GPTomics/bio-long-read-sequencing-basecalling

Version Compatibility

Reference examples tested with: Dorado 1.0+, pod5 0.3+, samtools 1.19+, chopper 0.7+.

Before using code patterns, verify installed versions match. If versions differ:

• CLI: <tool> --version then <tool> --help to confirm flags

Results depend on inputs that outlive the binary version - record them:

• The basecaller MODEL string (e.g. [email protected]) sets the entire error profile and must be propagated to every downstream tool. Pin it.
• Modified-base models carry a SECOND version ([email protected]_5mCG_5hmCG@v3); the mod version can lag the simplex version - check dorado download --list.
• R9.4.1 and RNA002 models were removed from Dorado v1.0 defaults; legacy data needs an archived model path.

If code throws an error, introspect the installed tool (dorado --help, dorado basecaller --help) and adapt the example to the actual API rather than retrying.

Nanopore Basecalling

"Basecall my Nanopore data" -> Convert raw signal (POD5) into reads with Dorado using the chemistry-matched model, deciding the accuracy tier and whether to call modifications now - because the model choice is baked irreversibly into the output.

• CLI: dorado basecaller sup pod5s/ > calls.bam (simplex), dorado basecaller sup,5mCG_5hmCG pod5s/ > calls.bam (with methylation), dorado duplex sup pod5s/ > duplex.bam (duplex)

PacBio note: PacBio "basecalling" (CCS -> HiFi reads) runs on-instrument/in SMRT Link; users receive HiFi BAMs already at Q20-Q30+. This skill is Oxford Nanopore / Dorado. HiFi assembly lives in genome-assembly/hifi-assembly.

The Single Most Important Modern Insight -- There Is No "The Reads," Only "The Reads As Called By This Model"

Basecalling is not fixed preprocessing that yields a neutral FASTQ. The model and version chosen are an analysis decision written permanently into the BAM, with three consequences a naive user misses:

1. Methylation is a basecalling decision, not a later analysis step. Modified bases are inferred from raw signal at basecall time by Remora models and emitted as MM/ML tags. A plain BAM/FASTQ with no MM/ML tags has thrown the signal away - mods CANNOT be recovered without re-basecalling from POD5. If methylation might ever matter, request it now (sup,5mCG_5hmCG) and KEEP the POD5. See nanopore-methylation.
2. Downstream polish/variant models must match the basecaller model+version. medaka and Clair3 ship per-model weights (Clair3 r1041_e82_400bps_sup_v500; medaka the dotted r1041_e82_400bps_sup_v5.2.0). A mismatched model silently degrades accuracy with no error. Propagate the basecaller model name to every downstream step.
3. Mixing model versions across a cohort is a batch effect. Different model versions have different identity and homopolymer-indel error profiles. Re-basecall the WHOLE cohort with ONE current model before joint or differential analysis.

Dorado Subcommand Taxonomy

Dorado (one GPU-first executable) replaced Guppy, which is end-of-life. Bonito is ONT's research/training basecaller (not production); Rerio hosts research-release models (niche mods, bacterial methylation).

| Subcommand | Purpose | Canonical invocation | |------------|---------|----------------------| | basecaller | simplex basecalling | dorado basecaller hac pod5s/ > calls.bam | | duplex | template+complement duplex | dorado duplex sup pod5s/ > duplex.bam | | demux | barcode classification/split | dorado demux --kit-name SQK-NBD114-24 --output-dir out/ calls.bam | | trim | standalone adapter/primer trim | dorado trim reads.bam > trimmed.bam | | aligner | minimap2 alignment (carries MM/ML) | dorado aligner ref.mmi reads.bam > aln.bam | | correct | HERRO single-read correction | dorado correct reads.fastq > corrected.fasta | | summary | sequencing-summary TSV from BAM | dorado summary calls.bam > summary.tsv | | download | model management | dorado download --model <name> / --list |

Model Naming Scheme (load-bearing)

Format {analyte}_{pore}_{chemistry}_{speed}@v{ver} + optional mod suffix, e.g. [email protected]_5mCG_5hmCG@v3.

| Token | Meaning | Examples | |-------|---------|----------| | analyte | molecule | dna, rna004 | | pore | flow-cell generation | r10.4.1 (current), r9.4.1 (legacy) | | chemistry | kit chemistry | e8.2 (Kit 14) | | speed | translocation speed -> sampling rate | 400bps (5 kHz DNA), 130bps (RNA004, 4 kHz) | | tier | model size/accuracy | fast, hac, sup | | version | model version | @v4.3.0, @v5.2.0, @v6.0.0 |

Passing the bare tier (sup) lets Dorado auto-detect chemistry from POD5 metadata and fetch the matching latest model; pin a version ([email protected]) or a full path for reproducibility. Append mods comma-separated (sup,5mCG_5hmCG,6mA); only one mod model per canonical base may be active.

