SAHA Data Access

Data schemas, metadata infrastructure, and access documentation for the Spatial Atlas of Human Anatomy (SAHA) — hosted on the AWS Open Data Program.

About SAHA

SAHA is a multimodal, subcellular-resolution spatial reference atlas of human tissues, profiling 15M+ cells from 100+ donors across 16 tissue types. It captures matched RNA (up to 19,000 genes), protein (67 markers), and histological data at ~50 nm resolution using CosMx, Xenium, GeoMx, and RNAscope spatial omics platforms. The atlas includes both healthy and disease tissues (colorectal cancer, pancreatic ductal adenocarcinoma, inflammatory bowel disease).

• Website: saha-project.org
• Preprint: bioRxiv 2025.06.16.658716
• Data deposits: zenodo.org/communities/saha

Repository Contents

data/             # Metadata Parquet tables (samples, donors, runs, panels)
schemas/          # JSON Schema definitions for each metadata table
scripts/          # Data standardization and validation scripts
docs/             # Data dictionary, quick start, access tiers, file formats
tutorials/        # Jupyter notebooks for querying and analyzing SAHA data
dataset.yaml      # AWS Open Data Registry submission file

Dataset Tally

Current cohort and run counts, with breakdowns by organ, status, assay, panel, and site: View the live tally → — built from publication Tables 1 & 2; regenerate with python scripts/generate_stats.py.

Metadata Tables

Four Parquet tables are provided in data/ and mirrored at s3://saha-open-data/metadata/.

Table	Rows	Description
samples.parquet	47	One row per sample — organ, condition, platform, donor, n_cells, QC status, S3 paths
donors.parquet	5	One row per donor — age group, sex, tissue source, consent level
runs.parquet	11	One row per acquisition run — instrument, institution, date, panel, protocol fields
panels.parquet	6	Panel configurations — platform, plex, gene/protein lists

Data Organization on S3

s3://saha-open-data/
├── processed/
│   ├── h5ad/
│   │   ├── cosmx/RNA/        # Organ-level CosMx RNA AnnData (e.g. SAHA_COL_RNA.h5ad)
│   │   ├── cosmx/PRT/        # Organ-level CosMx Protein AnnData
│   │   └── xenium/           # Sample-level Xenium AnnData
│   └── zarr/
│       ├── xenium/           # Xenium SpatialData zarr objects
│       └── cosmx/            # CosMx SpatialData zarr objects
├── raw/
│   ├── cosmx/{sample_id}/    # Raw CosMx flat files (controlled access)
│   └── xenium/{sample_id}/   # Raw Xenium output (controlled access)
├── metadata/                 # Parquet tables (samples, donors, runs, panels)
└── docs/                     # Documentation

CosMx processed data is aggregated to organ level (all samples from one organ in one h5ad). Xenium processed data is kept at sample level. Raw data requires registration — see docs/ACCESS_TIERS.md.

Quick Start

Load metadata locally (no AWS credentials needed):

import pandas as pd

samples = pd.read_parquet("data/samples.parquet")
colon = samples.query("organ == 'colon' and qc_pass == True")
print(colon[["sample_id", "donor_id", "n_cells", "s3_processed_path"]])

Load metadata from S3 (anonymous access):

import pandas as pd

samples = pd.read_parquet(
    "s3://saha-open-data/metadata/samples.parquet",
    storage_options={"anon": True},
)

Load a processed h5ad (anonymous access):

import scanpy as sc

adata = sc.read_h5ad(
    "s3://saha-open-data/processed/h5ad/cosmx/RNA/SAHA_COL_RNA.h5ad",
    backed="r",
)

Query metadata with Athena:

SELECT s.sample_id, s.organ, s.n_cells, d.consent_level
FROM saha.samples s
JOIN saha.donors d ON s.donor_id = d.donor_id
WHERE s.qc_pass = true
ORDER BY s.organ;

See tutorials/ for full worked examples.

Scripts

Script	Purpose
standardize_h5ad.py	Produce public h5ad files from internal objects (whitelist obs columns, add sample_id and donor_id)
create_cosmx_zarr.py	Build CosMx SpatialData zarr from raw flat files
ingest_samples.py	Validate and convert sample manifest → samples.parquet
ingest_donors.py	Validate and convert donor metadata → donors.parquet
ingest_runs.py	Validate and convert run metadata → runs.parquet
ingest_panels.py	Validate and convert panel definitions → panels.parquet
validate_release.py	Cross-check parquet referential integrity and S3 object existence
test_athena.py	Run validation queries against the Glue/Athena catalog
generate_data_dictionary.py	Auto-generate docs/DATA_DICTIONARY.md from schemas
generate_stats.py	Auto-generate docs/stats.html dataset tally from publication Tables 1 & 2

Contributing Data

Institutions contributing spatial omics data to SAHA should see docs/CONTRIBUTING.md for the submission process, required metadata fields, and validation instructions.

License

Code and schemas in this repository: MIT. SAHA datasets (open and registered tiers): CC-BY 4.0. Controlled-tier data are released under a signed DUA, not a public license. See docs/ACCESS_TIERS.md for details on the three access tiers and the SAHA Data Use Statement.

Citation

Park J, et al. Spatial Atlas of Human Anatomy (SAHA). in revision. Preprint: bioRxiv 2025.06.16.658716v3.

Contact

Jiwoon Park — saha-project.org