prescience-bio/dna-format-zoo

DNA Format Zoo

DNA Format Zoo is a collection of public genomic files in varied formats, encodings, reference builds, and producer-specific dialects. It is intended for testing parsers, converters, validators, and other bioinformatics tooling against real files with recorded provenance and expected characteristics.

The initial collection emphasizes HG002 (also known as NA24385), with additional publicly shared participants included when they provide useful format coverage.

Versioning

The collection version is recorded in the root VERSION file. main represents the latest working state; an immutable release exists only when a matching Git tag is published on Hugging Face. For example, VERSION value 0.1.0 corresponds to tag v0.1.0 once that release is created. Upstream dataset and file-format versions remain recorded separately in each artifact's path and metadata.

Before publishing a release, run tools/check-release-version.sh v0.1.0 (using the proposed tag) or run the script without an argument from an exactly tagged Git checkout.

Layout

Sample-derived artifacts use a format-first hierarchy:

data/<format>/<source-or-technology>/<participant>/<case>/

Some formats include a meaningful vendor release before the participant, such as data/snp/23andme/v2/<participant>/. Format specification versions such as VCF 4.2 are metadata rather than directory levels.

Files are classified by their semantic genomic format rather than only their filename extension. For example, Complete Genomics VAR files live under data/cga/complete-genomics/ while retaining their original .tsv.bz2 filenames. Their tab-delimited serialization and bzip2 compression are recorded separately in artifact.yaml.

Reference-sequence artifacts have a dedicated area within data/, organized by reference identity and sequence set:

data/references/<reference-id>/<sequence-set>/

For example, the hs37d5 FASTA and its indexes live in data/references/grch37/hs37d5/. The FASTA format and 1000 Genomes provenance remain metadata rather than directory levels.

Each leaf case contains:

• artifact.yaml: source URLs, reference information, participant, and file roles.
• Upstream payloads and sidecars directly beside the metadata, with their original basenames preserved.
• checksums.sha256: SHA-256 hashes of locally stored files.
• checksums.blake3: BLAKE3 hashes of locally stored files.

catalog.tsv provides a compact collection-wide index. participants.tsv records canonical dataset participant IDs, while participant-aliases.tsv maps source-specific identifiers such as HG002, NA24385, and huAA53E0 to the same participant.

Participant labels

Source-backed HG002 phenotype curation is in labels/HG002. Downloaded public phenotype and identity pages use data/phenotype/ with the same artifact provenance and checksum conventions as genomic files. Normalized labels distinguish donor history, HPO mappings and unresolved genetic hypotheses; they do not equate phenotype history with causal-variant or synthetic disease truth.

Integrity model

SHA-256 and BLAKE3 are calculated over the exact stored bytes. Publisher-provided checksums are recorded separately in artifact.yaml and must be recomputed locally before their verification status is marked match. A missing upstream checksum is recorded as unavailable, not as verified.

Reference-sequence identifiers (for example SAM M5 or GA4GH refget digests) are different from whole-file hashes and will be recorded separately when available.

Run tools/hash-artifacts.sh to regenerate case checksum files. Run tools/verify-artifacts.sh for complete SHA-256 and BLAKE3 verification, or tools/verify-artifacts-blake3.sh for a faster BLAKE3-only content pass. Both verifiers also check catalog completeness, metadata IDs and file lists, missing checksum manifests, and checksum coverage.

CRAM references

CRAM decoding requires the exact reference declared by the artifact. Use the repository-local path from artifact.yaml explicitly rather than relying on a conversion-time UR value retained in the CRAM header. For the included HG002 HiSeq X alignment:

samtools view \
  -T data/references/grch38/hs38dh/hs38DH.fa \
  data/cram/illumina/HG002/grch38-hiseqx-40x/HG002.hiseqx.pcr-free.40x.dedup.grch38.cram

Fast pipeline test fixtures

test/fixtures/hg002-chr20-100kb contains native-coordinate HG002 smoke-test inputs for approximately 100 kb on chromosome 20. It includes alignment-derived ONT and paired Illumina FASTQs, CRAM/CRAI slices, target-only BAM/BAI derivatives, GIAB truth VCF and high-confidence BED slices, and complete chromosome 20 references for GRCh37/hs37d5 and GRCh38/hs38DH.

The data-bearing files cover only the test interval. The complete chromosome sequence is retained in each reference so tools can decode CRAM and use ordinary genome coordinates without rewriting alignments or variants. See the fixture README for provenance, counts, caveats, and regeneration instructions.

test/fixtures/hg002-pacbio-hifi contains four sets of HG002 PacBio HiFi reads from the Revio run m84039_241001_220042_s2 (PacBio's 2024Q4 HG002_rep1), smallest first: 275 reads over the same chromosome 20 interval, 7,900 reads over 34 pharmacogene regions, a random 0.1% of the run (8,106 reads), and all of chromosome 20 (178,868 reads). Each set has the reads unaligned, as the run's own records, and the same reads as PacBio's HiFi-human-WGS-WDL v3.0.2 aligned them, so a pipeline's results can be compared stage by stage with PacBio's published results for the run. The GIAB truth slice above applies to the 100 kb set. The full run is not stored in this repository. See the fixture README for provenance and caveats.

Current scope

The HG002 HiSeq X and full haplotagged Oxford Nanopore BAMs have been converted losslessly to CRAM 3.1. Both are included with indexes, source provenance, exact reference identities, byte-level hashes, and decoded-record comparisons against their source BAMs. Their matching hs38DH and hs37d5 reference artifacts are also included for reproducible decoding.