pedigree-graph

Pedigree relationship extraction and kinship on a Rust core, with Python and R bindings.

A row-streaming engine classifies every relationship pair up to the fifth degree one individual at a time, so its working memory is linear in the pedigree size; pair lists, counts, and per-person burden all come from it. Each relationship type is parameterised by (up, down, n_ancestors):

• up: meioses from the first member (the junior role: offspring, descendant, niece_nephew, junior_cousin) up to the common ancestor(s); collateral categories are stored up ≥ down
• down: meioses from common ancestor(s) down to individual B
• n_ancestors: 1 (half/lineal) or 2 (full, i.e. mated pair)
• kinship = n_ancestors × (1/2)^(up + down + 1)

Install

pip install pedigree-graph

Binary wheels cover manylinux x86-64 and AArch64, macOS x86-64 and Apple Silicon, and Windows x86-64 (one abi3 wheel per platform for CPython >= 3.13). Elsewhere pip builds the sdist, which needs a Rust toolchain (rustc >= 1.85).

For development the locked pixi environment carries Python, Rust, and maturin and installs the package editable (the extension module is rebuilt with pixi run maturin develop --release after a Rust change):

git clone https://github.com/rwaples/pedigree-graph.git
cd pedigree-graph
pixi install --locked
pixi run test       # inner loop: everything but the `slow` tests, about 45 s
pixi run test-all   # before a commit: the full suite, about 100 s
pixi run test-rust  # the Rust core

Six of the ten slow tests are the random_30k integration gates, which run the kernels over the whole 30,300-row pedigree. One of them, the frozen v0.7.1 parity gate, takes most of the slow tier's time (about 80 s of 88 s measured on 2026-09-24); the rest take 2 to 14 s each. The other four are the N=2000 effective-size scaling tests in tests/test_effective_size_scaling.py.

Requires Python ≥ 3.13. Runtime deps: numpy and scipy. Construction, relationships, topology, pairwise kinship, the kinship matrices, the generation kinship summary, inbreeding, the lineage counts and the Ne prerequisites are compiled Rust (pedigree_graph._native, sources under crates/). numba is only a test extra, for the retired kernels the test suite keeps as oracles. Pandas is optional and only needed if you pass DataFrames to the constructors.

R

The R package pedigreegraph (under r/) runs the same Rust core. Each GitHub release from v0.10.0 carries its source tarball, which bundles every Rust dependency and builds offline given a Rust toolchain (rustc >= 1.85):

install.packages(
  "https://github.com/rwaples/pedigree-graph/releases/download/v0.10.0/pedigreegraph_0.10.0.tar.gz",
  repos = NULL, type = "source"
)
library(pedigreegraph)
pg <- pedigree_graph(data.frame(
  id = 1:6, mother = c(NA, NA, 1, 1, NA, 3), father = c(NA, NA, 2, 2, NA, 5)
))
pairs <- relationship_pairs(pg, max_degree = 3)    # one row per pair
pairs$kinship <- pair_kinship(pg, pairs$first, pairs$second)
K <- kinship_matrix(pg)                            # Matrix::dsCMatrix
F <- inbreeding(pg)

It covers construction, relationship pairs, pairwise kinship, the complete kinship matrix and inbreeding; views, counts, lineage and effective size are Python-only for now. For development, the r pixi environment carries R: pixi run -e r r-test installs r/ in place and runs its tests, pixi run -e r r-doc regenerates r/man/, and pixi run -e r tools/r_build_tarball.sh builds the source tarball.

Usage

import numpy as np
from pedigree_graph import RELATIONSHIPS, PedigreeGraph

# Construct from arrays (no pandas needed)
pg = PedigreeGraph.from_arrays(
    ids=np.array([0, 1, 2, 3, 4]),
    mother_ids=np.array([-1, -1, 0, 0, 0]),
    father_ids=np.array([-1, -1, 1, 1, 1]),
)

# Or from a table: a dict of columns (pandas-free), or any pandas/polars frame
pg = PedigreeGraph.from_frame(
    {
        "id": np.array([0, 1, 2, 3]),
        "mother": np.array([-1, -1, 0, 0]),
        "father": np.array([-1, -1, 1, 1]),
        "twin": np.array([-1, -1, -1, -1]),
        "sex": np.array([0, 1, 0, 1], dtype=np.int8),
        "generation": np.array([0, 0, 1, 1], dtype=np.int32),
    }
)
# pg = PedigreeGraph.from_frame(df)

# Pairs by relationship category, up to a given degree; each block holds
# read-only int32 first_rows / second_rows in the category's role orientation.
# execution="memory" returns the same blocks at the lowest peak memory
# (the result plus engine state) for about twice the wall time.
pairs = pg.relationship_pairs(max_degree=3)
first, second = pairs["FS"]  # full sibs
print(len(pairs["1C"]))  # 1st cousins (degree 3)
print(RELATIONSHIPS["FS"].nominal_kinship)  # 0.25

# Pedigree-expected kinship for those pairs, or for any two row arrays
kin = pg.pair_kinship(pairs)  # {code: float32 array}
kin_fs = pg.pair_kinship(first, second)

# Exact counts in O(N) memory (no pair lists), and the three kinship-matrix families
counts = pg.relationship_counts(max_degree=3)
K = pg.kinship_matrix()  # complete, CSC float32

# One O(N)-memory pass for full-graph category counts and per-person
# degree-1..5 burden, without materializing pair lists.
burden = pg.relationship_burden()
print(burden.category_counts["FS"])
print(burden.per_person[2, 0])  # degree-1 relatives of graph row 2

# Scalar counts for MZ, MO, FO, FS, MHS and PHS only, also exact
close = pg.close_relative_counts()  # no degree/category selector
assert close["FS"] == counts["FS"]
assert close["GP"] is None  # not computed, not zero

Absent optional columns read as absent: there is no sex default and no generation fallback. Effective-size estimators live in pedigree_graph.effective_size; the FrameLike protocol in pedigree_graph.typing. Migrating from 0.7.1: see the old-to-new table in CHANGELOG.md.

close_relative_counts() is full-graph only. Its requested and exact sets contain the six codes above; all other registry keys map to None. It counts parent edges and sibling groups without pair lists or adjacency powers. For grandparents, avuncular pairs, cousins, or a pedigree view, use relationship_counts() instead. The former estimate_relationship_counts method and the result fields approximate and clamped have been removed.

Relationship registry

Codes follow the convention up_down_n_anc:

See RELATIONSHIPS for the complete list; each RelationshipCategory carries code, label, degree, nominal_kinship, up, down, ancestor_count, and the two positional roles.

Architecture

For contributors: docs/architecture.md maps the module layout and the hidden contracts (coordinate space, exact count coverage, sparse-ID handling, default sex), each with its source of truth and regression test. The relationship/coordinate vocabulary is in CONTEXT.md; design decisions are in docs/adr/.

License

MIT — see LICENSE.