Skip to content

Broad-Allen Developing Mouse Brain E15.5: 3D

Loading 3.8 million cells…

Use DegaFiles or AnnData Zarr

Both hosted sources show the same cells, clusters, and alignments. This gallery defaults to compact Parquet DegaFiles; the visualization notebook uses AnnData Zarr. Use the source picker above to compare them. In Python, change only the source argument.

For this unreleased checkout, run npm run build and start the notebook kernel with CELLDEGA_LOCAL_ESM=1 before importing Celldega. This uses the new browser reader instead of an older published bundle.

import celldega as dega

options = dict(alignment="neighborhood", cluster_attr="leiden_1.0",
               rotation_x=20, rotation_orbit=-25, ini_zoom=-3.6, height=700)

from_zarr = dega.viz.CellCloud(
    zarr_url=(
        "https://huggingface.co/datasets/broadinstitute/Celldega_Supporting_Data/"
        "resolve/main/Allen-Broad_Dev-Mouse-Brain/E15.5/cellcloud_v2_1um.zarr"
    ),
    **options,
)

from_degafiles = dega.viz.CellCloud(
    base_url=(
        "https://raw.githubusercontent.com/cornhundred/celldega_xenium_E15.5_cell_cloud/"
        "8d3cd06fc1a478f386826c27ba0e900b40703a3a/xenium_E15.5_cell_cloud_compact"
    ),
    **options,
)

from_degafiles  # display this widget, or from_zarr

The compact DegaFiles are currently served from a fork while upstream PR #1 awaits merge. Both sources support cluster colors, gene-expression coloring, named alignments, and a linked Clustergram. See File Formats for their storage layout.

Measured loading performance

Compact Parquet was slightly faster in our hosted comparison, so it is the gallery default. These are medians from three loads and first Dcx selections per source, using the same browser renderer and bypassing browser HTTP caching:

Source Initial load First Dcx coloring Initial transfer / requests
Compact Parquet on GitHub 3.6 s 0.67 s 31.3 MB / 7
AnnData Zarr on Hugging Face 4.4 s 1.08 s 25.7 MB / 39

Parquet uses fewer requests, which can reduce time spent waiting on the network. Zarr transfers fewer bytes for the initial view. For Dcx, Parquet transfers 3.34 MB in one request, versus 3.43 MB across 28 requests for Zarr. We expect slightly faster loading with compact Parquet for this hosted example; network conditions and CDN caching can change the result. Timings include transfer, decoding, and rendering updates, and exclude the JavaScript bundle startup.

The complete exports are 494 MB for compact Parquet and 583 MB for Zarr, versus 4.59 GB for the original DegaFiles (9.3× and 7.9× smaller, respectively). Both use positions rounded to 1 µm and preserve original cell IDs and gene counts.

Prepare your own copy

The preparation script builds new exports of these same cells and alignments, rounding positions to the nearest 1 µm and using numeric row indices for repeated cell references. Original cell IDs, gene counts, cluster labels, and colors are retained. It writes a new local directory and does not upload or change the hosted data:

python scripts/prepare_allen_mouse_brain_cellcloud.py --download

Run this from the checkout to produce the new local exports. Existing hosted files are not modified. The complete output directories can also be served from another static host.