spatial-domains

spatial-domains

When to use

The user has a preprocessed spatial AnnData (obsm["X_pca"] and obsm["spatial"] populated) and wants tissue regions / niches identified per spot (obs["spatial_domain"]). Seven methods:

• leiden (default) — spatial-weighted Leiden (--resolution, --spatial-weight). No GPU.
• louvain — spatial-weighted Louvain. No GPU.
• spagcn — graph convolutional, fixed-K (--n-domains, --epochs, --spagcn-p). Requires torch + SpaGCN.
• stagate — graph attention with cell-type-aware regularisation (--stagate-alpha, --pre-resolution, --rad-cutoff / --k-nn). Requires torch + torch-geometric.
• graphst — graph-self-supervised (--epochs, --dim-output, --n-domains). Auto-detects 10x platform. Requires torch + GraphST.
• banksy — neighbourhood expression matrix + PCA (--lambda-param, --num-neighbours). 0.2 = cell-typing mode, 0.8 = domain mode.
• cellcharter — niche-graph clustering with auto-k (--auto-k, --auto-k-min/--auto-k-max, --n-layers). Requires cellcharter + pyro-ppl.

For spatially variable genes use spatial-genes; for spot-level cell-type labels use spatial-annotate.

Inputs & Outputs

Flow

1. Load AnnData (--input) or build a demo. Auto-compute obsm["X_pca"] if missing (logs a warning).
2. Default --n-domains to 7 for GNN methods (spagcn/stagate/graphst) and for cellcharter when --auto-k is off.
3. For graphst: infer --data-type from input metadata / path (10x Visium auto-detected).
4. Run the chosen method; write obs["spatial_domain"] (categorical) and method-specific obsm embedding.
5. Optionally refine domain assignments with neighbourhood smoothing (--refine).
6. Compute domain counts + proportions; per-domain neighbour-mixing summary.
7. Save tables, figures, processed.h5ad, report.md, result.json.

Gotchas

• --input missing → sys.exit(1) via print + sys.exit, NOT parser.error. spatial_domains.py:959-960 prints "ERROR: Provide --input or --demo" to stderr and sys.exit(1) — different from sibling skills' parser.error. Caller wrappers expecting parser.error (exit 2) get exit 1.
• obsm["X_pca"] is auto-computed when missing. spatial_domains.py:953-957 logs a warning and runs sc.pp.pca. The implicit PCA uses defaults (no HVG selection, no batch correction). For real data prefer spatial-preprocess upstream so the PCA reflects HVG-aware preprocessing.
• GNN methods auto-default --n-domains to 7. spatial_domains.py:962-968 silently sets args.n_domains = 7 for spagcn / stagate / graphst and for cellcharter (when --auto-k is off). Override explicitly or these K-fixed methods quietly target 7 clusters.
• obsm["spatial"] ↔ obsm["X_spatial"] sync. spatial_domains.py:77-79 ensures both keys exist (copies one to the other if missing). Some upstream skills only write one; this skill normalises.
• Per-method obsm embedding key. spatial_domains.py:110-113 records: STAGATE → obsm["X_stagate"], GraphST → obsm["X_graphst"], BANKSY → obsm["X_banksy_pca"], CellCharter → obsm["X_cellcharter"]. Leiden / Louvain / SpaGCN do NOT write a method-specific embedding.
• Performance warning at 30K cells. spatial_domains.py:548 logs a warning when n_cells > 30000 and method ∈ {graphst, spagcn, stagate}. The methods still run; consider downsampling or switching to leiden for large datasets.
• GraphST is unrecommended above 5K cells. spatial_domains.py:554 logs a separate warning specifically for graphst when n_cells is None or n_cells > 5000.

Key CLI

# Demo (synthetic spatial)
python omicsclaw.py run spatial-domains --demo --output /tmp/spatial_dom_demo

# Default leiden with spatial weighting
python omicsclaw.py run spatial-domains \
  --input preprocessed.h5ad --output results/ \
  --method leiden --resolution 1.0 --spatial-weight 0.3

# SpaGCN with explicit K
python omicsclaw.py run spatial-domains \
  --input preprocessed.h5ad --output results/ \
  --method spagcn --n-domains 8 --spagcn-p 0.5 --epochs 200

# STAGATE with cell-type-aware module
python omicsclaw.py run spatial-domains \
  --input preprocessed.h5ad --output results/ \
  --method stagate --rad-cutoff 150 --stagate-alpha 0.5 --n-domains 7

# CellCharter with auto-k
python omicsclaw.py run spatial-domains \
  --input preprocessed.h5ad --output results/ \
  --method cellcharter --auto-k --auto-k-min 4 --auto-k-max 12 --n-layers 3

See also

• references/parameters.md — every CLI flag, per-method tunables
• references/methodology.md — when each backend wins; cell-count rules of thumb
• references/output_contract.md — obs["spatial_domain"] + per-method obsm keys
• Adjacent skills: spatial-preprocess (upstream — produces obsm["X_pca"] / obsm["spatial"]), spatial-integrate (upstream — for multi-batch data, run integration first), spatial-genes (parallel — spatially variable gene ranking, NOT domain detection), spatial-annotate (parallel — spot-level cell-type labels, complementary to domain labels), spatial-de (downstream — DE between domains)