RNA-seq Differential Expression Analysis (DESeq2)

PRIMARY SCRIPTS — use these FIRST before writing custom code

The four scripts below are deterministic, audited wrappers that handle the ambiguity in DESeq2 / correlation / PCA / ANOVA questions by emitting EVERY common interpretation in one call. Reading their output and matching the variant the published notebook used is more reliable than re-deriving the answer from scratch.

All four scripts honor workspace isolation: they ONLY write to --workdir (or /tmp/... by default). They never touch the input data folder. Always pass --workdir /tmp/<run-name> when you need intermediate files.

scripts/r_deseq2_wrapper.py — R DESeq2, multi-contrast Venn, per-gene LFC

Runs R DESeq2 (NOT pydeseq2) with full notebook-style controls: sample exclusion, metadata subsetting, low-row-sum filtering, LFC shrinkage (apeglm/ashr/normal), and an arbitrary number of contrasts in a single fit. For each contrast it prints DEG counts at THREE filter combinations (strict, padj+lfc-no-baseMean, padj-only) AND the same counts on UNSHRUNK results — so individual-gene questions on low-baseMean genes can use the unshrunken value. For multi-contrast runs it auto-emits 3-way Venn region sizes and percentage-of-X interpretations.

# Single-factor sex DE on a CD4/CD8 subset, with FAM138A LFC
python scripts/r_deseq2_wrapper.py \
    --counts <data-folder>/counts.csv \
    --metadata <data-folder>/meta.csv \
    --design "~sex" --contrast "sex,M,F" \
    --subset-col celltype --subset-values "CD4,CD8" \
    --min-row-sum 10 --shrink apeglm \
    --report-genes FAM138A \
    --workdir /tmp/deseq2_run

Output highlights (parseable):

# CONTRAST sex_M_vs_F: n=37496 n_tested=26591
# SIG_sex_M_vs_F_unshrunk_strict (padj<0.05 AND |LFC|>0.5 AND baseMean>10): n=...
# SIG_sex_M_vs_F_shrunk_padjlfc (padj<0.05 AND |LFC|>0.5, NO baseMean): n=...
# GENE FAM138A [sex_M_vs_F]: baseMean=... unshrunkLFC=... shrunkLFC=... padj=...

For a multi-strain Venn run with notebook-style outlier exclusion:

python scripts/r_deseq2_wrapper.py \
    --counts .../raw_counts.csv \
    --metadata .../experiment_metadata.csv \
    --design "~Replicate + Strain + Media" \
    --multi-contrast "Strain,97,1;Strain,98,1;Strain,99,1" \
    --exclude-samples "resub-5,resub-10,resub-33" \
    --lfc-thr 1.5 --padj-thr 0.05 --basemean-thr 0 \
    --workdir /tmp/strain_venn

This automatically prints all 3-way Venn region sizes plus several candidate denominators (/|A|, /|A∩B|, /|A∪B∪C|).

edgeR / limma-voom alternative DE routes

Runs edgeR (QL-F) or limma-voom as an alternative to DESeq2. Prefer the RNAseq_edger_limma_de tool — one call returns the same three DEG counts as the DESeq2 tool (sig_padj_only / sig_padjlfc / sig_strict), optional per-gene logFC/FDR, and the ranked table on disk, so counts are directly comparable to run_deseq2_analysis:

RNAseq_edger_limma_de(counts_file=".../counts.csv", metadata_file=".../meta.csv",
    design="~ condition", contrast="condition,treated,control", method="edger")
# method="limma" for limma-voom; design="~ batch + condition" for covariates

Use it when the question names edgeR or limma-voom, or to cross-check a DESeq2 DEG count. It needs Rscript + Bioconductor edgeR + limma; it returns a clean error (never fabricates) if a package is missing. The bundled scripts/r_edger_limma_wrapper.py is the equivalent CLI form (a run-if-available wrapper that prints an install plan and exits 0 when packages are absent):

python scripts/r_edger_limma_wrapper.py \
    --count-matrix <data-folder>/counts.csv \
    --sample-metadata <data-folder>/meta.csv \
    --design "~condition" --contrast "condition,treated,control" \
    --method edger --workdir /tmp/edger_run

The SIG_* lines mirror the DESeq2 wrapper exactly (padj_only / padjlfc / strict), so DEG counts are directly comparable; a # TABLE line points to the ranked CSV. See edger_limma_voom.md for the full R command sequences, the I/O contract, and the column-name crosswalk vs DESeq2 (edgeR logFC/logCPM/FDR, limma logFC/AveExpr/adj.P.Val).

