Quick answer: DNA methylation analysis measures, per cytosine, what fraction of reads in a sample carry a methyl group — most commonly after bisulfite conversion, which turns unmethylated cytosines into uracil (read as thymine) while leaving methylated cytosines intact. The bioinformatics challenge is that conversion destroys three quarters of the sequence alphabet, so alignment needs specialized tools (Bismark, bwa-meth, BSmap), and every downstream number — methylation level, differentially methylated region, epigenetic claim — is only as honest as conversion efficiency, coverage, and cell-type composition allow.
If you are new to the sequencing layer itself, the NGS quality control guide covers the FASTQ-level checks that apply here too; the difference starts at alignment.
Table of Contents
What methylation is, and why the measurement is indirect
In mammals, the epigenetic mark of interest is almost always 5-methylcytosine at CpG dinucleotides — a methyl group on the cytosine in a 5′-CG context. Methylation at promoter CpGs generally represses transcription; enhancer and gene-body methylation carry different, context-dependent meanings. The mark is chemical, not sequence-level: a methylated and an unmethylated cytosine are the same base to a sequencer.
Bisulfite chemistry is the trick that converts chemistry into sequence:
- Unmethylated C → uracil → read as T.
- Methylated C → stays C.
- So a read coming from bisulfite-treated DNA is a reduced-alphabet version of the original, and the C/T pattern is the methylation signal.
Everything in this article follows from one consequence: after conversion, a genomic CpG site may appear in reads as either C or T, and methylation level at that site is simply the fraction of informative reads showing C.
Why bisulfite reads are hard to align
Conversion changes up to three quarters of cytosines, so a converted read no longer resembles the reference at ordinary mismatch tolerances. The specifics trip up every beginner:
- Reduced alphabet. Reads align only in the C→T (or G→A on the reverse strand) direction relative to the reference; a valid alignment can never show T where the reference has C (except at true SNPs or 5-hydroxymethylcytosine, see below).
- Genome size is effectively larger. Both strands convert separately, and aligners typically search a C→T-converted genome plus its reverse complement — indexes cost ~2× memory and time.
- SNPs masquerade as methylation (and vice versa). A C→T genetic variant is indistinguishable from an unmethylated cytosine in a single read; careful pipelines flag known variants or genotype them from the converted data.
- Clonal/PCR amplification duplicates inflate apparent coverage, which matters more here because coverage per site is often marginal to begin with.
The standard tools handle this by pre-converting the reference in silico: Bismark (aligns to a four-strand converted genome, then extracts methylation calls; the field’s long-standing default), bwa-meth or BSmap (align with an aligner tuned for bisulfite scoring), and for long reads, dedicated callers that skip conversion entirely (see the Oxford Nanopore section below).
The workflow, stage by stage
- FASTQ QC. As always: adapter check, quality distribution, and one bisulfite-specific metric — overall conversion rate, which some QC tools estimate from non-CpG cytosines. Below ~99% conversion, background unmethylation leaks into every estimate.
- Trim adapters and quality. Standard trimming, but be careful not to over-trim: bisulfite-converted ends have degraded quality profiles that interact with adapter detection.
- Align to the converted reference. Bismark or your chosen BS-aware aligner; output is an aligned BAM plus Bismark’s methylation-call report (
.txtwith per-cytosine C/T counts). - Extract per-cytosine calls.
bismark_methylation_extractorproduces per-site tables in three sequence contexts: CG, CHG, CHH (H = A, C, or T). CpG is the regulatory mark in mammals; CHH/CHG methylation is sparse in animals and mostly noise/context control, but it is the primary signal in plants. - Remove duplicates and filter. Deduplicate (Bismark has its own deduplication step), filter reads by alignment quality, and consider base-level filters (e.g., excluding sites near read ends or known SNPs).
- Aggregate and compare. Per-site tables become per-region summaries (promoters, CpG islands, enhancers) and then statistical comparisons between conditions — with the caveat that coverage thresholds decide which sites are even testable (commonly ≥10× per sample per site).
- Call differentially methylated regions (DMRs). Single noisy CpGs are rarely the unit of interest; DMR callers (methylKit, DSS, bsseq, dnmtools) model neighboring sites together and are where the statistics actually live.
Reading a methylation result honestly
- Methylation level is a ratio with a denominator. 90% at 3× coverage and 90% at 60× coverage are different claims. Coverage-aware sites-only analysis is not optional.
- Regions beat single sites. A DMR spanning several CpGs with consistent direction is far more credible than one site at p < 0.01 — smoothing/region methods exist precisely because adjacent CpGs are correlated (which also violates naive independence assumptions in per-site tests).
- Cell-type composition is the dominant confounder in tissue studies. Bulk tissue methylation differences routinely reflect shifts in cell proportions (more immune cells, fewer neurons) rather than within-cell epigenetic change. Reference-based deconvolution (e.g., Houseman-style methods with a suitable reference) or sorted/single-cell designs are the defenses.
