27-bp flanking sequence. We asked whether this size difference influenced our ability to align sequences to the reference genome. We truncated our sequence reads to 19 bp to mimic the MSCC read length and found that this caused a profound loss of ability to align reads unambiguously (Table S7 in Additional file 1). To compensate for the low alignment rate, the MSCC report described an ingenious strategy of alignment to the sequences immediately flanking the annotated HpaII sites in the reference genome 15, an approach sufficiently powerful that it generated the well-validated data that they described. However, it does not offer the possibility of identifying polymorphic HpaII sites at the high frequencies that we previously observed for our HELP-seq assay 10. We tested whether our longer sequences allowed the identification of loci at which an HpaII site is annotated in the reference genome but we obtain no sequence reads, and the opposite situation where we observed at least four MspI reads (the average number per annotated MspI/HpaII site) flanking a locus not annotated in the reference genome. In Table S8 in Additional file 1 we list approximately 6,600 candidate polymorphic HpaII sites, of which examples are shown in Figure 5, confirmed by targeted resequencing of those loci. The 6,600 loci were selected based on overlap with dbSNP entries, allowing us to evaluate the pattern of sequence variability at these loci. Approximately 80% of the SNPs are C:G to T:A transversions, consistent with deamination-mediated decay of methylcytosine being the cause of the polymorphism 23. Polymorphic CG dinucleotides are major potential sources of error not only for microarrays, which are designed to a consensus genomic sequence, but also for both bisulfite sequencing, which would read the C to T transversion as unmethylated, and mass spectrometry-based assays, requiring the development of specific analytical approaches such as we have described 24.

To test whether the HELP-tagging assay was generating data that are biologically plausible, we tested the methylation of different genomic sequence compartments as density plots of B angle values for the human ES cells used in these studies. In Figure S2a in Additional file 1 we show how promoters (defined as -2 kb to 2 kb from the transcription start site of RefSeq genes), gene bodies (the remaining region within the RefSeq gene) and intergenic (all other) sequences compare, finding the expected enrichment of hypomethylated loci with larger B angle values in promoter regions. When we compared unique with repetitive sequences, again we found the expected pattern of increased methylation of repetitive DNA compared with unique sequences (Figure S2b in Additional file 1). Combining these observations, we tested whether the transposable element component of annotated repetitive DNA sequences showed any tendency to unusual methylation near gene promoters. In Figure 6 we show that while transposable elements are generally methylated and are depleted near gene promoters, those that are proximal to promoters tend to be less methylated than those located more distally. While many types of transposable elements were represented in this promoter-proximal hypomethylated group, we found a subset to be the most markedly over-represented, as shown in Figure 6c.

The outcome of these studies was an improvement in the previously described MSCC 15 and HELP-seq 10 assays, not only by means of technical modifications such as the use of EcoP15I but also because of the concurrent use of MspI for normalization. The effect of these modifications was not only to increase the accuracy of the assay but also to enhance the ability to align sequences to the genome and thus identify polymorphic HpaII/MspI sites. The means of normalization of HpaII by MspI using an angular metric is an innovation that improved the data accuracy substantially and may have applications in other MPS assay normalization strategies. We were also able to discard reads that did not contain the expected adapter sequences, and created a straightforward data analytical pipeline that will facilitate processing of these HELP-tagging data by others.

The potential sources of systematic artifacts due to base composition or digestion product size were evaluated. Apart from a modest decrease in representation in regions above approximately 80% (G+C) content, base composition did not cause biases in representations, possibly in part due to our avoidance of a gel purification step in library preparation 22. Fragment length does influence the outcome, most likely due to effects on EcoP15I digestion 19, although the effects should be similar for both HpaII and MspI and should, therefore, largely cancel each other out in the normalization step. It is possible that endogeneous EcoP15I sites could influence the representations, but to have an effect they would have to be located within the 27 bp adjacent to HpaII/MspI sites and would cause digestion of the ligated adapter, causing those loci to be under-represented in both HpaII and MspI datasets. The most likely effect of these endogeneous sites is that they contribute to the proportion of loci at which we could not obtain sequence reads.

