pretty_name: VariantHound Research Snapshot
language:
- en
license: other
size_categories:
- 1M<n<10M
tags:
- biology
- genomics
- veterinary
- phenotype
- tabular
configs:
- config_name: benchmark_cases
data_files:
- split: test
path: data/benchmark_cases.jsonl
VariantHound Research Snapshot
VariantHound is an explainable, phenotype-aware research system for ranking candidate genes in canine inherited disease. This dataset repository contains the compact, redistributable artifacts generated from the public VariantHound source repository.
Research and educational use only. This snapshot is not a diagnostic dataset or medical device, and its rankings require expert review and experimental validation.
Current contents
The release contains:
- three literature-grounded canine benchmark cases;
- source and orthology provenance manifests;
- a checksum-pinned Ensembl 116 dog-human-mouse orthology snapshot in releases built by the publication workflow;
- a checksum pin for HPO 2026-09-01 and its matching human annotations, plus normalized JSONL in publication-workflow releases;
- a dated Monarch 2026-09-02 evidence snapshot with a checksum-pinned HGNC stable-ID crosswalk;
- a direct, retrieval-dated MGI genotype/phenotype snapshot with reconciled stable gene IDs;
- a processed canine OMIA snapshot with phenes, stable-ID gene evidence, causal variants, breeds, and publications;
- a compact PHENIO HP/MP/uPheno semantic-similarity closure with upstream ontology versions;
- content-addressed browser shards for terms, phenotype closure, stable-ID gene profiles, and orthology;
- machine-readable schemas for the published contracts;
- evaluation, data, orthology, HPO, Monarch, MGI, OMIA, and PHENIO documentation;
- a release manifest containing the source Git commit, byte sizes, and SHA-256 checksums.
data/benchmark_cases.jsonl is the viewer-friendly benchmark split. HPO and OMIA are published as normalized JSONL under their source directories; raw upstream assets are not mirrored. The snapshot does not include raw OMIA, Monarch, MGI, Ensembl, or Dog10K database exports.
Evaluation safety
The benchmark metadata defines direct evidence and disease masks for discovery simulation. Temporal evaluation is allowed only when a historical source snapshot predating the case publication has been independently materialized. Known-evidence retrieval results must not be reported as discovery performance.
Orthology status
The publication workflow materializes 41,504 Ensembl 116 protein-tree orthology edges: 20,279 dog-human and 21,225 dog-mouse links. It unions the dog, human, and mouse genome-specific exports because Ensembl partitions pairwise rows arbitrarily between those files. Every artifact is checked against its pinned byte count, official MD5, and SHA-256 before transformation. One-to-many and many-to-many relationships are retained and explicitly marked ambiguous.
HPO status
The publication workflow verifies three immutable GitHub release assets before materializing 20,482 ontology terms, 333,983 human gene-phenotype annotations, and 286,651 disease-phenotype annotations. Negated disease annotations remain explicit. The snapshot is marked current-snapshot-only and cannot be used as a historical temporal-evaluation source.
Monarch status
The dated Monarch KG subsets materialize 6,965 deduplicated human gene-disease associations and 296,123 mouse gene-phenotype associations. Human HGNC IDs are resolved through a separately pinned HGNC export: 4,799 of 4,800 genes have Ensembl IDs and all 4,800 have a stable NCBI or Ensembl join ID. Primary and aggregator provenance are retained so MGI-derived rows cannot be double-counted when direct MGI data is added.
Direct MGI status
The direct MGI snapshot contains 15,317 MP terms and 283,119 genotype-phenotype rows. Identifier reconciliation makes 276,606 rows rank-eligible. It retains but excludes 26 rows affected by conflicting NCBI mappings and 6,487 rows whose markers lack stable join IDs. Direct MGI and Monarch's MGI-derived view must be deduplicated by provenance before scoring.
OMIA status
The retrieval-dated canine transform contains 1,019 phenes, 435 gene-phene edges, and 594 variants linked to single-locus traits. OMIA internal gene IDs are resolved only through the separately pinned official export: 430 gene-phene rows and 591 variant rows are rank-eligible, while unresolved mappings remain available for audit. Each evidence row includes deterministic discovery mask IDs. Because OMIA provides a moving dump rather than historical releases, this snapshot is restricted to current-snapshot evaluation.
PHENIO status
The pinned PHENIO v2026-09-01 transform scans 82,128,289 materialized relations and retains 3,972,487 entailed reflexive subclass rows for 20,413 HPO, 15,317 MP, and 54,198 uPheno terms. This compact closure preserves the cross-species paths needed by deterministic semantic similarity while excluding unrelated integrated relations. The full upstream graph is not mirrored.
Reproducibility
Every uploaded release is built by pipelines/build_hf_release.py. Verify release-manifest.json before using the data. The manifest records every payload file and its SHA-256 digest.
Licensing and attribution
Original VariantHound code and documentation are Apache-2.0. Benchmark facts and future processed tables retain their upstream sources' terms and attribution requirements, so the aggregate dataset uses license: other. See docs/DATA.md, the per-case references, and the source manifest before redistribution or publication.
OMIA data should cite DOI 10.25910/2AMR-PV70 and acknowledge the Sydney Informatics Hub. OMIA asks users to contact its team about publications that rely on the MySQL dump or extended data.
Known limitations
- Three demonstration cases are not an adequate performance benchmark.
- The ontology vocabulary in the demo is intentionally compact and includes internal
VH:*terms. - Canonical gene symbols are not present in the Compara bulk export and remain null until a separately pinned identifier-enrichment source is applied; stable Ensembl gene IDs are the join keys.
- A learned ranking model is not supported by the current sample size.
See DATA_CONTRACT.md for the package layout and field-level expectations.