This dataset is licensed under CC-BY-NC-SA-4.0. Commercial use is not permitted unless explicit written permission is granted by the authors. By agreeing, you accept to share your contact information (email and username) with the repository authors and confirm that you will not use the dataset for harmful, unethical, or malicious purposes. Redistribution and derivatives are permitted under the same CC-BY-NC-SA-4.0 license, with attribution and non-commercial use only. Please allow 3-5 business days for our team to process your request. Requests with vague or insufficient descriptions of intended use will be declined.

HEST-1k Visium virtual single-cell spatial transcriptomics

Predicted virtual single-cell gene expression for HEST-1k Visium samples, produced with DeepSpot2Cell. DeepSpot2Cell is a Deep Sets model that predicts transcriptomic profiles at single-cell resolution from H&E images and CellViT cell segmentations, trained with spot-level Visium supervision only (no single-cell ground truth). At inference every detected cell in each Visium spot receives its own predicted expression vector.

Authors: Max Neuwinger, Kalin Nonchev, Glib Manaiev, Viktor Hendrik Koelzer, and Gunnar Raetsch.

Code: https://github.com/ratschlab/DeepSpot2Cell - the DeepSpot2Cell preprint is on bioRxiv (see the citation below).

Layout

Files are organized per organ; the virtual single-cell predictions are gzip AnnData (.h5ad) keyed by the HEST-1k sample id:

<organ>/virtual_sc/<sample_id>.h5ad        # e.g. bowel/virtual_sc/TENX147.h5ad
<organ>/thumbnails/<sample_id>_downscaled_fullres.jpeg
<organ>/tissue_seg/<sample_id>_contours.geojson  (+ _vis.jpg)
<organ>/spatial_plots/<sample_id>_spatial_plots.png
<organ>/metadata/<sample_id>.json
<organ>/{gene_list.json,gene_stats.csv,sample_manifest.csv}

Per-sample metadata is in metadata.csv (join key id); it merges the HEST-1k clinical/technical fields with this dataset's split, n_cells, um_per_px and patch_size_px. Raw H&E whole-slide images are not re-hosted here (they are identical to HEST-1k - see Download).

Contents (1 samples across 1 organs, 6,030 virtual cells)

Organ	#samples	#virtual cells
Breast	1	6,030

Auto-generated by the upload pipeline on 2026-07-17. Upload in progress - counts grow as files land.

Download

This is a gated dataset, so log in first (pip install huggingface_hub):

from huggingface_hub import login, snapshot_download, hf_hub_download
login(token="YOUR_HF_TOKEN")

repo = "ratschlab/HEST_Visium_virtual_single_cell_transcriptomics"
local_dir = "DS2C_data"

# one organ - virtual single-cell predictions only
snapshot_download(repo, repo_type="dataset", local_dir=local_dir,
                  allow_patterns="bowel/virtual_sc/*")

# one organ - everything (predictions + thumbnails + segmentations + metadata)
snapshot_download(repo, repo_type="dataset", local_dir=local_dir,
                  allow_patterns="lung/*")

# all virtual single-cell predictions across every organ
snapshot_download(repo, repo_type="dataset", local_dir=local_dir,
                  allow_patterns="*/virtual_sc/*")

# gene panels + per-organ manifests only
snapshot_download(repo, repo_type="dataset", local_dir=local_dir,
                  allow_patterns=["*/gene_list.json", "*/gene_stats.csv",
                                  "*/sample_manifest.csv"])

# specific samples (ids match HEST-1k)
snapshot_download(repo, repo_type="dataset", local_dir=local_dir,
                  allow_patterns=["bowel/virtual_sc/TENX147.h5ad",
                                  "lung/virtual_sc/TENX62.h5ad"])

Use the top-level metadata.csv (join key id) to select samples by organ, cancer type, disease state or train/val split before downloading:

import pandas as pd
meta = pd.read_csv(hf_hub_download(repo, "metadata.csv", repo_type="dataset"))
sel = meta.loc[(meta["organ"] == "Lung") & (meta["split"] == "val")]
snapshot_download(repo, repo_type="dataset", local_dir=local_dir,
                  allow_patterns=[f"{o.lower()}/virtual_sc/{i}.h5ad"
                                  for o, i in zip(sel["organ"], sel["id"])])

Raw H&E whole-slide images are not hosted here. They are byte-identical to HEST-1k; download the WSIs, full-resolution patches and CellViT segmentations from there using the same sample ids.

Loading and plotting

Each virtual-SC file is an AnnData (cells x genes):

import io, numpy as np, PIL.Image, scanpy as sc, squidpy as sq

adata = sc.read_h5ad("bowel/virtual_sc/TENX147.h5ad")
adata
# AnnData object with n_obs x n_vars = 97821 x 17956
#     obs: 'cell_id', 'spot_barcode', 'is_inside_spot'
#     var: 'gene'
#     uns: 'sample_id', 'organ', 'oncotree_code', 'disease_state', 'species',
#          'um_per_px', 'patch_size_px', 'expression_space', 'normalization',
#          'fullres_px_width', 'fullres_px_height', 'thumbnail_jpeg', 'thumbnail_shape'
#     obsm: 'spatial'   # (he_x, he_y) in full-resolution WSI pixels

