Bio Reporting Automated Qc Reports

Version Compatibility

Reference examples tested with: MultiQC 1.21+ (Plotly era), FastQC 0.12+, STAR 2.7.11+, Subread 2.0+, salmon 1.10+, samtools 1.19+, Picard 3.1+, fastp 0.23+

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

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

Config keys and defaults move between MultiQC releases (the plotting backend changed from HighCharts to Plotly at 1.20; flat-plot and AI thresholds shifted). When a default matters, confirm it against the installed version: python -c "import multiqc; print(multiqc.__version__)" and check that version's config_defaults.yaml.

If code throws an error, run multiqc --help and adapt flags to the installed version rather than retrying.

Automated QC Reports with MultiQC

"Aggregate QC results into one report" -> Walk a directory of tool outputs, parse the metrics those tools already wrote, and render one interactive HTML report plus a parseable multiqc_data/ directory.

• CLI: multiqc <dir> (scans for recognized tool outputs)

The Load-Bearing Idea: MultiQC Aggregates, It Does Not Measure

MultiQC computes nothing. It SCRAPES the log/metrics files that FastQC, STAR, Picard, salmon, bcftools, etc. already wrote, re-tabulates those numbers, and renders them. Every value in a report traces back to an upstream tool's output file. Four consequences that drive every real decision below:

• The report is a triage SNAPSHOT, not a pass/fail gate. There is no "MultiQC said FAIL -> stop the pipeline." MultiQC has no fail-on-threshold and exits 0 on bad QC. Green/amber/red is either the upstream tool's own status (FastQC writes PASS/WARN/FAIL; MultiQC just displays it) or a threshold a human configured. Gating is a SEPARATE step (see From Report to Gate).
• Garbage upstream becomes a clean-looking report. Run a tool with the wrong strandedness, wrong reference, or dedup on amplicon data, and MultiQC faithfully aggregates the wrong numbers into a polished HTML. Polish is not evidence of correctness.
• An empty report means "nothing matched," not "QC passed." Point MultiQC at the wrong directory or over-filter modules and it emits a near-empty report with a log warning and exit 0. Always check the sample count in the header against the roster expected.
• It is only as current as its parsers. Each module is a hand-written parser keyed to a specific log format. An upstream version bump that changes a header can silently drop a file or mis-map a column.

Basic Usage

multiqc results/ -o qc_report/            # scan results/, write qc_report/multiqc_report.html
multiqc results/ -n project_qc -o qc/     # custom report name
multiqc results/ -m fastqc -m star        # ONLY these modules (see scoping below)
multiqc results/ -c multiqc_config.yaml   # reproducible config-driven report

Supported Tools

MultiQC ships parsers for 100+ tools. Common assay groupings:

Module Detection Is Regex - Scope It

Detection runs off search_patterns.yaml: each module declares a filename glob/regex (fn/fn_re) and/or a file-content match (contents/contents_re, bounded by num_lines). Loose patterns (*.txt, *.log, *.json) in a messy directory cause FALSE module matches and PHANTOM samples - a file that is not really that tool's output gets parsed as one. A single file can also satisfy two modules.

Scope explicitly rather than trusting auto-detection across thousands of samples:

multiqc results/ --ignore "*_tmp/" --ignore "work/"   # drop paths from the search
multiqc results/ -m fastqc -m star -m salmon          # run ONLY named modules
multiqc results/ -e snippy -e custom_content          # run all EXCEPT named modules

Tighten an over-loose pattern by overriding sp: in the config (sp: {mytool: {fn: 'real_name_*.txt'}}). Production configs pin sp: and module_order instead of relying on detection.

Sample Names Are Derived, Not Declared

Sample names are NOT read from a manifest. MultiQC derives each name from the matched filename (or a sample column inside the file), then "cleans" it by trimming a ~100-entry default list of extensions (fn_clean_exts: .gz, .fastq, .bam, _fastqc, ...). This is how sampleA_R1.fastq.gz, sampleA.sorted.bam, and sampleA.salmon/ all collapse to one sampleA row gathering read, alignment, and quant metrics.

The same mechanism is the #1 large-cohort bug:

• Merge - two genuinely different inputs clean to the same name (two lanes both reduce to sampleA) and silently overwrite each other's metrics.
• Split - one sample appears as several rows because different tools cleaned its name differently (one kept _L001, another stripped it).

multiqc_data/multiqc_sources.txt maps every parsed file to the sample name it produced - read it first when diagnosing duplicate/missing rows. Controls:

General Statistics and Conditional Formatting Are Configured, Not Authoritative

The General Statistics table is one row per sample with columns each module contributes. Cell colors come from table_cond_formatting_rules (numeric gt/lt/eq/ge/le, string s_eq/s_contains/s_ne). A red ">10% duplication" cell is red because someone wrote that rule (or because a module ships a built-in default rule), not because biology says 10% is bad. Treat formatting as a configured convenience; absence of red is not a pass, and presence of red is not a biological verdict. Column visibility/order/naming are config too (table_columns_visible, table_columns_placement, table_columns_name).

Large Cohorts: MultiQC Downgrades Automatically

To keep the single HTML openable, MultiQC silently changes rendering as series counts grow. The exact thresholds have moved across versions - verify against the installed config_defaults.yaml - but the behaviors are:

Force a mode for reproducible visuals across cohort sizes: --flat / --interactive (config plots_force_flat / plots_force_interactive). At tens of thousands of samples, also scope with -m/--ignore or split into per-batch reports - MultiQC holds all parsed data in memory before rendering.