FASTQ Quality Control & Trimming Decisions

Run quality control on raw sequencing reads, interpret the report, and make an evidence-based decision about whether to trim — using real local command-line tools (FastQC, MultiQC, fastp, Cutadapt, seqkit).

Honesty contract (read first)

This skill drives real binaries. It must never fabricate QC numbers.

1. Preflight before anything. Check whether the required tools are on

PATH. If a required tool is missing, emit the install plan and STOP. Do not estimate, guess, or describe hypothetical QC results.

1. Never auto-trim. Trimming is a decision. QC-only is the default.

Only trim after inspecting adapter content / per-base quality, and only when the user has confirmed --mode trim.

1. Never overwrite raw FASTQs. All outputs go to a separate --workdir.

The input directory is read-only. Trimmed reads are written as NEW files.

1. If you cannot run, say so. "FastQC is not installed; here is the

install plan" is the correct answer — not a made-up PASS/FAIL table.

When to use vs. not

Use this skill when the user wants to:

• Run FastQC / fastp QC on one or more FASTQ (.fastq, .fq, .gz) files
• Interpret a FastQC report (per-base quality, adapter content, etc.)
• Decide whether adapter or quality trimming is needed before downstream work
• Summarize many samples into one MultiQC report
• Count reads, get length/GC stats, or subsample with seqkit
• Trim adapters/low-quality bases with fastp or Cutadapt (explicitly)

Do NOT use this skill for (route elsewhere):

• Differential expression / DEG / fold-change analysis -> tooluniverse-rnaseq-deseq2
• Read alignment, coverage depth, samtools, BWA -> tooluniverse-sequence-analysis
• Variant calling, VCF, VAF, mutation analysis -> tooluniverse-variant-analysis
• Single-cell / scRNA QC (per-cell metrics, scanpy) -> tooluniverse-single-cell

Essential inputs to confirm

Before running, confirm with the user (ask if unstated):

1. FASTQ paths — exact path(s). One file = single-end; an R1+R2 pair =

paired-end (e.g. *_R1.fastq.gz / *_R2.fastq.gz).

1. QC-only or trim? Default is QC-only. Only trim on explicit request.
2. Known adapters / primers? Standard Illumina adapters are auto-detected

by fastp; amplicon/primer sequences usually need explicit Cutadapt removal.

1. Organism — only needed if a contamination / over-representation screen

is requested (needs a reference; see Limitations).

1. Output directory — a --workdir SEPARATE from the input folder.
2. Read provenance — are these raw, already-trimmed, or UMI-tagged reads?

Already-trimmed reads should NOT be trimmed again; UMIs must be handled before trimming or you corrupt the UMI.

Preflight (do this first, every time)

The bundled script preflights for you, but the decision logic is:

import shutil
for tool in ("fastqc", "fastp", "seqkit"):
    print(tool, shutil.which(tool) or "MISSING")

command -v fastqc / shutil.which("fastqc") returning nothing means the tool is absent. If a required tool (FastQC for QC; FastQC+fastp for trim) is missing, emit:

mamba install -c bioconda -c conda-forge fastqc fastp seqkit multiqc
#   or
conda install -c bioconda -c conda-forge fastqc fastp seqkit multiqc

and stop. Do not proceed to fabricate output.

Tool roles

Rule of thumb: FastQC to diagnose, fastp to fix general adapter/quality, Cutadapt to fix a known primer/adapter precisely, seqkit to count/stat.

Bundled orchestration script

scripts/run_fastq_qc.py does the preflight + run-if-available + plan-if-missing flow, with workspace isolation built in.

