Imported from openai/plugins (plugins/ngs-analysis/skills/scrna-seq-qc/SKILL.md). Install upstream with npx skills add openai/plugins --skill scrna-seq-qc. Copyright stays with the author.

scRNA-seq QC

Start Here

Read references/qc-annotation-umap-heuristics.md before picking thresholds, annotation backends, or UMAP feature-selection rules.

Confirm what inputs exist before writing code:

• An AnnData object or equivalent with raw counts preserved.
• Per-sample, per-batch, or per-channel metadata, because QC and doublet detection should respect technical partitions.
• Organism, tissue, assay type, chemistry, and whether the data are whole-cell or single-nucleus.
• Whether a matched cell atlas or label-transfer reference exists for the tissue and species.

Preserve provenance in the output: package versions, thresholds, threshold-justification plots, counts removed or flagged at each filter, annotation backend and reference, marker-gene selection heuristic, and any manual cluster exclusions.

Workflow

1. Choose QC thresholds from the data, not from a fixed template.

   • Plot detected genes, total UMIs, mitochondrial fraction, and any tissue-specific nuisance signals overall and by batch.
   • Inspect all available QC metrics, but default filtering should use only the standard metrics: detected genes, total counts, and percent.mt.
   • Pick thresholds from the observed distributions and expected biology.
   • Save a plot with threshold lines and record why each threshold is appropriate for this dataset.
   • If another metric looks important enough to filter on, flag it as a dataset-specific issue, explain why, and consult the user before adding that extra filter.
   • Keep QC plots legible: do not overload a single panel with too many batches or categories when faceting, splitting, or summary views would communicate the result more clearly.

2. Run cell-level QC.

   • Remove or flag obvious low-quality barcodes using the chosen thresholds on detected genes, total counts, and percent.mt.
   • Use scDblFinder for doublet detection. Run it per batch or capture channel, and split very large batches before doublet calling.
   • Do not skip doublet calling or silently substitute another method. If scDblFinder cannot run in the environment, surface the blocker explicitly or get user approval before using a different caller.
   • Compute ambient-RNA style metrics and use them for filtering when the dataset and workflow support it.
   • Compute any other informative QC metrics when feasible, but do not turn those additional nonstandard metrics into hard filters without explicit user approval unless the user already asked for a stricter policy.
   • Prefer adding a passes_QC column instead of physically dropping cells when downstream provenance matters.

3. Build a latent space and inspect residual artifacts.

   • Decide whether scVI is warranted for this dataset and use case before training it.
   • Prefer a standard PCA/Scanpy workflow for smaller, simpler datasets with limited batch structure or when a conventional embedding answers the question cleanly.
   • Prefer scVI when integration across batches, donors, chemistries, or related datasets is important, or when the dataset is large and noisy enough that a learned latent space is likely to help.
   • Record why scVI or a conventional PCA workflow was chosen for this dataset.
   • Cluster and inspect low-quality, mixed-marker, or ambiguous clusters before downstream visualization.
   • Remove or flag artifact clusters only with explicit evidence, and record the rationale.

4. Annotate cells.

   • If a suitable Allen Brain Cell Atlas reference exists and the dataset is a compatible brain tissue and species, use MapMyCells or cell_type_mapper.
   • If no suitable Allen reference exists, use the closest matched reference for tissue, species, assay, and chemistry with an appropriate mapping tool.
   • If no reliable reference exists, annotate conservatively from canonical markers and cluster-level markers. Assign coarse labels first and leave uncertain clusters as unknown or ambiguous rather than overlabeling them.
   • Persist annotation confidence or probability fields when available, together with at least one coarse and one fine label.

5. Choose a general marker panel for global UMAP.

   • Do not rely on a perturbation-specific or brain-only marker panel.
   • Start from HVGs selected in a batch-aware way.
   • Add genes that distinguish major coarse compartments or high-confidence labels, for example top markers per coarse cluster or class.
   • Exclude nuisance-dominated genes if they swamp the embedding unless the biology requires them.
   • Document how the panel was chosen.

6. Generate UMAP visualizations.

   • For a global UMAP, use the learned latent space or the chosen informative marker panel, depending on which better matches the analytical goal and runtime constraints.
   • For per-group UMAPs, subset by a stable coarse label and use the latent representation unless there is a strong reason to rebuild on expression features.
   • Keep plotting separate from filtering so visualization choices do not mutate the core analysis object.
   • Make every plot legible. Use a reasonable number of categories per panel, prefer coarse labels on overview plots, and split or facet figures when fine labels, batches, or neighborhoods would otherwise make the figure unreadable.

7. Scale to large datasets without copying.

   • Keep matrices sparse whenever possible.
   • Avoid densifying whole matrices.
   • Avoid whole-object copies of AnnData or Seurat objects; use views, backed mode, chunked operations, and per-batch or per-group manifests instead.
   • When crossing Python and R boundaries, pass only the subset and metadata required for the step.
   • Write checkpoints after major stages so failures do not require restarting from raw ingest.

Deliverables

When implementing a pipeline, produce an auditable output set:

• Filtered .h5ad or equivalent object with raw counts preserved and QC or annotation fields in metadata.
• QC summary table with input cells, cells removed or flagged by each filter, final cells, and per-batch summaries.
• Threshold-justification plots for detected genes, UMIs, mitochondrial fraction, plus any additional QC metric that was inspected; clearly separate metrics that informed review from metrics that actually drove filtering.
• Parameter manifest with thresholds, package versions, annotation backend and reference, marker-panel heuristic, and any manual exclusions.
• UMAP coordinates and plots for global and per-group views when requested, with category counts and panel layouts chosen so the figures remain legible.

Embedded Runner

For 10x-style matrix bundles, a local runner is available:

python plugins/ngs-analysis/scripts/run_scrnaseq_post_count_qc.py --input-dir path/to/scrna_bundle

The input directory should contain matrix/, manifest.tsv, and dataset_metadata.json, unless explicit paths are supplied. Treat the runner as an auditable analysis surface: its marker-based fallback is PBMC-oriented when no matched reference is provided, so tissue-specific annotation and integration choices still require review.

The runner writes visualizations/index.html for portable artifact review, summary.md plus provenance/analysis_status.json for explicit completeness/blocker reporting, and auto-launches a localhost Marimo review app recorded in notebooks/marimo_server.json. It also writes notebooks/scrna_qc_review.marimo.py as a notebook backup over the generated PNG/CSV/H5AD outputs. Treat the notebook and review app as review layers, not as the source of truth; the run envelope and generated artifacts remain canonical.

Resources

• references/qc-annotation-umap-heuristics.md: Threshold-selection heuristics, annotation fallback strategy, general marker-panel selection rules, and large-dataset memory practices.