Decision Tree by Scenario

| Scenario | Recommended | Why | |----------|-------------|-----| | Any analysis (variant/assembly/methylation) | sup + matched model, pinned version | fast/hac error profile leaks into calls | | Live run / adaptive sampling / quick QC only | fast | speed; never for downstream analysis | | Routine work, compute-limited | hac | strong accuracy/compute balance (v5.2 closed much of the gap to sup) | | Methylation wanted now or maybe later | sup,5mCG_5hmCG (DNA), keep POD5 | mods are unrecoverable from a plain BAM -> nanopore-methylation | | Per-molecule accuracy, low input, phasing | dorado duplex sup | ~Q30 reads, but expect <10% duplex yield | | Diploid/phased T2T assembly from simplex | dorado correct (HERRO) before assembler | haplotype-aware Q22->Q40 -> genome-assembly/long-read-assembly | | Barcoded multiplexed run | basecall --no-trim, then dorado demux | trimming first strips barcodes before demux sees them | | Legacy R9.4.1 / RNA002 data | explicit archived model path | removed from Dorado v1.0 default downloads | | PacBio data | already HiFi; no Dorado step | CCS runs on-instrument -> genome-assembly/hifi-assembly |

Core Commands

# Simplex, super-accuracy, auto-detected chemistry-matched model (BAM is the default output)
dorado basecaller sup pod5s/ > calls.bam

# Pin the model version for reproducibility
dorado basecaller [email protected] pod5s/ > calls.bam

# Call methylation AT basecall time (CpG 5mC + 5hmC); KEEP pod5s/ - mods are unrecoverable later
dorado basecaller sup,5mCG_5hmCG pod5s/ > calls.bam
dorado basecaller sup,6mA pod5s/ > calls.bam               # all-context 6mA
# RNA004 direct RNA (cDNA CANNOT call mods - PCR erases the signal):
dorado basecaller [email protected],m6A_DRACH pod5s/ > rna_mods.bam

# FASTQ output and a per-read quality floor (relative filter, not a calibrated accuracy)
dorado basecaller sup pod5s/ --emit-fastq --min-qscore 10 > calls.fastq

# Duplex (needs raw POD5; cannot be recovered from simplex FASTQ); dx tag marks read types
dorado duplex sup pod5s/ > duplex.bam

# Demultiplex: basecall WITHOUT trimming, then demux (demux trims barcodes itself)
dorado basecaller sup pod5s/ --no-trim > calls.bam
dorado demux --kit-name SQK-NBD114-24 --output-dir demux/ calls.bam
dorado demux --kit-name SQK-NBD114-24 --barcode-both-ends --output-dir demux/ calls.bam  # stringent

# HERRO read correction for diploid/phased assembly (input FASTQ of HAC/SUP R10 reads >=10kb -> FASTA)
dorado download --model herro-v1
dorado correct reads.fastq > corrected.fasta

POD5 is ONT's default raw format (faster random access than FAST5). Convert FAST5 first:

pod5 convert fast5 raw/*.fast5 --output pod5s/    # FAST5 is legacy; basecalling it directly is slow
pod5 view pod5s/                                   # summary table (replaces deprecated `pod5 inspect reads`)
pod5 merge pod5s/*.pod5 --output merged.pod5

Per-Method Failure Modes

Methylation gone forever

Trigger: basecalling without a mod model, then wanting 5mC later. Mechanism: Remora infers mods from raw signal at basecall time; a plain BAM has only bases. Symptom: no MM/ML tags; modkit pileup returns nothing. Fix: re-basecall from POD5 with sup,5mCG_5hmCG; keep POD5 archives.

Barcodes land in unclassified

Trigger: default --trim all basecall, then a separate dorado demux. Mechanism: trimming removes the barcode before demux can read it. Symptom: most reads in unclassified.bam, low classification rate. Fix: basecall --no-trim, then demux (it trims barcodes itself).

