Pydeseq2

PyDESeq2

Use this skill for bulk RNA-seq negative-binomial GLMs, with explicit biological replicates, covariates, and comparisons. Cells, sequencing lanes, and repeated measurements from the same donor are not independent biological replicates. Donor-level pseudobulk can be appropriate when its aggregation and design match the scientific question; this is not a single-cell per-cell testing workflow.

Current PyDESeq2 0.5.4 provides Wald tests and apeGLM-style coefficient shrinkage. Do not assume every feature or numerical result of current R DESeq2 is implemented identically. See API and tested versions for supported calls, AnnData storage, and source/runtime verification limits.

Analysis contract

1. Establish the experimental unit, count provenance, sample exclusions, reference levels, and intended comparison before fitting. Use nonnegative integer counts in samples × genes order, with unique sample/gene IDs. TPM, FPKM, log data, VST data, and library-normalized counts are not model inputs. Integer checks alone cannot establish raw-count provenance. Approved estimated-count imports need an explicit upstream conversion and offset/length-correction policy.
2. Require equal sample sets and reorder metadata to the count index. Resolve unmatched IDs and missing design annotations explicitly; never silently take an intersection. Determine orientation from the file schema, not its shape.
3. Apply a prespecified gene prefilter after intentional sample exclusions. Total counts >=10 is a small example rule, not a universal biological threshold. Keep enough genes to estimate normalization and dispersion trends jointly.
4. Build a full-rank formula design with residual degrees of freedom. Explicitly encode categories/reference levels; numeric variables are continuous. Formula term order does not decide the tested effect when a contrast is explicit.
5. Fit size factors, gene-wise/trend/MAP dispersions, LFCs, and Cook diagnostics. Inspect warnings, convergence and normalization assumptions before interpreting results. Large size factors can reflect real library-depth differences.
6. Test an explicit contrast with the intended alpha, preserve the full Wald table, and optionally shrink the same single positive coefficient. Export counts/model metadata separately from statistical results.

Quick start

Install in an isolated environment; the repository test runner supplies the scientific dependencies separately from the project environment:

uv venv --python 3.13 .venv-pydeseq2
uv pip install --python .venv-pydeseq2/bin/python 'pydeseq2==0.5.4' 'anndata==0.13.4'

Paths below are illustrative; run from this skill directory with your own files. The bundled driver expects counts CSV as genes × samples by default, and metadata CSV as samples × annotations, each with IDs in its first column:

python scripts/run_deseq2_analysis.py \
  --counts counts.csv --metadata metadata.csv \
  --design '~batch + condition' --contrast condition treated control \
  --min-counts 10 --alpha 0.05 --n-cpus 1 --plots --output results/

Use --no-transpose for a counts file already in samples × genes order. The driver rejects duplicate source CSV headers, unequal sample sets, invalid count values, zero-count samples, missing contrast annotations, and unidentifiable designs. It sets the CLI contrast variable's reference level before fitting; numeric category codes supplied as the contrast variable are deliberately treated as categories. Resolve other design variables' types and missing values in the input metadata.

The driver performs Wald testing and shrinkage by default. --no-shrink exports only unshrunk effect estimates; --shrink-coeff can select only the coefficient that exactly matches the tested contrast. For interactions, continuous effects, or custom design matrices, use the Python patterns rather than guessing a CLI coefficient. The CLI does not expose thresholded tests or alternative normalization methods; use the API for those choices.

Outputs:

• deseq2_results.csv: current estimates (shrunken if requested) with original Wald stat, pvalue, and padj.
• deseq2_results_unshrunken.csv: preserved original Wald table.
• significant_genes.csv and results_sorted_by_padj.csv: the configured alpha selects the significant table; missing padj is never significant.
• deseq_dataset.h5ad: fitted AnnData snapshot for inspection, with counts and model fields; it is not a reconstituted DeseqDataSet or a DeseqStats object.
• analysis_manifest.json: contrast, alpha, shrinkage coefficient, and versions.
• With --plots: volcano and MA PNGs using the same alpha. Missing adjusted p-values stay missing; zero adjusted p-values are capped only for plotting.

Use a fresh output directory for each run. Record input provenance and sample/gene exclusion decisions alongside these files.

Statistical interpretation

Positive log2FoldChange for ['condition', 'treated', 'control'] means treated / control. A padj < alpha rule controls the selected multiple-testing procedure under its assumptions; it is not the probability that an individual result is false. BH adjustment within each contrast does not jointly adjust all contrasts in a multi-comparison analysis.

Separate a descriptive abs(log2FoldChange) > 1 filter from a formal test of an effect exceeding one log2 unit (lfc_null=1, alt_hypothesis='greaterAbs'). Thresholded or one-sided statistics are not automatically suitable signed preranking scores.

Missing p-values can result from all-zero genes or Cook filtering; independent filtering may leave a finite p-value with missing padj. Preserve these distinctions and report how many genes received finite tests. Do not replace missing values with zero or one in result tables.

For standard two-sided zero-null Wald tests, preranked enrichment commonly uses finite signed stat values from the unshrunken table, including eligible genes that are not significant. Inspect Cook-filtered rows (a statistic can remain finite while its p-value is missing), identifier mappings, and tied scores. Do not use an unsigned DESeq2 LRT statistic, absolute LFC, or a thresholded hit list as this signed ranking. ORA instead needs a prespecified hit rule and the tested/detectable mapped background. A product such as -log10(padj) * abs(LFC) is not a calibrated test or a substitute for signed Wald ranking.