Datasets:

vhoncharuk
/

DeepSpaceDB-Preview

license: mit
task_categories:
  - other
tags:
  - spatial-transcriptomics
  - xenium
  - spatial-omics
  - bioinformatics
pretty_name: DeepSpaceDB Xenium Preview

DeepSpaceDB Xenium Preview

A collection of processed 10x Genomics Xenium spatial transcriptomics samples, curated from public GEO depositions, stored as SpatialData zarr stores.

This is a small preview of the full DeepSpaceDB Xenium collection: 3 human breast cancer samples (DSIDX00001-DSIDX00003, GEO series GSE243168), each provided as an unfiltered data/ zarr and an analysis-ready processed/ zarr (see below). metadata.csv lists only these samples.

Structure

metadata.csv          # one row per sample: organ, condition, species, source, quality, etc.
data/
  DSIDX00001.zarr.zip  # one zipped SpatialData zarr store per sample (unfiltered)
  ...
processed/
  DSIDX00001.processed.zarr.zip  # same samples, filtered / normalized / clustered / annotated
  ...

Each <id>.zarr.zip (and <id>.processed.zarr.zip) is a SpatialData store containing (where available):

images: morphology_focus (DAPI/nuclear stain) and he_image (H&E, aligned via an Affine transform to the same coordinate space as the other elements)
labels: cell_labels, nucleus_labels (raster segmentation masks)
shapes: cell_boundaries, nucleus_boundaries, cell_circles
points: transcripts (per-molecule detections)
tables: table (cell x gene raw count matrix, AnnData)

`data/` vs `processed/`

Both folders hold the same samples as SpatialData zarr stores with the same images, labels, shapes and transcripts. They differ in the tables/table AnnData (cell x gene table):

	`data/`	`processed/`
Cells	all cells from the Xenium output (DSIDX00001: 167,780)	after cell-level QC filtering (DSIDX00001: 164,000)
Expression	raw counts in `X` only	`X` plus `layers`: `counts`, `normalized`, `log1p`
Embeddings / graph	none	`obsm`: `X_pca`, `X_umap`; `varm`: `PCs`; `obsp`: `connectivities`, `distances`
Clustering	none	`obs.leiden` (Leiden clusters), `obs.banksy_domain` (BANKSY spatial domains)
Cell types	none	`obs.predicted_cell_type`, `obs.main_cell_type`, `obs.cell_type_correlation`
Per-cell QC	Xenium metrics only	adds `n_counts`, `n_genes_by_counts`, `log1p_n_genes_by_counts`, `log1p_total_counts`, `x`, `y`
Per-gene stats	none	`var`: `mean_counts`, `log1p_mean_counts`, `total_counts`, `log1p_total_counts`, `n_cells`, `n_cells_by_counts`, `pct_dropout_by_counts`
`uns`	`spatialdata_attrs`	adds `leiden`, `log1p`, `pca`, `umap`

So data/ is the minimal, unfiltered Xenium output converted to SpatialData, and processed/ is the analysis-ready version: filtered, normalized, embedded, clustered and annotated. Cell types are assigned per Leiden cluster (by correlation to a reference), so all cells in a cluster share one label. morphology_mip is omitted from both. Files are named processed/<id>.processed.zarr.zip and load exactly like data/ (extract the zip, then sd.read_zarr).

Loading a sample

spatialdata.read_zarr() cannot currently read directly from a .zip (this is an open upstream limitation - see scverse/spatialdata#1141) - extract first, then read normally:

import zipfile
import spatialdata as sd

zipfile.ZipFile("DSIDX00001.zarr.zip").extractall("DSIDX00001.zarr")
sdata = sd.read_zarr("DSIDX00001.zarr")

Selecting samples by metadata

metadata.csv lets you filter before downloading any zarr data:

import pandas as pd
from huggingface_hub import hf_hub_download

meta = pd.read_csv(hf_hub_download(repo_id="vhoncharuk/DeepSpaceDB-Preview", repo_type="dataset", filename="metadata.csv"))
breast_samples = meta[meta.organ == "breast"]

for dsidx_id in breast_samples["id"]:
    hf_hub_download(repo_id="vhoncharuk/DeepSpaceDB-Preview", repo_type="dataset", filename=f"data/{dsidx_id}.zarr.zip")

Metadata columns

id (internal accession), sample_id (GEO GSM or source sample name), series_id (GEO GSE, or source collection), organism, condition / condition_detailed (+ Disease Ontology IDs), organ / organ_detailed (+ UBERON IDs), sex, age_or_stage, ethnicity_or_strain, description, source, URL, PMID, date_published, has_manual_annotation, platform, quality.