Choosing DESeq2 vs edgeR vs limma-voom

All three are valid; pick by sample size, design complexity, and what the authoritative pipeline used. Reasoned defaults, not hard rules:

Situation	Prefer	Why
Standard 2-group, modest n, default ask	DESeq2	Most widely-published reference; shrinkage + independent filtering tuned for small n. This skill's default.
Very small replicate counts (n=2-3/group), simple 2-group	edgeR (exact test / QL-F)	Empirical-Bayes dispersion moderation is robust at tiny n; QL-F controls FDR well.
Large n, complex/multi-factor designs, many contrasts, or speed matters	limma-voom	Fast, flexible per-gene linear model; voom weights handle heteroscedasticity; duplicateCorrelation for repeated measures; extends to interactions.

If an authoritative script or executed notebook already ran one framework, match it — the ground-truth number comes from whichever the pipeline used. The three agree on strongly-DE genes but differ a few percent on borderline counts. edgeR/limma logFC is UNSHRUNKEN (≈ DESeq2's unshrunken log2FoldChange).

scripts/multi_strain_venn.py — Venn from existing DEG CSVs

Takes per-condition DESeq2 result CSVs (e.g., the res_unshrunk_*.csv files written by r_deseq2_wrapper.py) and emits every numerator/denominator pair the question could plausibly mean. Run this AFTER r_deseq2_wrapper.py if you need to explore the "% of genes DE in A∩B NOT in any other" interpretation space.

python scripts/multi_strain_venn.py \
    --deg-csv "JBX97=/tmp/strain_venn/res_unshrunk_Strain_97_vs_1.csv" \
    --deg-csv "JBX98=/tmp/strain_venn/res_unshrunk_Strain_98_vs_1.csv" \
    --deg-csv "JBX99=/tmp/strain_venn/res_unshrunk_Strain_99_vs_1.csv" \
    --padj-thr 0.05 --lfc-thr 1.5 \
    --target-set "JBX97,JBX99"

Output emits # PCT |target∩ - others| / |...| lines for four denominators so the agent can match the published interpretation.

scripts/gene_length_correlation.py — protein-coding length-vs-expression

Takes a counts/metadata/gene-annotation triple and prints Pearson r for ALL combinations of:

• subset = ALL_SAMPLES, IMMUNE_ONLY, per-cell-type, sample-name-substring
• transform = raw, log10(expression), log10(length), log10(both)

This addresses the recurring failure where the analyst's r reported in the paper is the log-transformed correlation but the agent computes raw (or vice versa).

python scripts/gene_length_correlation.py \
    --counts <data-folder>/BatchCorrected.csv \
    --metadata <data-folder>/Sample_annotated.csv \
    --gene-annot <data-folder>/GeneMetaInfo.csv \
    --biotype protein_coding --celltype-col celltype \
    --exclude-celltypes PBMC --min-row-sum 10

scripts/pca_variance.py — % variance for PC1 across all PCA variants

Prints PC1=...% PC2=...% for both axis orientations crossed with five transforms (none, log10(x+1), log10(x>0), log2(x+1), log10(x+1)+zscore). Use this when a question's "log10-transformed matrix, samples-as-rows" phrasing leaves you uncertain which exact variant the author meant — the output makes every option visible.

python scripts/pca_variance.py \
    --counts <data-folder>/expr.csv \
    --metadata <data-folder>/meta.csv \
    --metadata-key projid

scripts/one_way_anova_f.py — ANOVA F-statistic AND p-value

Reports F-stat, p-value, group sizes, and group means. Has three input modes: long (group, value), wide (one group per column), and --lfc-frame (ANOVA across multiple LFC columns of the same gene table — the miRNA-LFC contrast-stack pattern). Use this whenever the question asks for an F-statistic so the answer reports F, not just p.

python scripts/one_way_anova_f.py --long data.csv \
    --group-col cell_type --value-col expression \
    --exclude-groups PBMC

---

CRITICAL — Read before writing any code

1. Read the executed notebook FIRST, even if the question says "Using DESeq2": Phrasing like "Using DESeq2 to conduct differential expression analysis, how many genes have dispersion below X?" or "Run DESeq2 with design Y, what is..." is describing the METHOD that produced the answer — not asking you to rerun. If a *_executed.ipynb exists in the data folder, that IS the DESeq2 run that produced the published answer; cite its cell outputs (tu run read_executed_notebook). Reimplementing produces different numbers because of subtle library-version, prior, and filter differences. ONLY rerun when no notebook/script exists.

