Version Compatibility

Reference examples tested with: medaka 2.2+, minimap2 2.28+, samtools 1.19+.

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 medaka MODEL must match the basecaller (pore + chemistry + speed + mode + version), e.g. r1041e82400bpssupv5.2.0. A mismatch silently degrades output. Prefer auto-detection from the BAM; verify with medaka tools list_models.
• Default models advance with each release (consensus ...supv5.2.0, variant ...supvariantv5.0.0 at time of writing); confirm with medaka tools listmodels.
• medaka v2 renamed subcommands (consensus->inference, stitch->sequence, variant->vcf) and moved the backend to PyTorch; v1 tutorials fail.

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

Medaka Polishing

"Polish my Nanopore assembly" -> Run one medaka consensus pass directly on the assembler output, with the model that matches the basecaller - because a mismatched model silently makes the consensus worse.

• CLI: medaka_consensus -i reads.fq -d draft.fa -o out/ -t 8 (model auto-detected from the basecaller annotation)

medaka is an Oxford Nanopore tool. For PacBio (HiFi/CLR) it is the wrong tool entirely - route to genome-assembly/assembly-polishing.

The Single Most Important Modern Insight -- A Mismatched Model Silently Degrades; HiFi Must Never Be Fed to Medaka; Prove It on Held-Out Data

medaka is a basecaller-model-specific neural consensus net trained on one exact stack (pore + motor enzyme + speed + basecaller mode + basecaller version). Three consequences:

1. The model must match the basecaller, and a mismatch fails silently. Fed reads from a different stack, medaka applies corrections calibrated for an error fingerprint that is not there and misses the real one - the consensus gets WORSE, but medaka exits 0, writes a FASTA, and prints no warning. This is the #1 ONT-polishing footgun, sprung by ordinary acts (re-basecalling with newer Dorado, copying a 2020 model name, polishing a public assembly with the default). Prefer auto-detection (medaka tools resolvemodel --automodel consensus reads.bam); treat a stale model name as a reason to re-basecall, not to proceed.
2. HiFi (and CLR) must never be fed to medaka. It has no PacBio models; an ONT error-model net "corrects" HiFi toward errors HiFi does not make, and HiFi is already QV40+. If the reads are PacBio, medaka is simply wrong -> genome-assembly/assembly-polishing.
3. Success is only real on held-out data. medaka maximizes agreement between the consensus and its input pileup, so grading it on those same reads is circular and always looks good. medaka's "N changes" is a risk signal, not a success signal. Measure with reference-free Merqury QV before vs after on held-out / different-platform k-mers (design deferred to genome-assembly/assembly-polishing).

What medaka Is For (three modes, same model rule)

Decision Tree by Scenario

medaka_consensus Mechanics

The wrapper runs three steps: align (mini_align, a thin veil over minimap2 -x map-ont), infer (medaka inference, the neural net over the pileup), and stitch (medaka sequence, regions -> consensus FASTA).

# Canonical modern usage - model auto-detected from the basecaller annotation in the reads
medaka_consensus -i reads.fastq -d draft.fa -o medaka_out/ -t 8
# medaka_out/consensus.fasta is the polished assembly

# Native bacterial isolate: use the methylation-aware bacterial model
medaka_consensus -i reads.fastq -d draft.fa -o medaka_out/ -t 8 --bacteria

# Resolve / list models (do this when auto-detection cannot pick)
medaka tools resolve_model --auto_model consensus reads.bam
medaka tools list_models

medaka runs directly on the assembler (Flye) output as a SINGLE pass - do NOT pre-run Racon (contemporary models are trained on raw assembler output; v2 removed the bundled racon wrapper) and do NOT run medaka twice (iteration was racon's role; a second pass risks flipping correct bases).

Haploid variant calling (v2 names)

medaka_variant emits a VCF only (no consensus FASTA); apply it to the reference with bcftools consensus to get a haploid consensus sequence.

