STP-Bench

STP-Bench is a comprehensive benchmark dataset for evaluating deep learning models on spatial transcriptomics prediction from histopathology whole-slide images (WSIs).

The dataset accompanies the paper introducing STP-Bench, a unified benchmark designed to standardize evaluation across multiple spatial transcriptomics technologies, tissue types, and experimental protocols.

Unlike previous studies that evaluate on a single cohort or technology, STP-Bench aggregates diverse spatial transcriptomics datasets into a unified benchmark, providing standardized data organization and evaluation protocols for robust and reproducible benchmarking.

---

📊 Dataset Statistics

STP-Bench consists of two benchmark settings:

• STP-Bench Internal: used for model training and internal evaluation.
• STP-Bench External: independent test cohorts used to assess generalization.

This release contains 316 slides: the public part of the internal cohort, all 14 external test sets, and 11 additional slides used only in the data-scaling experiments (see below).

Benchmark	Dataset	Source	Cancer Type	Platform	Slides	Patients	ST Spots
Internal	WUSTL-BRCA-Visium	WUSTL	IDC	Visium	47	21	157,138
	SNU-GBM-Visium	SNU	GBM	Visium	29	17	115,354
	🔒 MGB-PRAD-Visium	MGB	PRAD	Visium	22	14	97,491
	NCCHE-LUAD-Visium	NCCHE	LUAD	Visium	43	30	96,954
	HEST-CCRCC-Visium	Meylan	RCC	Visium	24	24	74,220
	WUSTL-PDAC-Visium	WUSTL	PAAD	Visium	19	14	66,447
	🔒 AUGSBURG-CRC-Xenium	Augsburg, 10x Genomics	CRC	Xenium	15	15	54,105
	NCCHE-LUAD-Xenium	NCCHE	LUAD	Xenium	16	16	34,519
	LAUSANNE-BRCA-Xenium	Lausanne	BRCA	Xenium	15	14	22,196
External	HEST-PRAD-Visium	KTH, MGB, 10x Genomics	PRAD	Visium	34	13	105,674
	HEST-BRCA-Xenium	10x Genomics	BRCA	Xenium	9	7	101,434
	HEST-CRC-Visium	Roche, 10x Genomics	CRC	Visium	21	15	65,448
	WUSTL-RCC-Visium	WUSTL	RCC	Visium	10	10	33,090
	HEST-BRCA-Visium	10x Genomics	BRCA	Visium	7	6	26,218
	HEST-GBM-Visium	10x Genomics	GBM	Visium	3	3	17,814
	HEST-PAAD-Visium	Ruijin	PAAD	Visium	4	4	13,001
	HEST-LUNG-Visium	10x Genomics	Lung	Visium	2	2	10,053
	HEST-LUNG-Visium HD	10x Genomics	Lung	Visium HD	5	5	9,980
	HEST-PAAD-Xenium	10x Genomics	PAAD	Xenium	3	3	9,793
	HEST-CRC-Visium HD	10x Genomics	CRC	Visium HD	4	4	8,959
	LAUSANNE-LUNG-Xenium	Lausanne	Lung	Xenium	6	6	7,516
	HEST-BRCA-Visium HD	10x Genomics	BRCA	Visium HD	3	3	7,209
	HEST-CRC-Xenium	10x Genomics	CRC	Xenium	1	1	4,167

🔒 = reported in the paper but not included in this public release (see the note below).

Internal cohort: 230 slides, 165 patients, and 718,424 spatial transcriptomics spots in 9 datasets. Of these, the 7 datasets without 🔒 are released here (193 slides, 136 patients, 566,828 spots).

External cohort: 112 slides, 82 patients, and 420,356 spatial transcriptomics spots in 14 test sets.

Data-scaling slides: 11 breast Visium slides (11 patients, 12,430 spots; GSM6433*, from the adult human breast cell atlas) are used only as additional training slides in the multi-institution data-scaling experiments.

🔒 Not included in this release. Two in-house datasets used in the accompanying paper are not distributed due to institutional data-sharing restrictions: MGB-PRAD (Visium, 22 slides) and AUGSBURG-CRC (Xenium, 15 slides). Consequently, the paper's experiments that use these two datasets as the training cohort in the external evaluations cannot be fully reproduced from this release. Four HEST-CRC-Xenium slides (TENX111, TENX147, TENX148, TENX149) whose H&E images are identical to AUGSBURG-CRC slides are also withheld, so the two cross-tissue tests that evaluate on HEST-CRC-Xenium can be run only on the one remaining slide.

---

📁 Dataset Structure

STP-Bench/
├── metadata/    <sample_id>.json     per-slide annotations (organ, disease, technology, pixel size, ...)
├── wsis/        <sample_id>.tif      whole-slide H&E images (.tiff for some slides)
├── patches/     <sample_id>.h5       224x224 patches centered on the ST spots (barcode, coords, img)
└── st/          <sample_id>.h5ad     spatial transcriptomics measurements (AnnData)

Directory	Description
metadata	Slide annotations
wsis	Whole-slide H&E images
patches	Spot-centered image patches for each slide
st	Spatial transcriptomics measurements for each slide

---

💻 Supporting Code

Benchmark code is available at

GitHub

https://github.com/NEXGEM/STP-Bench

The repository contains

• preprocessing pipeline
• data loading
• evaluation scripts
• baseline implementations
• benchmark splits
• training examples

---

⬇️ Loading Example

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="nexgem/STP-Bench",
    repo_type="dataset",
    local_dir="./STP-Bench"
)

📖 Citation

If you use STP-Bench, please cite

@article{stpbench2026,
  title={STP-BENCH: A Unified Systematic Benchmark for Virtual Spatial Transcriptomics from Histopathology Images},
  author={Chung, Youngmin and Ha, Ji Hun and Song, Andrew H. and Almagro-P{\'e}rez, Cristina and Seo, Chaeyoung and Suh, Won Jun and Beom, Jeong Won and Oh, Kyoung Bin and Ruppin, Eytan and Mahmood, Faisal and Lee, Joo Sang},
  journal={arXiv preprint arXiv:2609.05956},
  year={2026},
  doi={10.48550/arXiv.2609.05956}
}

---

📜 License

Please refer to the original licenses of each constituent dataset.

The benchmark organization and preprocessing pipeline are released under the CC BY 4.0 License unless otherwise specified.

---

📬 Contact

Youngmin Chung

• Sungkyunkwan University, Suwon, Republic of Korea
• ymblue@skku.edu

Ji Hun Ha

• Sungkyunkwan University, Suwon, Republic of Korea
• gkwlgns323@skku.edu