Long-Read Basecalling and Model Dependence: Why the Same Reads Recalled Later Produce Different Variants
Bioinformatics Insight Series
Long-Read Basecalling and Model Dependence: Why the Same Reads Recalled Later Produce Different Variants
In short-read sequencing, the FASTQ is close to a measurement. In nanopore sequencing, it is an inference — a neural network's best guess at what sequence produced an electrical trace. Swap the model and the trace has not changed, but the bases have.
A collaborator sends you a nanopore dataset from 2023 and a variant list generated at the time. You reprocess it with your current pipeline — same POD5 files, same aligner, same variant caller settings — and the variant list comes back different. Some homopolymer indels have disappeared. A handful of substitutions that were previously called are gone. One or two new ones have appeared.
Nothing in your pipeline is broken. The reads were basecalled with a newer model, and the newer model read the same electrical signal differently.
This is a category of dependency that most short-read practitioners never have to think about, because short-read basecalling is effectively frozen: the instrument converts fluorescence images to bases on-board, the conversion is stable across years, and nobody re-derives FASTQ from raw images. Long-read platforms invert that assumption. The raw signal is archived precisely because the interpretation is expected to improve, and every improvement is a silent change to what your data says.
The FASTQ is a model output, not a measurement
A nanopore run measures the disruption in ionic current as a strand translocates through a protein pore. That trace — the "squiggle" — is the actual observation. Converting it into A, C, G, and T is a machine learning problem, solved by a neural network trained on signal-to-sequence pairs.
The architecture is not incidental. Oxford Nanopore ships accuracy tiers that differ in model size and compute cost, and the vendor's own guidance ties them to different applications: fast basecalling for real-time insight when compute is limited, high-accuracy (HAC) for high-throughput variant analysis, and super-accuracy (SUP) for de novo assembly and low-frequency variant work such as somatic and single-cell applications. Choosing a tier is a scientific decision disguised as a resource decision.
Model versions move too. In an independent benchmark of Dorado's v5.0.0 and v5.2.0 DNA models on bacterial isolates with high-quality references, Ryan Wick found the architectures had diverged between releases: the HAC model grew from five alternating-direction LSTM layers to seven, a 37% increase in parameter count, while the SUP architecture appeared unchanged, so its differences came from training data and weights instead. Two models with the same name and different version suffixes are not minor revisions of each other.
Figure 1. The nanopore FASTQ sits downstream of a learned transformation. Because the raw signal is retained, the same molecule can be re-interpreted indefinitely — and each re-interpretation is a new dataset, not a refreshed copy of the old one. Homopolymer length and modification-adjacent bases are where the two interpretations diverge most readily.
Homopolymers: where the signal genuinely runs out
A homopolymer is a run of identical bases. As the strand translocates, a run of identical bases produces a stretch of current with no distinguishing features — the pore reports roughly the same value whether five or six identical bases are passing through. Length must be inferred from dwell time, which is stochastic.
This is the failure mode the chemistry has been chasing for a decade, and it is closer to solved than it once was without being solved. In Wick's benchmark, the residual error profile of the newest SUP assemblies was dominated by exactly this: across 30 assemblies he counted 99 bp of error at 79 loci, of which 44 were homopolymer-length errors and 12 were other single-base indels. Homopolymers were the single largest remaining category after everything else had been optimized away.
Longer runs degrade further. A 2026 systematic evaluation across six R10.4.1 mitochondrial datasets reported that basecalling accuracy declined progressively with homopolymer length, with a marked drop beyond roughly 10 bp, and that both simplex and duplex read accuracy frequently fell below 50% in tracts longer than 10 bases. The same work noted the relationship is probabilistic rather than deterministic — short tracts still produced occasional errors, and long tracts did not invariably produce indels.
The practical consequence is that a homopolymer indel call in long-read data carries a model-version dependency that a substitution elsewhere in the genome does not. If a clinically relevant variant sits in or adjacent to a homopolymer tract, the call is partly a property of which basecaller produced the FASTQ.
