`cell_types`¶

Cell-type assignment via Leiden clustering on the sample-removed cell embedding (Z_clust by default). When n_target_clusters is set, the resolution is increased recursively until the Leiden output reaches the target; if it overshoots, a dendrogram-based aggregation merges the closest clusters (by centroid distance) until the count matches. Optionally computes and saves a UMAP. The resulting labels are written to adata.obs["cell_type"] as stringified integers.

Source: preparation/cell_type_cpu.py:12 (GPU variant: preparation/cell_type_gpu.py, function cell_types_gpu)

Signature¶

def cell_types(
    anndata_cell,
    cell_type_column="cell_type",
    existing_cell_types=False,
    n_target_clusters=None,
    umap=True,
    save=False,
    output_dir=None,
    defined_output_path=None,
    leiden_cluster_resolution=0.8,
    cell_embedding_column=None,
    cell_embedding_num_PCs=20,
    verbose=True,
    umap_plots=True,
    _recursion_depth=0,
) -> AnnData

Parameters¶

Name	Type	Default	Description
`anndata_cell`	AnnData	—	Cell-level object from `preprocess` carrying `obsm["Z_clust"]` (sample-removed embedding).
`cell_type_column`	str	`"cell_type"`	Column name where labels are stored in `.obs`.
`existing_cell_types`	bool	`False`	If `True`, reuse labels already present in `cell_type_column` instead of re-clustering (and aggregate via dendrogram if above `n_target_clusters`).
`n_target_clusters`	int, optional	`None`	Target number of clusters. Triggers recursive resolution tuning + dendrogram aggregation.
`umap`	bool	`True`	Whether to compute a UMAP.
`save`	bool	`False`	Write the updated AnnData to disk.
`output_dir`	str, optional	`None`	Directory for the UMAP PNG, the `cell_type.csv`, and, with `save=True`, the updated h5ad.
`defined_output_path`	str, optional	`None`	Override the exact h5ad path that is written.
`leiden_cluster_resolution`	float	`0.8`	Starting Leiden resolution.
`cell_embedding_column`	str, optional	`None`	Which embedding to cluster on. When `None`, defaults to `"Z_clust"`.
`cell_embedding_num_PCs`	int	`20`	Number of components used in the neighborhood graph (RNA).
`verbose`	bool	`True`	Print progress.
`umap_plots`	bool	`True`	Save the UMAP visualization as a PNG.
`_recursion_depth`	int	`0`	Internal — do not set manually.

Returns¶

AnnData — the updated anndata_cell with .obs["cell_type"] populated and, if requested, X_umap in .obsm.

API change

The previous two-argument form (anndata_sample plus defined_sample_output_path) and the (anndata_cell, anndata_sample) tuple return are gone. cell_types now operates on and returns a single cell-level AnnData; the sample-level object is materialized later by the sample-embedding step.

Output files¶

{output_dir}/preprocess/umap_*.png (when umap_plots=True, umap=True, and output_dir is set).
{output_dir}/preprocess/cell_type.csv (cell-id → cell-type table, written whenever output_dir is set).
{output_dir}/preprocess/adata_preprocessed.h5ad (or defined_output_path) when save=True.

Usage¶

from sampledisco.preparation.cell_type_cpu import cell_types

adata_cluster = cell_types(
    anndata_cell=adata_cluster,
    leiden_cluster_resolution=0.99,
    n_target_clusters=None,
    umap=True,
    save=True,
    output_dir="sampledisco_demo_output/rna",
)

ATAC has its own function

ATAC cell typing is handled by cell_types_atac in preparation/ATAC_cell_type.py (GPU: cell_types_atac_gpu), which works on the ATAC diffusion-map / LSI embedding and uses a cosine neighborhood metric. The RNA cell_types above does detect ATAC inputs by the presence of X_lsi in .obsm and switches to a cosine metric for the neighborhood graph, but the dedicated cell_types_atac entry point is preferred for ATAC.

cell_types¶