spatial-domains

spatial-domains

When to use

Find tissue domains from expression and coordinates. Use spatial-annotate for named cell labels and spatial-genes for variable genes.

Use from a step

from skills._sdk.notebook import load_skill, read_input, write_output
library = load_skill("spatial-domains")
data = read_input("input.h5ad")
data = library.identify(data, method="leiden", spatial_weight=0.3)
write_output(library.domain_counts(data), "tables/results.csv")

Run examples/example_step.py through the step runner.

API

<!-- api:begin generated from _api.py; regenerate with run.py api <skill dir> --write -->

identify(adata, *, method: str='leiden', resolution: float=1.0, spatial_weight: float=0.3, refine: bool=False, random_state: int | None=None, **parameters)

Identify domains in place and return the same AnnData.

Reads log-normalized X, X_pca and spatial coordinates; graph methods reuse existing expression neighbors. SpaGCN, STAGATE and BANKSY results vary between runs because their training wrappers do not expose every RNG.

:param adata: Preprocessed spatial AnnData; expression values are retained. :param method: CLI default leiden, or louvain/spagcn/stagate/graphst/banksy/cellcharter. :param resolution: Graph-clustering resolution, CLI default 1.0. :param spatial_weight: Spatial graph weight for Leiden/Louvain, CLI default 0.3. :param refine: False by default; True smooths labels using spatial KNN. :param random_state: None preserves CLI defaults: backend seed 42 for STAGATE, GraphST and CellCharter, otherwise 0; PCA uses 0. An explicit integer overrides both PCA and supported backend seeds. Existing PCA is reused. :param parameters: Backend options listed in references/parameters.md; fixed-K methods use n_domains=7 unless supplied. :returns: The same AnnData with spatial_domain and JSON run diagnostics. :raises ValueError: Unsupported method or invalid graph parameters. :raises ImportError: A backend is missing; use install_skill_deps with the named package.

run_info(adata, *, keep: bool=True) -> dict

Read the method, domain sizes, refinement status and effective seeds.

pca_random_state is None when identify reused existing PCA coordinates.

:param adata: AnnData returned by identify. :param keep: True retains diagnostics; False removes them before CLI serialization. :returns: Diagnostic dict, or an empty dict before analysis.

domain_counts(adata)

Count observations and percentages per domain.

:param adata: AnnData with spatial_domain labels; X is not read. :returns: DataFrame with domain, n_cells and proportion (percent). :raises KeyError: Domain labels are absent.

domain_figure(adata)

Plot domain labels in spatial coordinates.

:param adata: AnnData with spatial_domain and spatial coordinates; X is not read. :returns: A matplotlib Figure without writing files. :raises KeyError: Domain labels are absent.

<!-- api:end -->

Methods and parameters

Leiden/Louvain combine expression and spatial graphs. SpaGCN, STAGATE, GraphST, BANKSY and CellCharter remain optional. Fixed-K methods default to seven domains. See parameters and methodology.

Gotchas

• identify reuses expression neighbors. SpaGCN, STAGATE and BANKSY results vary between runs. spatial_weight=0 uses expression-only graph clustering.
• run_info(keep=False) removes library diagnostics before CLI serialization.

Inputs and outputs

The library returns AnnData, DataFrames or Figures without file writes. The CLI keeps reports, result.json, tables and conditional gallery outputs. See the complete output contract for filenames and conditions.

CLI

python skills/spatial/spatial-domains/spatial_domains.py --input data.h5ad --output results/spatial-domains

See also

• Methodology
• Parameters
• Output contract

Dependencies

anndata, cellcharter, GraphST, igraph, louvain, matplotlib, numpy, pandas, pybanksy, scanpy, scikit-learn, scipy, seaborn, SpaGCN, squidpy, STAGATE-pyG, torch