Targeted bisulfite sequencing of the dynamic DNA methylome
© The Author(s) 2016
Received: 14 July 2016
Accepted: 10 November 2016
Published: 3 December 2016
The ability to measure DNA methylation precisely and efficiently continues to drive our understanding of this modification in development and disease. Whole genome bisulfite sequencing has the advantage of theoretically capturing all cytosines in the genome at single-nucleotide resolution, but it has a number of significant practical drawbacks that become amplified with increasing sample numbers. All other technologies capture only a fraction of the cytosines that show dynamic regulation across cell and tissue types.
Here, we present a novel hybrid selection design focusing on loci with dynamic methylation that captures a large number of differentially methylated gene-regulatory elements. We benchmarked this assay against matched whole genome data and profiled 25 human tissue samples to explore its ability to detect differentially methylated regions.
Our target capture design fills a major gap left by all other assays that exist to map DNA methylation. It maintains the ability to link cytosine methylation to genetic differences, the single-base resolution and the analysis of neighboring cytosines while notably reducing the cost per sample by focusing the sequencing effort on the most informative and relevant regions of the genome.
KeywordsDNA methylation Bisulfite sequencing Target enrichment
DNA methylation, most commonly at cytosines in the CpG dinucleotide, plays an important role in gene and genome regulation [1, 2]. Despite decades of elegant work, we continue to learn more about how and in what context DNA methylation functions through an ever-increasing collection of data. New insights are frequently enabled by technical advances in our ability to effectively map and quantify DNA methylation . While numerous technologies exist, the most widespread are all based on the principle of sodium bisulfite-induced selective deamination of unmethylated cytosine to uracil  with either microarrays or sequencing as read-out. Whole genome bisulfite sequencing (WGBS) has the advantage of theoretically capturing all cytosines in the genome at single-nucleotide resolution , but it has also a number of practical drawbacks that become amplified with increasing sample numbers. These include the sequencing cost to achieve sufficient coverage, data storage and computing time as well as the fact that most of the genome is depleted of CpGs and hence many reads lack any relevant information. Moreover, the majority of CpGs are static and do not change their methylation state across cell and tissue types, decreasing the information content of WGBS reads even further .
The triage process for differentially methylated regions (DMRs) and dynamic CpGs from the two main data sources [5, 12] is outlined in Additional file 1: Figure S1a-d. Our final list of 119,809 DyMe-Seq targets encompasses 91,039,504 bp of which approximately 90 Mb harboring 2.3 M million CpGs constitute legitimate capture space of the corresponding SeqCap Epi probes (Additional file 2). The chosen ~3% of the genome does not comprise all targetable dynamically methylated loci. Nor is 90 Mb the upper limit of the targeting technology. Rather, our curated DyMe-Seq target list represents a carefully balanced compromise between information content, number of capture probes as well as sequencing cost per sample. By design, our set of targets is predominantly enriched for differentially methylated putative regulatory regions and TF binding sites (TFBS) with >90% carrying H3K27ac or H3K4me3 annotations, while maintaining a representative coverage of classic genomic features such as promoters, CpG Islands (CGIs) and CGI shores (Fig. 1b,c). Importantly, our target set captures on average 44% of all putative enhancer regions that are enriched for H3K27ac chromatin marks across 87 distinct cell types or tissues (Additional file 1: Figure S1e). Our final list of prime targets still includes ~40% of our unranked initial list of candidate dynamic CpGs that were filtered solely by repeat content (≤60%), number of CpGs (≥2) and length (≥100 bp) of target regions, irrespective of overlap with genomic features (Additional file 1: Figure S1d).
The gold standard in the field remains WGBS, but as noted its cost and inefficiency are limiting its broad use despite decreasing sequencing costs. The advantage of sequencing only 90 Mb of highly informative targets instead of WGBS becomes clear when comparing the reads required for each sample to achieve 30× coverage across an increasing number of samples: 74 million versus 1.7 billion 100-base reads for 100 samples, covering each CpG in at least 80% of all samples (Additional file 1: Figure S1f). As would be expected, our design always captures a fraction of CpGs that are also covered by other common platforms including methylation bead arrays (Illumina), reduced representation bisulfite sequencing  and off-the-shelf targeted bisulfite-sequencing assays (Roche, Agilent). However, the overlap with any single one is less than 25% (Fig. 1d). Hence, DyMe-Seq fills an important gap between existing targeted assays and comprehensive WGBS (Additional file 1: Figure S1g, h).
