Long-Read Methylation Calling: Why the Methylome You Get Depends on Which Model Read the Signal
Long-Read Methylation Calling: Why the Methylome You Get Depends on Which Model Read the Signal
Bisulfite sequencing converts unmethylated cytosine chemically, so the methylation state is written into the sequence itself and read out like any other base. Long-read platforms don't convert anything. They record the raw current as a molecule passes through the pore and hand it to a neural network that returns a probability the base was modified. That is a different kind of claim. The methylome you report is a model output, and swapping the model — a version bump, an accuracy tier, a different modification set — changes it without changing the DNA.
The basecalling piece in this series covered how the same raw signal yields different sequences depending on which model interprets it, and why a FASTQ is a model output rather than a direct reading. Modified-base calling is the same architecture applied to a harder question, and the consequences are sharper — because for a canonical base there is at least a fixed four-letter answer the model is aiming at, while for a modification the model is asked whether something is present that leaves only a subtle imprint on the current.
The bisulfite piece covered a different failure mode entirely: incomplete conversion, DNA degradation, and the mapping difficulties of a three-letter genome. Those are properties of a chemical assay. Everything here is a property of an inference. Both technologies produce a number between 0 and 100 that gets called "methylation," and the two numbers are not produced by comparable processes.
Where the modification call actually happens
The mechanism is worth stating precisely, because the structure of the problem follows from it. When a basecaller runs a modified-base model, this runs as an extra pass on each read after normal basecalling: the basecalled sequence and the raw-signal context are given to another neural network, which returns a modification probability for each candidate base, stored in the read's MM and ML tags.1
Two things follow. First, the modification call is downstream of the basecall, so it inherits whatever the sequence model decided. Second, and more consequentially, the output is a probability per base per read — not a methylation state. Everything a downstream analysis treats as "the methylome" is produced by aggregating those probabilities, and aggregation requires a second set of choices.
The models themselves are modest and specific. In one independent evaluation, the modified-base models tested each had roughly 3.2 million parameters and shared an architecture: convolutional processing of signal and sequence context, a merge convolution, and two LSTM layers.2 A small network, trained on particular data, asked to detect a chemical modification from an electrical trace.
The model is a parameter, and it changes
The clearest evidence that this is not a settled measurement comes from what happens when the model changes.
Users comparing two modified-base model versions on the same control datasets — E. coli with known methylation at defined motifs — found that under one model version, tens of thousands of sites passed as highly methylated, while under a newer version on the same data the count for the equivalent analysis collapsed to zero, with the confidence threshold shifting from roughly 0.73 to 0.99 between versions.3 Same DNA, same instrument, same downstream tool: a model version boundary changed the answer categorically.
That is an edge case involving non-CpG contexts, and it was raised as a bug rather than expected behaviour. But it illustrates the general property cleanly: the model version is not a footnote to a methylation result, it is part of the measurement.
The more routine version of this appeared in a July 2026 independent evaluation. The vendor had stated of its newest high-accuracy model that it delivers the most accurate modified basecalling ever produced, outperforming the super-accuracy tier on every single model.2 Tested on bacterial genomes with motif-based methylation as an internal check, the results did not support that ordering: the best-performing model depended on the metric, the newer high-accuracy model did not always outperform its predecessor, and the older high-accuracy model sometimes outperformed super-accuracy.2
For 5mC specifically, the super-accuracy model had the best low-peak and high-peak means but noticeably poorer recall at the expected CCWGG motif, a result consistent across all three genomes carrying that motif.2 The author's practical conclusion is the useful one: for analyses focused on 5mC where sensitivity matters, the high-accuracy model might be the better choice — the opposite of the general recommendation.2
The author also offers a candid explanation worth carrying: the newest model may have been developed with a focus on human sequencing, and results on human data might look more favourable.2 Model performance is a function of what the model was trained and tuned for.
Training data determines what the model can see
This is the deepest version of the problem, and it is not a defect of any particular tool. A modified-base model learns the signal signature of a modification in the sequence contexts represented in its training set. Contexts that were rare in training are called poorly, and contexts that were absent are not called at all.
The plant case makes this concrete. Animal 5mC occurs mainly at CpG sites, but plant 5mC occurs in three contexts — CpG, CHG, and CHH — and CHH methylation, though less abundant, matters for silencing transposable elements.4 Vendor models perform well for CpG and CHG in plants but struggle with CHH, most likely due to the limited availability of positive CHH samples for training.4
A purpose-built model addressing exactly that gap — trained on a dataset covering 97.2% of possible 9-mers centred on CHH motifs — achieved whole-genome CHH methylation frequency correlations of 0.705 to 0.838 against bisulfite data across nine species, an improvement of 23.4% to 117.6% over the general-purpose model.4
Read the upper end of that range carefully. A correlation more than doubling because someone assembled better training data for a specific sequence context is not a small tuning gain. It means the original number, for that context, was substantially a statement about training coverage.
The same logic applies to platform generations. Substantial changes in pore structure and signal collection frequency between chemistry versions meant that methylation detection tools developed for the earlier platform could not be directly applied to the later one.4 The signal representation changed, so the models had to be rebuilt.
