epistasis-v2

High-performance Python library for fitting high-order epistatic interactions in genotype-phenotype maps. A clean-break rewrite of harmslab/epistasis.

Status: alpha. Phase 1 port, Phase 2 Rust kernel, the Phase 3 Walsh-Hadamard OLS fast path, sparse design matrices for high-order Lasso/ElasticNet, the power and spline nonlinear variants, a monotonic global-epistasis variant (MAVE-NN style), and the remaining classifiers (LDA / QDA / GP / GMM) are all in.

A multi-page Streamlit showcase lives under examples/ and is published at epistasis-v2.streamlit.app.

What changed from v1

• Rust hot-path kernels via PyO3 (epistasis._core) instead of a shipped Cython .c blob.
• uv + maturin build. pyproject.toml only; no setup.py.
• Python 3.10 through 3.13. Older interpreters dropped.
• Type hints on the public API; mypy --strict in CI.
• Composition over @use_sklearn MRO injection. Concrete models hold an sklearn estimator as an attribute and forward calls explicitly, which unlocks modern sklearn (>=1.2) that broke the v1 trick when normalize= was removed.
• Walsh-Hadamard fast-path for Hadamard-encoded OLS fits: O(n log n) closed-form solve, no dense design matrix. Auto-engaged in EpistasisLinearRegression.fit when the attached GPM is a full-order biallelic library under global encoding; everything else falls back to the sklearn path.
• Sparse design-matrix path for Lasso / ElasticNet via scipy.sparse.csc_matrix. sparse="auto" (default) engages for model_type="local" where the per-site product columns are 0/1; pass sparse=True / False to override. This is the memory fix for L >= 20 where the dense float64 design matrix used to OOM.
• Monotonic global-epistasis variant EpistasisMonotonicGE modeled as a sum of K tanh sigmoids with b_k, c_k >= 0, following Tareen et al. 2022 (MAVE-NN). Identifiable by construction; modern alternative to the power transform when the nonlinearity isn't a clean Box-Cox shape.
• Coordinated rewrite of the gpmap dependency as gpmap-v2. Consumes binary_packed (uint8 2D) and encoding_table with site_index instead of the deprecated genotype_index.
• No backward compatibility with v1. Pin the v1 package if you need that behavior.

Repository layout

epistasis-v2/
├── pyproject.toml          uv + maturin build, ruff + mypy + pytest config
├── Cargo.toml              Rust workspace
├── python/epistasis/       Python source (installed as `epistasis`)
├── crates/epistasis-core/  Rust crate, exposed as `epistasis._core`
├── tests/                  pytest suite
├── benches/                pytest-benchmark suites (matrix kernels + FWHT)
├── docs/                   Sphinx docs (Phase 5)
├── .github/workflows/      CI (lint, test, matrix) + release (semantic-release, maturin wheels, PyPI OIDC)
├── CHANGELOG.md            generated by python-semantic-release
└── CONTRIBUTING.md         commit conventions, dev workflow

Installation (dev)

Requires Python >= 3.10 and a Rust toolchain. gpmap-v2 is pulled from PyPI.

uv sync
uv run maturin develop --release
uv run pytest

For lint and type-check:

uv run ruff check .
uv run ruff format --check .
uv run mypy python/epistasis

Plotting

Plotting support is optional. For matplotlib:

pip install "epistasis-v2[plot]"

epistasis.pyplot.plot_coefs draws fitted coefficients as a bar chart colored by interaction order, with a site-participation grid underneath (the signature figure from v1):

from epistasis.models.linear import EpistasisLinearRegression
from epistasis.pyplot import plot_coefs

model = EpistasisLinearRegression(order=3).add_gpm(gpm).fit()
fig, (bar_axis, grid_axis) = plot_coefs(model)

Current progress

Phase 0 (scaffold), Phase 1 (port), Phase 2 (Rust kernels), and Phase 3 (FWHT fast path + sparse design matrices for Lasso/ElasticNet) are complete.

Ported modules:

• epistasis.mapping (sites, coefficients, EpistasisMap)
• epistasis.matrix (encoded vectors and design matrix; Rust-backed)
• epistasis.exceptions (EpistasisError, XMatrixError, FittingError)
• epistasis.utils (genotypes_to_X)
• epistasis.models.base (AbstractEpistasisModel, EpistasisBaseModel)
• epistasis.models.linear (EpistasisLinearRegression with analytic coefficient standard errors and a Walsh-Hadamard fast path for full-order biallelic fits, EpistasisRidge, EpistasisLasso and EpistasisElasticNet with an auto-engaged scipy.sparse design-matrix path)
• epistasis.models.nonlinear (EpistasisNonlinearRegression, FunctionMinimizer, EpistasisPowerTransform (Sailer & Harms 2017), EpistasisSpline (smoothing spline via scipy.interpolate.UnivariateSpline), EpistasisMonotonicGE (monotone tanh-sum global epistasis, Tareen et al. 2022))
• epistasis.models.classifiers (EpistasisLogisticRegression, EpistasisLDA, EpistasisQDA, EpistasisGaussianProcess, EpistasisGaussianMixture)
• epistasis.simulate (simulate_linear_gpm, simulate_random_linear_gpm)
• epistasis.stats (Pearson, R^2, RMSD, SS residuals, AIC, split_gpm)
• epistasis.validate (k_fold, holdout)
• epistasis.sampling.bayesian (BayesianSampler via emcee 3)
• epistasis.fast (fwht_ols_coefficients: closed-form OLS via FWHT)

Rust hot-path kernels in epistasis._core:

• encode_vectors (uint8 binary_packed to int8 Hadamard/local encoding)
• build_model_matrix (parallel site-product over genotype rows; flat ragged sites layout)
• fwht (iterative butterfly Fast Walsh-Hadamard Transform)

Benchmarks vs v1

Measured on Windows 11 against epistasis==0.7.5 + gpmap==0.7.0. Full biallelic space (AT alphabet), timeit best-of-5. See benchmarks/vs_v1.py for reproducible scripts and setup instructions.

Note on v1 times: the Cython extension in epistasis 0.7.5 requires MSVC to compile and produced no pre-built Windows wheel; times below use the pure-Python fallback, which is slower than actual v1+Cython. Even so, the FWHT fast path in v2 is orders of magnitude faster at full order.

fit() order=1 (sklearn lstsq path in both versions)

fit() full order (v1: dense lstsq, v2: FWHT O(N log N))

Rust kernel vs NumPy reference (internal; release build, 16 threads; see benches/)

Contributing

See CONTRIBUTING.md. Commits follow Conventional Commits; releases and the changelog are automated by python-semantic-release.

License

Unlicense (public domain). See UNLICENSE.