Our exploration of the distribution of cytosine methylation in the same human ES cell line studied by Lister et al. 6 showed consistent results, with hypomethylation of transcription start sites and methylation of transposable elements, as expected from long-standing observations in the field. We furthermore discovered a limited subset of transposable elements that is hypomethylated when in close proximity to transcription start sites. When this subset was studied to determine whether certain types of transposable elements were disproportionately over-represented, we found two broad classes, one of transposable element fossils with no innate capacity to replicate themselves (the ancient DNA, long interspersed repetitive elements (LINEs) and short interspersed repetitive elements (SINEs) shown in Figure 6c) and younger ERV1 long terminal repeat retroelements. Loss of methylation of functionally inactive transposable elements is likely to be of no negative consequence to the host genome, consistent with the host defense hypothesis 25, while the young ERV long terminal repeats represent a group of transposons whose function has been harnessed as promoters of endogeneous genes 2627. This observation demonstrates the value of a high-resolution, genome-wide assay like HELP-tagging to define potential functional elements in an unbiased manner.

H1 human ES cells (NIH code WA01 from Wicell Research Institute, Madison, WI, USA) were cultured on matrigel (BD Biosciences, San Diego, CA, USA), at 37°C, 5% O₂ and 5% CO₂. Amplified human ES cell pluripotency was assessed by flow cytometry with SSEA4, CD24 and Oct4 markers. To extract DNA, the cells were suspended in 10 ml of a solution of 10 mM Tris-HCl (pH 8.0), 0.1 M EDTA and 1 ml of 10% SDS to which 10 μl of RNase A (20 mg/ml) was added. After incubation for 1 hour at 37°C, 50 μl of proteinase K (20 mg/ml) was added and the solution was gently mixed and incubated in a 50°C water bath overnight. To purify the lysate, it was extracted three times using saturated phenol, then twice with chloroform, and dialyzed for 16 hours at 4°C against three changes of 0.2× SSC. Following dialysis, the DNA was concentrated by coating the dialysis bags in polyethylene glycol (molecular weight 20,000). The purity and final concentration of the purified DNA was checked by spectrometry (Nanodrop, Wilmington, DE, USA).

The sample preparation steps are illustrated in Figure 1. Two custom adapters were created for HELP-tagging, referred to as AE and AS. As well as an Illumina adapter sequence, adapter AE contains an EcoP15I recognition site and a T7 promoter sequence. Adapter AS contains an Illumina sequencing primer sequence. The adapter and primer sequences for library preparation are listed in Table S9 in Additional file 1. Genomic DNA (5 μg) was digested with HpaII and MspI in separate 200 μl reactions and purified by phenol/chloroform extraction followed by ethanol precipitation. The digested genomic DNA was ligated to adapter AE using a New England Biolabs Quick Ligation Kit (25 μl of 2× Quick ligase buffer, 3 μg of HpaII-digested DNA or 1 μg of MspI-digested DNA, 0.1 μl of Adapter AE (1 μM), 3 μl of Quick Ligase in a final volume of 50 μl). After AE ligation, the products were purified using Agencourt AMpure beads (Beckman Coulter, Brea CA, USA), then digested with EcoP15I (New England Biolabs). The restriction fragments were end-repaired to inhibit to dimerization of adapters, and tailed with a single dA, at the 3' end. After the dA tailing reaction, adapter AS was ligated to the dA-tailed fragments using a New England Biolabs Quick Ligation Kit (25 μl of 2× Quick ligase buffer, 2.5 μl of adapter AS (10 μM), 2.5 μl of Quick Ligase in a final volume 50 μl). After ligation, products were purified using the MinElute PCR purification kit (Qiagen, Hilden, Germany) and in vitro-transcribed using the Ambion MEGAshortscriptkit (Life Technologies, Carlsbad, CA, USA). Following in vitro transcription, products were purified with an RNeasy clean-up kit (Qiagen) before reverse transcription was performed using the Invitrogen SuperScript III kit (Life Technologies). The first strand cDNA produced was used as a template for PCR using the following conditions: 96°C for 2 minutes, then 18 cycles of 96°C for 15 seconds and 72°C for 15 seconds followed by 5 minutes at 72°C for the final extension. After PCR, the library was purified using a QIAQuick PCR clean-up kit (Qiagen).