# Wrapper form (renamed from medaka_haploid_variant in v2) - haploid samples only
medaka_variant -i reads.fastq -r reference.fa -o variant_out/

# Manual form - note v2 subcommand names and the hdf -> ref -> out argument order
minimap2 -ax map-ont reference.fa reads.fq | samtools sort -o aln.bam && samtools index aln.bam
medaka inference aln.bam probs.hdf --model r1041_e82_400bps_sup_variant_v5.0.0
medaka vcf probs.hdf reference.fa variants.vcf

# Optional: turn the VCF into a haploid consensus FASTA
bgzip variants.vcf && tabix -p vcf variants.vcf.gz
bcftools consensus -f reference.fa variants.vcf.gz > consensus.fasta

Per-Method Failure Modes

Silent model mismatch

Trigger: running medaka with a model that does not match the basecaller chemistry/version. Mechanism: the net corrects toward the wrong error fingerprint. Symptom: lower held-out QV; medaka exits 0 with no warning. Fix: auto-detect from the BAM; if forced to pick, derive from the actual basecaller and confirm in list_models; treat a stale name as a reason to re-basecall.

HiFi fed to medaka

Trigger: polishing a PacBio assembly with medaka. Mechanism: ONT-only error model, no PacBio support, on already-QV40+ data. Symptom: degraded/homogenized consensus. Fix: do not; route to genome-assembly/assembly-polishing.

Racon-first off-distribution

Trigger: running Racon before medaka out of habit. Mechanism: contemporary models are trained on raw assembler output; racon-polished input is off the training distribution. Symptom: no gain or mild harm. Fix: run medaka directly on the Flye output; one pass.

Missing plasmid poisons the chromosome

Trigger: an assembly missing a small replicon (~80% identical to a chromosomal region). Mechanism: the absent plasmid's reads misalign onto the chromosome, and medaka "corrects" toward that spurious evidence. Symptom: clustered changes that introduce real errors. Fix: make the assembly structurally complete first; inspect medaka's changes for clustering (clustered = mapping artifact, not scattered homopolymer fixes).

Validating on the polishing reads

Trigger: judging the polish by medaka's change count or by re-mapping the same reads. Mechanism: medaka optimizes agreement with its input pileup. Symptom: "improvement" that is circular. Fix: reference-free Merqury QV before vs after on held-out / different-platform k-mers.

Quantitative Thresholds

Common Errors

References

• medaka. Oxford Nanopore Technologies. https://github.com/nanoporetech/medaka (no journal paper; cite the repository).
• Zheng Z, Li S, Su J, Leung AW, Lam TW, Luo R. 2022. Symphonizing pileup and full-alignment for deep learning-based long-read variant calling (Clair3). Nat Comput Sci 2:797-803.
• Vaser R, Sović I, Nagarajan N, Šikić M. 2017. Fast and accurate de novo genome assembly from long uncorrected reads (Racon). Genome Res 27:737-746.
• Wick RR, Judd LM, Holt KE. 2023. Assembling the perfect bacterial genome using Oxford Nanopore and Illumina sequencing. PLoS Comput Biol 19(3):e1010905.
• Rhie A, Walenz BP, Koren S, Phillippy AM. 2020. Merqury: reference-free quality, completeness, and phasing assessment for genome assemblies. Genome Biol 21:245.
• Wick RR. 2024. Medaka v2: progress and potential pitfalls. https://rrwick.github.io/2024/10/17/medaka-v2.html (blog; source of the missing-plasmid footgun).

Related Skills

• basecalling - The basecaller model+version medaka's model must match
• clair3-variants - ONT small-variant (diploid/germline) calling; medaka diploid is deprecated
• long-read-alignment - minimap2 map-ont, the alignment medaka's mini_align wraps
• genome-assembly/assembly-polishing - Polishing strategy authority (HiFi doctrine, hybrid tiers, Merqury QV design)
• genome-assembly/long-read-assembly - Produces the Flye draft medaka polishes
• genome-assembly/assembly-qc - Merqury QV / BUSCO before-vs-after measurement