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. dna_r10.4.1_e8.2_400bps_sup@v5.2.0) sets the entire error profile and must be propagated to every downstream tool. Pin it.
• Modified-base models carry a SECOND version (..._sup@v5.0.0_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).

Model Naming Scheme (load-bearing)

Format {analyte}_{pore}_{chemistry}_{speed}@v{ver} + optional mod suffix, e.g. dna_r10.4.1_e8.2_400bps_sup@v5.2.0_5mCG_5hmCG@v3.

Passing the bare tier (sup) lets Dorado auto-detect chemistry from POD5 metadata and fetch the matching latest model; pin a version (sup@v5.2.0) 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

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 dna_r10.4.1_e8.2_400bps_sup@v5.2.0 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 rna004_130bps_sup@v5.1.0,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).