xdxtools-genomes

Reference genome index builds for the OTTER bisulfite, RNA-seq, and PDX workflows.

Cloning a reference release here is generally faster than rebuilding it: bismark, bowtie2, and STAR index generation takes hours on a full mammalian genome, and the result is identical between machines.

Layout

The dataset mirrors the registry layout that OTTER's reference contract declares — see reference migration — so an extracted release is a working registry entry:

genomes/<id>/<release>/
├── indexes/
│   ├── <id>_<release>_bismark.tar.gz     root: indexes/bismark/
│   ├── <id>_<release>_bowtie2.tar.gz     root: indexes/bowtie2/
│   └── <id>_<release>_star.tar.gz        root: indexes/star/
├── fasta.tar.gz                          root: fasta/
├── annotations.tar.gz                    root: annotations/
├── reference.yaml
├── manifest.json
└── checksums.sha256

Each archive's root matches the path reference.yaml records for it — indexes/bismark for an index, fasta for the sequence — so extracting into the release directory restores the structure the contract describes.

The three files outside the archives are what make a release resolvable: reference.yaml declares the assembly, aliases, and per-asset digests along with the tool that built each index; manifest.json lists every file in the release with its digest; and checksums.sha256 carries the same set in sha256sum form.

Releases

id	release	assembly	organism	dataset size
hg19	GRCh37.p13-gencode-v19	GRCh37.p13	Homo sapiens	41.2 GB
hg38	GRCh38-gencode-v44	GRCh38	Homo sapiens	40.9 GB
mm10	GRCm38-gencode-M25	GRCm38	Mus musculus	36.7 GB
mm39	GRCm39-gencode-vM39	GRCm39	Mus musculus	37.5 GB
mm9	NCBIM37-gencode-M1	NCBIM37	Mus musculus	35.9 GB

A release is always named and selected as <id>@<release>; the identifier alone does not name an immutable artifact. The sizes above are the stored archives; an extracted release is larger, since the indexes are held uncompressed on disk.

Every release is named <assembly>-gencode-<annotation-version>, and the annotations are GENCODE primary assemblies throughout.

Fetching

otter-install downloads and extracts a release into a local registry:

otter-install -reference-fetch

Configured through the environment:

variable	default	purpose
OTTER_REFERENCE_FETCH_RELEASES	—	comma-separated <id>@<release> selections
OTTER_REFERENCE_FETCH_ASSETS	all	restrict to a subset of bismark,bowtie2,star,fasta,annotations
OTTER_REFERENCE_FETCH_REGISTRY_ROOT	~/.otter/references	where genomes/ is written
OTTER_REFERENCE_FETCH_REPO	this dataset	the dataset to read
OTTER_REFERENCE_FETCH_BASE_URL	https://huggingface.co	a mirror can be substituted
OTTER_REFERENCE_FETCH_REVISION	main	branch, tag, or commit

Fetching needs no credentials, because the dataset is public and ungated. That is a requirement rather than an incidental property: a gate would make the anonymous fetch fail even while the dataset still reported itself as public.

Plain curl works too, if you would rather assemble a registry by hand:

base=https://huggingface.co/datasets/fallingstar10/xdxtools-genomes/resolve/main
curl -fL -o mm39.tar.gz \
  "$base/genomes/mm39/GRCm39-gencode-vM39/indexes/mm39_GRCm39-gencode-vM39_bowtie2.tar.gz"

Verification

manifest.json lists every file in the release with its digest, and checksums.sha256 carries the same set in sha256sum form. Either can be checked directly:

cd genomes/mm39/GRCm39-gencode-vM39
python3 - <<'PY'
import hashlib, json, pathlib
root = pathlib.Path(".")
entries = json.loads((root / "manifest.json").read_text())
mismatched = 0
for entry in entries:
    digest = hashlib.sha256((root / entry["path"]).read_bytes()).hexdigest()
    if "sha256:" + digest != entry["sha256"]:
        mismatched += 1
        print("MISMATCH", entry["path"])
print(f"{len(entries)} files checked, {mismatched} mismatched")
PY

Or, once the archives are extracted, with the provider's own tooling:

sha256sum --check checksums.sha256

A release is worth verifying before use rather than after: an index that disagrees with its recorded digest is one that was corrupted in transit or mixed up with another release.

Provenance

Each release records the tool and parameters that produced every index, so a build is auditable rather than merely assumed — for example STAR 2.7.11b with sjdbOverhang: 149 for GRCm39-gencode-vM39. Indexes are not portable across STAR major versions, so read reference.yaml before mixing a fetched index with a differently-pinned pipeline.

Sequences and annotations are redistributed under their providers' terms; the GENCODE FASTA and GTF retain the licences of their respective releases.

Scope

This dataset holds reference data only. Benchmarks, fixtures, and run evidence live in fallingstar10/otter-data.