Methylation-aware models change the canonical bases too
The second dependency is less intuitive. Base modifications alter the current signature — that is precisely why nanopore can detect methylation directly from native DNA without bisulfite conversion. But it also means that a modification the model was never trained on presents as an unfamiliar signal, and the network resolves it into whichever canonical base fits best.
The clearest demonstration comes from a set of Listeria monocytogenes isolates sequenced by the Taiwan CDC and characterized in Communications Biology. Twelve strains were sequenced identically; seven assembled to high quality between Q47 and Q60, while five came in unexpectedly low at Q26 to Q32, and the low-quality genomes did not improve on repeat sequencing. The failure was not the usual suspect: the low-quality genomes carried 1,533 to 5,670 mismatch errors versus 0 to 40 in the high-quality set, while homopolymer indels were comparable between groups and were not the source of the deficit.
The cause turned out to be epigenetic. Whole-genome amplification, which erases modifications, raised one isolate from Q26 to Q53 and cut its mismatches from 5,670 to 16 — the same DNA, stripped of methylation, basecalled cleanly. The authors traced the affected loci to lineage-specific pentanucleotide motifs and concluded that the errors arose from modification types the basecalling model had not been trained on. Notably, the problem persisted across basecaller versions and accuracy tiers they tested, including HAC versus SUP, which is the point: this is a training-coverage gap, not a compute-budget gap.
That last property is what makes these errors dangerous. Because they recur at the same position across the pileup, consensus and read-depth-based polishing cannot remove them — systematic errors are recurrent basecalling errors at the same locus, not correctable by the consensus of read pileups. They pass every filter that assumes errors are independent across reads. And the downstream damage propagates: the same study found that phylogeny built from the uncorrected ONT-only genomes was distorted by these errors, while the WGA-demodified and computationally corrected genomes clustered correctly with the hybrid assemblies.
The corollary for methylation-aware basecalling is that calling modifications is not a free add-on to a canonical run. It is a different model, trained on different data, and the canonical bases it emits are not guaranteed to match the ones a canonical-only model would have emitted from the same signal.
What actually changes between versions
Model updates are usually announced in terms of read accuracy, which understates their effect on downstream calls. Wick's benchmark is useful here because it measured both. Between the two HAC versions, median read identity rose from Q16.3 to Q17.0 — about 15% fewer read errors — but median assembly errors dropped from 37.5 to 13, a 65% reduction. A modest change in per-read accuracy produced a four-fold-larger change in the consensus.
The same benchmark also shows that newer is not uniformly better on every axis. For the SUP tier, the newer model's reads had a slightly lower median identity than the older model's, though the difference was small, while assembly accuracy still improved modestly. Version numbers are not a scalar quality ranking.
| Dependency | What changes | Where it surfaces |
|---|---|---|
| Accuracy tier | Model size and compute; fast, HAC, and SUP are distinct networks | Low-VAF and somatic calls, de novo assembly consensus |
| Model version | Architecture and/or training weights, under the same tier name | Consensus accuracy, disproportionate to the read-level change |
| Modification model | Which modifications are represented in training | Systematic substitutions at modified motifs; unaffected by depth |
| Chemistry and flow cell | Pore and translocation conditions the model was trained against | Homopolymer resolution; model–chemistry mismatch |
| Simplex vs. duplex | One strand read once, or both strands reconciled | Per-read accuracy at the cost of yield |
What this means for practice
The uncomfortable implication is that a long-read variant call is reproducible only if the basecalling configuration is reproducible, and most projects do not record it with the seriousness they apply to aligner and caller versions.
Record the basecaller as a pipeline parameter
Basecaller name, version, model string including its version suffix, accuracy tier, and any modification models — recorded alongside the aligner and caller versions, not in a lab notebook. The full model identifier is the meaningful unit; a run described only as "SUP" is underspecified.
