grus

A meaning-only pedigree IR (protobuf), a deterministic renderer that draws it as a Bennett/NSGC-standard SVG pedigree, and importers from the pedigree file formats people already have. Python 3.13, MIT.

uv add grus            # or: pip install grus
grus import family.fam -o family.pbtxt      # PLINK/LINKAGE PED, kinship2 tables, Phenopackets Family, Open Pedigree JSON
grus validate family.pbtxt
grus render family.pbtxt -o family.svg      # --png with the `raster` extra (needs libcairo)

from grus import convert, ir, render

ps = convert.import_file("family.fam")            # -> PedigreeSet, validated
svg = render.render_set_svg(ps)                   # one SVG for the figure; render.render_svg for one pedigree
text = ir.dump_set_pbtxt(ps)                      # canonical, diffable text form

What the IR is

Who is in the pedigree, their gender and clinical status, and how they relate — never coordinates. Identity is Position{generation, index} (the drawn "II-3"); relationships are Mating{partner_a?, partner_b?, offspring[]} so single parents, founder sibships, consanguinity, separated couples and childlessness are first-class; clinical status is a per-condition list (affected / carrier / presymptomatic, with inheritance and onset when stated); twins, adoption, donor parentage, pregnancy outcomes and collapsed sibships live where the standard draws them. The schema covers the whole Bennett 2008 + 2022 vocabulary and evolves additively. Contract: schema/proto/grus/models/pedigree.proto; rationale and the assessment of Phenopackets / GA4GH Pedigree as alternatives: docs/design/ir.md.

What the renderer does

Ranks generations, orders each row to minimise crossings while keeping couples adjacent and sibships contiguous, solves x on a constraint model, and draws standard symbols and connectors — deterministically, so the same IR always yields the same bytes. Consanguineous loops, multiple mates and cross-lineage marriages route as edges rather than failing; the few shapes it still cannot draw raise DeferredFeatureError instead of drawing something wrong. docs/design/renderer.md, docs/design/layout-v2.md.

Importers

grus.convert reads PLINK .fam/.ped/.psam and LINKAGE .pre, kinship2/Pedixplorer tables (with the relation matrix), GA4GH Phenopackets v2 Family JSON, and Open Pedigree / PhenoTips simple JSON. Generation and index are synthesised (no data format carries them), the source id is kept in external_id, and nothing is inferred that the source did not state. Per-format mapping tables: docs/design/convert.md.

Where it comes from

grus is built at the Centre for Population Genomics to round-trip pedigree figures in papers: figure → IR (a vision model) → figure′ (this renderer), judged for semantic equivalence. That programme — the extraction prompt, corpus mining, evaluation harness and review UI — is not part of this package; the IR and renderer are its stable core. docs/PRODUCT.md; terms in GLOSSARY.md. Docs and docstrings here occasionally cite design notes that stay in the private tree (architecture.md, eval.md, corpus.md, docs/plans/); those pointers are provenance, not links.

Developing

uv sync, then uv run pytest, uv run --group lint pyright, pre-commit install. CI runs the same gates (.github/workflows/): pre-commit, pytest (with and without the raster extra), stub freshness, and buf breaking against the base branch. A v* tag matching the pyproject version builds, publishes to PyPI (trusted publisher) and creates the GitHub release (release.yml). Protos are the source of truth; after editing one, buf lint, buf breaking --against .git#branch=main, and uv run --group codegen python -m tools.schema.regen (needs buf).