ambitious-bio/corpus-synthetic-demo

Corpus by Ambitious Bio: synthetic multi-omic structure demo

This package is the structural companion to Corpus by Ambitious Bio, a donor-matched human tissue multi-omic dataset for biological AI. It shows how Corpus deliveries are organized, how molecular assay files connect to donors, specimens, clinical encounters and residence histories, and how to load and join the tables, using entirely fabricated records so that it can be shared without restriction. The real dataset is described at ambitious.bio/corpus and is available by request through the gated ambitious-bio/corpus repository.

A small, entirely fabricated dataset showing how molecular assay files can connect to donors, specimens, clinical encounters and residence histories. It is intended for evaluating data structure, navigation and joins before a data partnership. It is not a biological reference standard or a model-performance benchmark.

All donor and specimen identities, measurements, clinical histories, dates and locations are invented. Synthetic feature identifiers and tiny molecular fixtures demonstrate file structure; they do not establish biological findings or production assay performance. No real donor records or measured values are included.

Two example cohorts

Source	Donors	Sampling pattern
SYN_TISSUE_ATLAS	SYN_D001, SYN_D002	Liver and lung specimens from each donor, with selected tissue sections for spatial profiling.
SYN_LONGITUDINAL_FLUID	SYN_D003, SYN_D004	Plasma specimens at days 0, 30 and 90 from each donor.

Both cohorts include fabricated clinical records and residence histories. The fluid cohort demonstrates repeated collection from the same person. The tissue cohort demonstrates measurements across tissues from the same donor. Cohort membership does not imply that every donor or specimen has every assay; use the assay manifest to check availability.

The example contains 4 donors, 12 specimens or sections, 18 assays, 8 clinical encounters and 24 clinical events. Its molecular content is intentionally uneven: one tissue proteomics file, four bulk RNA files, six timed plasma proteomics files and seven WGS, single-cell and spatial assay examples. This makes missing-modality handling visible.

What the files demonstrate

Layer	Example form
Bulk tissue proteomics	A compact TSV preserving the 12 columns of the existing protein-group Panel export.
Plasma proteomics	The same Panel layout, used as an illustrative extension for longitudinal plasma measurements. Its use here is not a claim that a separate plasma export was verified.
Bulk transcriptomics	Gene-level TSV with sample_id, gene_id, gene_symbol, raw_counts and tpm.
Germline WGS	Short-read and long-read processed variant fixtures attached to a donor; the source DNA specimen is unspecified in this demo.
Somatic WGS	Short-read and long-read processed variant fixtures attached to a tissue specimen and linked to its donor.
Single-cell transcriptomics	A small Chromium-style MEX count matrix, with cell identifiers scoped to its assay.
Spatial transcriptomics	Visium-style and Xenium-style MEX count matrices, coordinates and illustrative images linked to tissue sections; Xenium-style records also use Parquet.
Clinical records	Encounters and associated diagnoses, medications and laboratory observations.
Residence history	Dated intervals with deliberately fictitious postal codes for temporal joins.

This package contains processed examples. It does not contain raw sequencing reads, instrument spectra or a complete vendor pipeline run. The RNA, variant, clinical and linkage layouts are proposed demonstration designs. The proteomics column layout is the part directly matched to the existing source export. File names resembling vendor outputs indicate the format being illustrated, not vendor certification.

Start here

1. Read metadata/donors.csv and metadata/specimens.csv to see donor and specimen relationships.
2. Use metadata/assays.csv and metadata/files.csv to locate measurements and their files.
3. Open clinical/ehr_encounters.csv, clinical/ehr_events.csv and metadata/residence_history.csv to inspect the time-linked clinical examples.
4. Consult DATA_MODEL.md for join rules and DATA_DICTIONARY.md for flat-table and native-file field definitions. data_dictionary.csv provides the flat-table definitions in a machine-readable form.

Treat each CSV table as a separate table. Do not concatenate donors, specimens, assays and encounters into one row type. Molecular files under assays/ retain their own file conventions and attach through the manifests.