Next, we assessed key performance properties of DyMe-Seq including genomic biases, coverage distribution, input requirements, off-target and PCR-duplicate rates (Additional file 3: Figure S2). This analysis revealed minimal differences in GC content distribution of the captured targets among technical replicates (Additional file 3: Figure S2b). Lowering the amount of input DNA increased the PCR-duplication rates from 3% for 1 microgram to 6% and 15% for 500 and 250 ng, respectively (Additional file 3: Figure S2c), when using a conventional library preparation strategy where DNA fragments are ligated to sequencing adapters prior to bisulfite conversion (“Methods” section). This result prompted us to continue with an input amount of 500 ng for most subsequent analysis. However, we also evaluated the possibility of lowering the input amounts further by switching to a “bisulfite-first” pre-capture library protocol (Swift Biosystems), yielding dramatically lower PCR-duplication rates essentially constant across an input range from 500 ng down to 160 ng (Additional file 3: Figure S2d). Finally, evaluation of our off-target rate revealed that on average 30% of all reads map outside of the targeted regions (Additional file 3: Figure S2e, f). Based on these analyses, we conclude that 30× coverage requires only 48 million reads for a single library (assuming a 10% duplication rate and 30% off-target reads). However, as noted above, comparing 100 samples where each CpG is covered 30× in at least 80 samples will require more reads (~74 M) per library.
To meet the continuously growing need for more efficient ways to capture DNA methylation information, we designed a hybrid selection target set for a carefully curated list of the most dynamically regulated CpGs (the set also captures 20.8 million CpAs, the predominant non-CpG methylation sites in selected cell types) and demonstrate its application across 20 distinct cell and tissue types. Our approach is substantially more economical than WGBS in terms of sequencing cost per sample, while providing higher coverage of many of the same, relevant regions. The savings in sequencing costs will often far outweigh the cost of performing the additional enrichment step which can be multiplexed by pooling barcoded libraries prior to hybrid selection . However, the choice of the appropriate DNAme profiling assay will ultimately depend on the specific scientific question. If it is critical to detect as many DMRs as possible, high-coverage WGBS may be the only feasible solution.
Our cost-effective DyMe-Seq assay will be a valuable tool not only for the methylation and developmental biology community, but specifically for many areas of clinical research, including prognostic and biomarker discovery as well as the emerging field of epigenome wide association studies (EWAS) .
Identification of target set
We considered two main sources of DMRs: those identified in Ziller et al. 2013 (minus those hypomethylated in sperm)  and those defined in Schultz et al.  using the same analytical strategy. We then filtered the Schultz et al. list for those already found in Ziller et al. 2013 and used these two sets as our starting DMR set. In addition, we included a subset of tissue specific DMRs (T-DMRs) that were not included in the aforementioned sets . We then proceeded with the analysis of these DMR sets separately for several steps, excluding the sex chromosomes from our analysis. In particular, we filtered these lists stringently in order to identify regions suitable for a hybrid capture approach. To that end, we only included DMRs with a repeat content below 60% based on the repeat masker annotation and at least two dynamic CpGs based on a minimum observed methylation difference of at least 30% between any two samples included in each DMR set. Subsequently, we computed a region-level score based on: 1. the density of dynamic CpGs, 2. the maximum observed methylation difference, 3. the overlap with a putative enhancer regions, defined as the union peak set of H3K27ac peaks across 70 distinct cell types from the ENCODE and Epigenome Roadmap Project, or an annotated RefSeq promoter (±1 kb of TSS). These three scores were each rescaled to the unit interval and then simply added for each DMR. We then ordered the DMRs within each set according to this score. In order to avoid that DMRs originating from one cell type dominate the set of selected regions, we next determined for each DMR the sample with its minimum methylation level. We then selected 53 Mb of DMRs from the Ziller et al. set according to the previously computed score, but maintaining a balanced representation of DMRs originating from all 23 cell types in the collection. This effectively skips regions with higher scores, if their most hypomethylated sample condition was already higher represented in the selected capture set than other samples. Since one purpose of this DMR set is the application to cohorts of many individual such as GTEx, we also selected all DMRs from the remainder of the ranked list that overlapped with eQTLs identified by the GTEx consortium  in the following tissues: blood, lung, subcutaneous adipose tissue, skeletal muscle and stomach. In a similar fashion, we selected 30 Mb from Schultz et al. DMR set and identified a 30-Mb set of DMRs. Finally, we also included a set of DMRs that were identified with the Illumina methylation 450 K array across multiple tissues, exhibiting a methylation difference of at least 30% and were not yet present in our final selected DMR set. Together, the union of these DMR lists constituted our initial candidate region set for the DyMe-Seq assay, comprising a total of 95 Mb. Next, this set was further optimized and filtered by Roche-NimbleGen in order to remove regions not suitable for capture, leaving us with 91,039,504 bp and 119,809 target regions (Additional file 2).