- Batch and conversion efficiency differences between samples mimic global hypomethylation. Randomize libraries across batches and include conversion-rate checks per sample.
- CpG islands, shores, and shelves behave differently. Islands are usually unmethylated and gain methylation when a promoter silences; shores (0–2 kb flanking) often carry the most informative differential signal; interpret by context, not by a single genome coordinate.
- Methylation is not expression. Promoter hypermethylation often — but not always — coincides with reduced transcription. Pairing methylation with RNA-seq from the same samples is how a methylation claim moves from correlation toward mechanism.
Beyond bisulfite: the measurement alternatives
Worth knowing because the choice shapes the entire analysis:
- WGBS (whole-genome bisulfite sequencing) — genome-wide single-CpG resolution; expensive, high duplication risk, needs deep coverage.
- RRBS — enzymatically enriches CpG-dense regions first; cheap promoter-focused profiling; no genome-wide coverage.
- Infinium arrays (450K/850K) — fixed probe sites, cheap, superb for cohort-scale epidemiology; limited to preselected positions and cross-hybridization caveats.
- Oxford Nanopore / PacBio direct detection — reads native methylation without conversion, phasing methylation with haplotypes and long-range context in one pass; accuracy models are improving but validation expectations differ from bisulfite.
- 5-hydroxymethylcytosine (5hmC) — bisulfite cannot distinguish 5mC from 5hmC (both protected); oxidative-bisulfite or TAB-seq variants separate them when the distinction matters (notably in brain tissue).
Key takeaways
- Bisulfite conversion turns an epigenetic mark into a sequence difference: C/T at each cytosine is the measurement.
- Alignment needs bisulfite-aware tools (Bismark, bwa-meth, BSmap) because the reduced alphabet breaks ordinary aligners’ assumptions.
- Per-cytosine methylation level = C reads / informative reads, and coverage thresholds determine which claims are possible.
- Differentially methylated regions from methods that model neighboring CpGs are the standard evidence unit, not single-site p-values.
- Cell-type composition, batch, and conversion efficiency are the confounders that manufacture epigenetic findings in bulk tissue.
- Arrays, RRBS, WGBS, and long-read direct detection trade cost, resolution, and scope — the choice precedes every analysis decision.
Challenge
Download a small public bisulfite dataset (SRA hosts many WGBS/RRBS runs) and run the extraction stage yourself, or use the per-cytosine report if you are not ready to align. Pick one promoter CpG island: report its mean coverage, its methylation level in each sample, and three flanking-shore sites. Then try to state one biological claim the data support and one they cannot support because of coverage or cell-mixture limits. Holding both sentences at once is the working posture of methylation analysis.
Want to run an epigenetics analysis with proper confounder handling? Ampersand Academy offers one-to-one mentoring in bioinformatics — part of Mahadhi, a development and digital marketing company.
FAQ
What does methylation level at a CpG site mean numerically?
The fraction of aligned, informative reads at that position that show a cytosine after bisulfite conversion: 0.0 means no read was methylated, 1.0 means all were, and the number in between is a proportion of sampled molecules — whose reliability depends entirely on coverage depth.
Why can’t a normal aligner handle bisulfite-converted reads?
Conversion changes up to three quarters of cytosines, so true converted reads mismatch the reference constantly and normal scoring rejects them. Bisulfite-aware aligners search a C→T-converted reference in the correct mismatch direction so that converted reads align with realistic scores.
What is the difference between WGBS and RRBS?
WGBS bisulfites the whole genome, giving unbiased single-CpG coverage everywhere at high cost. RRBS enriches CpG-dense fragments (islands, promoters) enzymatically first, so you get deep coverage of regulatory regions at a fraction of the price but no genome-wide view.
How much coverage do I need per site?
As a practical floor, around 10× per site per sample for per-CpG comparisons; region-level callers can work with less because they borrow signal across neighboring sites. Below that, a methylation level like 50% at 4× is essentially an anecdote.
How do I know the bisulfite conversion worked?
Non-CpG cytosines in mammalian samples should be almost fully converted; the residual methylation rate there (from conversion reports or QC tools) estimates failure. Below roughly 99% efficiency, unmethylated positions will falsely register as methylated across every site in the sample.
Can bisulfite sequencing tell 5mC apart from 5hmC?
No — both are protected from conversion, so standard bisulfite data reports them jointly as methylated. When the distinction matters (brain tissue especially), use oxidative-bisulfite or TAB-seq protocols, or long-read direct-detection methods that model the modifications separately.
Why might a differential methylation result be a cell-composition artifact?
Bulk tissue is a mixture of cell types with very different methylation landscapes. If case samples contain more of one cell type than controls, thousands of CpGs shift together without any within-cell epigenetic change — which is why deconvolution, sorted cells, or single-cell methods are the standard checks before a biological claim.