Overlay predicted expression on the embedded H&E thumbnail:

img = np.asarray(PIL.Image.open(io.BytesIO(adata.uns["thumbnail_jpeg"].tobytes())))
lib = str(adata.uns["sample_id"])
adata.uns["spatial"] = {lib: {"images": {"hires": img / 255.0},
                              "scalefactors": {"tissue_hires_scalef": 1.0,
                                               "spot_diameter_fullres": 1.0}}}

# obsm['spatial'] is (he_x, he_y) in full-res pixels; scale to the thumbnail and
# swap to (row, col) for squidpy.
scale = img.shape[1] / float(adata.uns["fullres_px_width"])
adata.obsm["spatial_plot"] = (adata.obsm["spatial"] * scale)[:, [1, 0]]

sq.pl.spatial_scatter(adata, basis="spatial_plot",
                      color=[None, "CEACAM5", "CD3E", "COL1A1"],
                      title=["H&E", "CEACAM5 (tumour)", "CD3E (T cell)", "COL1A1 (stroma)"],
                      size=4, img_alpha=0.9, ncols=2, library_id=lib)

Data organization

Field	Description
`adata.X`	Dense float16 matrix (cells x genes), log1p-normalized expression, clipped >= 0, rounded to 3 decimals (see note).
`adata.obs['cell_id']`	Unique cell identifier, traceable to the CellViT segmentation.
`adata.obs['spot_barcode']`	Visium spot the cell belongs to (`patch_<row>_<col>_<sample_id>`).
`adata.obs['is_inside_spot']`	1 if the nucleus centroid falls inside the 55 um Visium spot, else 0.
`adata.var['gene']`	Gene names (organ-specific panel; see `gene_list.json`).
`adata.obsm['spatial']`	Cell centroid `(he_x, he_y)` in full-resolution WSI pixels.
`adata.uns['sample_id']`	HEST-1k sample id.
`adata.uns['um_per_px']`	Microns per pixel of the source WSI.
`adata.uns['patch_size_px']`	Inference patch crop size (pixels).
`adata.uns['expression_space']`	`"log1p_norm10k"`.
`adata.uns['thumbnail_jpeg']`	JPEG-encoded H&E overview (1-D uint8); `thumbnail_shape` is `[H, W, 3]`.
`adata.uns['organ']`, `['oncotree_code']`, `['disease_state']`, ...	HEST-1k metadata fields.

Expression space: log1p(10000 * raw_counts / total_counts_in_spot) - standard scanpy normalize_total + log1p. Recover approximate normalized counts with np.expm1().

Rounding note: to keep the atlas compact, X is rounded to 3 decimal places in log1p space before storage (absolute error below 1e-3, mean 2.4e-4; the values are float16 model predictions, so this is within their intrinsic precision, and it roughly halves the download).

Companion assets (per organ)

Alongside virtual_sc/, each organ ships:

thumbnails/<id>_downscaled_fullres.jpeg - H&E overview.
tissue_seg/<id>_contours.geojson (+ <id>_vis.jpg) - CellViT tissue/cell contours.
spatial_plots/<id>_spatial_plots.png - spot-overlay QC visualization.
metadata/<id>.json - per-sample HEST-1k metadata.
gene_list.json, gene_stats.csv, sample_manifest.csv - panel + sample summary.

Gene panels

Panels use a hybrid strategy: Xenium anchor genes present in >=50% of Visium training samples come first (for cross-platform validation), then Visium HVGs ranked by consensus votes across training samples.

	Bowel	Breast	Lung
Xenium anchor genes	409	393	286
Visium filler genes	17,547	18,029	13,929
Total	17,956	18,422	14,215

Performance

Accuracy is evaluated on held-out Xenium test samples (true single-cell measurements) by Pearson correlation between predicted and measured per-gene expression.

Model

Architecture - DeepSpot2Cell: a Deep Sets model with multi-head cell attention and neighbourhood context.
Foundation model - H-Optimus-1 (patch + cell embeddings).
Cell segmentation - CellViT++ (from HEST-1k).
Supervision - spot-level Visium only; no single-cell ground truth is seen during training.
Loss - MSE + Pearson-correlation loss on log1p-normalized spot expression.

How to cite

If you use this dataset, please cite:

@article{nonchev2025deepspot2cell,
  title   = {DeepSpot2Cell: Predicting Virtual Single-Cell Spatial Transcriptomics
             from H&E Images using Spot-Level Supervision},
  author  = {Nonchev, Kalin and Manaiev, Glib and Koelzer, Viktor H. and Raetsch, Gunnar},
  journal = {bioRxiv},
  year    = {2025},
  doi     = {10.1101/2025.09.23.678121}
}

@article{jaume2024hest,
  title   = {HEST-1k: A Dataset for Spatial Transcriptomics and Histology Image Analysis},
  author  = {Jaume, Guillaume and others},
  journal = {Advances in Neural Information Processing Systems},
  year    = {2024}
}

NB

This dataset is a companion to HEST-1k; sample ids match HEST-1k exactly. The H&E images and Visium data originate from HEST-1k - please also respect HEST-1k's terms of access. Computational analysis was performed at Leonhard Med, the secure trusted research environment at ETH Zurich.

Downloads last month: 10

Total file size:

112 MB