Raw reads were aligned to the human genome (hg19) using bsMap 2.7  with the following parameters bsmap -v 0.1 -s 16 -q 20 -w 100 -S 1 -u –R. Subsequently, we used picard tools (http://picard.sourceforge.net) version 1.139 to further process and QC the aligned data files. In particular, we used MarkDuplicates with standard parameter settings to mark and remove likely PCR duplicates, CollectAlignmentMetrics to compute basic alignment statistics, and CalculateHsMetrics with Additional file 2 to calculate all hybrid capture-related metrics, including the on target rate. In order to determine the methylation state of all CpGs captured and assess the bisulfite conversion rate, we used the mcall module in the MOABS  software suite with standard parameter settings. Finally, we converted the resulting CpG level files to bigBed files for visualization in the IGV , filtering out all CpGs that were covered with less than five reads.
Analysis of DyMe-Seq data was conducted in R using the methylKit  package and a 200-bp tiling of the target capture set. To that end, we imported the CpG level methylation call files from mcall into R using the methylKit function read and then computed the weighted methylation mean across for each 200-bp tile using the function getData, weighting the methylation level of each CpG with its coverage. We then merged the tile level methylation information across all samples and retained only those tiles covered with more than 10 reads in 70% or more of all samples. To compute differentially methylated tiles, we deliberately decided to choose a simple approach and performed Fisher’s exact test on all pairs of samples for each tile (see Additional file 5: R script). Subsequently, we corrected the resulting p-values using Benjamini–Hochberg  correction and defined regions with a q value ≤0.05 and an absolute methylation difference ≥0.3 as differentially methylated. Finally, we merged differentially methylated tiles between two samples into larger DMRs if they were less than 400 bp apart. The results of this analysis are displayed in Fig. 3c.
For all genomic features considered in this study, we defined them as overlapping with any of our regions if there was at least 1 shared base. Promoters were defined as RefSeq gene transcription start sites ±1 kb. CpG islands and CpG island shores were defined as previously  using the CpG island hunter. H3K27ac and H3K4me3 peak sets were defined as previously described . Briefly, IDR  peaks were identified for each cell or tissue type from the ENCODE  core cell lines and Roadmap Epigenome Project  using the IDR framework. The resulting peak sets were then merged, taking the union of all peaks and defined as the H3K27ac and H3K4me3 reference set. For the transcription factor union set, we used the transcription factor binding site cluster track provided by the ENCODE consortium .
Pathway and TERA analysis
Differentially methylated regions identified using the simple strategy outlined above between the GTEX samples artery coronary and nerve tibial were used an example for pathway and transcription factor binding analysis. For the pathway analysis, we selected all regions hypomethylated (according to the criteria listed above, see Additional file 6: Table S2) and used the web-tool GREAT to identify associated biological themes. The results were filtered according to the GREAT  standard criteria and a subset plotted in Fig. 3d. For the transcription factor binding site analysis, we used the ERA approach  to determine transcription factor motifs that are associated with differential methylation of 200-bp tiles of differentially methylated target sites across the entire GTEX. We then compute the differential ERA scores between the artery coronary and nerve tibial samples and plot the top 15 motifs associated with each condition in Fig. 3e. In the figure, we replaced the motif name with one representative transcription factors associated with each motif.
Comparison of WGBS and DyMe-Seq
In order to compare the consistency of WGBS and our approach, we selected two cell types for which high-quality WGBS data were available, HUES64 embryonic stem cells and CD8+ T cells, and performed DyMe-Seq with our capture set on the same samples. Next, we performed DMR discovery across the entire genome, now using a state-of-the-art beta-binomial model approach implemented in the DSS  package. More specifically, we first ran the dmlTest function on the comparison HUES64 versus CD8 for WGBS and DyMe-Seq separately, using one sample per group and smoothing turned on. Subsequently, we identified differentially methylated DMLs using the callDML function with a p value threshold of 0.001. Finally, we merged DMLs and identified DMRs using the callDMR function with the following parameters: delta = 0.3, p.threshold = 0.01, minCG = 2, dis.merge = 200 and otherwise default parameters. For the DyMe-Seq comparison, we then only considered those DMRs that were within 300 bp of one of our target regions. The results of this analysis are displayed in Fig. 2d–f.