The second decision: the confidence filter
Suppose the model is appropriate and its probabilities are well calibrated. A per-read probability still has to become a per-site percentage, and that conversion has its own free parameter.
The standard tool filters low-confidence calls before tabulating. By default it estimates a threshold by sampling predictions from a subset of reads, defining confidence as the maximum predicted probability for each call, and setting the threshold at the tenth percentile of those confidence values.5 Calls below that value are not counted.5
The important property is that this threshold is estimated from the dataset's own confidence distribution rather than from prediction accuracy.6 Two consequences follow directly.
It is not comparable across samples. A sample whose calls are broadly confident and one whose calls are broadly uncertain both lose their bottom decile, at different absolute cutoffs. The filter's strictness is set by the data being filtered.
And because the percentile is not accuracy-based, datasets dominated by low-confidence calls — as with rare modifications — get assigned an insufficiently stringent threshold, producing elevated false discovery rates.6 The rarer the modification, the weaker the default protection.
Practitioners find this in the obvious way: working on a species with no known methylation profile, one user found the default threshold landed around 0.4–0.5 and that raising the filter threshold raised the average methylation, with 0.7 giving figures comparable to related species and 0.9 giving extreme values.7 The methylation percentage moved with a filtering parameter, and the way to choose the parameter was to see which output looked biologically plausible.
There is also a subtler interaction. Isolating a single modification type eliminates the possibility of misclassification between modified base types, resulting in higher accuracy.8 A model asked to distinguish 5mC from 5hmC can confuse them; a model asked only about 5mC cannot. Which modifications you request changes the accuracy of the ones you get.
| Choice | Why it moves the answer | Visible in the output? |
|---|---|---|
| Model version | Retrained networks assign different probabilities; version boundaries have produced categorical changes on control data3 | No — the percentage looks identical in form |
| Accuracy tier | Higher-accuracy tiers do not uniformly win; the best model depends on the metric2 | No |
| Modification set requested | Isolating one modification removes cross-type misclassification and raises accuracy8 | Only implicitly, via which codes appear |
| Sequence-context coverage in training | Under-represented contexts are called poorly; absent contexts are not called4 | No — absence of calls looks like absence of methylation |
| Confidence percentile filter | Threshold is estimated per dataset, not from accuracy6 | No — filtered calls simply do not appear |
| Coverage at the site | Concordance with orthogonal methods improves with depth, notably beyond about 20×9 | Usually yes, if depth is reported alongside |
Effects are drawn from the cited studies and tool documentation; specific values reflect those datasets, platforms, and versions.
Why this is not the bisulfite problem in new clothes
It would be easy to fold this into the general observation that methylation assays disagree. The disagreement is real — but the reasons differ, and that difference determines what you can do about it.
Bisulfite's characteristic errors are chemical and physical. Incomplete conversion is widely recognized as critical because it leads to overestimation of methylation levels, and the conversion process itself can degrade DNA, potentially losing reads carrying unmethylated cytosines.9 These are properties of a reaction. They are addressed with controls, conversion-rate estimates, and chemistry changes.
Long-read methylation errors are properties of an inference: training coverage, model version, architecture, calibration, and threshold selection. They are addressed by knowing which model ran and validating it in your context. No amount of improving the sequencing chemistry fixes a model that never saw your sequence context.
The empirical agreement between the two is good but not identity. Across platforms, correlation coefficients around 0.8 are typical, with higher concordance in GC-rich regions and at greater depth, and improvement beyond about 20× coverage.9 Systematic offsets exist in a consistent direction: bisulfite reports higher average methylation levels, while long reads detect more methylated cytosines overall, particularly in repetitive elements and regions where bisulfite coverage is poor.9 In a broader comparison, enzymatic conversion showed the highest concordance with bisulfite — unsurprising given similar chemistry — while nanopore showed lower agreement with both yet captured certain loci uniquely.10
Those last clauses are the point. Long reads disagree with bisulfite partly because they can see places bisulfite cannot. A correlation of 0.8 between a chemical assay and a model inference, measured over the subset of the genome where both work, is not a statement that one is 80% right.
What is genuinely strong here
The counterpoint matters, because none of this argues against long-read methylation.
Accuracy on defined material is high: basecalling of 5mC on synthetic strands with known composition reaches precision, recall, and F1 all above 99%, and vendor figures cite 99.6% for 6mA and 98.6% for 4mC/5mC on synthetic truth sets with current models.11 The capability is real and improving.
The unique capabilities are substantial and have no bisulfite equivalent. Methylation and sequence come from the same molecule with no separate library, so methylation is phased with genotype automatically. Duplex sequencing permits determination of strand modification symmetry and asymmetry by examining both strands of the same molecule, revealing that strand-asymmetric modification is widespread across the genome but reduced at imprinting control regions and CTCF binding sites.12 That measurement is not available from a bisulfite workflow at all.
And in independent benchmarking against orthogonal data, nanopore detection of 5mC and 5hmC was accurate relative to compared techniques.12 The technology works. The argument is about what has to travel alongside the number.
