DNA Methylation Calling from Bisulfite Sequencing: Why the Chemistry That Reveals Methylation Also Corrupts the Data

Zetobit · Bioinformatics Insight Series

DNA Methylation Calling from Bisulfite Sequencing: Why the Chemistry That Reveals Methylation Also Corrupts the Data

Bisulfite conversion is an elegant trick: it rewrites unmethylated cytosines as thymines and leaves methylated ones untouched, turning an invisible epigenetic mark into a readable base change. The same reaction also fragments the DNA, collapses the alphabet, and manufactures false positives — and every methylation call has to survive that damage.

By Kanna Nandakumar, PhD Zetobit LLC

DNA methylation — the addition of a methyl group to cytosine, mostly at CpG dinucleotides — is one of the most consequential marks in the genome. It silences genes, defines cell identity, drives cancer through promoter hypermethylation, and underlies the epigenetic clocks used to estimate biological age. Reading it at single-base resolution is the job of bisulfite sequencing, and the core idea is beautiful in its simplicity: treat DNA with sodium bisulfite, and unmethylated cytosines deaminate to uracil (read as thymine) while methylated cytosines are protected and stay cytosine.¹ Methylation becomes a C-versus-T readout at every cytosine.

The problem is that this conversion is not a gentle relabeling. It is a harsh chemical assault that damages the molecule while marking it, and it collapses the information content of the sequence in the process. The result is that a bisulfite methylation pipeline spends most of its effort not measuring methylation but defending against the artifacts of the reaction that made methylation visible. Understanding those artifacts is the difference between a methylation value and a number that only looks like one.

The alphabet collapses, and alignment suffers

Because the vast majority of cytosines in the genome are unmethylated, converting them to thymine means most Cs become Ts. The four-letter genome effectively collapses to three letters, with far fewer distinguishing features per kilobase.² A standard aligner cannot use such reads directly, and the reduced complexity makes reads harder to place uniquely — a serious problem given that many CpG islands, the regulatory regions of greatest interest, sit in or near repetitive sequence.³

Worse, once cytosines are converted, the two DNA strands are no longer complementary — a C-to-T change on one strand is not mirrored by a matching change on the other.⁴ Purpose-built bisulfite aligners like Bismark handle this by converting everything to a common alphabet: reads are transformed to C-to-T (and G-to-A) versions and aligned against equivalently converted versions of the reference, in parallel, so the aligner can recover which of the (up to four) bisulfite strands a read came from before making any methylation call.⁴ A directional library sequences only two of those strands; a non-directional one can populate all four, and getting the library type wrong in the aligner settings silently discards or misassigns reads.⁵ None of this is a knob on a normal DNA pipeline — it is a different alignment problem.

The readout is asymmetric: an unmethylated cytosine that escapes conversion is indistinguishable, base for base, from a truly methylated one. The two are separable only statistically, using a measured conversion rate — which is why the control that quantifies incomplete conversion is not optional.

Incomplete conversion: a false positive by construction

Here is the asymmetry that makes methylation calling treacherous. The evidence for methylation is a cytosine that stayed a cytosine. But an unmethylated cytosine that simply escaped the reaction also stays a cytosine — and is now indistinguishable, base for base, from a genuinely methylated site. Every cytosine that resists deamination becomes a false methylation call.² Because conversion is never perfectly complete, a naïve caller that flags any residual cytosine as methylated has an extremely high false-positive rate; this is a documented and well-characterized failure of simple thresholding.⁶

The standard defense is a spiked-in, fully unmethylated control — most commonly the lambda-phage genome added to each sample before conversion. Since every cytosine in that control is known to be unmethylated, any cytosine still read as C after conversion measures the incomplete-conversion rate directly, giving a per-sample false-positive floor that the analysis must account for.⁷ Without this control, there is no way to know whether a modest methylation signal is real biology or unconverted background. And it cuts both ways: pushing conversion to completion requires harsher conditions that degrade more DNA, so reducing false positives can increase DNA loss — a direct trade-off, not a free tuning parameter.⁸

The reaction destroys the template it reads. The heat and low pH needed for complete conversion fragment DNA — by depyrimidination — cutting average fragment length to roughly 200–500 bp and reducing the intact template available for library prep.² For already-limited clinical inputs like FFPE tissue and cell-free DNA, this degradation is often the dominant constraint, which is exactly why enzymatic alternatives were developed.