Approximation of coverage requirements
To approximate the total per sample coverage required to capture each CpG with 30X, we followed the Lander–Waterman theory . In particular, we computed the required genome coverage to capture each CpG at 30× in 80% of each sample with more than 95% probability using Poisson statistics. Since the capture of any CpG in one particular sample is independent of the capture in a different experiment, we simply computed the probability to capture a CpG with 30 reads in N samples as the Nth product of the Poisson cumulative distribution function as a function of the lambda parameter. Subsequently, we determined the lowest lambda parameter for which the CpG was captured with 95% probability in 80% of the N samples and computed this number for N = 1–100. The resulting lambda then represents an approximation of the required read coverage after performing all filtering steps. This approximation does not take into account assay-specific biases or enrichment steps and likely represents more an upper bound on the required coverage. We used this modeling approach to estimate the required coverage for both WGBS and DyMe-Seq. For WGBS, we assumed an effective genome size of 2.7 × 109 and for our assay 90 Mb. To incorporate the effects of PCR duplicates, we assumed a duplication rate of 20% for WGBS and 10% for our assay, based on our empirical observations across many samples. For DyMe-Seq, we additionally assumed an off-target rate of 30% based on our empirical observations. The approximated raw coverage per base was then multiplied by the respective target genome sizes, divided by the read length (100 bp) and adjusted for duplicate and off-target effects. The results are then plotted in Additional file 1: Figure S1f.
Library preparation and sequencing
All human DNA samples used for this study are listed in Additional file 7: Table S3. Hybrid-selected sequencing libraries were prepared using a custom SeqCap probe pool (Roche) essentially following the manufacturer’s SeqCap Epi protocol except that we lowered the amount of input DNA from 1 to 0.5, 0.25 or 0.16 µg per sample and pooled 2–4 PCR-amplified indexed libraries prior to hybrid selection.
Unless otherwise noted in the main text, pre-capture libraries were constructed following the “adapter-ligation-first” protocol of the Kapa Biosystems kit included in the SeqCap Epi reagents from Roche. Briefly, we sheared input genomic DNA (0.25–0.5 µg in 130 µl in Covaris microTUBES) for 3 min on a LE220 sonicator set to duty factor 30%, peak incident power 140 W and 200 cycler per burst. The sheared DNA (mode ~200 bp) was concentrated with 1.7 volumes of Agencourt AMPure XP beads (Beckman Coulter). Beads were resuspended in 70 µl end repair master mix (Kapa Biosystems). End repair, A-tailing, ligation to indexed 5-methyl-C modified adapters, and dual size selection on AMPure beads to narrow the size distribution of the fragment library was performed according to the SeqCap Epi protocol. After EpiTect Fast bisulfite conversion (Qiagen; extending the two 60 °C cycles to 20 min.), the entire eluate from the spin column (20 µl) was PCR-amplified for 12 cycles in 80 µl of HiFi HotStart Uracil+ ReadyMix (Kapa Biosystems). AMPure cleaned-up PCR products were quantified by Qubit (Thermo Fisher).
To lower input DNA amounts, we tested the “bisulfite-conversion first” library construction protocol of the Accel-NGS Methyl-Seq kit (Swift Biosciences) following the kit instructions with the following exceptions: (1) To minimize the off-target sequencing rate, we sheared the input DNA to ~200 bp fragments using the LE220 settings described above instead of shearing to ~400 bp fragments as recommended for Accel-NGS WGBS libraries; (2) we doubled the PCR volume and used 8 PCR cycles for the pre-capture library amplification in 1× HiFi HotStart ReadyMix (Kapa Biosystems).