# QC only (default) — never modifies reads
python scripts/run_fastq_qc.py \
    --fastq reads/sample_R1.fastq.gz reads/sample_R2.fastq.gz \
    --workdir /tmp/fastq_qc_run

# QC + trim (explicit) — fastp writes NEW trimmed files into --workdir
python scripts/run_fastq_qc.py \
    --fastq reads/sample_R1.fastq.gz reads/sample_R2.fastq.gz \
    --workdir /tmp/fastq_qc_run \
    --mode trim

Behavior:

• Preflights FastQC (+ fastp in trim mode) and seqkit. If a required

tool is missing it prints the install plan and exits 0 — no fabricated QC.

• Runs FastQC (always) + seqkit stats (if present) into --workdir.
• In --mode trim, runs fastp writing *.trimmed.fastq.gz into

--workdir/trimmed/ — raw inputs are never touched.

• Refuses to run if --workdir equals an input directory (overwrite guard).

For a project-level summary after FastQC, run MultiQC over the workdir:

multiqc /tmp/fastq_qc_run -o /tmp/fastq_qc_run/multiqc

INTERPRETATION — FastQC module -> meaning -> action

This table is the core value-add. Map each FastQC module to what PASS/WARN/FAIL means and what to actually do. (See references/fastqc_interpretation.md for the long form with thresholds and worked cases.)

Decision summary for "do I need to trim?"

• Adapter content FAIL/WARN with a 3' adapter ramp -> yes, adapter-trim.
• Per-base quality FAIL at the 3' tail -> yes, quality-trim that tail.
• Overrepresented = adapter/primer-dimer -> yes, trim; overrepresented =

biology (rRNA, abundant transcript) -> no, proceed.

• High duplication / GC anomaly / N-spike -> investigate, not a trimming fix.
• Everything PASS -> proceed without trimming.

Workflow

1. Confirm inputs (paths, pairing, mode, adapters, provenance).
2. Preflight tools. If missing -> install plan, STOP.
3. Run QC (--mode qc): FastQC + seqkit -> read the report.
4. Interpret each flagged module with the table above.
5. Decide trim vs investigate vs proceed. State the decision and why.
6. (If trimming chosen) run --mode trim (fastp) or Cutadapt for precise

primer removal; re-run FastQC on the trimmed output to confirm the fix.

1. (Optional) MultiQC for a multi-sample summary.
2. Report: per-module status, the trim decision + rationale, and the exact

commands run. Never report numbers a tool did not actually produce.

Limitations (honest)

• Requires local binaries. FastQC/fastp/seqkit/Cutadapt/MultiQC must be

installed (bioconda). This is not a cloud service; with no tools installed the skill can only emit an install plan, not QC results.

• Large files. Whole-lane FASTQs can be many GB; FastQC/fastp are

single-pass and memory-light but still I/O-bound. Use seqkit sample to subsample for a quick look on huge files.

• Contamination / cross-species screening is NOT included by default. GC

anomalies and "is this the right organism" need a reference index (e.g. FastQ Screen + bowtie2 indexes, or Kraken2) — extra setup beyond this skill's bundled tools.

• No deduplication of raw reads. PCR-duplicate removal is an alignment-stage

decision (Picard/samtools markdup); FastQC duplication is diagnostic only.

• UMI-aware trimming needs UMI extraction first (umi_tools); naive trimming

corrupts UMIs.

Completeness checklist

• [ ] Inputs confirmed (paths, single/paired, raw vs trimmed, adapters)
• [ ] Tools preflighted; install plan emitted if any required tool missing
• [ ] QC run with outputs in a workdir separate from inputs (raw preserved)
• [ ] Each flagged FastQC module interpreted (meaning + action)
• [ ] Explicit trim/investigate/proceed decision with rationale
• [ ] Trimming (if done) was opt-in, wrote new files, raw FASTQs untouched
• [ ] Post-trim FastQC re-run to confirm the fix (if trimmed)
• [ ] No QC numbers reported that a tool did not actually produce

References

• references/fastqc_interpretation.md — full module-by-module thresholds + cases
• references/tools_and_install.md — install commands, tool flags, command recipes
• references/trimming_decisions.md — when/how to trim (fastp vs Cutadapt), pitfalls