Datasetpublic

xiao233333/asi-fimsa-workshop-2026

ASI-FIMSA 2026 spatial omics workshop — staged data Small, Colab-sized artifacts derived from the public 10x Genomics Atera whole-transcriptome Xenium preview dataset of FFPE human breast cancer, prepared for a two-hour hands-on workshop at the ASI-FIMSA meeting. Notebooks and build scripts: https://github.com/xiao233333/ASI-FIMSA-workshop-2026 Attribution Derived from Preview Data: Atera In Situ Gene Expression, FFPE Human Breast Cancer. Data © 10x Genomics, used… See the full description on the dataset page: https://huggingface.co/datasets/xiao233333/asi-fimsa-workshop-2026.

sourceHugging Facecc-by-4.0updated 1mo agoView on Hugging Face

0likes18downloads

Dataset Card

ASI-FIMSA 2026 spatial omics workshop — staged data

Small, Colab-sized artifacts derived from the public 10x Genomics Atera whole-transcriptome Xenium preview dataset of FFPE human breast cancer, prepared for a two-hour hands-on workshop at the ASI-FIMSA meeting.

Notebooks and build scripts: https://github.com/xiao233333/ASI-FIMSA-workshop-2026

Attribution

Derived from Preview Data: Atera In Situ Gene Expression, FFPE Human Breast Cancer. Data © 10x Genomics, used under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). These files are modified from the original — see What changed below.

Files

file	size	contents
`atera_wholeslide_cells.h5ad`	19.9 MB	All 170,057 cells. Aligned centroids in `obsm['spatial']`, vendor `graphclust` and `kmeans_10`, `obsm['X_umap']`, `obsm['X_pca']`, our `cell_type` labels, and a 69-gene marker counts matrix. No images.
`atera_crop.zarr.zip`	18.0 MB	A `SpatialData` store for a 2,000 µm window: `he` (3, 1826, 1826), `he_zoom` (3, 1096, 1096) at native 0.2738 µm/px, 16,006 cell boundary polygons, and a (16006, 69) table. One `"global"` coordinate system, in micrometres.
`atera_crop_lr.h5ad`	5.4 MB	The same 16,006 Crop cells as `atera_crop.zarr.zip`, with the same `cell_type` labels and the same `obsm['spatial']`, but carrying a different panel: raw sparse counts for 1,673 connectomeDB2020 ligands and receptors that Atera measures. Built for ligand–receptor analysis, which the 69-gene panel cannot support. No images.
`cci_panel_genes.txt`	—	The 1,673 gene symbols in that panel, and how many complete ligand–receptor pairs each database contributes.
`atera_cluster_annotation.csv`	—	The 34 vendor clusters → 12 named cell types, with scores, margins and `low_confidence` flags.
`crop_window.json`	—	Exactly which window was chosen and why, including its cell-type composition.
`manifest.json`	—	Sizes, sha256 checksums and origins.

Loading

python

from huggingface_hub import hf_hub_download
import anndata as ad, spatialdata as sd, zipfile, pathlib

REPO = "xiao233333/asi-fimsa-workshop-2026"

adata = ad.read_h5ad(hf_hub_download(REPO, "atera_wholeslide_cells.h5ad", repo_type="dataset"))

z = hf_hub_download(REPO, "atera_crop.zarr.zip", repo_type="dataset")
with zipfile.ZipFile(z) as f:          # unzip first -- see the note below
    f.extractall("atera_crop")
sdata = sd.read_zarr(next(pathlib.Path("atera_crop").glob("*.zarr")))

# The ligand-receptor view of the same cells.
lr = ad.read_h5ad(hf_hub_download(REPO, "atera_crop_lr.h5ad", repo_type="dataset"))
lr.obs["imagecol"] = lr.obsm["spatial"][:, 0]   # stlearn reads positions from
lr.obs["imagerow"] = lr.obsm["spatial"][:, 1]   # these, not from obsm["spatial"]

`spatialdata.read_zarr()` cannot open a `.zarr.zip` in place (its store resolver handles LocalStore/FsspecStore and then wants store.root, which ZipStore lacks). Extract first.

Why there are two views of the same 16,006 cells

atera_crop.zarr.zip carries a 69-gene panel chosen to name cell types. Intersected with connectomeDB2020 that leaves four complete literature-supported ligand–receptor pairs, so no interaction analysis is possible on it. Atera is whole-transcriptome, so the genes were measured all along — atera_crop_lr.h5ad carries 1,673 of them and supports 2,168 pairs. The two files describe identical cells in identical coordinates and can be joined on obs['cell_id'].

What changed from the original

The original bundle is ~65 GB. From it we produced ~38 MB by: subsetting to one 2,000 µm spatial window for the imaging artifact; restricting the expression matrix to a 69-gene marker panel; downsampling the H&E overview; and adding cell-type labels. Coordinates are in the H&E-aligned frame (0.27377667386671845 µm/px), not the Xenium instrument frame — the two differ in scale and by roughly a 90° rotation, and are not interconvertible.

Caveats — please read before reusing

—This is pre-release data. 18,028-target "Human WTA (pre-release)" panel, Atera v1 chemistry, Gen2 prototype instrument. 10x notes preview data may be superseded.
—The cell-type labels are ours, not 10x's, and they are teaching material rather than a reference annotation. They were derived automatically from the vendor's per-cluster differential expression and then reviewed. 14 of 34 clusters carry low_confidence; one is Mixed (plasma+mast) and three are Unassigned because the evidence did not support a call.
—B cells are not resolved as their own population. 9.75% of cells express at least one of MS4A1/CD79A/CD79B/BANK1, but they sit inside the large lymphocyte cluster labelled T cell rather than forming their own. All four genes are in the panel, so the question can be re-asked.
—The 69-gene panel is a marker panel chosen for teaching. For anything quantitative, go back to the full 18,028-gene matrix in the original dataset.