SeqCap Epi hybridization reactions contained a total of 1 µg of a pool of 2–4 PCR-amplified pre-capture libraries, a total of 1 nmol of 2–4 index-specific blocking oligonucleotides, and the custom SeqCap probe pool designed for the DyMe-Seq targets listed in Additional file 2. After hybridization (typically 70 h), bead capture, low- and high-stringency washes, the entire bead-bound captured material was amplified by 12 cycles of PCR. Hybrid-selected DyMe-Seq libraries were sequenced on an Illumina HiSeq 2500 instrument in fast mode together with a 10% spike-in of a non-indexed PhiX174 library to generate a median of 65 million indexed 100-base purity-filtered paired reads per library. Alignment rates ranged from 90 to 96%. Mean target coverage ranged from 26× to 63×. The bisulfite conversion rate of cytosines in non-CpG context was 99.6% on average. Standard performance metrics for each library are available in Additional file 7: Table S3.
dynamic methylation sequencing
whole genome bisulfite sequencing
differentially methylated region
transcription factor binding site
MZ and AM conceived the study. MZ performed region selection and all computational analysis. ES, HG and AG implemented the DyMe-Seq and performed all experiments. MZ, AG and AM wrote the paper. All authors read and approved the final manuscript.
We would like to thank all the members of the Meissner lab. A.M. is a New York Stem Cell Foundation Robertson Investigator.
The authors declare that they have no competing interests.
Availability of data and materials
Data accession: All processed hybrid capture data are available under the Datasets download tab on the GTEx portal (http://www.gtexportal.org) or https://owww.molgen.mpg.de/~meissner/DyMe-Seq/, and the described WGBS data are available through the Roadmap Epigenomics Project (http://www.roadmapepigenomics.org). Materials: The DyMe-Seq capture probes will be available for purchase from Roche.
This work was funded by NHGRI (U01HG007610) and the New York Stem Cell Foundation.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
- Smith ZD, Meissner A. DNA methylation: roles in mammalian development. Nat Rev Genet. 2013;14(3):204–20.View ArticlePubMedGoogle Scholar
- Schubeler D. Function and information content of DNA methylation. Nature. 2015;517(7534):321–6.View ArticlePubMedGoogle Scholar
- Laird PW. Principles and challenges of genome-wide DNA methylation analysis. Nat Rev Genet. 2010;11(3):191–203.View ArticlePubMedGoogle Scholar
- Frommer M, McDonald LE, Millar DS, Collis CM, Watt F, Grigg GW, Molloy PL, Paul CL. A genomic sequencing protocol that yields a positive display of 5-methylcytosine residues in individual DNA strands. Proc Natl Acad Sci USA. 1992;89(5):1827–31.View ArticlePubMedPubMed CentralGoogle Scholar
- Ziller MJ, Gu H, Muller F, Donaghey J, Tsai LT, Kohlbacher O, De Jager PL, Rosen ED, Bennett DA, Bernstein BE, et al. Charting a dynamic DNA methylation landscape of the human genome. Nature. 2013;500(7463):477–81.View ArticlePubMedPubMed CentralGoogle Scholar
- Lee EJ, Pei L, Srivastava G, Joshi T, Kushwaha G, Choi JH, Robertson KD, Wang X, Colbourne JK, Zhang L, et al. Targeted bisulfite sequencing by solution hybrid selection and massively parallel sequencing. Nucl Acids Res. 2011;39(19):e127.View ArticlePubMedPubMed CentralGoogle Scholar
- Li Q, Suzuki M, Wendt J, Patterson N, Eichten SR, Hermanson PJ, Green D, Jeddeloh J, Richmond T, Rosenbaum H, et al. Post-conversion targeted capture of modified cytosines in mammalian and plant genomes. Nucl Acids Res. 2015;43(12):e81.View ArticlePubMedPubMed CentralGoogle Scholar
- Allum F, Shao X, Guenard F, Simon MM, Busche S, Caron M, Lambourne J, Lessard J, Tandre K, Hedman AK, et al. Characterization of functional methylomes by next-generation capture sequencing identifies novel disease-associated variants. Nat Commun. 2015;6:7211.View ArticlePubMedPubMed CentralGoogle Scholar
