CyteBase datasets

CyteBase publishes public single-cell datasets, converted from CZ CELLxGENE Discover, as CyteArc DataStores in public Hugging Face buckets. CyteArc opens them like any other remote store: by URI, reading metadata and count blocks over the network as an operation needs them, without downloading the dataset.

Dataset URIs¶

Each public CyteBase release is a read-only bucket named by year and month, such as Nygen/cytebase-2610 for the October 2026 release. A dataset’s store lives at:

hf://buckets/Nygen/cytebase-<release>/datasets/<cytebase_id>/data.zarr

cytebase_id is the dataset’s CyteBase ID, for example mcglade_2021_scrna_data_human_fallopian_tube_30f5e171. Keep the release in the URIs you record, so that your analysis names the exact build it used.

To find datasets, search the catalog with the separate cytebase package. Its results include each dataset’s URI in the zarr_uri column:

# Install the catalog client first: pip install cytebase
import cytebase

# Search the October 2026 release anonymously and take the first match's URI.
catalog = cytebase.Catalog(version="2610", token=False)
matches = catalog.search("fallopian tube", limit=5)
uri = matches[0]["zarr_uri"]

The CyteBase documentation covers catalog search, facet filters, SQL queries, and releases.

Explore a dataset read-only¶

import cytearc

# Choose a dataset in the October 2026 release.
uri = (
    "hf://buckets/Nygen/cytebase-2610/datasets/"
    "mcglade_2021_scrna_data_human_fallopian_tube_30f5e171/data.zarr"
)

# Open the published store read-only; its counts stay in the bucket.
ds = cytearc.DataStore(
    uri, zarr_mode="r", storage_options={"token": False}, min_features_per_cell=-1
)

# Read the published cell annotations for the active cells.
meta = ds.cells.to_pandas_dataframe(["ids", "cell_type"], key="I")

# Plot the authors' UMAP, which the build imports when the source provides one.
umap = ds.artifacts.find("embedding", imported="X_umap", from_assay="RNA")
ds.plots.embedding(layout=umap, color_by="cell_type")

A read-only open writes nothing: an operation that would compute a new result raises PermissionError. min_features_per_cell=-1 keeps the published cell selection, and storage_options={"token": False} reads anonymously; omit it to use your Hugging Face login. Plotting needs CyteArc’s plotting dependencies, cytearc[extra].

Analyze a dataset¶

Mount the dataset to save your own results. A mount is a local, writable store whose counts stay in the bucket:

# Create a local analysis store backed by the remote counts.
analysis = cytearc.mount_datastore(
    uri,
    at="fallopian_tube.zarr",
    storage_options={"token": False},
    min_features_per_cell=-1,
)

# Run the RNA pipeline; its results are saved in fallopian_tube.zarr.
run = analysis.pipeline.run(label="baseline")

# Color the computed UMAP by the selected clustering.
analysis.plots.embedding(run=run, color_by="clusters")

Reopen the mount later with cytearc.DataStore("fallopian_tube.zarr", min_features_per_cell=-1). Its counts still resolve from the bucket, so later reads need network access. Explore a remote dataset explains mounted analysis in detail, and Joint reanalysis of CyteBase datasets analyzes several CyteBase datasets together without copying their counts.

Verified opens with the cytebase package¶

The cytebase package opens the same stores with Catalog.open_datastore and Catalog.mount_datastore. These also check the published dataset record and build receipt against the opened arrays, and a catalog mount writes a .cytebase.json receipt that pins the build it used. See the CyteBase documentation when you need those checks.