What this means for a pipeline
Record the model, not just the tool
A methylation result is reproducible only with the full model identifier — basecaller version, basecalling model and version, and the specific modified-base model and its version. Given that model boundaries have produced categorical differences on control data,3 "called with Dorado" is not a method description. This is the same requirement the series has made for benchmark versions, splice-predictor thresholds, and germline filtering strategies.
Do not compare methylomes across model versions
If a cohort was basecalled over a period during which models changed, model version is a batch variable correlated with time. Either re-basecall everything with one model — the raw signal is retained precisely so this is possible — or treat version as a covariate and expect it to matter. Re-basecalling is the cleaner option and is one of the genuine advantages of the platform.
Set the confidence threshold explicitly and report it
The default percentile is data-dependent and therefore not comparable across samples.6 For cross-sample work, set an explicit threshold so every sample is filtered identically, and record the value. Where the modification is rare, the default is known to be insufficiently stringent.6
Validate in your organism and sequence context
The plant CHH case shows that general-purpose models can be substantially wrong in contexts that were sparse in training,4 and this generalizes to any non-model organism or unusual context. Use control material with known modification status where available; tooling exists specifically to validate modified-base calls against ground-truth datasets.8
Request only the modifications you need
Since isolating a modification type removes cross-type misclassification,8 asking for 5hmC when the question is 5mC costs accuracy on 5mC. Match the model to the question.
Treat absence of calls as unmeasured, not unmethylated
A model that does not cover a sequence context reports nothing there. Downstream, nothing is indistinguishable from zero. Any conclusion of the form "no methylation in region X" needs the prior question answered: was that context in the model's scope at all?
The shape of the problem
The basecalling piece made the point that a FASTQ is not a reading of a molecule but a model's best reconstruction of one. Modified-base calling extends that by one step, and the extension is where the trouble concentrates: the sequence model has a fixed alphabet and a strong prior from decades of reference genomes, while the modification model is asked to detect a subtle chemical state from the same trace, using training data that necessarily covers some contexts well and others barely.
What arrives in the analysis is a percentage. It looks like the bisulfite percentage. It goes into the same differential-methylation tools, the same clocks, the same classifiers. But one number came from a chemical conversion and the other from a neural network's probability, thresholded at a percentile of its own confidence distribution, and neither the model nor the threshold travels with the value.
The methylome is real. The number is a reading of it made by a specific model on a specific day, and that model is a parameter of the result.
References
- Oxford Nanopore Technologies. Dorado basecaller documentation — modified basecalling and the
MM/MLtag mechanism. https://github.com/nanoporetech/dorado (tool documentation, cited for the mechanism and available modification models) - Wick RR. Dorado v2.0.0 part 3: methylation. 8 July 2026. https://rrwick.github.io/2026/07/08/dorado-v2-methylation.html (independent evaluation on bacterial genomes; archived with DOI 10.5281/zenodo.21254032 — cited for the model-comparison results and the quoted vendor claim)
- v5.2 DNA 5mC model not reporting non-CpG methylation. Dorado issue #1440, nanoporetech/dorado. https://github.com/nanoporetech/dorado/issues/1440 (user-reported issue with control-dataset tables; cited as a documented instance of version-boundary behaviour, not as a peer-reviewed result)
- Accurate cross-species 5mC detection for Oxford Nanopore sequencing in plants with DeepPlant. https://pmc.ncbi.nlm.nih.gov/articles/PMC11971355/
- Oxford Nanopore Technologies. Modkit documentation — filtering and threshold estimation. https://github.com/nanoporetech/modkit/blob/master/filtering.md (tool documentation, cited for the default percentile mechanism)
- ModkitOpt: systematic optimisation of modkit parameters. bioRxiv. 2025. https://www.biorxiv.org/content/10.64898/2025.12.19.695383v1.full.pdf (preprint; cited for the observation that the percentile threshold is not accuracy-based and for the rare-modification consequence)
- Which thresholds to choose? Modkit issue #245, nanoporetech/modkit. https://github.com/nanoporetech/modkit/issues/245 (community discussion, cited as an illustration of threshold sensitivity in practice)
- Oxford Nanopore Technologies. Modified base best practices and benchmarking. EPI2ME. https://epi2me.nanoporetech.com/mod-validation-data/ (vendor documentation, cited for the isolation-improves-accuracy point and the existence of validation tooling)
- A comparison of DNA methylation detection between HiFi sequencing and whole genome bisulfite sequencing in monozygotic twins with Down syndrome. PLOS ONE. 2025;20(8):e0329593. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0329593
- Comparison of current methods for genome-wide DNA methylation profiling. Epigenetics & Chromatin. 2025;18:616. https://link.springer.com/article/10.1186/s13072-025-00616-3
- Oxford Nanopore Technologies. Sequencing accuracy — modified base accuracy figures on synthetic truth sets. https://nanoporetech.com/platform/accuracy (vendor documentation; accuracy claims are the manufacturer's own)
- Double and single stranded detection of 5-methylcytosine and 5-hydroxymethylcytosine with nanopore sequencing. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11830040/