Imported from openai/plugins (plugins/ngs-analysis/skills/ngs-bulk-rnaseq-differential-expression/SKILL.md). Install upstream with npx skills add openai/plugins --skill ngs-bulk-rnaseq-differential-expression. Copyright stays with the author.

Bulk RNA-seq Differential Expression

Use this skill when the user has raw counts or a count-generation output and wants differential expression, contrasts, QC plots, or ranked gene tables.

Essential Inputs

Confirm:

• raw count matrix path and sample metadata path
• gene ID type and annotation mapping requirement
• biological conditions, replicates, batch variables, donor pairing, covariates, and exclusions
• exact contrasts and baseline levels
• preferred statistical framework: DESeq2, edgeR, limma-voom, or existing lab standard
• output needs: normalized counts, PCA, sample distance, volcano plots, heatmaps, ranked tables, GSEA-ready lists

Preconditions

Do not start differential expression until:

• raw counts are preserved
• each requested contrast has enough biological replication
• sample metadata row names match count matrix columns
• batch/covariate choices are explicit
• exploratory PCA/sample-distance plots do not reveal obvious swaps or failed libraries

Route

For most count matrices, use DESeq2 or edgeR. Use limma-voom when the study design or lab standard favors it. Keep the analysis in R when using Bioconductor unless the user specifically asks for a Python-only workflow.

The plugin-owned local runner is:

python plugins/ngs-analysis/scripts/run_bulk_rnaseq_de.py \
  --count-matrix count_matrix.tsv \
  --sample-metadata sample_metadata.tsv \
  --contrasts contrasts.tsv \
  --execute

Use --method auto unless the user or lab standard specifies DESeq2, edgeR, or limma_log2. Auto mode uses DESeq2 when integer-like counts and the package are available, falls back to edgeR for integer-like counts, and uses limma_log2 for non-integer expression matrices.

Use --input-mode to declare whether the matrix is raw_counts, normalized_expression, or log_expression. When --input-mode auto is used, the runner infers the mode and records a warning if normalization is skipped because the matrix is already transformed.

Preflight command:

python plugins/ngs-analysis/scripts/ngs_preflight.py --pipeline bulk_rnaseq_differential_expression --emit-install-plan

Decision Points

• Never compare groups without stating the design formula and contrast.
• Treat batch correction in modeling separately from visual batch removal.
• Do not filter genes using post-hoc knowledge of the contrast.
• For paired or repeated-measures designs, model subject/donor explicitly.
• Report genes with effect size, uncertainty, adjusted p-value, and filtering status.

Outputs

Produce:

• design formula and contrast manifest
• QC plots: library size, detected genes, PCA/sample distance, mean-variance trend, and outlier review
• input-mode-aware matrix exports plus the modeling/log-scale matrix used for DE
• differential-expression tables per contrast
• explicit .not_tested.tsv stubs for contrasts blocked by insufficient replication or confounding
• auto-launched localhost Marimo review app recorded in notebooks/marimo_server.json
• caveats for small n, confounded designs, failed samples, or batch variables that cannot be estimated
• standard run envelope: run_manifest.json, config.json, validation/, logs/, versions/, visualizations/, notebooks/, artifact_index.json, and summary.md