Hugging Face's logo

Datasets:
deeeech
/
ST-aggregate

Modalities:
Tabular
Text
Formats:
parquet
Size:
10K - 100K
Tags:
spatial-transcriptomics
single-cell
transcriptomics
biology
tabular
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
ST-aggregate / README.md
deeeech's picture
deeeech
Publish streamlined manually gated ST-aggregate catalog v2
2ccc1c6 verified
preview code
|
Raw
History Blame Contribute Delete
4.98 kB
metadata 
pretty_name: ST-aggregate
license: other
tags:
  - spatial-transcriptomics
  - single-cell
  - transcriptomics
  - biology
  - tabular
configs:
  - config_name: datasets
    data_files: data/datasets.parquet
    default: true
  - config_name: open_matrix_candidates
    data_files: data/open_matrix_candidates.parquet
  - config_name: dataset_sources
    data_files: data/dataset_sources.parquet
  - config_name: sources
    data_files: data/sources.parquet
  - config_name: rights_statements
    data_files: data/rights_statements.parquet
  - config_name: source_rights
    data_files: data/source_rights.parquet
  - config_name: gene_axis
    data_files: data/gene_axis.parquet
  - config_name: distribution_summary
    data_files: data/distribution_summary.parquet

ST-aggregate

ST-aggregate is a rights-aware catalog of human single-cell spatial transcriptomics studies. The public analysis table contains 300 source-backed human datasets representing 200,860,787 cells across CosMx, Xenium, MERFISH, Visium, and related platforms.

Important: this v2 repository is a manually gated federated catalog and rebuild index. It does not host expression matrices, per-cell metadata, coordinates, or images. Obtain those objects from the recorded original sources under their respective terms.

Public views

View Datasets Represented cells
datasets 300 200,860,787
Human assay-aware primary 299 200,361,575

Retired empty records, reserved enums, and three organism-unresolved quarantine records (IDs 15, 68, and 281) are not included. Dataset 133 (499,212 cells) remains in datasets but has assay_aware_primary_candidate=false because its official archive does not expose a complete measured-feature axis.

The main table intentionally omits internal lifecycle, release-disposition, row-availability, tombstone/replacement, sparse ontology, model-system, panel, and empty publication-link fields. Detailed provenance and rights remain in normalized relationship tables.

What is hosted

  • compact dataset-level assay, organism, tissue, disease, and cell-count metadata;
  • 351 links from the 300 datasets to 295 original repository records;
  • source URLs, accessions, titles, and organism evidence;
  • 246 rights statements and a complete 295-row source-to-rights relation;
  • the exact ordered 10,763-gene feature axis;
  • CSV mirrors, checksums, schema documentation, and sealed-release provenance.

Distribution tiers

Tier Datasets Represented cells Public v2 behavior
open_matrix_candidate 90 57,498,711 Source pointer only; eligible for a separately reviewed future mirror
pointer_rebuild_only 204 141,456,391 Source pointer and rights metadata only
pointer_rebuild_only_pending_biostudies_object_equivalence 6 1,905,685 Pointer only pending equivalence evidence

open_matrix_candidate is an eligibility classification, not a claim that the matrix is present here. Every row in matrix_hosted_on_huggingface is false in this release. Manual Hub approval controls access to the catalog but does not replace the original sources' terms.

Load 

from datasets import load_dataset

datasets = load_dataset("deeeech/ST-aggregate", "datasets", split="train")
sources = load_dataset("deeeech/ST-aggregate", "sources", split="train")
rights = load_dataset("deeeech/ST-aggregate", "rights_statements", split="train")
genes = load_dataset("deeeech/ST-aggregate", "gene_axis", split="train")

Join datasets to dataset_sources by dataset_id, then join to sources and source_rights by source_id. Join source_rights to rights_statements by rights_id.

Rights and responsible use

There is no blanket license for the third-party source data. The Hub card uses license: other because each source retains its own terms. Public availability is not treated as an open-data license: NOASSERTION and GEO repository policy records remain pointer/rebuild-only. Review docs/RIGHTS.md and the rights tables before downloading or redistributing upstream data.

Publication links and citation

All dataset records link to an original repository source. Scholarly-paper DOI and PMID curation is not yet complete, so empty publication-link columns are not included in v2. Repository accessions such as Zenodo record DOIs must not be mistaken for paper citations.

Until ST-aggregate receives a DOI, cite the exact Hugging Face repository revision and cite every upstream study/source used in an analysis.

Integrity

This catalog derives from the immutable canonical-v17 publication candidate. Its root-manifest SHA-256 is 15e25994d6438c38784df785694a84c6b885e7643cfc1e9ccc9cef2f3dcb7797. The ordered gene-axis semantic SHA-256 is 44d473c84980b905f96349c9d66eeda0db515d34d667f1969d0a0b0cadf087db.

See release_manifest.json, CHECKSUMS.sha256, docs/SCHEMA.md, docs/PROVENANCE.md, and docs/LIMITATIONS.md for the complete contract.