Experimental design and QC

The statistics downstream are only as good as the design and the QC gates. Decide design before sequencing; apply QC before, during, and after quantification. This is what makes a bulk RNA-seq result defensible.

Experimental design

Replication

• Use biological replicates (independent samples), not technical (same library re-sequenced). Technical replicates measure machine noise, not biological variability, and don't license generalization.
• ≥3 per group is a planning starting point, not a power guarantee; choose sample size from expected dispersion, effect sizes and the intended contrast. Two per group can be fitted but give fragile inference. The bundled validator uses a conservative policy of at least two per group and recommends three; this is not a universal DESeq2 mathematical requirement.
• Once adequate depth is reached, biological replication often adds more power than deeper sequencing. Evaluate the tradeoff for the assay and target genes.

Depth, length, layout

• ~20–30M mapped reads/sample is a rough planning example for many mammalian gene-level studies, not a QC cutoff. Push higher (50M+) for lowly expressed genes, novel transcripts, or isoform-level work.
• Paired-end and longer reads help mapping/isoforms but aren't required for gene-level DE; single-end is fine if that's what you have.
• Keep layout, read length, kit, and depth consistent across all samples in a comparison.

Avoid confounding (the design killer)

• A batch is anything technical that varies across samples: processing day, sequencing lane/flowcell, kit lot, operator, RNA extraction round.
• If a batch is perfectly aligned with your condition (all treated processed Monday, all controls Tuesday), the biological effect is mathematically unrecoverable. No analysis fixes this.
• Defenses: randomize sample-to-batch assignment, and balance so every batch contains every condition. Record all batch variables in the metadata.

Design formulas (hand to PyDESeq2)

• A conventional additive design is: ~batch + condition.
• Continuous covariate: ~age + condition (ensure it's numeric).
• Interaction (does the treatment effect differ by genotype?): ~genotype + condition + genotype:condition.
• The design matrix must be full rank with positive residual degrees of freedom. A crosstab is an initial diagnostic, not proof of estimability; check the complete encoded matrix including subjects, interactions and covariates. Do not discard a confounded batch and then claim an isolated treatment effect. Set reference levels and an explicit contrast: column order alone does not specify the scientific comparison.

QC gates

The ranges below are investigation triggers, not universal pass/fail limits. Interpret them against organism, library protocol, RNA integrity, reference and study-wide distributions.

Raw-read QC (FastQC / MultiQC)

• Per-base quality — bulk of bases ≥ Q30; some drop at read ends is normal (trimming/soft-clipping handles it).
• Adapter content — flagged adapters → trim (Path B step 2; Path A does it automatically).
• Over-represented sequences — adapters, rRNA, or highly expressed transcripts. Persistent rRNA suggests poor depletion.
• GC content — a bimodal/odd distribution can indicate contamination.
• Sequence duplication — high duplication is expected in RNA-seq (highly expressed genes); see below.

Alignment / quantification QC

• STAR uniquely-mapped % — typically >70–80% for a good library/reference. Low → wrong/old reference, contamination, or degraded RNA.
• Salmon mapping rate (logs/salmon_quant.log) — usually >70%. Low → investigate reference mismatch, contamination, read quality and library type. Decoys improve assignment specificity and can reduce transcript mapping; their absence is not a simple explanation for a low rate.
• featureCounts assigned % — low "assigned" with high "unassigned_NoFeatures" often means wrong strandedness (-s).
• rRNA fraction — high rRNA wastes reads; note it, and consider --remove_ribo_rna on Path A.
• Verify strandedness matches across tools (see upstream-manual.md).

Don't deduplicate for standard DE

PCR/optical duplicates look alarming but in RNA-seq mostly reflect genuine high expression. Standard gene-level DE (DESeq2) does not remove duplicates. Only consider dedup with UMIs (use the UMI, not coordinate dedup).

Post-quantification QC (before trusting DE)

Always do this on the counts, ideally on variance-stabilized/log values:

• PCA — do biological replicates cluster? Does the main axis separate your condition, or a batch? A technical batch dominating PC1 warrants checking design and data quality; include it only when estimable and scientifically justified. An obvious outlier may be a swap/failure.
• Sample-distance heatmap / hierarchical clustering — confirms grouping and exposes mislabeled or swapped samples.
• If a batch clearly structures the data, add it to the design (~batch + condition); if it's unknown, consider surrogate-variable / RUV approaches (out of scope here — note it).

After DE: p-value histogram

• Under a calibrated continuous null, null p-values are uniform; real signal may add a peak near zero. Discrete low-count tests and filtering can change the appearance, so interpret the histogram with model diagnostics.
• A peak near one or a U-shape is not by itself proof of a broken design. Inspect count distributions, outliers, filtering and covariates. Do not tune the model or exclude samples simply to obtain the desired histogram or more discoveries.

Quick gate checklist

[ ] >=3 biological replicates per group
[ ] batch recorded and NOT confounded with condition
[ ] raw FastQC reviewed; adapters trimmed
[ ] mapping/assignment rate acceptable; strandedness verified
[ ] PCA + sample-distance heatmap inspected; outliers/swaps resolved
[ ] design full-rank, residual degrees of freedom positive, contrast explicit
[ ] p-value histogram sane after DE
[ ] versions pinned (pipeline -r, tools, genome+annotation release)

Official statistical reference: DESeq2 vignette, including design rank, contrasts and transformed-count QC.