Bisulfite conversion and MassArray (Sequenom, San Diego, CA, USA) were performed using an aliquot of the same sample of DNA as was used for the high-throughput assays described above. Bisulfite conversion was performed with an EZ DNA Methylation kit (Zymo Research, Orange, CA, USA). Bisulfite primers were designed using MethPrimer 28, specifying the desired product length (250 to 450 bp), primer length (23 to 29 bp) and primer Tm (56 to 62°C). PCR was performed using FastStart High Fidelity Taq polymerase (Roche, Basel, Switzerland) with the following conditions: 95°C for 10 minutes, then 42 cycles of 95°C for 30 seconds, primer-specific Tm for 30 seconds and 72°C for 1 minute, followed by 72°C for 10 minutes for the final extension. Primer-specific Tm and sequence information are provided in Table S6 in Additional file 1. Bisulfite MassArray assays were performed by the institutional Genomics Core Facility. The data were analyzed using the analytical pipeline we have previously described 24.

Four lanes of sequencing were performed using an Illumina GA IIx Sequencer at the institutional Epigenomics Shared Facility. Three lanes were used for technical replicates of MspI, for the methylation-insensitive reference dataset. Images generated by the Illumina sequencer were analyzed by Illumina pipeline software (versions 1.3 to 1.4). Initial data processing was performed using the default read length of 36 bp, after which we isolated the sequences in which we found adapter sequences on the 3'-end, replaced the adapter sequence with a poly(N) sequence of the same length, and re-ran the Illumina ELAND pipeline again on these sequences with the sequence length set at 27 bp (the 2 to 28 bp subsequence). The data within the ELAND_extended.txt files were used for counting the number of aligned sequences adjacent to each CCGG (HpaII/MspI) site annotated in the hg18 freeze of the human genome at the UCSC genome browser. We permitted up to two mismatches in each sequence, and allowed a sequence to align to up to a maximum of 10 locations within the genome. For non-unique alignments, a sequence was assigned a partial count for each alignment location amounting to 1/n, where n represents the total number of aligned positions. To normalize the data between experiments, the number of sequences associated with each HpaII site was divided by the total number of sequences (including partial counts) aligning to all HpaII sites in the same sample. We refer to this figure as the fixed count below.

To examine an influence of (G+C) mononucleotide content on counts of sequences obtained, we extracted the (G+C) annotation from the hg18 freeze of the human genome at UCSC and examined the distribution of sequence counts according to (G+C) content. Annotated percentages of (G+C) content were available for adjacent 5-bp windows. For each annotated HpaII site, we calculated the mean percentage (G+C) for a 50-bp region centered at the restriction site. Counts of sequences associated with HpaII sites were obtained for 50 sequential non-overlapping windows of 2% (G+C) (the minimum possible in a sample of 50-bp regions). These data were then normalized as a proportion of the total number of fragments. Comparisons were made to the expected frequencies, which for each 2% (G+C) bin was represented by the counts of HpaII sites falling within a range relative to the total number of HpaII sites in the genome. This analysis was performed on both HpaII and MspI-digested DNA for comparison.

The potential effect of distance between HpaII sites on sequences counts obtained at each HpaII site was measured by summing the counts of sequences aligning within each restriction fragment, and normalizing the result with respect to total sequence count. As with the (G+C) analysis above, this was performed for both MspI and HpaII digested restriction fragments. The data were compared with the expected distribution determined by performing virtual restriction digestion using genomic HpaII site coordinates, and normalizing the number of virtual fragments of each size with respect to the total number of these virtual fragments.