PCR bias tilts the numbers further

A subtler distortion enters at amplification. Converted, unmethylated sequence is AT-rich, while methylated regions retain more cytosines and remain relatively GC-rich — and the two amplify with different efficiency. Unmethylated (AT-rich) fragments tend to amplify more readily than methylated ones, so PCR can underrepresent methylated regions and skew the measured methylation level.³ The methylation ratio a pipeline reports is therefore shaped not only by the true biology and the conversion rate but by amplification chemistry, which is why amplification-free and enzymatic library methods exist specifically to reduce this bias.³

The enzymatic alternative — and why it doesn't erase the problem

Enzymatic methyl sequencing (EM-seq) reaches the same C-versus-T readout through a gentler enzymatic route rather than harsh chemistry, and in head-to-head clinical comparisons it delivers higher library yield, longer inserts, better coverage of GC-rich regions, and lower duplication — advantages that matter most for degraded and low-input samples.⁹ But EM-seq shares the deep structural features of the problem: it still collapses the alphabet to three letters, still depends on conversion completeness, and can itself suffer incomplete conversion on low-input material.⁸ Switching chemistries improves the inputs to the bioinformatics; it does not remove the need for conversion controls, bisulfite-aware alignment, or careful statistics.

What a real methylation pipeline requires

Turning converted reads into trustworthy methylation calls is a chain of context-specific steps, each guarding against a distinct artifact:

• Bisulfite-aware alignment with the correct library type. Use a converted-alphabet aligner (e.g., Bismark, bwa-meth) and set directional vs. non-directional correctly, or reads are silently lost or misassigned.⁴
• A conversion-rate control. Spike in an unmethylated genome (lambda) and quantify incomplete conversion per sample; treat it as the false-positive floor for every call.⁷
• Model-based calling, not thresholds. Account for conversion rate, sequencing error, and coverage with a statistical model rather than flagging any residual C as methylated.⁶
• Context and strand separation. Distinguish CpG, CHG, and CHH contexts and handle each of the up-to-four bisulfite strands, since pooling them blindly conflates distinct biology.⁴
• Coverage-aware interpretation. A methylation ratio from a handful of reads is a noisy estimate; require adequate depth per cytosine before reporting a level, and weight PCR-duplicate handling accordingly.¹⁰

As with the other blind spots in this series, the honest output declares its own foundations: the measured conversion rate, the aligner and library type, the per-site coverage, and the statistical model behind each call. A methylation percentage reported without its conversion control is not a measurement — it is a cytosine count wearing a lab coat.

The takeaway

Bisulfite sequencing is a genuinely clever way to make an invisible epigenetic mark legible, and it remains a workhorse of clinical and research epigenomics. But the conversion that reveals methylation also degrades the DNA, collapses the sequence alphabet, breaks strand complementarity, and — most dangerously — makes an unconverted cytosine impossible to distinguish from a methylated one without a control. The methylation value at the end of the pipeline is only as good as the conversion rate behind it and the bisulfite-aware analysis around it. Reading the mark correctly means accounting, at every step, for the damage done in making it readable.

References

1. Bisulfite sequencing: introduction, features, workflow, and applications (bisulfite converts unmethylated C→U/T; methylated C protected). CD Genomics. cd-genomics.com
2. Targeted bisulfite sequencing — reduced complexity (three-letter code), destructive conversion (200–500 bp fragments), incomplete conversion false positives. CD Genomics. 2025. cd-genomics.com
3. Bisulfite sequencing in DNA methylation analysis (PCR amplification bias against methylated/GC-rich fragments; amplification-free and EM-seq alternatives). Abcam. abcam.com
4. Krueger F, Andrews SR. Bismark: a flexible aligner and methylation caller for Bisulfite-Seq (C-to-T / G-to-A conversion of reads and genome; strand recovery; CpG/CHG/CHH contexts). Bioinformatics. 2011. academic.oup.com
5. Bismark User Guide — directional vs non-directional libraries (OT/OB vs OT/CTOT/OB/CTOB); --directional discards CTOT/CTOB alignments. Babraham Bioinformatics. babraham.ac.uk
6. A Bayesian framework to identify methylcytosines from BS-seq data (incomplete conversion; extremely high false-positive rate for simple thresholding). PLoS Comput Biol. PMC4177668
7. A practical guide to the measurement and analysis of DNA methylation (lambda-phage spike-in to measure conversion; converted-genome alignment). PMC. PMC6775954
8. Ultra-mild bisulfite sequencing / UBS-seq context — conversion-completeness vs DNA degradation trade-off; EM-seq incomplete conversion on low input. Nat Commun. 2025. nature.com
9. Comprehensive comparison of enzymatic and bisulfite DNA methylation analysis in clinically relevant samples (EM-seq yield/insert/GC-coverage advantages; bisulfite DNA damage by depyrimidination). Clin Epigenetics. 2025. s13148-025-01959-0
10. Variable performance of bisulfite sequencing methods and read-mapping software (coverage/depth differences; aligner handling of converted cytosines). PMC. 2025. PMC11957057

Zetobit LLC · Bioinformatics Insight Series · Lexington, KY · zetobit.com