Sc Grn

sc-grn

Use from a step

grn = load_skill("sc-grn")
adata = read_input("expression.h5ad")
edges = grn.infer_adjacencies(adata, tfs=["SPI1", "IRF8", "STAT1"],
                              method="correlation", n_top=50)
regulons = grn.regulons_from_adjacencies(edges)
scores = grn.score_regulons(adata, regulons, method="mean")
write_output(edges, "tables/adjacencies.csv")
write_output(scores, "tables/mean_target_expression.csv")

Functions do not annotate the input automatically. The PBMC example in examples/example_step.py uses caller-supplied TFs and explicitly labeled mean-expression scores.

API

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

infer_adjacencies(adata, *, tfs, method: str='grnboost2', layer: str | None=None, random_state: int=42, n_top: int=50, n_jobs: int=4) -> pd.DataFrame

Return TF, target and importance columns using the caller's TF list.

grnboost2 retains the CLI's fallback to absolute Spearman correlation when its backend fails or returns no edges. Correlation excludes supplied TFs from candidate targets and keeps n_top targets per TF. No motif validation occurs here; a warning and run_info report any fallback.

:param tfs: TF names; only names present in the selected matrix are used. :param method: grnboost2 (default) or correlation. :param layer: Explicit expression layer; None prefers counts, aligned raw, then X. :param random_state: GRNBoost2 seed, default 42. Correlation is deterministic. :param n_top: Correlation targets per TF, default 50. :param n_jobs: GRNBoost2 workers, default 4. :returns: A new DataFrame; input AnnData is not modified. :raises ValueError: The method or budget is invalid, or no TF overlaps.

run_info(adjacencies: pd.DataFrame, *, keep: bool=True) -> dict

Return an independent backend/fallback record; keep=False removes it from attrs.

prune_regulons(adjacencies: pd.DataFrame, *, database_glob: str, motif_annotations: str, n_top: int=50, rank_threshold: int=5000, auc_threshold: float=0.05, nes_threshold: float=3.0, n_jobs: int=4) -> list[dict]

Return motif-pruned regulon dictionaries using the existing cisTarget bridge.

Requires pySCENIC and caller-provided databases and motif annotations. Thresholds retain the CLI defaults: rank 5000, AUC 0.05, NES 3 and 50 targets. Backend and resource errors propagate; no data are downloaded.

regulons_from_adjacencies(adjacencies: pd.DataFrame, *, n_top: int=50) -> list[dict]

Group the strongest edges per TF without motif validation; default 50 targets.

score_regulons(adata, regulons: list[dict], *, method: str='aucell', random_state: int=42, n_jobs: int=4) -> pd.DataFrame

Return per-cell regulon scores without annotating the input.

aucell requires pySCENIC and ranks count-layer/aligned-raw/X expression. mean averages each regulon's targets in X, matching the old simplified CLI. It is not AUCell, an enrichment statistic, or motif validation.

:param method: aucell (default) or mean; no automatic scoring fallback. :param random_state: AUCell seed, default 42; mean is deterministic. :param n_jobs: AUCell workers, default 4. :returns: DataFrame indexed by cell; attrs names the scoring_method and is_aucell. :raises ValueError: The scoring method is unknown. :raises ImportError: AUCell's optional backend is unavailable.

regulon_heatmap_figure(scores: pd.DataFrame, *, groups: pd.Series | None=None)

Return a score heatmap, optionally averaged over aligned cell groups.

<!-- api:end -->