Silent accuracy loss downstream

Trigger: polishing/calling with a medaka/Clair3 model that doesn't match the basecaller model+version. Mechanism: per-model neural weights expect a specific error profile. Symptom: no error, just quietly worse consensus/calls. Fix: propagate the basecaller model name; use medaka tools resolve_model --auto_model; pick the matching Clair3 model dir.

Duplex double-counting

Trigger: treating every read in a duplex BAM as an independent molecule. Mechanism: a simplex parent and its duplex offspring both appear. Symptom: inflated coverage/allele counts. Fix: the dx:i:-1 tag marks simplex parents of duplex reads - filter them when counting molecules (dx:i:1 = duplex, dx:i:0 = simplex-only).

Cohort batch effect

Trigger: runs basecalled with different model versions joined for analysis. Mechanism: version-specific identity/indel error profiles confound a technical batch with biology. Symptom: spurious between-run differences. Fix: re-basecall the whole cohort with one model version.

Quantitative Thresholds

| Threshold | Source | Rationale | |-----------|--------|-----------| | sup for any analysis | ONT model guidance | fast/hac error profiles contaminate variant/assembly/methylation calls | | R10.4.1 SUP modal accuracy ~Q20 (99%) | Sereika 2022 | dual-reader head fixes homopolymers; enables nanopore-only near-finished genomes | | Duplex read ~Q30; yield typically <10% of reads | community benchmarks | duplex is library-prep/loading-limited, not free accuracy | | A "Q20" base errs at ~Q12.5 empirically | Delahaye 2021 | nanopore qscores >Q10 are overconfident posteriors; use for relative filtering only | | HERRO input reads >=10 kbp, HAC/SUP R10 | Dorado correct docs | HERRO operates on 4096-bp chunks; shorter reads dropped | | --min-qscore 10 as a permissive QC floor | convention | Q10 ~ 90% nominal; a starting filter, not a hard rule |

Common Errors

| Error / symptom | Cause | Solution | |-----------------|-------|----------| | "Failed to determine sequencing chemistry from data" | R9/RNA002 or non-standard kit; bare tier can't auto-resolve | pass an explicit model path; for legacy chemistry use an archived model | | No MM/ML tags in BAM | basecalled without a mod model | re-basecall from POD5 with sup,5mCG_5hmCG | | Most reads unclassified after demux | trimmed before demux | basecall --no-trim, then demux | | --model sup errors | model is the positional arg, not a flag | dorado basecaller sup pod5s/ | | dorado correct reads.bam fails | input is FASTQ(.gz), output FASTA | dorado correct reads.fastq > corrected.fasta | | Out of GPU memory | batch too large for VRAM (sup is heaviest) | lower --batchsize; or drop to hac | | cDNA m6A calling returns nothing | PCR erased native modifications | use direct RNA (RNA004), not cDNA |

References

• Sereika M, Kirkegaard RH, Karst SM, et al. 2022. 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. Nat Methods 19:823-826.
• Stanojević D, Lin D, Nurk S, Florez de Sessions P, Šikić M. 2026. Telomere-to-telomere assembly using HERRO-corrected Nanopore simplex reads. Nature (online ahead of print). DOI 10.1038/s41586-026-10563-y.
• Wick RR, Judd LM, Holt KE. 2019. Performance of neural network basecalling tools for Oxford Nanopore sequencing. Genome Biol 20:129.
• Pagès-Gallego M, de Ridder J. 2023. Comprehensive benchmark and architectural analysis of deep learning models for nanopore sequencing basecalling. Genome Biol 24:71.
• Delahaye C, Nicolas J. 2021. Sequencing DNA with nanopores: troubles and biases. PLoS ONE 16(10):e0257521.
• Gamaarachchi H, Samarakoon H, et al. 2025. The enduring advantages of the SLOW5 file format for raw nanopore sequencing data. GigaScience giaf118.

Related Skills

• long-read-qc - Assess read length/quality and run health after basecalling
• nanopore-methylation - Pile up the MM/ML tags this skill must request at basecall time
• long-read-alignment - Map the reads; use -y to carry MM/ML tags through alignment
• medaka-polishing - Consensus model that must match this basecaller model+version
• clair3-variants - Variant model that must match this basecaller model+version
• genome-assembly/long-read-assembly - Assemble the reads (HERRO-corrected for diploid/T2T)
• genome-assembly/hifi-assembly - PacBio HiFi (basecalled on-instrument, not here)
• epitranscriptomics/m6anet-analysis - ONT direct-RNA m6A from signal
• workflows/longread-sv-pipeline - End-to-end basecall -> align -> SV call