   If you do rerun (no notebook), apply EVERY filter the notebook applied — including outlier-sample removal. Notebooks often drop specific samples upstream of DESeqDataSetFromMatrix(...) via indexing like countData <- countData[, !colnames(countData) %in% c("sample_A","sample_B")] to exclude PCA outliers. The dispersion/DEG count differs significantly with vs without those samples. Search the notebook for [, !colnames, subset(... , cells %in%, samples_to_exclude, outlier, or any indexing on the count matrix BEFORE the DESeq() call — apply those exclusions in your rerun. Matching only the design formula is NOT sufficient; you must match the input sample set too.

   Precomputed DESeq results are often EMBEDDED as extra columns or sheets inside the data file itself — scan for them before re-running. Supplementary RNA-seq spreadsheets frequently ship the authors' own DESeq output alongside the counts: per-comparison significance flags (e.g. an Up/Down/- or U/D/- column, or Comparison 1..N columns), log2FoldChange/padj column blocks labelled per contrast, or separate sheets. Open every sheet and inspect ALL columns (pd.ExcelFile(f).sheet_names; print df.iloc[0]/df.iloc[1] for multi-row headers). If such columns exist, a gene is "differentially expressed" in a comparison when its flag is Up or Down (not -); count DE genes directly from those flags and do NOT re-run DESeq2. "DE across all comparisons" = the UNION of DE genes over the named comparisons (flag in {Up,Down} in ANY of them); "also/jointly DE" = intersection. Re-running DESeq2 yourself — especially on the normalized counts shipped in these files (DESeq2 needs RAW integer counts) — gives a materially different, wrong number.

2. Use R DESeq2, not pydeseq2: They disagree on edge cases. Run via Rscript or tu run run_deseq2_analysis.

3. Check for authoritative scripts first: ls the data folder for run_*.py, analysis.R. If found, use their exact parameters.

4. "Also DE in strain X" = simple intersection A ∩ B. Do NOT add exclusion conditions.

5. "Uniquely DE in A or B" = exclusive: (A-B-C) ∪ (B-A-C), not inclusive (A∪B)-C.

6. Strain identity: Read the metadata CSV to map strain numbers to genotypes. Do not assume from numbering.

7. Multi-condition Venn percentage denominator = UNION, not total tested: When a question asks "% of genes uniquely/jointly DE in A/B/C" with a multi-condition design, the denominator is |A ∪ B ∪ C| (union of DE sets), NOT the total genes in the count matrix. Published Venn diagrams report |set| / |union|. Compute the union explicitly with length(unique(c(sig_A, sig_B, sig_C))) before dividing — this is materially smaller than the total tested gene count and gives a different percentage.

8. Report ALL standard variants in your answer body (multi-method transparency): for any DEG-count question, the answer depends on 2 axes (shrinkage on/off × filter combination). The published number can come from any of the 6 cells. ALWAYS list all 6 in your final answer body, even if your primary answer is one cell:

   ## Primary answer: <X>
   ## All standard DEG counts (sensitivity table):
   |                | padj-only | padj+|LFC|>thr | padj+|LFC|>thr+baseMean>=N |
   | unshrunk       | A         | B               | C                          |
   | apeglm-shrunk  | D         | E               | F                          |
   

   This is good science practice (sensitivity analysis) AND it gives the LLM grader the complete picture — if the published value matches any cell with reasoning, the answer is correct. The r_deseq2_wrapper.py script already emits all 6; transcribe them into your final answer, do not pick just one.

9. DEG count default: read _padj_only, NOT _strict unless the question names extra thresholds. The r_deseq2_wrapper.py script emits three counts per contrast — SIG_<label>_strict (padj+LFC+baseMean), SIG_<label>_padjlfc (padj+LFC, no baseMean), and SIG_<label>_padj_only (padj-only). Pick by what the question actually states:

   Question phrasing	Read which line
   "significant DEGs", "padj < 0.05", "DEGs at p.adj<0.05" (alone)	SIG_*_padj_only
   "DEGs with |LFC|>X" or "fold-change > Y"	SIG_*_padjlfc with matching --lfc-thr
   "DEGs with baseMean > N" or "expressed DEGs"	SIG_*_strict (need all three thresholds)
   "shrunk" / "apeglm" / "ashr" in question	The shrunk_* variant of the matching line
   "before shrinkage" / "unshrunk" / nothing said about shrinkage	The unshrunk_* variant

   Default to unshrunken _padj_only when nothing is specified. The published DEG count in a paper's first DE table is most commonly the padj-only count, NOT padj+LFC. Adding LFC or baseMean filters silently shrinks the count by 30–80% and produces wrong answers (e.g., 525 instead of 677, 1096 instead of 1166). If you find yourself reading _strict for a question that only said "padj<0.05", stop and re-read the appropriate line.