Load the tabular examples

The dataset card defines ten CSV/TSV configurations: files (the default), donors, specimens, assays, residence_history, ehr_encounters, ehr_events, tissue_proteomics, plasma_proteomics and bulk_transcriptomics. Sample files with the same schema are grouped within their assay configuration.

from pathlib import Path
from datasets import load_dataset

# From the Hugging Face Hub
files = load_dataset("ambitious-bio/corpus-synthetic-demo", "files", split="train")
donors = load_dataset("ambitious-bio/corpus-synthetic-demo", "donors", split="train")
plasma = load_dataset("ambitious-bio/corpus-synthetic-demo", "plasma_proteomics", split="train")

# Or from a local copy of the repository
root = Path("corpus-synthetic-demo").resolve()
files = load_dataset(str(root), "files", split="train")

For a dependency-free look at the residence table, Python's CSV reader preserves every field as text, including leading zeros:

import csv

with (root / "metadata/residence_history.csv").open(newline="") as handle:
    residences = list(csv.DictReader(handle))

The CSV reader leaves missing fields as empty strings. Parse dates and numeric fields explicitly when calculating or joining by time.

From the package root, run the standard-library example to follow plasma specimens through donor and collection-date joins to the matching residence interval:

python3 examples/load_demo.py

In the Hugging Face loader, train is a container name; it is not a benchmark training partition. Explicit field types preserve identifiers and postal codes as strings. Blank CSV/TSV fields load as null values; this configuration does not distinguish an empty string from a missing value. collection_day loads as a nullable float for CSV compatibility, with integer-valued days when present. Native scientific assets, including Parquet files, are downloadable through the file manifest and require their own format readers. If the hosted Hugging Face viewer reports an error for a configuration, open a discussion on the repository.

Interpretation rules

• Germline results belong to a donor. Their specimen_id is blank because their analytical scope is the donor; this package leaves the source DNA specimen unspecified.
• Somatic results belong to an assayed tissue specimen. They must not be copied to every specimen from that donor.
• Single-cell barcodes and spatial identifiers must be combined with their assay identifier before joining across files.
• An EHR encounter does not require a matched specimen collection. Join through the donor and time; use a specimen link only when explicitly provided.
• All dates are invented absolute ISO dates or timestamps. Day 0, 30 and 90 describe timing within the fluid demonstration, not a real study schedule.
• Postal codes such as 00001 are fictional strings, not usable geography. Preserve leading zeros and join residence intervals using the observation or collection time.
• Annotation fields in the proteomics fixture are synthetic placeholders, not evidence that a pathway is active in that sample.

Planned Atera extension

Atera is a separate spatial output family. Its current official specification includes an experiment.spatial JSON manifest, cell-feature data in AnnData Zarr, transcript and segmentation data in zipped Zarr, and morphology images in OME-TIFF. A future example should add those assets and their schema explicitly while preserving the same donor, specimen, section and assay links. The current Xenium and Visium fixtures should retain their own platform labels. See 10x Genomics Atera output specification.

Scientific format references

The native fixtures follow selected structures from VCF 4.3, Cell Ranger MEX matrices, Space Ranger spatial outputs and Xenium 3.2 outputs. Each MEX matrix uses matrix.mtx.gz, features.tsv.gz and barcodes.tsv.gz. The package includes small representative subsets, not complete vendor output bundles. Its VCF reference is a fabricated contig, and its spatial images are geometric fixtures with no biological content.

License and citation

This synthetic package is released under the Creative Commons Attribution 4.0 International license. You may use, modify and redistribute it with attribution to Ambitious Bio. The license covers only this fabricated demonstration package; Corpus itself is licensed commercially and is not covered by this license.

If you use the package, cite it as: Ambitious Bio. Corpus by Ambitious Bio: synthetic multi-omic structure demo. Hugging Face, 2026. https://huggingface.co/datasets/ambitious-bio/corpus-synthetic-demo

About Ambitious Bio

Ambitious Bio (Biological Data, Inc., Cambridge, Massachusetts) produces Corpus, a prospectively generated human tissue dataset for teams training the next generation of biological AI. Website: ambitious.bio · Contact: hello@ambitious.bio · Main dataset repository: ambitious-bio/corpus