Anonymous-HD2Spot-NIPs-2026/HD2Spot-dataset

HD2Spot

A multi-tissue benchmark resource for spatial transcriptomics cell-type deconvolution, built from real high-resolution spatial measurements (Xenium and Visium HD) rather than scRNA-seq random mixing.

Anonymous release for double-blind peer review (NeurIPS 2026 Datasets and Benchmarks Track).

Resource summary

• 17 datasets spanning 5 tissues (breast, colon, kidney, ovary, brain)
• 2 species (human, mouse)
• 2 platforms (10x Genomics Visium HD, 10x Genomics Xenium)
• ~238K simulated Visium-style spots with cell-type composition labels derived from cellular-resolution annotations
• Total compressed size: ~25 GB

Distribution format

Each dataset is shipped as a single ZIP archive at the repository root. Unzip the archive of interest, and you get a folder containing exactly four files:

File	Description
st.h5ad	Simulated Visium-style spatial transcriptomics object (AnnData). .X is spot-level gene counts; .obs retains QC fields only — all cell-type count columns have been removed to enforce alignment with annotation.csv.
sc.h5ad	Reference single-cell RNA-seq AnnData for the same tissue. .obs carries the cell-type annotation used to construct the simulated spots.
he.tif	Companion H&E (or post-staining) image, aligned to the spatial coordinate system.
annotation.csv	Annotated cell-type counts per spot (rows = spots in row-order matching st.h5ad.obs_names; columns = cell types).

Dataset list

Dataset	Compressed	Platform	Species	Tissue	Annotated cell type
visiumhd_human_breast_cancer	0.9 GB	Visium HD	Human	Breast	29
visiumhd_human_colon_cancer_01	3.6 GB	Visium HD	Human	Colon	38
visiumhd_human_colon_cancer_02	2.9 GB	Visium HD	Human	Colon	38
visiumhd_human_colon_cancer_03	3.8 GB	Visium HD	Human	Colon	38
visiumhd_human_ovary_cancer	2.9 GB	Visium HD	Human	Ovary	16
visiumhd_mouse_brain_01	0.5 GB	Visium HD	Mouse	Brain	59
visiumhd_mouse_brain_02	0.5 GB	Visium HD	Mouse	Brain	59
visiumhd_mouse_kidney_01	0.4 GB	Visium HD	Mouse	Kidney	13
visiumhd_mouse_kidney_02	0.6 GB	Visium HD	Mouse	Kidney	13
xenium_human_breast_cancer_01	0.7 GB	Xenium	Human	Breast	29
xenium_human_breast_cancer_02	2.1 GB	Xenium	Human	Breast	29
xenium_human_breast_cancer_03	1.8 GB	Xenium	Human	Breast	29
xenium_human_breast_cancer_04	1.5 GB	Xenium	Human	Breast	29
xenium_human_colon_cancer	1.1 GB	Xenium	Human	Colon	38
xenium_human_ovary_cancer_01	0.8 GB	Xenium	Human	Ovary	16
xenium_human_ovary_cancer_02	1.4 GB	Xenium	Human	Ovary	16
xenium_mouse_brain	0.6 GB	Xenium	Mouse	Brain	59

Tissue coverage: breast (5), colon (4), ovary (3), kidney (2, mouse only), brain (3).

Spot alignment guarantee

For every dataset, the row count of annotation.csv matches st.h5ad.n_obs, with rows aligned by position. The cell-type column names in annotation.csv correspond to cell-type categories present in sc.h5ad.obs.

Loading example

import anndata as ad
import pandas as pd
import zipfile
from huggingface_hub import hf_hub_download

zip_path = hf_hub_download(
    repo_id="Anonymous-HD2Spot-NIPs-2026/HD2Spot-dataset",
    repo_type="dataset",
    filename="visiumhd_mouse_kidney_01.zip",
)

with zipfile.ZipFile(zip_path) as z:
    z.extractall(".")

st = ad.read_h5ad("visiumhd_mouse_kidney_01/st.h5ad")
sc = ad.read_h5ad("visiumhd_mouse_kidney_01/sc.h5ad")
gt = pd.read_csv("visiumhd_mouse_kidney_01/annotation.csv")
print(st.shape, sc.shape, gt.shape)

Citation

Withheld for double-blind review.

License

Released under CC-BY-4.0. The underlying primary data sources follow their respective original licenses.