- Roadmap Epigenomics Consortium, Kundaje A, Meuleman W, Ernst J, Bilenky M, Yen A, Heravi-Moussavi A, Kheradpour P, Zhang Z, Wang J, et al. Integrative analysis of 111 reference human epigenomes. Nature. 2015;518(7539):317–30.View ArticlePubMed CentralGoogle Scholar
- Bernstein BE, Birney E, Dunham I, Green ED, Gunter C, Snyder M. An integrated encyclopedia of DNA elements in the human genome. Nature. 2012;489(7414):57–74.View ArticleGoogle Scholar
- Gifford CA, Ziller MJ, Gu H, Trapnell C, Donaghey J, Tsankov A, Shalek AK, Kelley DR, Shishkin AA, Issner R, et al. Transcriptional and epigenetic dynamics during specification of human embryonic stem cells. Cell. 2013;153(5):1149–63.View ArticlePubMedPubMed CentralGoogle Scholar
- Schultz MD, He Y, Whitaker JW, Hariharan M, Mukamel EA, Leung D, Rajagopal N, Nery JR, Urich MA, Chen H, et al. Human body epigenome maps reveal noncanonical DNA methylation variation. Nature. 2015;523(7559):212–6.View ArticlePubMedPubMed CentralGoogle Scholar
- Meissner A, Gnirke A, Bell GW, Ramsahoye B, Lander ES, Jaenisch R. Reduced representation bisulfite sequencing for comparative high-resolution DNA methylation analysis. Nucl Acids Res. 2005;33(18):5868–77.View ArticlePubMedPubMed CentralGoogle Scholar
- GTEX-consortium: Human genomics. The genotype-tissue expression (GTEx) pilot analysis: multitissue gene regulation in humans. Science. 2015;348(6235):648–60.View ArticleGoogle Scholar
- Ziller MJ, Edri R, Yaffe Y, Donaghey J, Pop R, Mallard W, Issner R, Gifford CA, Goren A, Xing J, et al. Dissecting neural differentiation regulatory networks through epigenetic footprinting. Nature. 2014;518(7539):355–9.View ArticlePubMedPubMed CentralGoogle Scholar
- Rakyan VK, Down TA, Balding DJ, Beck S. Epigenome-wide association studies for common human diseases. Nat Rev Genet. 2011;12(8):529–41.View ArticlePubMedPubMed CentralGoogle Scholar
- Irizarry RA, Ladd-Acosta C, Wen B, Wu Z, Montano C, Onyango P, Cui H, Gabo K, Rongione M, Webster M, Ji H, Potash JB, Sabunciyan S, Feinberg AP. The human colon cancer methylome shows similar hypo- and hypermethylation at conserved tissue-specific CpG island shores. Nat Genet. 2009;41(2):178–86. doi:10.1038/ng.298.View ArticlePubMedPubMed CentralGoogle Scholar
- Xi Y, Li W. BSMAP: whole genome bisulfite sequence MAPping program. BMC Bioinform. 2009;10:232.View ArticleGoogle Scholar
- Sun D, Xi Y, Rodriguez B, Park HJ, Tong P, Meong M, Goodell MA, Li W. MOABS: model based analysis of bisulfite sequencing data. Genome Biol. 2014;15(2):R38.View ArticlePubMedPubMed CentralGoogle Scholar
- Thorvaldsdottir H, Robinson JT, Mesirov JP. Integrative genomics viewer (IGV): high-performance genomics data visualization and exploration. Brief Bioinform. 2012;14:178–92.View ArticlePubMedPubMed CentralGoogle Scholar
- Akalin A, Kormaksson M, Li S, Garrett-Bakelman FE, Figueroa ME, Melnick A, Mason CE. methylKit: a comprehensive R package for the analysis of genome-wide DNA methylation profiles. Genome Biol. 2012;13(10):R87.View ArticlePubMedPubMed CentralGoogle Scholar
- Benjamini Y, Hochberg Y. Controlling the false discovery rate—a practical and powerful approach to multiple testing. J R Stat Soc Ser B Methodol. 1995;57(1):289–300.Google Scholar
- Li QH, Brown JB, Huang HY, Bickel PJ. Measuring reproducibility of high-throughput experiments. Ann Appl Stat. 2011;5(3):1752–79.View ArticleGoogle Scholar
- McLean CY, Bristor D, Hiller M, Clarke SL, Schaar BT, Lowe CB, Wenger AM, Bejerano G. GREAT improves functional interpretation of cis-regulatory regions. Nat Biotechnol. 2010;28(5):495–501.View ArticlePubMedPubMed CentralGoogle Scholar
- Feng H, Conneely KN, Wu H. A Bayesian hierarchical model to detect differentially methylated loci from single nucleotide resolution sequencing data. Nucl Acids Res. 2014;42(8):e69.View ArticlePubMedPubMed CentralGoogle Scholar
- Lander ES, Waterman MS. Genomic mapping by fingerprinting random clones: a mathematical analysis. Genomics. 1988;2(3):231–9.View ArticlePubMedGoogle Scholar