HESRT

Human-reviewed, accession-level spatial-omics artifacts with tissue images, expression matrices and source provenance.

Samples (GSMs)	Studies (GSEs)	Expression observations	Compressed packages
4,468	507	37,054,005	413.94 GB

Each sample is a self-contained, checksummed tar.zst package. Browse the catalog, select the accessions you need, then download those samples. The full collection contains approximately 647.72 GB of uncompressed member data; do not clone the repository to try one sample.

The counts describe sample artifacts and source-defined observations (spots, cells or bins), not independent donors. Every included package has a recorded human visual-review PASS. Source interpretation and outstanding verification annotations remain available in the catalog and known issues.

See the data

These nine examples span human cortex, mouse brain, spinal cord, pancreas, Visium HD, FFPE, CytAssist, legacy spatial transcriptomics and Xenium labels. They are selected illustrations, not a representative sample or a new quality assessment. Each thumbnail comes from a released package, with source hashes and display details in preview provenance. The visual_preview Dataset Viewer configuration pairs tissue thumbnails with the package's existing alignment-QC graphic.

Full preview gallery · Usage guide · Schema · Ontology mapping · Known issues · Reuse and citation

Start with a 3.2 MB sample

Python 3.11+. The examples resolve a single revision, verify checksums, extract to disk and open the matrix in backed mode.

python -m pip install "huggingface_hub>=0.34,<2"
python -c "from huggingface_hub import snapshot_download; snapshot_download('Biogod/HESRT', repo_type='dataset', revision='2026-09-17-v1.3', allow_patterns=['examples/*'], local_dir='hesrt_tools', token=False, max_workers=4)"
python -m pip install -r hesrt_tools/examples/requirements.txt
python hesrt_tools/examples/quickstart.py --output hesrt_demo

This loads GSM8915497 / GSE294669, verifies every package member, and reports matrix dimensions, coordinate shape and image format. For filtering, complete-study downloads, local verification and evaluation-split examples, see the usage guide.

Browse and filter

from huggingface_hub import HfApi, hf_hub_download
import pandas as pd

revision = HfApi(token=False).dataset_info(
    "Biogod/HESRT", revision="2026-09-17-v1.3"
).sha
catalog = pd.read_parquet(hf_hub_download(
    "Biogod/HESRT", "releases/2026-09-17/catalog.parquet",
    repo_type="dataset", revision=revision, token=False
))
human = catalog.loc[catalog.organism.eq("Homo sapiens")]
print(human[["gsm_id", "tissue", "technology", "n_obs", "archive_bytes"]].head())

The native Viewer has three configurations: reviewed_catalog (4,468 rows, 96 columns), studies (507 rows), and visual_preview (nine examples). These are lightweight indexes and previews, not streamed full-resolution matrices. reviewed is an inventory split, not a train/test assignment.

Composition

Exact catalog label	Samples
Homo sapiens	2,736
Mus musculus	1,533
10x Genomics Visium	3,943
Spatial Transcriptomics	267
10x Genomics Visium for FFPE	108
10x Genomics Visium HD	57
10x Genomics Visium CytAssist	22
10x Genomics Xenium In Situ	8

Technology and tissue labels are source-derived. Labels can be heterogeneous, and null disease values mean unknown. The counts above use exact catalog labels; related technologies are not silently merged.

Ontology harmonization

The catalog now includes a conservative machine-readable ontology layer while preserving every original metadata field. Organisms map to NCBI Taxonomy; animal anatomy maps to UBERON; plant anatomy maps to Plant Ontology; disease maps to MONDO. No fuzzy match is accepted automatically.

Current coverage is 4,468/4,468 organism records, 2,844/4,468 tissue labels, and 1,800/2,304 non-null disease labels. A further 105 disease labels are explicitly classified as control/reference/non-disease states rather than being forced into MONDO. Compound tumor/specimen phrases that do not identify an ontology term with sufficient confidence stay unresolved.

Use *_ontology_status and *_ontology_match_type to distinguish exact mappings, source-validated aliases, controls, ambiguity and unresolved values. See the ontology mapping protocol and audit.

Package contents

GSM....tar.zst
  spatial.h5ad               Expression matrix and spatial coordinates
  wsi.tif                    Selected tissue image; inspect actual format
  metadata.json              Source-associated sample metadata
  metadata.normalized.json   Normalized values, raw fields and statuses
  provenance.json            Source and processing provenance
  review.json                Recorded visual-review decision
  alignment_qc.png           Existing QC graphic, when available
  alignment_strict_qc.png    Additional QC graphic, when available
  manifest.json              Per-member sizes and SHA256 hashes

Expression/image binaries are unchanged by lossless packaging of the selected standardized artifacts. HESRT does not claim that every X matrix contains source-verified raw UMI counts. A release-wide numerical audit identified eight samples whose X matrices contain clear non-integer values: GSM7781044 (GSE243225) and GSM8243007–GSM8243013 (GSE266244). These source-defined matrices are retained as deposited/standardized rather than reverse-engineered into synthetic counts. Analyses that require integer raw counts should exclude or source-verify these eight samples. The remaining matrices are integer-typed or numerically integer-like in the audit, but that alone is not proof of raw-count provenance. Coordinate frame, units and expression representation should be interpreted using provenance. The name wsi.tif does not guarantee a TIFF file, a specific stain, or a full-resolution whole-slide scan.

Release files

The immutable release namespace is releases/2026-09-17/. Sample catalog and CSV fallback support selective download. Study index, schema, metadata coverage, ontology sidecar, ontology mapping summary, package manifest, SHA256SUMS, and release record provide machine-readable provenance. Prior repository paths remain intact.

For evaluation, keep studies and shared content together, resolve donor/cohort overlap and freeze the exact revision. The provided split script is an illustrative leakage-aware starting point, not an official benchmark protocol.

Reuse and citation

Underlying source materials retain their source-specific terms; no blanket redistribution license is asserted. Cite the source GEO accessions/publications and HESRT release 2026-09-17-v1.3, recording its resolved Hub commit. See reuse and citation and release notes. This resource is intended for research, not clinical decision making.