Retain the raw signal, and understand what retaining it commits you to
The field's standing practice is to keep POD5 because interpretation improves. As one recent nanopore data descriptor puts it, frequent updates to basecalling models mean only access to raw signal permits reanalysis with newer algorithms — and methylation-based basecalling errors are the specific example they cite as motivation, given the effect on genome reconstruction and genotyping. Retention is correct. But it converts your dataset into something that has no single canonical sequence, only a sequence-as-of-a-model. One community resource has responded by treating submissions as living data, re-basecalling all POD5 reads on a quarterly cycle and releasing versioned BAM files — an honest acknowledgment that the basecall is a derived, perishable product.
Re-basecall the whole cohort, not the new samples
The most common way this becomes a real problem is a longitudinal study where early samples were called with one model and later samples with another. The model version then correlates with sample batch, and any systematic difference between models becomes a confounder indistinguishable from biology. If you upgrade, upgrade everything.
Treat homopolymer and modification-adjacent calls as a separate confidence class
Both categories fail systematically rather than randomly, so depth does not rescue them. In clinical contexts, a variant in a long homopolymer tract or at a methylated motif deserves orthogonal confirmation regardless of how many reads support it.
Check that the model matches the chemistry
Models are trained against specific pore and translocation conditions. Running a model against data from a chemistry it was not trained on is a silent quality loss rather than an error, and mismatches are easy to introduce when reprocessing archived runs.
The short version
Short-read pipelines can treat the FASTQ as the starting point of analysis. Long-read pipelines cannot: the FASTQ is already an analytical result, produced by a versioned model with its own training-data limitations and its own failure modes. Provenance has to start one step earlier than most people start it.
A different kind of version pinning
Bioinformatics has largely absorbed the lesson that tool versions belong in the methods section. Basecalling models have not fully made that transition, partly because they sit upstream of where most analysts consider their pipeline to begin, and partly because the vendor relationship frames them as instrument firmware rather than analytical software.
They are analytical software. A neural network that converts a physical measurement into a discrete sequence, with a documented error profile that varies by version, is doing inference — and every property we demand of inference elsewhere in the pipeline applies. The signal in the POD5 file will not change. What we believe it says will keep changing, and that is worth knowing about your own data before someone else discovers it.
Zetobit builds and validates CAP/CLIA-compliant NGS pipelines, including long-read workflows where basecalling provenance and model-version control are part of the validation record. If you are standing up long-read analysis or reconciling call sets across model versions, we're happy to talk.
References
- Wick RR. Dorado v1.0.0 and the v5.2.0 basecalling models. Ryan Wick's bioinformatics blog, 27 May 2025. doi:10.5281/zenodo.15524362.
- Chiou C-S, Chen B-H, Wang Y-W, Kuo N-T, Chang C-H, Huang Y-T. Correcting modification-mediated errors in nanopore sequencing by nucleotide demodification and reference-based correction. Communications Biology 6:1215 (2023). doi:10.1038/s42003-023-05605-4.
- Oxford Nanopore Technologies. Nanopore sequencing accuracy. Platform documentation, accessed July 2026.
- A practical framework for detecting and correcting Nanopore homopolymer errors in animal mitochondrial genome assembly. Research Square preprint (2026). doi:10.21203/rs.3.rs-9835208/v1. Preprint — not yet peer reviewed.
- A whole-genome sequencing dataset of nanopore raw signals for bacterial genotyping and methylation analysis. Scientific Data (2025). doi:10.1038/s41597-025-06319-4.
- SquiDBase: a community resource of raw nanopore data from microbes. NAR Genomics and Bioinformatics 8(1):lqaf213 (2026). doi:10.1093/nargab/lqaf213.
- Sereika M, Kirkegaard RH, Karst SM, et al. Oxford Nanopore R10.4 long-read sequencing enables the generation of near-finished bacterial genomes from pure cultures and metagenomes without short-read or reference polishing. Nature Methods 19:823–826 (2022). doi:10.1038/s41592-022-01539-7.
- Huang Y-T, Liu P-Y, Shih P-W. Homopolish: a method for the removal of systematic errors in nanopore sequencing by homologous polishing. Genome Biology 22:95 (2021). doi:10.1186/s13059-021-02282-6.