sc-metacell

When to use

The user has a normalised scRNA AnnData with a low-D embedding (obsm["X_pca"] by default) and wants compact metacell aggregates suitable for downstream GRN inference, slow-RNA-velocity analysis, or robust cluster-level statistics that need fewer but higher-coverage units. Two methods:

• seacells (default) — kernel archetypal analysis with iterative refinement (--min-iter / --max-iter). Aggregates similar cells into archetypes. Auto-falls back to kmeans if the SEACells package isn't installed.
• kmeans — KMeans clustering on the embedding; faster, no SEACells dependency.

Output is a new AnnData (madata) where each row is one metacell. For per-cluster marker discovery on the original AnnData, use sc-markers. For ordering cells, use sc-pseudotime.

Inputs & Outputs

<!-- AUTO-GENERATED from skill.yaml (interface) — do not edit by hand. Regenerate: python scripts/generate_skill_md.py <skill_dir> -->

Inputs

• Modalities: scrna
• File types: .h5ad
• Requires a preprocessed AnnData (X normalised, PCA/neighbours present)

Outputs

• tables/cell_metadata.csv
• tables/cell_to_metacell.csv
• tables/centroid_points.csv
• tables/embedding_points.csv
• tables/metacell_summary.csv
• figures/metacell_centroids.png
• figures/metacell_size_distribution.png
• figures/r_embedding_discrete.png
• analysis_summary.txt
• metacells.h5ad
• metacells_annotated.h5ad
• processed.h5ad
• report.md
• result.json
• Processed AnnData (saves_h5ad) — adds obs: metacell

Flow

1. SEACells fallback check: try import SEACells; if it fails, silently switch --method to kmeans.
2. Load AnnData (--input) or build a demo.
3. Preflight: obsm[--use-rep] exists; --n-metacells is < n_cells and ≥ 2; warn if no layers["counts"].
4. For SEACells: run kernel archetypal analysis with --min-iter / --max-iter; for KMeans: cluster the embedding.
5. Aggregate cell-level expression into metacell-level expression (sum over layers["counts"] if present, else .X).
6. Build per-metacell summary (size, dominant --celltype-key label).
7. Save metacell AnnData, tables, figures, report.md, result.json.

Gotchas

• --method seacells silently auto-falls back to kmeans if SEACells isn't installed. sc_metacell.py:326-332 catches ImportError from import SEACells and logs a warning, then sets args.method = "kmeans". The actually-used method is recorded in result.json["method"] (the post-fallback value), but the request-vs-execute distinction isn't preserved as separate keys here. Inspect logs / report.md to confirm.
• All preflight failures raise SystemExit(1), not ValueError. sc_metacell.py:113 raises SystemExit(1) when obsm[--use-rep] is missing (with available-embedding hint); :123 raises when --n-metacells >= n_cells (with a sane suggested value); :128 raises when --n-metacells < 2; :340 raises SystemExit("Provide --input or use --demo"). Wrappers expecting ValueError need to catch SystemExit.
• processed.h5ad is the cell-level AnnData with metacell labels — NOT the metacell-aggregated AnnData. sc_metacell.py:463 saves the original adata (with obs["metacell"] added) to processed.h5ad; sc_metacell.py:393 writes the aggregated metacell-shape madata to tables/metacells.h5ad. Downstream skills like sc-de / sc-enrichment (advertised in the next-step block at sc_metacell.py:534-535) consume processed.h5ad, not the aggregated shape — pass tables/metacells.h5ad explicitly when you actually need per-metacell rows.
• Aggregation prefers layers["counts"] over .X. The preflight warns if layers["counts"] is absent (sc_metacell.py:131-134); when it's missing the script aggregates from .X (typically log-normalised) which is mathematically less defensible than summing raw counts.
• --celltype-key defaults to leiden. sc_metacell.py:77 defaults to leiden; if your AnnData has labels under a different key (e.g., cell_type), pass --celltype-key cell_type so the per-metacell "dominant cell type" column is meaningful.

Key CLI

# Demo
python omicsclaw.py run sc-metacell --demo --output /tmp/sc_metacell_demo

# Default SEACells, 30 metacells
python omicsclaw.py run sc-metacell \
  --input clustered.h5ad --output results/

# KMeans (fast, no SEACells dependency)
python omicsclaw.py run sc-metacell \
  --input clustered.h5ad --output results/ \
  --method kmeans --n-metacells 50

# Use Harmony-corrected embedding + custom celltype key
python omicsclaw.py run sc-metacell \
  --input integrated.h5ad --output results/ \
  --use-rep X_harmony --celltype-key cell_type --n-metacells 100

# More refinement iterations for SEACells (slower, smoother archetypes)
python omicsclaw.py run sc-metacell \
  --input clustered.h5ad --output results/ \
  --min-iter 20 --max-iter 60

See also

• references/parameters.md — every CLI flag, SEACells vs KMeans tunables
• references/methodology.md — when metacells help; SEACells archetypal math
• references/output_contract.md — madata.obs schema; cell-to-metacell map columns
• Adjacent skills: sc-clustering (upstream — produces obs["leiden"] for the celltype-key default), sc-batch-integration (upstream — produces X_harmony embedding for --use-rep), sc-grn (downstream — GRN inference is more stable on metacells than single cells), sc-de (parallel — sample-aware DE between conditions; complements per-metacell aggregation)