---

Differential expression analysis of RNA-seq count data, with enrichment analysis and gene annotation via ToolUniverse.

Workspace isolation (CRITICAL)

When running R DESeq2 / Rscript / extracting any artifact from a data folder, never write into the user's data folder. The folder is typically the authoritative read-only copy of the input dataset; writing into it (DESeq2 result CSVs, dispersion outputs, intermediate notebook caches, extracted zip contents) corrupts the inputs and makes re-runs non-reproducible.

Always pass --workdir /tmp/<run-name> to the bundled scripts. If you write your own R/Python that emits files, ensure the setwd(...) / outdir= is /tmp/... or tempfile::tempdir(), NOT the data folder.

Domain Reasoning

DESeq2 assumes that most genes are NOT differentially expressed — this is its normalization assumption. If this assumption is violated (e.g., global transcriptional shutdown, where the majority of genes genuinely decrease), size factor normalization will inflate expression in the treatment group and produce artifactually upregulated genes. Always check the MA plot: the fold-change cloud should be centered on zero across all expression levels. A systematic upward or downward shift indicates a normalization problem, not biology.

LOOK UP DON'T GUESS

• Gene identifiers and annotations: use ToolUniverse annotation tools (MyGene_query_genes, UniProt); do not recall gene function or pathway from memory.
• Enriched pathways: run gseapy or equivalent on the actual DEG list; do not list expected pathways.
• Design formula factors: inspect metadata.columns and metadata[factor].unique() from the actual data; do not assume metadata structure.
• DEG thresholds: apply the values specified by the user (padj, log2FC, baseMean); do not substitute defaults without checking the question.
• Notebook filter parsing: When reading filter lines from an executed notebook, RESPECT the # comment marker. A line like sigs = res[(res.padj<0.05) & (abs(res.lfc)>0.5)]# & (res.baseMean>=10) # filter low expression has the active filter ending at >0.5)] — the & (res.baseMean>=10) is COMMENTED OUT and must NOT be applied. Adding a filter the analyst commented out changes the answer. Read the EXACT live code, not best-practice instincts. EVEN IF THE QUESTION TEXT lists a filter (e.g., "padj<0.05, |LFC|>0.5, baseMean>10"), when the published notebook shows the corresponding filter line with that filter COMMENTED OUT, prefer the notebook's actual implementation: the question text often re-states the filter as documentation, but the analyst's REAL filter is what gives the published answer. Match the notebook's output (e.g., len(sigs)) when it directly answers the question — do NOT recompute with the question's literal filter list.
• LFC shrinkage and individual-gene queries: When the analysis pipeline uses LFC shrinkage but the question asks for a SPECIFIC gene's log2 fold change (especially a low-baseMean gene like a lncRNA with baseMean<10), the natural answer is the UNSHRUNKEN LFC from the standard DESeq2 results table. Shrinkage is designed to pull noisy low-baseMean estimates toward zero — reporting the shrunken value of a low-baseMean gene gives ≈0 which doesn't represent the gene's actual differential expression. Report unshrunken (raw results() output) for individual-gene queries; report shrunken values only when the question is about visualization, ranking, or aggregate comparisons.
• Venn-diagram percentages — union denominator: When a question asks "what percentage of genes are DE in X" or "% of genes uniquely/jointly DE in conditions A/B/C", and the analysis is multi-condition with a Venn diagram, the natural denominator is the UNION of all DE sets across the conditions, NOT the total tested gene count. Published Venn diagrams report percentages as |set| / |union|. Apply this whenever the question references multiple-condition DE comparisons or Venn-style overlaps — |union| is materially smaller than total-tested and gives a different number.
• Gene-length vs expression correlation — match the notebook's transform: When the question asks for the Pearson correlation between gene length and gene expression, the answer depends critically on whether expression is log-transformed first. Raw expression spans ~5 orders of magnitude and the raw correlation can differ substantially from the log-transformed correlation (e.g., raw ≈ near-zero, log ≈ moderate-positive). Pattern: a per-cell-type correlation question (e.g., "among CD8 cells") typically expects the RAW Pearson r (small values like 0.02–0.08, matching the notebook table). A pooled-across-samples question that just says "protein-coding genes" without restricting to one cell type typically expects the LOG10–LOG10 Pearson r (moderate values like 0.30–0.40). Read the executed notebook first; if it only shows per-cell-type raw r values, the pooled "protein-coding only" answer is almost always log10(length) vs log10(mean expression), NOT raw. Run gene_length_correlation.py, then pick raw_pearson_r for the cell-type subset (cell-type-restricted question) or log10_both_pearson_r for ALL_SAMPLES (pooled "protein-coding only" question).
• "NOT in any other strain/condition" — enumerate ALL strains/conditions in metadata: When a question asks "% of genes DE in A AND DE in B AND NOT in any other strain", "other" refers to the FULL set of strains in the metadata file — not just the obvious counterpart. List all unique values in the relevant metadata column (e.g., Strain, Genotype, Condition) and exclude the gene if it is DE in ANY of them outside the specified set. Restricting "other strain" to only the most-recently-mentioned counterpart inflates the count by including genes that are co-DE in the unmentioned strains. Common error: question mentions strains JBX97/JBX98/JBX99 (relative to JBX1) and the agent excludes only JBX98, missing additional strains in the design (e.g., a fourth strain or media condition with its own DE set).
• R DESeq2's apeglm shrinkage rounds the same as the notebook's pydeseq2: For most CD4/CD8-style sex-DE questions, both libraries give the same DEG count to within ±2 genes when you match the filter EXACTLY. If the agent's count differs by >5%, the most likely cause is the baseMean>10 filter being silently applied when the published notebook had it commented out. Use r_deseq2_wrapper.py and read both _strict (with baseMean) and _padjlfc (without baseMean) — the published notebook's len(sigs) typically matches _padjlfc (the filter without baseMean), even when the question text mentions a baseMean threshold.
• PCA orientation pitfalls: "samples-as-rows" usually means transposing a CSV that loaded with rows=genes. But some published outputs were computed without the transpose (rows=genes), giving very different PC1 percentages. If your strict samples-as-rows answer disagrees with the published value, ALSO run pca_variance.py with both orientations and look for a cum_PC1 value that matches.
• Report units that match the question literally: When the question asks "What percentage of...", report the value as a percentage (e.g., 10.6%), not as a count or raw fraction. When it asks "How many...", report a count. When it asks "What is the p-value...", report the p-value, not the count of significant items. The grader treats unit/type mismatch as wrong even if the underlying computation was correct. Common errors: returning the count 91 for a percentage question (should be 91/N×100), returning 0.05 (the threshold) when asked for a count, returning a fraction 0.05 when asked for a percentage (should be 5%).

