hgvs-weaver-data

Reference data for hgvs-weaver: a builder that turns a publisher's release into a local store, and the DataProvider that feeds weaver from it.

An HGVS engine is only as correct as what it is told about each transcript. A RefSeq transcript's sequence is not always the genome's — about one record in forty differs — and where they differ by an insertion or deletion, where the difference sits is a choice. NCBI publishes its choice: the alignment of every RefSeq transcript to the assembly. The builder takes that alignment exon by exon rather than deriving one, so the positions weaver computes are the ones ClinVar's names and VariantValidator's projections follow.

Every placement NCBI publishes is kept — on a chromosome, an alternate locus or a patch — so a gene the primary assembly lacks is served where it is. An Ensembl transcript is placed at its exons, after a check that its record is the genome spliced there. docs/design/placements.md has the reasons. A retired RefSeq version — the one an older variant name cites — is served from NCBI's historical set, stacked under the current release (docs/design/retired-versions.md).

A store is one record per gene — every transcript, its protein, its alignments and the sequences they cite — in bagz files with sorted key and interval tables that are read into memory at open. A lookup is a search over bytes in memory and one ranged read, from local disk or gs://.

weaver-data-build refseq ...     # one RefSeq release's bundles, as a shard
weaver-data-build historical ... # NCBI's historical set of RefSeq versions, each with its Entrez status
                                 # (scripts/fetch_refseq_status.py), as a shard to stack under a release's
weaver-data-build ensembl ...    # one Ensembl release's, placed at its exons and checked against the genome
weaver-data-build index ...      # the index and manifest over an ordered list of shards
weaver-data-build genome ...     # the assembly, cut into compressed blocks

The at-rest format is defined by the protos under proto/, and buf breaking gates every change: stores sit in other projects' buckets, so the format only grows.

Installing

pip install hgvs-weaver-data; the module is weaver_data_provider.

Platforms

bagz publishes wheels for Linux x86_64 and macOS arm64; anywhere else, installing this package builds bagz from source.

Development

uv sync, then uv run pytest. The Python stubs under src/ are generated from proto/ by uv run python scripts/regen.py and committed.