bio-reads-qc-mapping

Ingest, QC, and map reads with reproducible outputs. Use for raw read processing and coverage stats.

SKILL.md

GitHub

Bio Reads QC Mapping

Validate a read sample sheet, run read QC and trimming, and map reads to an optional reference with a restartable driver.

Instructions

Tool guides and versions: docs/README.md.

  1. Validate sample_sheet.tsv against schemas/sample-sheet.schema.json and inspect the plan before running anything:

    uv run --script skills/bio-reads-qc-mapping/scripts/run_reads_qc_mapping.py \
      sample_sheet.tsv --out results/bio-reads-qc-mapping --threads 4
    # Inspect run_manifest.json, then execute the same plan on a compute node:
    uv run --script skills/bio-reads-qc-mapping/scripts/run_reads_qc_mapping.py \
      sample_sheet.tsv --out results/bio-reads-qc-mapping --threads 16 --execute
    

    read_type must be paired_short, single_short, or long. Mapping runs only for rows with a non-empty reference; a missing reference is not a mapping failure. The driver reuses a stage only when its declared outputs are non-empty and the stage's .done marker exists. Run --execute through the scheduler (sbatch) for real data, not on a login node.

  2. Know what the driver runs, and adapt the plan when the defaults do not fit:

    • Short reads: fastp v1.3.3+ with default adapter and quality trimming. Use BBDuk (BBTools container) instead when you need k-mer contaminant, spike-in, or host removal (docs/bbduk.md).
    • Long reads: filtlong with no thresholds, which passes every read through. Add --min_length, --keep_percent, or --target_bases for the project (docs/filtlong.md).
    • Short-read mapping: bwa-mem2 mem. Build the index first with bwa-mem2 index reference.fasta; the driver does not.
    • Long-read mapping: minimap2 -ax map-ont. For PacBio HiFi use map-hifi, and for PacBio CLR use map-pb.
    • The driver writes SAM. Sort and index it (samtools sort, samtools index), then compute per-reference coverage with samtools coverage or CoverM.
  3. For long reads, use basecaller-aware QC first. For ONT, prefer Dorado trimming during basecalling or demultiplexing when starting from POD5 or BAM. For FASTQ-only input, use chopper for quality, length, and end trimming, or filtlong v0.3.1 when selecting reads for assembly (v0.3.0 renamed the long options to --short_1 / --short_2). Use Pychopper for full-length cDNA. Use Porechop_ABI only as a documented fallback for adapter discovery, and record why.

    • For very large ONT FASTQ inputs, do not spend the first full read pass on gzip -t or raw seqkit stats unless asked. Record raw stat metadata, optionally check a small sample, make the first full pass the filtering step, then run seqkit stats on the outputs.
    • Pychopper can write plain FASTQ even when the output path ends in .gz. Give its outputs plain .fastq names unless you compress them yourself, and check gzip magic bytes before running gzip -t. Rename mislabeled legacy outputs to .fastq before resuming.
    • Pychopper report plotting can fail after the reads are written. If the classified, unclassified, rescued, and read-stats outputs exist and are non-empty, resume downstream from them instead of rerunning Pychopper.
  4. Faster mappers, when hardware allows:

    • Short reads on a GPU node: NVIDIA Parabricks fq2bam (GPU BWA-MEM with sorting and duplicate marking).
    • Long reads on a GPU node: Parabricks minimap2 (presets map-ont, map-hifi, lr:hq).
    • The CPU and GPU forks mm2-fast and mm2-gb are built on minimap2 v2.24, not current v2.30+. Use them only when the older base version is acceptable.
  5. Record each tool, version, thread count, and any GPU device in the run log and tasks/METHODS.md.

Input Requirements

Prerequisites:

  • Tools declared in the project's pinned pixi environment; BBTools runs from the bryce911/bbtools container. See docs/README.md.
  • A bwa-mem2 index next to each short-read reference. Inputs:
  • sample_sheet.tsv with exactly the columns sample_id, read_type, read1, read2, reference; relative paths resolve against the sheet's directory.
  • reads/.fastq.gz or reads/.fastq
  • reference.fasta (optional)

Output

  • results/bio-reads-qc-mapping/run_manifest.json (validated rows, planned commands, and per-stage status after --execute)
  • results/bio-reads-qc-mapping/<sample_id>/ with trimmed reads (reads.fastq, or reads_R1.fastq and reads_R2.fastq), fastp.json, fastp.html, mapped.sam, and the qc.done and mapping.done markers
  • results/bio-reads-qc-mapping/mapping_stats.tsv (mapping_status is planned before --execute, then completed or reused, and not_requested for rows with no reference)
  • stdout: the last line is one JSON envelope {ok, skill, out, manifest, warnings} (driver stdout contract in AGENTS.md)
  • Sample sheet contract: schemas/sample-sheet.schema.json

Quality Gates

  • The sample sheet validates, and the plan covers every row exactly once.
  • Post-QC read counts are plausible against the raw counts (fastp.json or seqkit stats).
  • Mapping rate meets project thresholds, applied only to rows that supplied a reference.
  • On execution failure, preserve logs and report the failed command; retry only after diagnosing the cause and recording the changed parameters. Report unmet biological thresholds as results; never tune parameters solely to pass a gate.
  • Long-read QC records where trimming happened: basecaller or demultiplexer, chopper, filtlong, Pychopper, or a documented Porechop_ABI fallback.
  • Huge ONT inputs have raw size and mtime recorded and no redundant full-file raw preflight.
  • File type is checked by content, not suffix, before tools that expect gzip.
  • Before accepting reused outputs downstream, run a light content check (seqkit stats, FASTQ header sniff, or gzip magic).

Examples

The runnable fixture at fixtures/sample_sheet.tsv covers paired-end, single-end, and long reads, with mapping requested for two of the three rows.

Troubleshooting

Issue: Pychopper failed during report or stat plotting but output FASTQs exist Solution: Treat it as a recoverable post-processing failure. Confirm the classified FASTQ is non-empty and readable, fix any misleading .gz suffix, run seqkit stats, and resume downstream from the existing Pychopper outputs.

Issue: bwa-mem2 mem fails immediately on a new reference Solution: The index is missing. Run bwa-mem2 index reference.fasta once, then rerun with --execute; completed QC stages are reused.