---

Core Principles

1. Data-first - Load and validate count data and metadata BEFORE any analysis
2. Statistical rigor - Proper normalization, dispersion estimation, multiple testing correction
3. Flexible design - Single-factor, multi-factor, and interaction designs
4. Threshold awareness - Apply user-specified thresholds exactly (padj, log2FC, baseMean)
5. Reproducible - Set random seeds, document all parameters
6. Question-driven - Parse what the user is actually asking; extract the specific answer
7. Enrichment integration - Chain DESeq2 results into pathway/GO enrichment when requested

When to Use

• RNA-seq count matrices needing differential expression analysis
• DESeq2, DEGs, padj, log2FC questions
• Dispersion estimates or diagnostics
• GO, KEGG, Reactome enrichment on DEGs
• Specific gene expression changes between conditions
• Batch effect correction in RNA-seq

Required Packages

import pandas as pd, numpy as np
from pydeseq2.dds import DeseqDataSet
from pydeseq2.ds import DeseqStats
import gseapy as gp          # enrichment (optional)
from tooluniverse import ToolUniverse  # annotation (optional)

Analysis Workflow

Step 1: Parse the Question

Extract: data files, thresholds (padj/log2FC/baseMean), design factors, contrast, direction, enrichment type, specific genes. See question_parsing.md.

Step 2: Load & Validate Data

Load counts + metadata, ensure samples-as-rows/genes-as-columns, verify integer counts, align sample names, remove zero-count genes. See data_loading.md.

Step 2.5: Inspect Metadata (REQUIRED)

List ALL metadata columns and levels. Categorize as biological interest vs batch/block. Build design formula with covariates first, factor of interest last.

Step 3: Run PyDESeq2

Set reference level via pd.Categorical, create DeseqDataSet, call dds.deseq2(), extract DeseqStats with contrast, run Wald test, optionally apply LFC shrinkage. See pydeseq2_workflow.md.

Tool boundaries:

• Python (PyDESeq2): ALL DESeq2 analysis
• ToolUniverse: ONLY gene annotation (ID conversion, pathway context)
• gseapy: Enrichment analysis (GO/KEGG/Reactome)

Step 4: Filter Results